2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
37 int hugepages_treat_as_movable
;
39 int hugetlb_max_hstate __read_mostly
;
40 unsigned int default_hstate_idx
;
41 struct hstate hstates
[HUGE_MAX_HSTATE
];
43 * Minimum page order among possible hugepage sizes, set to a proper value
46 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
48 __initdata
LIST_HEAD(huge_boot_pages
);
50 /* for command line parsing */
51 static struct hstate
* __initdata parsed_hstate
;
52 static unsigned long __initdata default_hstate_max_huge_pages
;
53 static unsigned long __initdata default_hstate_size
;
54 static bool __initdata parsed_valid_hugepagesz
= true;
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 /* minimum size accounting */
149 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
150 if (delta
> spool
->rsv_hpages
) {
152 * Asking for more reserves than those already taken on
153 * behalf of subpool. Return difference.
155 ret
= delta
- spool
->rsv_hpages
;
156 spool
->rsv_hpages
= 0;
158 ret
= 0; /* reserves already accounted for */
159 spool
->rsv_hpages
-= delta
;
164 spin_unlock(&spool
->lock
);
169 * Subpool accounting for freeing and unreserving pages.
170 * Return the number of global page reservations that must be dropped.
171 * The return value may only be different than the passed value (delta)
172 * in the case where a subpool minimum size must be maintained.
174 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
182 spin_lock(&spool
->lock
);
184 if (spool
->max_hpages
!= -1) /* maximum size accounting */
185 spool
->used_hpages
-= delta
;
187 /* minimum size accounting */
188 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
189 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
192 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
194 spool
->rsv_hpages
+= delta
;
195 if (spool
->rsv_hpages
> spool
->min_hpages
)
196 spool
->rsv_hpages
= spool
->min_hpages
;
200 * If hugetlbfs_put_super couldn't free spool due to an outstanding
201 * quota reference, free it now.
203 unlock_or_release_subpool(spool
);
208 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
210 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
213 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
215 return subpool_inode(file_inode(vma
->vm_file
));
219 * Region tracking -- allows tracking of reservations and instantiated pages
220 * across the pages in a mapping.
222 * The region data structures are embedded into a resv_map and protected
223 * by a resv_map's lock. The set of regions within the resv_map represent
224 * reservations for huge pages, or huge pages that have already been
225 * instantiated within the map. The from and to elements are huge page
226 * indicies into the associated mapping. from indicates the starting index
227 * of the region. to represents the first index past the end of the region.
229 * For example, a file region structure with from == 0 and to == 4 represents
230 * four huge pages in a mapping. It is important to note that the to element
231 * represents the first element past the end of the region. This is used in
232 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
234 * Interval notation of the form [from, to) will be used to indicate that
235 * the endpoint from is inclusive and to is exclusive.
238 struct list_head link
;
244 * Add the huge page range represented by [f, t) to the reserve
245 * map. In the normal case, existing regions will be expanded
246 * to accommodate the specified range. Sufficient regions should
247 * exist for expansion due to the previous call to region_chg
248 * with the same range. However, it is possible that region_del
249 * could have been called after region_chg and modifed the map
250 * in such a way that no region exists to be expanded. In this
251 * case, pull a region descriptor from the cache associated with
252 * the map and use that for the new range.
254 * Return the number of new huge pages added to the map. This
255 * number is greater than or equal to zero.
257 static long region_add(struct resv_map
*resv
, long f
, long t
)
259 struct list_head
*head
= &resv
->regions
;
260 struct file_region
*rg
, *nrg
, *trg
;
263 spin_lock(&resv
->lock
);
264 /* Locate the region we are either in or before. */
265 list_for_each_entry(rg
, head
, link
)
270 * If no region exists which can be expanded to include the
271 * specified range, the list must have been modified by an
272 * interleving call to region_del(). Pull a region descriptor
273 * from the cache and use it for this range.
275 if (&rg
->link
== head
|| t
< rg
->from
) {
276 VM_BUG_ON(resv
->region_cache_count
<= 0);
278 resv
->region_cache_count
--;
279 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
281 list_del(&nrg
->link
);
285 list_add(&nrg
->link
, rg
->link
.prev
);
291 /* Round our left edge to the current segment if it encloses us. */
295 /* Check for and consume any regions we now overlap with. */
297 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
298 if (&rg
->link
== head
)
303 /* If this area reaches higher then extend our area to
304 * include it completely. If this is not the first area
305 * which we intend to reuse, free it. */
309 /* Decrement return value by the deleted range.
310 * Another range will span this area so that by
311 * end of routine add will be >= zero
313 add
-= (rg
->to
- rg
->from
);
319 add
+= (nrg
->from
- f
); /* Added to beginning of region */
321 add
+= t
- nrg
->to
; /* Added to end of region */
325 resv
->adds_in_progress
--;
326 spin_unlock(&resv
->lock
);
332 * Examine the existing reserve map and determine how many
333 * huge pages in the specified range [f, t) are NOT currently
334 * represented. This routine is called before a subsequent
335 * call to region_add that will actually modify the reserve
336 * map to add the specified range [f, t). region_chg does
337 * not change the number of huge pages represented by the
338 * map. However, if the existing regions in the map can not
339 * be expanded to represent the new range, a new file_region
340 * structure is added to the map as a placeholder. This is
341 * so that the subsequent region_add call will have all the
342 * regions it needs and will not fail.
344 * Upon entry, region_chg will also examine the cache of region descriptors
345 * associated with the map. If there are not enough descriptors cached, one
346 * will be allocated for the in progress add operation.
348 * Returns the number of huge pages that need to be added to the existing
349 * reservation map for the range [f, t). This number is greater or equal to
350 * zero. -ENOMEM is returned if a new file_region structure or cache entry
351 * is needed and can not be allocated.
353 static long region_chg(struct resv_map
*resv
, long f
, long t
)
355 struct list_head
*head
= &resv
->regions
;
356 struct file_region
*rg
, *nrg
= NULL
;
360 spin_lock(&resv
->lock
);
362 resv
->adds_in_progress
++;
365 * Check for sufficient descriptors in the cache to accommodate
366 * the number of in progress add operations.
368 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
369 struct file_region
*trg
;
371 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
372 /* Must drop lock to allocate a new descriptor. */
373 resv
->adds_in_progress
--;
374 spin_unlock(&resv
->lock
);
376 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
382 spin_lock(&resv
->lock
);
383 list_add(&trg
->link
, &resv
->region_cache
);
384 resv
->region_cache_count
++;
388 /* Locate the region we are before or in. */
389 list_for_each_entry(rg
, head
, link
)
393 /* If we are below the current region then a new region is required.
394 * Subtle, allocate a new region at the position but make it zero
395 * size such that we can guarantee to record the reservation. */
396 if (&rg
->link
== head
|| t
< rg
->from
) {
398 resv
->adds_in_progress
--;
399 spin_unlock(&resv
->lock
);
400 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
406 INIT_LIST_HEAD(&nrg
->link
);
410 list_add(&nrg
->link
, rg
->link
.prev
);
415 /* Round our left edge to the current segment if it encloses us. */
420 /* Check for and consume any regions we now overlap with. */
421 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
422 if (&rg
->link
== head
)
427 /* We overlap with this area, if it extends further than
428 * us then we must extend ourselves. Account for its
429 * existing reservation. */
434 chg
-= rg
->to
- rg
->from
;
438 spin_unlock(&resv
->lock
);
439 /* We already know we raced and no longer need the new region */
443 spin_unlock(&resv
->lock
);
448 * Abort the in progress add operation. The adds_in_progress field
449 * of the resv_map keeps track of the operations in progress between
450 * calls to region_chg and region_add. Operations are sometimes
451 * aborted after the call to region_chg. In such cases, region_abort
452 * is called to decrement the adds_in_progress counter.
454 * NOTE: The range arguments [f, t) are not needed or used in this
455 * routine. They are kept to make reading the calling code easier as
456 * arguments will match the associated region_chg call.
458 static void region_abort(struct resv_map
*resv
, long f
, long t
)
460 spin_lock(&resv
->lock
);
461 VM_BUG_ON(!resv
->region_cache_count
);
462 resv
->adds_in_progress
--;
463 spin_unlock(&resv
->lock
);
467 * Delete the specified range [f, t) from the reserve map. If the
468 * t parameter is LONG_MAX, this indicates that ALL regions after f
469 * should be deleted. Locate the regions which intersect [f, t)
470 * and either trim, delete or split the existing regions.
472 * Returns the number of huge pages deleted from the reserve map.
473 * In the normal case, the return value is zero or more. In the
474 * case where a region must be split, a new region descriptor must
475 * be allocated. If the allocation fails, -ENOMEM will be returned.
476 * NOTE: If the parameter t == LONG_MAX, then we will never split
477 * a region and possibly return -ENOMEM. Callers specifying
478 * t == LONG_MAX do not need to check for -ENOMEM error.
480 static long region_del(struct resv_map
*resv
, long f
, long t
)
482 struct list_head
*head
= &resv
->regions
;
483 struct file_region
*rg
, *trg
;
484 struct file_region
*nrg
= NULL
;
488 spin_lock(&resv
->lock
);
489 list_for_each_entry_safe(rg
, trg
, head
, link
) {
491 * Skip regions before the range to be deleted. file_region
492 * ranges are normally of the form [from, to). However, there
493 * may be a "placeholder" entry in the map which is of the form
494 * (from, to) with from == to. Check for placeholder entries
495 * at the beginning of the range to be deleted.
497 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
503 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
505 * Check for an entry in the cache before dropping
506 * lock and attempting allocation.
509 resv
->region_cache_count
> resv
->adds_in_progress
) {
510 nrg
= list_first_entry(&resv
->region_cache
,
513 list_del(&nrg
->link
);
514 resv
->region_cache_count
--;
518 spin_unlock(&resv
->lock
);
519 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
527 /* New entry for end of split region */
530 INIT_LIST_HEAD(&nrg
->link
);
532 /* Original entry is trimmed */
535 list_add(&nrg
->link
, &rg
->link
);
540 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
541 del
+= rg
->to
- rg
->from
;
547 if (f
<= rg
->from
) { /* Trim beginning of region */
550 } else { /* Trim end of region */
556 spin_unlock(&resv
->lock
);
562 * A rare out of memory error was encountered which prevented removal of
563 * the reserve map region for a page. The huge page itself was free'ed
564 * and removed from the page cache. This routine will adjust the subpool
565 * usage count, and the global reserve count if needed. By incrementing
566 * these counts, the reserve map entry which could not be deleted will
567 * appear as a "reserved" entry instead of simply dangling with incorrect
570 void hugetlb_fix_reserve_counts(struct inode
*inode
, bool restore_reserve
)
572 struct hugepage_subpool
*spool
= subpool_inode(inode
);
575 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
576 if (restore_reserve
&& rsv_adjust
) {
577 struct hstate
*h
= hstate_inode(inode
);
579 hugetlb_acct_memory(h
, 1);
584 * Count and return the number of huge pages in the reserve map
585 * that intersect with the range [f, t).
587 static long region_count(struct resv_map
*resv
, long f
, long t
)
589 struct list_head
*head
= &resv
->regions
;
590 struct file_region
*rg
;
593 spin_lock(&resv
->lock
);
594 /* Locate each segment we overlap with, and count that overlap. */
595 list_for_each_entry(rg
, head
, link
) {
604 seg_from
= max(rg
->from
, f
);
605 seg_to
= min(rg
->to
, t
);
607 chg
+= seg_to
- seg_from
;
609 spin_unlock(&resv
->lock
);
615 * Convert the address within this vma to the page offset within
616 * the mapping, in pagecache page units; huge pages here.
618 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
619 struct vm_area_struct
*vma
, unsigned long address
)
621 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
622 (vma
->vm_pgoff
>> huge_page_order(h
));
625 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
626 unsigned long address
)
628 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
630 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
633 * Return the size of the pages allocated when backing a VMA. In the majority
634 * cases this will be same size as used by the page table entries.
636 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
638 struct hstate
*hstate
;
640 if (!is_vm_hugetlb_page(vma
))
643 hstate
= hstate_vma(vma
);
645 return 1UL << huge_page_shift(hstate
);
647 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
650 * Return the page size being used by the MMU to back a VMA. In the majority
651 * of cases, the page size used by the kernel matches the MMU size. On
652 * architectures where it differs, an architecture-specific version of this
653 * function is required.
655 #ifndef vma_mmu_pagesize
656 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
658 return vma_kernel_pagesize(vma
);
663 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
664 * bits of the reservation map pointer, which are always clear due to
667 #define HPAGE_RESV_OWNER (1UL << 0)
668 #define HPAGE_RESV_UNMAPPED (1UL << 1)
669 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
672 * These helpers are used to track how many pages are reserved for
673 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
674 * is guaranteed to have their future faults succeed.
676 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
677 * the reserve counters are updated with the hugetlb_lock held. It is safe
678 * to reset the VMA at fork() time as it is not in use yet and there is no
679 * chance of the global counters getting corrupted as a result of the values.
681 * The private mapping reservation is represented in a subtly different
682 * manner to a shared mapping. A shared mapping has a region map associated
683 * with the underlying file, this region map represents the backing file
684 * pages which have ever had a reservation assigned which this persists even
685 * after the page is instantiated. A private mapping has a region map
686 * associated with the original mmap which is attached to all VMAs which
687 * reference it, this region map represents those offsets which have consumed
688 * reservation ie. where pages have been instantiated.
690 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
692 return (unsigned long)vma
->vm_private_data
;
695 static void set_vma_private_data(struct vm_area_struct
*vma
,
698 vma
->vm_private_data
= (void *)value
;
701 struct resv_map
*resv_map_alloc(void)
703 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
704 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
706 if (!resv_map
|| !rg
) {
712 kref_init(&resv_map
->refs
);
713 spin_lock_init(&resv_map
->lock
);
714 INIT_LIST_HEAD(&resv_map
->regions
);
716 resv_map
->adds_in_progress
= 0;
718 INIT_LIST_HEAD(&resv_map
->region_cache
);
719 list_add(&rg
->link
, &resv_map
->region_cache
);
720 resv_map
->region_cache_count
= 1;
725 void resv_map_release(struct kref
*ref
)
727 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
728 struct list_head
*head
= &resv_map
->region_cache
;
729 struct file_region
*rg
, *trg
;
731 /* Clear out any active regions before we release the map. */
732 region_del(resv_map
, 0, LONG_MAX
);
734 /* ... and any entries left in the cache */
735 list_for_each_entry_safe(rg
, trg
, head
, link
) {
740 VM_BUG_ON(resv_map
->adds_in_progress
);
745 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
747 return inode
->i_mapping
->private_data
;
750 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
752 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
753 if (vma
->vm_flags
& VM_MAYSHARE
) {
754 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
755 struct inode
*inode
= mapping
->host
;
757 return inode_resv_map(inode
);
760 return (struct resv_map
*)(get_vma_private_data(vma
) &
765 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
767 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
768 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
770 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
771 HPAGE_RESV_MASK
) | (unsigned long)map
);
774 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
776 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
777 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
779 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
782 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
784 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
786 return (get_vma_private_data(vma
) & flag
) != 0;
789 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
790 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
792 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
793 if (!(vma
->vm_flags
& VM_MAYSHARE
))
794 vma
->vm_private_data
= (void *)0;
797 /* Returns true if the VMA has associated reserve pages */
798 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
800 if (vma
->vm_flags
& VM_NORESERVE
) {
802 * This address is already reserved by other process(chg == 0),
803 * so, we should decrement reserved count. Without decrementing,
804 * reserve count remains after releasing inode, because this
805 * allocated page will go into page cache and is regarded as
806 * coming from reserved pool in releasing step. Currently, we
807 * don't have any other solution to deal with this situation
808 * properly, so add work-around here.
810 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
816 /* Shared mappings always use reserves */
817 if (vma
->vm_flags
& VM_MAYSHARE
) {
819 * We know VM_NORESERVE is not set. Therefore, there SHOULD
820 * be a region map for all pages. The only situation where
821 * there is no region map is if a hole was punched via
822 * fallocate. In this case, there really are no reverves to
823 * use. This situation is indicated if chg != 0.
832 * Only the process that called mmap() has reserves for
835 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
837 * Like the shared case above, a hole punch or truncate
838 * could have been performed on the private mapping.
839 * Examine the value of chg to determine if reserves
840 * actually exist or were previously consumed.
841 * Very Subtle - The value of chg comes from a previous
842 * call to vma_needs_reserves(). The reserve map for
843 * private mappings has different (opposite) semantics
844 * than that of shared mappings. vma_needs_reserves()
845 * has already taken this difference in semantics into
846 * account. Therefore, the meaning of chg is the same
847 * as in the shared case above. Code could easily be
848 * combined, but keeping it separate draws attention to
849 * subtle differences.
860 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
862 int nid
= page_to_nid(page
);
863 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
864 h
->free_huge_pages
++;
865 h
->free_huge_pages_node
[nid
]++;
868 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
872 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
873 if (!is_migrate_isolate_page(page
))
876 * if 'non-isolated free hugepage' not found on the list,
877 * the allocation fails.
879 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
881 list_move(&page
->lru
, &h
->hugepage_activelist
);
882 set_page_refcounted(page
);
883 h
->free_huge_pages
--;
884 h
->free_huge_pages_node
[nid
]--;
888 /* Movability of hugepages depends on migration support. */
889 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
891 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
892 return GFP_HIGHUSER_MOVABLE
;
897 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
898 struct vm_area_struct
*vma
,
899 unsigned long address
, int avoid_reserve
,
902 struct page
*page
= NULL
;
903 struct mempolicy
*mpol
;
904 nodemask_t
*nodemask
;
905 struct zonelist
*zonelist
;
908 unsigned int cpuset_mems_cookie
;
911 * A child process with MAP_PRIVATE mappings created by their parent
912 * have no page reserves. This check ensures that reservations are
913 * not "stolen". The child may still get SIGKILLed
915 if (!vma_has_reserves(vma
, chg
) &&
916 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
919 /* If reserves cannot be used, ensure enough pages are in the pool */
920 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
924 cpuset_mems_cookie
= read_mems_allowed_begin();
925 zonelist
= huge_zonelist(vma
, address
,
926 htlb_alloc_mask(h
), &mpol
, &nodemask
);
928 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
929 MAX_NR_ZONES
- 1, nodemask
) {
930 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
931 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
935 if (!vma_has_reserves(vma
, chg
))
938 SetPagePrivate(page
);
939 h
->resv_huge_pages
--;
946 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
955 * common helper functions for hstate_next_node_to_{alloc|free}.
956 * We may have allocated or freed a huge page based on a different
957 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
958 * be outside of *nodes_allowed. Ensure that we use an allowed
959 * node for alloc or free.
961 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
963 nid
= next_node_in(nid
, *nodes_allowed
);
964 VM_BUG_ON(nid
>= MAX_NUMNODES
);
969 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
971 if (!node_isset(nid
, *nodes_allowed
))
972 nid
= next_node_allowed(nid
, nodes_allowed
);
977 * returns the previously saved node ["this node"] from which to
978 * allocate a persistent huge page for the pool and advance the
979 * next node from which to allocate, handling wrap at end of node
982 static int hstate_next_node_to_alloc(struct hstate
*h
,
983 nodemask_t
*nodes_allowed
)
987 VM_BUG_ON(!nodes_allowed
);
989 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
990 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
996 * helper for free_pool_huge_page() - return the previously saved
997 * node ["this node"] from which to free a huge page. Advance the
998 * next node id whether or not we find a free huge page to free so
999 * that the next attempt to free addresses the next node.
1001 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1005 VM_BUG_ON(!nodes_allowed
);
1007 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1008 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1013 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1014 for (nr_nodes = nodes_weight(*mask); \
1016 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1019 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1020 for (nr_nodes = nodes_weight(*mask); \
1022 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1025 #if (defined(CONFIG_X86_64) || defined(CONFIG_S390)) && \
1026 ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1027 defined(CONFIG_CMA))
1028 static void destroy_compound_gigantic_page(struct page
*page
,
1032 int nr_pages
= 1 << order
;
1033 struct page
*p
= page
+ 1;
1035 atomic_set(compound_mapcount_ptr(page
), 0);
1036 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1037 clear_compound_head(p
);
1038 set_page_refcounted(p
);
1041 set_compound_order(page
, 0);
1042 __ClearPageHead(page
);
1045 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1047 free_contig_range(page_to_pfn(page
), 1 << order
);
1050 static int __alloc_gigantic_page(unsigned long start_pfn
,
1051 unsigned long nr_pages
)
1053 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1054 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
1057 static bool pfn_range_valid_gigantic(struct zone
*z
,
1058 unsigned long start_pfn
, unsigned long nr_pages
)
1060 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1063 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1067 page
= pfn_to_page(i
);
1069 if (page_zone(page
) != z
)
1072 if (PageReserved(page
))
1075 if (page_count(page
) > 0)
1085 static bool zone_spans_last_pfn(const struct zone
*zone
,
1086 unsigned long start_pfn
, unsigned long nr_pages
)
1088 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1089 return zone_spans_pfn(zone
, last_pfn
);
1092 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1094 unsigned long nr_pages
= 1 << order
;
1095 unsigned long ret
, pfn
, flags
;
1098 z
= NODE_DATA(nid
)->node_zones
;
1099 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1100 spin_lock_irqsave(&z
->lock
, flags
);
1102 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1103 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1104 if (pfn_range_valid_gigantic(z
, pfn
, nr_pages
)) {
1106 * We release the zone lock here because
1107 * alloc_contig_range() will also lock the zone
1108 * at some point. If there's an allocation
1109 * spinning on this lock, it may win the race
1110 * and cause alloc_contig_range() to fail...
1112 spin_unlock_irqrestore(&z
->lock
, flags
);
1113 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1115 return pfn_to_page(pfn
);
1116 spin_lock_irqsave(&z
->lock
, flags
);
1121 spin_unlock_irqrestore(&z
->lock
, flags
);
1127 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1128 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1130 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1134 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1136 prep_compound_gigantic_page(page
, huge_page_order(h
));
1137 prep_new_huge_page(h
, page
, nid
);
1143 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1144 nodemask_t
*nodes_allowed
)
1146 struct page
*page
= NULL
;
1149 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1150 page
= alloc_fresh_gigantic_page_node(h
, node
);
1158 static inline bool gigantic_page_supported(void) { return true; }
1160 static inline bool gigantic_page_supported(void) { return false; }
1161 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1162 static inline void destroy_compound_gigantic_page(struct page
*page
,
1163 unsigned int order
) { }
1164 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1165 nodemask_t
*nodes_allowed
) { return 0; }
1168 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1172 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1176 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1177 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1178 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1179 1 << PG_referenced
| 1 << PG_dirty
|
1180 1 << PG_active
| 1 << PG_private
|
1183 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1184 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1185 set_page_refcounted(page
);
1186 if (hstate_is_gigantic(h
)) {
1187 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1188 free_gigantic_page(page
, huge_page_order(h
));
1190 __free_pages(page
, huge_page_order(h
));
1194 struct hstate
*size_to_hstate(unsigned long size
)
1198 for_each_hstate(h
) {
1199 if (huge_page_size(h
) == size
)
1206 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1207 * to hstate->hugepage_activelist.)
1209 * This function can be called for tail pages, but never returns true for them.
1211 bool page_huge_active(struct page
*page
)
1213 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1214 return PageHead(page
) && PagePrivate(&page
[1]);
1217 /* never called for tail page */
1218 static void set_page_huge_active(struct page
*page
)
1220 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1221 SetPagePrivate(&page
[1]);
1224 static void clear_page_huge_active(struct page
*page
)
1226 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1227 ClearPagePrivate(&page
[1]);
1230 void free_huge_page(struct page
*page
)
1233 * Can't pass hstate in here because it is called from the
1234 * compound page destructor.
1236 struct hstate
*h
= page_hstate(page
);
1237 int nid
= page_to_nid(page
);
1238 struct hugepage_subpool
*spool
=
1239 (struct hugepage_subpool
*)page_private(page
);
1240 bool restore_reserve
;
1242 set_page_private(page
, 0);
1243 page
->mapping
= NULL
;
1244 VM_BUG_ON_PAGE(page_count(page
), page
);
1245 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1246 restore_reserve
= PagePrivate(page
);
1247 ClearPagePrivate(page
);
1250 * A return code of zero implies that the subpool will be under its
1251 * minimum size if the reservation is not restored after page is free.
1252 * Therefore, force restore_reserve operation.
1254 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1255 restore_reserve
= true;
1257 spin_lock(&hugetlb_lock
);
1258 clear_page_huge_active(page
);
1259 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1260 pages_per_huge_page(h
), page
);
1261 if (restore_reserve
)
1262 h
->resv_huge_pages
++;
1264 if (h
->surplus_huge_pages_node
[nid
]) {
1265 /* remove the page from active list */
1266 list_del(&page
->lru
);
1267 update_and_free_page(h
, page
);
1268 h
->surplus_huge_pages
--;
1269 h
->surplus_huge_pages_node
[nid
]--;
1271 arch_clear_hugepage_flags(page
);
1272 enqueue_huge_page(h
, page
);
1274 spin_unlock(&hugetlb_lock
);
1277 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1279 INIT_LIST_HEAD(&page
->lru
);
1280 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1281 spin_lock(&hugetlb_lock
);
1282 set_hugetlb_cgroup(page
, NULL
);
1284 h
->nr_huge_pages_node
[nid
]++;
1285 spin_unlock(&hugetlb_lock
);
1286 put_page(page
); /* free it into the hugepage allocator */
1289 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1292 int nr_pages
= 1 << order
;
1293 struct page
*p
= page
+ 1;
1295 /* we rely on prep_new_huge_page to set the destructor */
1296 set_compound_order(page
, order
);
1297 __ClearPageReserved(page
);
1298 __SetPageHead(page
);
1299 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1301 * For gigantic hugepages allocated through bootmem at
1302 * boot, it's safer to be consistent with the not-gigantic
1303 * hugepages and clear the PG_reserved bit from all tail pages
1304 * too. Otherwse drivers using get_user_pages() to access tail
1305 * pages may get the reference counting wrong if they see
1306 * PG_reserved set on a tail page (despite the head page not
1307 * having PG_reserved set). Enforcing this consistency between
1308 * head and tail pages allows drivers to optimize away a check
1309 * on the head page when they need know if put_page() is needed
1310 * after get_user_pages().
1312 __ClearPageReserved(p
);
1313 set_page_count(p
, 0);
1314 set_compound_head(p
, page
);
1316 atomic_set(compound_mapcount_ptr(page
), -1);
1320 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1321 * transparent huge pages. See the PageTransHuge() documentation for more
1324 int PageHuge(struct page
*page
)
1326 if (!PageCompound(page
))
1329 page
= compound_head(page
);
1330 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1332 EXPORT_SYMBOL_GPL(PageHuge
);
1335 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1336 * normal or transparent huge pages.
1338 int PageHeadHuge(struct page
*page_head
)
1340 if (!PageHead(page_head
))
1343 return get_compound_page_dtor(page_head
) == free_huge_page
;
1346 pgoff_t
__basepage_index(struct page
*page
)
1348 struct page
*page_head
= compound_head(page
);
1349 pgoff_t index
= page_index(page_head
);
1350 unsigned long compound_idx
;
1352 if (!PageHuge(page_head
))
1353 return page_index(page
);
1355 if (compound_order(page_head
) >= MAX_ORDER
)
1356 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1358 compound_idx
= page
- page_head
;
1360 return (index
<< compound_order(page_head
)) + compound_idx
;
1363 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1367 page
= __alloc_pages_node(nid
,
1368 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1369 __GFP_REPEAT
|__GFP_NOWARN
,
1370 huge_page_order(h
));
1372 prep_new_huge_page(h
, page
, nid
);
1378 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1384 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1385 page
= alloc_fresh_huge_page_node(h
, node
);
1393 count_vm_event(HTLB_BUDDY_PGALLOC
);
1395 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1401 * Free huge page from pool from next node to free.
1402 * Attempt to keep persistent huge pages more or less
1403 * balanced over allowed nodes.
1404 * Called with hugetlb_lock locked.
1406 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1412 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1414 * If we're returning unused surplus pages, only examine
1415 * nodes with surplus pages.
1417 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1418 !list_empty(&h
->hugepage_freelists
[node
])) {
1420 list_entry(h
->hugepage_freelists
[node
].next
,
1422 list_del(&page
->lru
);
1423 h
->free_huge_pages
--;
1424 h
->free_huge_pages_node
[node
]--;
1426 h
->surplus_huge_pages
--;
1427 h
->surplus_huge_pages_node
[node
]--;
1429 update_and_free_page(h
, page
);
1439 * Dissolve a given free hugepage into free buddy pages. This function does
1440 * nothing for in-use (including surplus) hugepages.
1442 static void dissolve_free_huge_page(struct page
*page
)
1444 spin_lock(&hugetlb_lock
);
1445 if (PageHuge(page
) && !page_count(page
)) {
1446 struct hstate
*h
= page_hstate(page
);
1447 int nid
= page_to_nid(page
);
1448 list_del(&page
->lru
);
1449 h
->free_huge_pages
--;
1450 h
->free_huge_pages_node
[nid
]--;
1451 update_and_free_page(h
, page
);
1453 spin_unlock(&hugetlb_lock
);
1457 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1458 * make specified memory blocks removable from the system.
1459 * Note that start_pfn should aligned with (minimum) hugepage size.
1461 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1465 if (!hugepages_supported())
1468 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << minimum_order
));
1469 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
)
1470 dissolve_free_huge_page(pfn_to_page(pfn
));
1474 * There are 3 ways this can get called:
1475 * 1. With vma+addr: we use the VMA's memory policy
1476 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1477 * page from any node, and let the buddy allocator itself figure
1479 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1480 * strictly from 'nid'
1482 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1483 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1485 int order
= huge_page_order(h
);
1486 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1487 unsigned int cpuset_mems_cookie
;
1490 * We need a VMA to get a memory policy. If we do not
1491 * have one, we use the 'nid' argument.
1493 * The mempolicy stuff below has some non-inlined bits
1494 * and calls ->vm_ops. That makes it hard to optimize at
1495 * compile-time, even when NUMA is off and it does
1496 * nothing. This helps the compiler optimize it out.
1498 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1500 * If a specific node is requested, make sure to
1501 * get memory from there, but only when a node
1502 * is explicitly specified.
1504 if (nid
!= NUMA_NO_NODE
)
1505 gfp
|= __GFP_THISNODE
;
1507 * Make sure to call something that can handle
1510 return alloc_pages_node(nid
, gfp
, order
);
1514 * OK, so we have a VMA. Fetch the mempolicy and try to
1515 * allocate a huge page with it. We will only reach this
1516 * when CONFIG_NUMA=y.
1520 struct mempolicy
*mpol
;
1521 struct zonelist
*zl
;
1522 nodemask_t
*nodemask
;
1524 cpuset_mems_cookie
= read_mems_allowed_begin();
1525 zl
= huge_zonelist(vma
, addr
, gfp
, &mpol
, &nodemask
);
1526 mpol_cond_put(mpol
);
1527 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nodemask
);
1530 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1536 * There are two ways to allocate a huge page:
1537 * 1. When you have a VMA and an address (like a fault)
1538 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1540 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1541 * this case which signifies that the allocation should be done with
1542 * respect for the VMA's memory policy.
1544 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1545 * implies that memory policies will not be taken in to account.
1547 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1548 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1553 if (hstate_is_gigantic(h
))
1557 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1558 * This makes sure the caller is picking _one_ of the modes with which
1559 * we can call this function, not both.
1561 if (vma
|| (addr
!= -1)) {
1562 VM_WARN_ON_ONCE(addr
== -1);
1563 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1566 * Assume we will successfully allocate the surplus page to
1567 * prevent racing processes from causing the surplus to exceed
1570 * This however introduces a different race, where a process B
1571 * tries to grow the static hugepage pool while alloc_pages() is
1572 * called by process A. B will only examine the per-node
1573 * counters in determining if surplus huge pages can be
1574 * converted to normal huge pages in adjust_pool_surplus(). A
1575 * won't be able to increment the per-node counter, until the
1576 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1577 * no more huge pages can be converted from surplus to normal
1578 * state (and doesn't try to convert again). Thus, we have a
1579 * case where a surplus huge page exists, the pool is grown, and
1580 * the surplus huge page still exists after, even though it
1581 * should just have been converted to a normal huge page. This
1582 * does not leak memory, though, as the hugepage will be freed
1583 * once it is out of use. It also does not allow the counters to
1584 * go out of whack in adjust_pool_surplus() as we don't modify
1585 * the node values until we've gotten the hugepage and only the
1586 * per-node value is checked there.
1588 spin_lock(&hugetlb_lock
);
1589 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1590 spin_unlock(&hugetlb_lock
);
1594 h
->surplus_huge_pages
++;
1596 spin_unlock(&hugetlb_lock
);
1598 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1600 spin_lock(&hugetlb_lock
);
1602 INIT_LIST_HEAD(&page
->lru
);
1603 r_nid
= page_to_nid(page
);
1604 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1605 set_hugetlb_cgroup(page
, NULL
);
1607 * We incremented the global counters already
1609 h
->nr_huge_pages_node
[r_nid
]++;
1610 h
->surplus_huge_pages_node
[r_nid
]++;
1611 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1614 h
->surplus_huge_pages
--;
1615 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1617 spin_unlock(&hugetlb_lock
);
1623 * Allocate a huge page from 'nid'. Note, 'nid' may be
1624 * NUMA_NO_NODE, which means that it may be allocated
1628 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1630 unsigned long addr
= -1;
1632 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1636 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1639 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1640 struct vm_area_struct
*vma
, unsigned long addr
)
1642 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1646 * This allocation function is useful in the context where vma is irrelevant.
1647 * E.g. soft-offlining uses this function because it only cares physical
1648 * address of error page.
1650 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1652 struct page
*page
= NULL
;
1654 spin_lock(&hugetlb_lock
);
1655 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1656 page
= dequeue_huge_page_node(h
, nid
);
1657 spin_unlock(&hugetlb_lock
);
1660 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1666 * Increase the hugetlb pool such that it can accommodate a reservation
1669 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1671 struct list_head surplus_list
;
1672 struct page
*page
, *tmp
;
1674 int needed
, allocated
;
1675 bool alloc_ok
= true;
1677 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1679 h
->resv_huge_pages
+= delta
;
1684 INIT_LIST_HEAD(&surplus_list
);
1688 spin_unlock(&hugetlb_lock
);
1689 for (i
= 0; i
< needed
; i
++) {
1690 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1695 list_add(&page
->lru
, &surplus_list
);
1700 * After retaking hugetlb_lock, we need to recalculate 'needed'
1701 * because either resv_huge_pages or free_huge_pages may have changed.
1703 spin_lock(&hugetlb_lock
);
1704 needed
= (h
->resv_huge_pages
+ delta
) -
1705 (h
->free_huge_pages
+ allocated
);
1710 * We were not able to allocate enough pages to
1711 * satisfy the entire reservation so we free what
1712 * we've allocated so far.
1717 * The surplus_list now contains _at_least_ the number of extra pages
1718 * needed to accommodate the reservation. Add the appropriate number
1719 * of pages to the hugetlb pool and free the extras back to the buddy
1720 * allocator. Commit the entire reservation here to prevent another
1721 * process from stealing the pages as they are added to the pool but
1722 * before they are reserved.
1724 needed
+= allocated
;
1725 h
->resv_huge_pages
+= delta
;
1728 /* Free the needed pages to the hugetlb pool */
1729 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1733 * This page is now managed by the hugetlb allocator and has
1734 * no users -- drop the buddy allocator's reference.
1736 put_page_testzero(page
);
1737 VM_BUG_ON_PAGE(page_count(page
), page
);
1738 enqueue_huge_page(h
, page
);
1741 spin_unlock(&hugetlb_lock
);
1743 /* Free unnecessary surplus pages to the buddy allocator */
1744 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1746 spin_lock(&hugetlb_lock
);
1752 * When releasing a hugetlb pool reservation, any surplus pages that were
1753 * allocated to satisfy the reservation must be explicitly freed if they were
1755 * Called with hugetlb_lock held.
1757 static void return_unused_surplus_pages(struct hstate
*h
,
1758 unsigned long unused_resv_pages
)
1760 unsigned long nr_pages
;
1762 /* Uncommit the reservation */
1763 h
->resv_huge_pages
-= unused_resv_pages
;
1765 /* Cannot return gigantic pages currently */
1766 if (hstate_is_gigantic(h
))
1769 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1772 * We want to release as many surplus pages as possible, spread
1773 * evenly across all nodes with memory. Iterate across these nodes
1774 * until we can no longer free unreserved surplus pages. This occurs
1775 * when the nodes with surplus pages have no free pages.
1776 * free_pool_huge_page() will balance the the freed pages across the
1777 * on-line nodes with memory and will handle the hstate accounting.
1779 while (nr_pages
--) {
1780 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1782 cond_resched_lock(&hugetlb_lock
);
1788 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1789 * are used by the huge page allocation routines to manage reservations.
1791 * vma_needs_reservation is called to determine if the huge page at addr
1792 * within the vma has an associated reservation. If a reservation is
1793 * needed, the value 1 is returned. The caller is then responsible for
1794 * managing the global reservation and subpool usage counts. After
1795 * the huge page has been allocated, vma_commit_reservation is called
1796 * to add the page to the reservation map. If the page allocation fails,
1797 * the reservation must be ended instead of committed. vma_end_reservation
1798 * is called in such cases.
1800 * In the normal case, vma_commit_reservation returns the same value
1801 * as the preceding vma_needs_reservation call. The only time this
1802 * is not the case is if a reserve map was changed between calls. It
1803 * is the responsibility of the caller to notice the difference and
1804 * take appropriate action.
1806 enum vma_resv_mode
{
1811 static long __vma_reservation_common(struct hstate
*h
,
1812 struct vm_area_struct
*vma
, unsigned long addr
,
1813 enum vma_resv_mode mode
)
1815 struct resv_map
*resv
;
1819 resv
= vma_resv_map(vma
);
1823 idx
= vma_hugecache_offset(h
, vma
, addr
);
1825 case VMA_NEEDS_RESV
:
1826 ret
= region_chg(resv
, idx
, idx
+ 1);
1828 case VMA_COMMIT_RESV
:
1829 ret
= region_add(resv
, idx
, idx
+ 1);
1832 region_abort(resv
, idx
, idx
+ 1);
1839 if (vma
->vm_flags
& VM_MAYSHARE
)
1841 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1843 * In most cases, reserves always exist for private mappings.
1844 * However, a file associated with mapping could have been
1845 * hole punched or truncated after reserves were consumed.
1846 * As subsequent fault on such a range will not use reserves.
1847 * Subtle - The reserve map for private mappings has the
1848 * opposite meaning than that of shared mappings. If NO
1849 * entry is in the reserve map, it means a reservation exists.
1850 * If an entry exists in the reserve map, it means the
1851 * reservation has already been consumed. As a result, the
1852 * return value of this routine is the opposite of the
1853 * value returned from reserve map manipulation routines above.
1861 return ret
< 0 ? ret
: 0;
1864 static long vma_needs_reservation(struct hstate
*h
,
1865 struct vm_area_struct
*vma
, unsigned long addr
)
1867 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1870 static long vma_commit_reservation(struct hstate
*h
,
1871 struct vm_area_struct
*vma
, unsigned long addr
)
1873 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1876 static void vma_end_reservation(struct hstate
*h
,
1877 struct vm_area_struct
*vma
, unsigned long addr
)
1879 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1882 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1883 unsigned long addr
, int avoid_reserve
)
1885 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1886 struct hstate
*h
= hstate_vma(vma
);
1888 long map_chg
, map_commit
;
1891 struct hugetlb_cgroup
*h_cg
;
1893 idx
= hstate_index(h
);
1895 * Examine the region/reserve map to determine if the process
1896 * has a reservation for the page to be allocated. A return
1897 * code of zero indicates a reservation exists (no change).
1899 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
1901 return ERR_PTR(-ENOMEM
);
1904 * Processes that did not create the mapping will have no
1905 * reserves as indicated by the region/reserve map. Check
1906 * that the allocation will not exceed the subpool limit.
1907 * Allocations for MAP_NORESERVE mappings also need to be
1908 * checked against any subpool limit.
1910 if (map_chg
|| avoid_reserve
) {
1911 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
1913 vma_end_reservation(h
, vma
, addr
);
1914 return ERR_PTR(-ENOSPC
);
1918 * Even though there was no reservation in the region/reserve
1919 * map, there could be reservations associated with the
1920 * subpool that can be used. This would be indicated if the
1921 * return value of hugepage_subpool_get_pages() is zero.
1922 * However, if avoid_reserve is specified we still avoid even
1923 * the subpool reservations.
1929 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1931 goto out_subpool_put
;
1933 spin_lock(&hugetlb_lock
);
1935 * glb_chg is passed to indicate whether or not a page must be taken
1936 * from the global free pool (global change). gbl_chg == 0 indicates
1937 * a reservation exists for the allocation.
1939 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
1941 spin_unlock(&hugetlb_lock
);
1942 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
1944 goto out_uncharge_cgroup
;
1945 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
1946 SetPagePrivate(page
);
1947 h
->resv_huge_pages
--;
1949 spin_lock(&hugetlb_lock
);
1950 list_move(&page
->lru
, &h
->hugepage_activelist
);
1953 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1954 spin_unlock(&hugetlb_lock
);
1956 set_page_private(page
, (unsigned long)spool
);
1958 map_commit
= vma_commit_reservation(h
, vma
, addr
);
1959 if (unlikely(map_chg
> map_commit
)) {
1961 * The page was added to the reservation map between
1962 * vma_needs_reservation and vma_commit_reservation.
1963 * This indicates a race with hugetlb_reserve_pages.
1964 * Adjust for the subpool count incremented above AND
1965 * in hugetlb_reserve_pages for the same page. Also,
1966 * the reservation count added in hugetlb_reserve_pages
1967 * no longer applies.
1971 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
1972 hugetlb_acct_memory(h
, -rsv_adjust
);
1976 out_uncharge_cgroup
:
1977 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1979 if (map_chg
|| avoid_reserve
)
1980 hugepage_subpool_put_pages(spool
, 1);
1981 vma_end_reservation(h
, vma
, addr
);
1982 return ERR_PTR(-ENOSPC
);
1986 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1987 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1988 * where no ERR_VALUE is expected to be returned.
1990 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1991 unsigned long addr
, int avoid_reserve
)
1993 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1999 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
2001 struct huge_bootmem_page
*m
;
2004 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2007 addr
= memblock_virt_alloc_try_nid_nopanic(
2008 huge_page_size(h
), huge_page_size(h
),
2009 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
2012 * Use the beginning of the huge page to store the
2013 * huge_bootmem_page struct (until gather_bootmem
2014 * puts them into the mem_map).
2023 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2024 /* Put them into a private list first because mem_map is not up yet */
2025 list_add(&m
->list
, &huge_boot_pages
);
2030 static void __init
prep_compound_huge_page(struct page
*page
,
2033 if (unlikely(order
> (MAX_ORDER
- 1)))
2034 prep_compound_gigantic_page(page
, order
);
2036 prep_compound_page(page
, order
);
2039 /* Put bootmem huge pages into the standard lists after mem_map is up */
2040 static void __init
gather_bootmem_prealloc(void)
2042 struct huge_bootmem_page
*m
;
2044 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2045 struct hstate
*h
= m
->hstate
;
2048 #ifdef CONFIG_HIGHMEM
2049 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2050 memblock_free_late(__pa(m
),
2051 sizeof(struct huge_bootmem_page
));
2053 page
= virt_to_page(m
);
2055 WARN_ON(page_count(page
) != 1);
2056 prep_compound_huge_page(page
, h
->order
);
2057 WARN_ON(PageReserved(page
));
2058 prep_new_huge_page(h
, page
, page_to_nid(page
));
2060 * If we had gigantic hugepages allocated at boot time, we need
2061 * to restore the 'stolen' pages to totalram_pages in order to
2062 * fix confusing memory reports from free(1) and another
2063 * side-effects, like CommitLimit going negative.
2065 if (hstate_is_gigantic(h
))
2066 adjust_managed_page_count(page
, 1 << h
->order
);
2070 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2074 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2075 if (hstate_is_gigantic(h
)) {
2076 if (!alloc_bootmem_huge_page(h
))
2078 } else if (!alloc_fresh_huge_page(h
,
2079 &node_states
[N_MEMORY
]))
2082 h
->max_huge_pages
= i
;
2085 static void __init
hugetlb_init_hstates(void)
2089 for_each_hstate(h
) {
2090 if (minimum_order
> huge_page_order(h
))
2091 minimum_order
= huge_page_order(h
);
2093 /* oversize hugepages were init'ed in early boot */
2094 if (!hstate_is_gigantic(h
))
2095 hugetlb_hstate_alloc_pages(h
);
2097 VM_BUG_ON(minimum_order
== UINT_MAX
);
2100 static char * __init
memfmt(char *buf
, unsigned long n
)
2102 if (n
>= (1UL << 30))
2103 sprintf(buf
, "%lu GB", n
>> 30);
2104 else if (n
>= (1UL << 20))
2105 sprintf(buf
, "%lu MB", n
>> 20);
2107 sprintf(buf
, "%lu KB", n
>> 10);
2111 static void __init
report_hugepages(void)
2115 for_each_hstate(h
) {
2117 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2118 memfmt(buf
, huge_page_size(h
)),
2119 h
->free_huge_pages
);
2123 #ifdef CONFIG_HIGHMEM
2124 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2125 nodemask_t
*nodes_allowed
)
2129 if (hstate_is_gigantic(h
))
2132 for_each_node_mask(i
, *nodes_allowed
) {
2133 struct page
*page
, *next
;
2134 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2135 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2136 if (count
>= h
->nr_huge_pages
)
2138 if (PageHighMem(page
))
2140 list_del(&page
->lru
);
2141 update_and_free_page(h
, page
);
2142 h
->free_huge_pages
--;
2143 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2148 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2149 nodemask_t
*nodes_allowed
)
2155 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2156 * balanced by operating on them in a round-robin fashion.
2157 * Returns 1 if an adjustment was made.
2159 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2164 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2167 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2168 if (h
->surplus_huge_pages_node
[node
])
2172 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2173 if (h
->surplus_huge_pages_node
[node
] <
2174 h
->nr_huge_pages_node
[node
])
2181 h
->surplus_huge_pages
+= delta
;
2182 h
->surplus_huge_pages_node
[node
] += delta
;
2186 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2187 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2188 nodemask_t
*nodes_allowed
)
2190 unsigned long min_count
, ret
;
2192 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2193 return h
->max_huge_pages
;
2196 * Increase the pool size
2197 * First take pages out of surplus state. Then make up the
2198 * remaining difference by allocating fresh huge pages.
2200 * We might race with __alloc_buddy_huge_page() here and be unable
2201 * to convert a surplus huge page to a normal huge page. That is
2202 * not critical, though, it just means the overall size of the
2203 * pool might be one hugepage larger than it needs to be, but
2204 * within all the constraints specified by the sysctls.
2206 spin_lock(&hugetlb_lock
);
2207 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2208 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2212 while (count
> persistent_huge_pages(h
)) {
2214 * If this allocation races such that we no longer need the
2215 * page, free_huge_page will handle it by freeing the page
2216 * and reducing the surplus.
2218 spin_unlock(&hugetlb_lock
);
2219 if (hstate_is_gigantic(h
))
2220 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2222 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2223 spin_lock(&hugetlb_lock
);
2227 /* Bail for signals. Probably ctrl-c from user */
2228 if (signal_pending(current
))
2233 * Decrease the pool size
2234 * First return free pages to the buddy allocator (being careful
2235 * to keep enough around to satisfy reservations). Then place
2236 * pages into surplus state as needed so the pool will shrink
2237 * to the desired size as pages become free.
2239 * By placing pages into the surplus state independent of the
2240 * overcommit value, we are allowing the surplus pool size to
2241 * exceed overcommit. There are few sane options here. Since
2242 * __alloc_buddy_huge_page() is checking the global counter,
2243 * though, we'll note that we're not allowed to exceed surplus
2244 * and won't grow the pool anywhere else. Not until one of the
2245 * sysctls are changed, or the surplus pages go out of use.
2247 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2248 min_count
= max(count
, min_count
);
2249 try_to_free_low(h
, min_count
, nodes_allowed
);
2250 while (min_count
< persistent_huge_pages(h
)) {
2251 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2253 cond_resched_lock(&hugetlb_lock
);
2255 while (count
< persistent_huge_pages(h
)) {
2256 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2260 ret
= persistent_huge_pages(h
);
2261 spin_unlock(&hugetlb_lock
);
2265 #define HSTATE_ATTR_RO(_name) \
2266 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2268 #define HSTATE_ATTR(_name) \
2269 static struct kobj_attribute _name##_attr = \
2270 __ATTR(_name, 0644, _name##_show, _name##_store)
2272 static struct kobject
*hugepages_kobj
;
2273 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2275 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2277 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2281 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2282 if (hstate_kobjs
[i
] == kobj
) {
2284 *nidp
= NUMA_NO_NODE
;
2288 return kobj_to_node_hstate(kobj
, nidp
);
2291 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2292 struct kobj_attribute
*attr
, char *buf
)
2295 unsigned long nr_huge_pages
;
2298 h
= kobj_to_hstate(kobj
, &nid
);
2299 if (nid
== NUMA_NO_NODE
)
2300 nr_huge_pages
= h
->nr_huge_pages
;
2302 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2304 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2307 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2308 struct hstate
*h
, int nid
,
2309 unsigned long count
, size_t len
)
2312 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2314 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2319 if (nid
== NUMA_NO_NODE
) {
2321 * global hstate attribute
2323 if (!(obey_mempolicy
&&
2324 init_nodemask_of_mempolicy(nodes_allowed
))) {
2325 NODEMASK_FREE(nodes_allowed
);
2326 nodes_allowed
= &node_states
[N_MEMORY
];
2328 } else if (nodes_allowed
) {
2330 * per node hstate attribute: adjust count to global,
2331 * but restrict alloc/free to the specified node.
2333 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2334 init_nodemask_of_node(nodes_allowed
, nid
);
2336 nodes_allowed
= &node_states
[N_MEMORY
];
2338 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2340 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2341 NODEMASK_FREE(nodes_allowed
);
2345 NODEMASK_FREE(nodes_allowed
);
2349 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2350 struct kobject
*kobj
, const char *buf
,
2354 unsigned long count
;
2358 err
= kstrtoul(buf
, 10, &count
);
2362 h
= kobj_to_hstate(kobj
, &nid
);
2363 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2366 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2367 struct kobj_attribute
*attr
, char *buf
)
2369 return nr_hugepages_show_common(kobj
, attr
, buf
);
2372 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2373 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2375 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2377 HSTATE_ATTR(nr_hugepages
);
2382 * hstate attribute for optionally mempolicy-based constraint on persistent
2383 * huge page alloc/free.
2385 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2386 struct kobj_attribute
*attr
, char *buf
)
2388 return nr_hugepages_show_common(kobj
, attr
, buf
);
2391 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2392 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2394 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2396 HSTATE_ATTR(nr_hugepages_mempolicy
);
2400 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2401 struct kobj_attribute
*attr
, char *buf
)
2403 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2404 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2407 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2408 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2411 unsigned long input
;
2412 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2414 if (hstate_is_gigantic(h
))
2417 err
= kstrtoul(buf
, 10, &input
);
2421 spin_lock(&hugetlb_lock
);
2422 h
->nr_overcommit_huge_pages
= input
;
2423 spin_unlock(&hugetlb_lock
);
2427 HSTATE_ATTR(nr_overcommit_hugepages
);
2429 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2430 struct kobj_attribute
*attr
, char *buf
)
2433 unsigned long free_huge_pages
;
2436 h
= kobj_to_hstate(kobj
, &nid
);
2437 if (nid
== NUMA_NO_NODE
)
2438 free_huge_pages
= h
->free_huge_pages
;
2440 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2442 return sprintf(buf
, "%lu\n", free_huge_pages
);
2444 HSTATE_ATTR_RO(free_hugepages
);
2446 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2447 struct kobj_attribute
*attr
, char *buf
)
2449 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2450 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2452 HSTATE_ATTR_RO(resv_hugepages
);
2454 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2455 struct kobj_attribute
*attr
, char *buf
)
2458 unsigned long surplus_huge_pages
;
2461 h
= kobj_to_hstate(kobj
, &nid
);
2462 if (nid
== NUMA_NO_NODE
)
2463 surplus_huge_pages
= h
->surplus_huge_pages
;
2465 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2467 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2469 HSTATE_ATTR_RO(surplus_hugepages
);
2471 static struct attribute
*hstate_attrs
[] = {
2472 &nr_hugepages_attr
.attr
,
2473 &nr_overcommit_hugepages_attr
.attr
,
2474 &free_hugepages_attr
.attr
,
2475 &resv_hugepages_attr
.attr
,
2476 &surplus_hugepages_attr
.attr
,
2478 &nr_hugepages_mempolicy_attr
.attr
,
2483 static struct attribute_group hstate_attr_group
= {
2484 .attrs
= hstate_attrs
,
2487 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2488 struct kobject
**hstate_kobjs
,
2489 struct attribute_group
*hstate_attr_group
)
2492 int hi
= hstate_index(h
);
2494 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2495 if (!hstate_kobjs
[hi
])
2498 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2500 kobject_put(hstate_kobjs
[hi
]);
2505 static void __init
hugetlb_sysfs_init(void)
2510 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2511 if (!hugepages_kobj
)
2514 for_each_hstate(h
) {
2515 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2516 hstate_kobjs
, &hstate_attr_group
);
2518 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2525 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2526 * with node devices in node_devices[] using a parallel array. The array
2527 * index of a node device or _hstate == node id.
2528 * This is here to avoid any static dependency of the node device driver, in
2529 * the base kernel, on the hugetlb module.
2531 struct node_hstate
{
2532 struct kobject
*hugepages_kobj
;
2533 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2535 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2538 * A subset of global hstate attributes for node devices
2540 static struct attribute
*per_node_hstate_attrs
[] = {
2541 &nr_hugepages_attr
.attr
,
2542 &free_hugepages_attr
.attr
,
2543 &surplus_hugepages_attr
.attr
,
2547 static struct attribute_group per_node_hstate_attr_group
= {
2548 .attrs
= per_node_hstate_attrs
,
2552 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2553 * Returns node id via non-NULL nidp.
2555 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2559 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2560 struct node_hstate
*nhs
= &node_hstates
[nid
];
2562 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2563 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2575 * Unregister hstate attributes from a single node device.
2576 * No-op if no hstate attributes attached.
2578 static void hugetlb_unregister_node(struct node
*node
)
2581 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2583 if (!nhs
->hugepages_kobj
)
2584 return; /* no hstate attributes */
2586 for_each_hstate(h
) {
2587 int idx
= hstate_index(h
);
2588 if (nhs
->hstate_kobjs
[idx
]) {
2589 kobject_put(nhs
->hstate_kobjs
[idx
]);
2590 nhs
->hstate_kobjs
[idx
] = NULL
;
2594 kobject_put(nhs
->hugepages_kobj
);
2595 nhs
->hugepages_kobj
= NULL
;
2600 * Register hstate attributes for a single node device.
2601 * No-op if attributes already registered.
2603 static void hugetlb_register_node(struct node
*node
)
2606 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2609 if (nhs
->hugepages_kobj
)
2610 return; /* already allocated */
2612 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2614 if (!nhs
->hugepages_kobj
)
2617 for_each_hstate(h
) {
2618 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2620 &per_node_hstate_attr_group
);
2622 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2623 h
->name
, node
->dev
.id
);
2624 hugetlb_unregister_node(node
);
2631 * hugetlb init time: register hstate attributes for all registered node
2632 * devices of nodes that have memory. All on-line nodes should have
2633 * registered their associated device by this time.
2635 static void __init
hugetlb_register_all_nodes(void)
2639 for_each_node_state(nid
, N_MEMORY
) {
2640 struct node
*node
= node_devices
[nid
];
2641 if (node
->dev
.id
== nid
)
2642 hugetlb_register_node(node
);
2646 * Let the node device driver know we're here so it can
2647 * [un]register hstate attributes on node hotplug.
2649 register_hugetlbfs_with_node(hugetlb_register_node
,
2650 hugetlb_unregister_node
);
2652 #else /* !CONFIG_NUMA */
2654 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2662 static void hugetlb_register_all_nodes(void) { }
2666 static int __init
hugetlb_init(void)
2670 if (!hugepages_supported())
2673 if (!size_to_hstate(default_hstate_size
)) {
2674 default_hstate_size
= HPAGE_SIZE
;
2675 if (!size_to_hstate(default_hstate_size
))
2676 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2678 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2679 if (default_hstate_max_huge_pages
) {
2680 if (!default_hstate
.max_huge_pages
)
2681 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2684 hugetlb_init_hstates();
2685 gather_bootmem_prealloc();
2688 hugetlb_sysfs_init();
2689 hugetlb_register_all_nodes();
2690 hugetlb_cgroup_file_init();
2693 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2695 num_fault_mutexes
= 1;
2697 hugetlb_fault_mutex_table
=
2698 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2699 BUG_ON(!hugetlb_fault_mutex_table
);
2701 for (i
= 0; i
< num_fault_mutexes
; i
++)
2702 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2705 subsys_initcall(hugetlb_init
);
2707 /* Should be called on processing a hugepagesz=... option */
2708 void __init
hugetlb_bad_size(void)
2710 parsed_valid_hugepagesz
= false;
2713 void __init
hugetlb_add_hstate(unsigned int order
)
2718 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2719 pr_warn("hugepagesz= specified twice, ignoring\n");
2722 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2724 h
= &hstates
[hugetlb_max_hstate
++];
2726 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2727 h
->nr_huge_pages
= 0;
2728 h
->free_huge_pages
= 0;
2729 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2730 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2731 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2732 h
->next_nid_to_alloc
= first_memory_node
;
2733 h
->next_nid_to_free
= first_memory_node
;
2734 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2735 huge_page_size(h
)/1024);
2740 static int __init
hugetlb_nrpages_setup(char *s
)
2743 static unsigned long *last_mhp
;
2745 if (!parsed_valid_hugepagesz
) {
2746 pr_warn("hugepages = %s preceded by "
2747 "an unsupported hugepagesz, ignoring\n", s
);
2748 parsed_valid_hugepagesz
= true;
2752 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2753 * so this hugepages= parameter goes to the "default hstate".
2755 else if (!hugetlb_max_hstate
)
2756 mhp
= &default_hstate_max_huge_pages
;
2758 mhp
= &parsed_hstate
->max_huge_pages
;
2760 if (mhp
== last_mhp
) {
2761 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2765 if (sscanf(s
, "%lu", mhp
) <= 0)
2769 * Global state is always initialized later in hugetlb_init.
2770 * But we need to allocate >= MAX_ORDER hstates here early to still
2771 * use the bootmem allocator.
2773 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2774 hugetlb_hstate_alloc_pages(parsed_hstate
);
2780 __setup("hugepages=", hugetlb_nrpages_setup
);
2782 static int __init
hugetlb_default_setup(char *s
)
2784 default_hstate_size
= memparse(s
, &s
);
2787 __setup("default_hugepagesz=", hugetlb_default_setup
);
2789 static unsigned int cpuset_mems_nr(unsigned int *array
)
2792 unsigned int nr
= 0;
2794 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2800 #ifdef CONFIG_SYSCTL
2801 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2802 struct ctl_table
*table
, int write
,
2803 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2805 struct hstate
*h
= &default_hstate
;
2806 unsigned long tmp
= h
->max_huge_pages
;
2809 if (!hugepages_supported())
2813 table
->maxlen
= sizeof(unsigned long);
2814 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2819 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2820 NUMA_NO_NODE
, tmp
, *length
);
2825 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2826 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2829 return hugetlb_sysctl_handler_common(false, table
, write
,
2830 buffer
, length
, ppos
);
2834 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2835 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2837 return hugetlb_sysctl_handler_common(true, table
, write
,
2838 buffer
, length
, ppos
);
2840 #endif /* CONFIG_NUMA */
2842 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2843 void __user
*buffer
,
2844 size_t *length
, loff_t
*ppos
)
2846 struct hstate
*h
= &default_hstate
;
2850 if (!hugepages_supported())
2853 tmp
= h
->nr_overcommit_huge_pages
;
2855 if (write
&& hstate_is_gigantic(h
))
2859 table
->maxlen
= sizeof(unsigned long);
2860 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2865 spin_lock(&hugetlb_lock
);
2866 h
->nr_overcommit_huge_pages
= tmp
;
2867 spin_unlock(&hugetlb_lock
);
2873 #endif /* CONFIG_SYSCTL */
2875 void hugetlb_report_meminfo(struct seq_file
*m
)
2877 struct hstate
*h
= &default_hstate
;
2878 if (!hugepages_supported())
2881 "HugePages_Total: %5lu\n"
2882 "HugePages_Free: %5lu\n"
2883 "HugePages_Rsvd: %5lu\n"
2884 "HugePages_Surp: %5lu\n"
2885 "Hugepagesize: %8lu kB\n",
2889 h
->surplus_huge_pages
,
2890 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2893 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2895 struct hstate
*h
= &default_hstate
;
2896 if (!hugepages_supported())
2899 "Node %d HugePages_Total: %5u\n"
2900 "Node %d HugePages_Free: %5u\n"
2901 "Node %d HugePages_Surp: %5u\n",
2902 nid
, h
->nr_huge_pages_node
[nid
],
2903 nid
, h
->free_huge_pages_node
[nid
],
2904 nid
, h
->surplus_huge_pages_node
[nid
]);
2907 void hugetlb_show_meminfo(void)
2912 if (!hugepages_supported())
2915 for_each_node_state(nid
, N_MEMORY
)
2917 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2919 h
->nr_huge_pages_node
[nid
],
2920 h
->free_huge_pages_node
[nid
],
2921 h
->surplus_huge_pages_node
[nid
],
2922 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2925 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
2927 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
2928 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
2931 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2932 unsigned long hugetlb_total_pages(void)
2935 unsigned long nr_total_pages
= 0;
2938 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2939 return nr_total_pages
;
2942 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2946 spin_lock(&hugetlb_lock
);
2948 * When cpuset is configured, it breaks the strict hugetlb page
2949 * reservation as the accounting is done on a global variable. Such
2950 * reservation is completely rubbish in the presence of cpuset because
2951 * the reservation is not checked against page availability for the
2952 * current cpuset. Application can still potentially OOM'ed by kernel
2953 * with lack of free htlb page in cpuset that the task is in.
2954 * Attempt to enforce strict accounting with cpuset is almost
2955 * impossible (or too ugly) because cpuset is too fluid that
2956 * task or memory node can be dynamically moved between cpusets.
2958 * The change of semantics for shared hugetlb mapping with cpuset is
2959 * undesirable. However, in order to preserve some of the semantics,
2960 * we fall back to check against current free page availability as
2961 * a best attempt and hopefully to minimize the impact of changing
2962 * semantics that cpuset has.
2965 if (gather_surplus_pages(h
, delta
) < 0)
2968 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2969 return_unused_surplus_pages(h
, delta
);
2976 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2979 spin_unlock(&hugetlb_lock
);
2983 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2985 struct resv_map
*resv
= vma_resv_map(vma
);
2988 * This new VMA should share its siblings reservation map if present.
2989 * The VMA will only ever have a valid reservation map pointer where
2990 * it is being copied for another still existing VMA. As that VMA
2991 * has a reference to the reservation map it cannot disappear until
2992 * after this open call completes. It is therefore safe to take a
2993 * new reference here without additional locking.
2995 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2996 kref_get(&resv
->refs
);
2999 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3001 struct hstate
*h
= hstate_vma(vma
);
3002 struct resv_map
*resv
= vma_resv_map(vma
);
3003 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3004 unsigned long reserve
, start
, end
;
3007 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3010 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3011 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3013 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3015 kref_put(&resv
->refs
, resv_map_release
);
3019 * Decrement reserve counts. The global reserve count may be
3020 * adjusted if the subpool has a minimum size.
3022 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3023 hugetlb_acct_memory(h
, -gbl_reserve
);
3028 * We cannot handle pagefaults against hugetlb pages at all. They cause
3029 * handle_mm_fault() to try to instantiate regular-sized pages in the
3030 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3033 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3039 const struct vm_operations_struct hugetlb_vm_ops
= {
3040 .fault
= hugetlb_vm_op_fault
,
3041 .open
= hugetlb_vm_op_open
,
3042 .close
= hugetlb_vm_op_close
,
3045 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3051 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3052 vma
->vm_page_prot
)));
3054 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3055 vma
->vm_page_prot
));
3057 entry
= pte_mkyoung(entry
);
3058 entry
= pte_mkhuge(entry
);
3059 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3064 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3065 unsigned long address
, pte_t
*ptep
)
3069 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3070 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3071 update_mmu_cache(vma
, address
, ptep
);
3074 static int is_hugetlb_entry_migration(pte_t pte
)
3078 if (huge_pte_none(pte
) || pte_present(pte
))
3080 swp
= pte_to_swp_entry(pte
);
3081 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3087 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3091 if (huge_pte_none(pte
) || pte_present(pte
))
3093 swp
= pte_to_swp_entry(pte
);
3094 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3100 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3101 struct vm_area_struct
*vma
)
3103 pte_t
*src_pte
, *dst_pte
, entry
;
3104 struct page
*ptepage
;
3107 struct hstate
*h
= hstate_vma(vma
);
3108 unsigned long sz
= huge_page_size(h
);
3109 unsigned long mmun_start
; /* For mmu_notifiers */
3110 unsigned long mmun_end
; /* For mmu_notifiers */
3113 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3115 mmun_start
= vma
->vm_start
;
3116 mmun_end
= vma
->vm_end
;
3118 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3120 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3121 spinlock_t
*src_ptl
, *dst_ptl
;
3122 src_pte
= huge_pte_offset(src
, addr
);
3125 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3131 /* If the pagetables are shared don't copy or take references */
3132 if (dst_pte
== src_pte
)
3135 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3136 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3137 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3138 entry
= huge_ptep_get(src_pte
);
3139 if (huge_pte_none(entry
)) { /* skip none entry */
3141 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3142 is_hugetlb_entry_hwpoisoned(entry
))) {
3143 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3145 if (is_write_migration_entry(swp_entry
) && cow
) {
3147 * COW mappings require pages in both
3148 * parent and child to be set to read.
3150 make_migration_entry_read(&swp_entry
);
3151 entry
= swp_entry_to_pte(swp_entry
);
3152 set_huge_pte_at(src
, addr
, src_pte
, entry
);
3154 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3157 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3158 mmu_notifier_invalidate_range(src
, mmun_start
,
3161 entry
= huge_ptep_get(src_pte
);
3162 ptepage
= pte_page(entry
);
3164 page_dup_rmap(ptepage
, true);
3165 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3166 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3168 spin_unlock(src_ptl
);
3169 spin_unlock(dst_ptl
);
3173 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3178 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3179 unsigned long start
, unsigned long end
,
3180 struct page
*ref_page
)
3182 struct mm_struct
*mm
= vma
->vm_mm
;
3183 unsigned long address
;
3188 struct hstate
*h
= hstate_vma(vma
);
3189 unsigned long sz
= huge_page_size(h
);
3190 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3191 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3193 WARN_ON(!is_vm_hugetlb_page(vma
));
3194 BUG_ON(start
& ~huge_page_mask(h
));
3195 BUG_ON(end
& ~huge_page_mask(h
));
3197 tlb_start_vma(tlb
, vma
);
3198 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3200 for (; address
< end
; address
+= sz
) {
3201 ptep
= huge_pte_offset(mm
, address
);
3205 ptl
= huge_pte_lock(h
, mm
, ptep
);
3206 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3211 pte
= huge_ptep_get(ptep
);
3212 if (huge_pte_none(pte
)) {
3218 * Migrating hugepage or HWPoisoned hugepage is already
3219 * unmapped and its refcount is dropped, so just clear pte here.
3221 if (unlikely(!pte_present(pte
))) {
3222 huge_pte_clear(mm
, address
, ptep
);
3227 page
= pte_page(pte
);
3229 * If a reference page is supplied, it is because a specific
3230 * page is being unmapped, not a range. Ensure the page we
3231 * are about to unmap is the actual page of interest.
3234 if (page
!= ref_page
) {
3239 * Mark the VMA as having unmapped its page so that
3240 * future faults in this VMA will fail rather than
3241 * looking like data was lost
3243 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3246 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3247 tlb_remove_tlb_entry(tlb
, ptep
, address
);
3248 if (huge_pte_dirty(pte
))
3249 set_page_dirty(page
);
3251 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3252 page_remove_rmap(page
, true);
3255 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3257 * Bail out after unmapping reference page if supplied
3262 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3263 tlb_end_vma(tlb
, vma
);
3266 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3267 struct vm_area_struct
*vma
, unsigned long start
,
3268 unsigned long end
, struct page
*ref_page
)
3270 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3273 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3274 * test will fail on a vma being torn down, and not grab a page table
3275 * on its way out. We're lucky that the flag has such an appropriate
3276 * name, and can in fact be safely cleared here. We could clear it
3277 * before the __unmap_hugepage_range above, but all that's necessary
3278 * is to clear it before releasing the i_mmap_rwsem. This works
3279 * because in the context this is called, the VMA is about to be
3280 * destroyed and the i_mmap_rwsem is held.
3282 vma
->vm_flags
&= ~VM_MAYSHARE
;
3285 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3286 unsigned long end
, struct page
*ref_page
)
3288 struct mm_struct
*mm
;
3289 struct mmu_gather tlb
;
3293 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3294 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3295 tlb_finish_mmu(&tlb
, start
, end
);
3299 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3300 * mappping it owns the reserve page for. The intention is to unmap the page
3301 * from other VMAs and let the children be SIGKILLed if they are faulting the
3304 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3305 struct page
*page
, unsigned long address
)
3307 struct hstate
*h
= hstate_vma(vma
);
3308 struct vm_area_struct
*iter_vma
;
3309 struct address_space
*mapping
;
3313 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3314 * from page cache lookup which is in HPAGE_SIZE units.
3316 address
= address
& huge_page_mask(h
);
3317 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3319 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
3322 * Take the mapping lock for the duration of the table walk. As
3323 * this mapping should be shared between all the VMAs,
3324 * __unmap_hugepage_range() is called as the lock is already held
3326 i_mmap_lock_write(mapping
);
3327 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3328 /* Do not unmap the current VMA */
3329 if (iter_vma
== vma
)
3333 * Shared VMAs have their own reserves and do not affect
3334 * MAP_PRIVATE accounting but it is possible that a shared
3335 * VMA is using the same page so check and skip such VMAs.
3337 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3341 * Unmap the page from other VMAs without their own reserves.
3342 * They get marked to be SIGKILLed if they fault in these
3343 * areas. This is because a future no-page fault on this VMA
3344 * could insert a zeroed page instead of the data existing
3345 * from the time of fork. This would look like data corruption
3347 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3348 unmap_hugepage_range(iter_vma
, address
,
3349 address
+ huge_page_size(h
), page
);
3351 i_mmap_unlock_write(mapping
);
3355 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3356 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3357 * cannot race with other handlers or page migration.
3358 * Keep the pte_same checks anyway to make transition from the mutex easier.
3360 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3361 unsigned long address
, pte_t
*ptep
, pte_t pte
,
3362 struct page
*pagecache_page
, spinlock_t
*ptl
)
3364 struct hstate
*h
= hstate_vma(vma
);
3365 struct page
*old_page
, *new_page
;
3366 int ret
= 0, outside_reserve
= 0;
3367 unsigned long mmun_start
; /* For mmu_notifiers */
3368 unsigned long mmun_end
; /* For mmu_notifiers */
3370 old_page
= pte_page(pte
);
3373 /* If no-one else is actually using this page, avoid the copy
3374 * and just make the page writable */
3375 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3376 page_move_anon_rmap(old_page
, vma
);
3377 set_huge_ptep_writable(vma
, address
, ptep
);
3382 * If the process that created a MAP_PRIVATE mapping is about to
3383 * perform a COW due to a shared page count, attempt to satisfy
3384 * the allocation without using the existing reserves. The pagecache
3385 * page is used to determine if the reserve at this address was
3386 * consumed or not. If reserves were used, a partial faulted mapping
3387 * at the time of fork() could consume its reserves on COW instead
3388 * of the full address range.
3390 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3391 old_page
!= pagecache_page
)
3392 outside_reserve
= 1;
3397 * Drop page table lock as buddy allocator may be called. It will
3398 * be acquired again before returning to the caller, as expected.
3401 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3403 if (IS_ERR(new_page
)) {
3405 * If a process owning a MAP_PRIVATE mapping fails to COW,
3406 * it is due to references held by a child and an insufficient
3407 * huge page pool. To guarantee the original mappers
3408 * reliability, unmap the page from child processes. The child
3409 * may get SIGKILLed if it later faults.
3411 if (outside_reserve
) {
3413 BUG_ON(huge_pte_none(pte
));
3414 unmap_ref_private(mm
, vma
, old_page
, address
);
3415 BUG_ON(huge_pte_none(pte
));
3417 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3419 pte_same(huge_ptep_get(ptep
), pte
)))
3420 goto retry_avoidcopy
;
3422 * race occurs while re-acquiring page table
3423 * lock, and our job is done.
3428 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3429 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3430 goto out_release_old
;
3434 * When the original hugepage is shared one, it does not have
3435 * anon_vma prepared.
3437 if (unlikely(anon_vma_prepare(vma
))) {
3439 goto out_release_all
;
3442 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3443 pages_per_huge_page(h
));
3444 __SetPageUptodate(new_page
);
3445 set_page_huge_active(new_page
);
3447 mmun_start
= address
& huge_page_mask(h
);
3448 mmun_end
= mmun_start
+ huge_page_size(h
);
3449 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3452 * Retake the page table lock to check for racing updates
3453 * before the page tables are altered
3456 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3457 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3458 ClearPagePrivate(new_page
);
3461 huge_ptep_clear_flush(vma
, address
, ptep
);
3462 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3463 set_huge_pte_at(mm
, address
, ptep
,
3464 make_huge_pte(vma
, new_page
, 1));
3465 page_remove_rmap(old_page
, true);
3466 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3467 /* Make the old page be freed below */
3468 new_page
= old_page
;
3471 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3477 spin_lock(ptl
); /* Caller expects lock to be held */
3481 /* Return the pagecache page at a given address within a VMA */
3482 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3483 struct vm_area_struct
*vma
, unsigned long address
)
3485 struct address_space
*mapping
;
3488 mapping
= vma
->vm_file
->f_mapping
;
3489 idx
= vma_hugecache_offset(h
, vma
, address
);
3491 return find_lock_page(mapping
, idx
);
3495 * Return whether there is a pagecache page to back given address within VMA.
3496 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3498 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3499 struct vm_area_struct
*vma
, unsigned long address
)
3501 struct address_space
*mapping
;
3505 mapping
= vma
->vm_file
->f_mapping
;
3506 idx
= vma_hugecache_offset(h
, vma
, address
);
3508 page
= find_get_page(mapping
, idx
);
3511 return page
!= NULL
;
3514 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3517 struct inode
*inode
= mapping
->host
;
3518 struct hstate
*h
= hstate_inode(inode
);
3519 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3523 ClearPagePrivate(page
);
3525 spin_lock(&inode
->i_lock
);
3526 inode
->i_blocks
+= blocks_per_huge_page(h
);
3527 spin_unlock(&inode
->i_lock
);
3531 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3532 struct address_space
*mapping
, pgoff_t idx
,
3533 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3535 struct hstate
*h
= hstate_vma(vma
);
3536 int ret
= VM_FAULT_SIGBUS
;
3544 * Currently, we are forced to kill the process in the event the
3545 * original mapper has unmapped pages from the child due to a failed
3546 * COW. Warn that such a situation has occurred as it may not be obvious
3548 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3549 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3555 * Use page lock to guard against racing truncation
3556 * before we get page_table_lock.
3559 page
= find_lock_page(mapping
, idx
);
3561 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3564 page
= alloc_huge_page(vma
, address
, 0);
3566 ret
= PTR_ERR(page
);
3570 ret
= VM_FAULT_SIGBUS
;
3573 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3574 __SetPageUptodate(page
);
3575 set_page_huge_active(page
);
3577 if (vma
->vm_flags
& VM_MAYSHARE
) {
3578 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3587 if (unlikely(anon_vma_prepare(vma
))) {
3589 goto backout_unlocked
;
3595 * If memory error occurs between mmap() and fault, some process
3596 * don't have hwpoisoned swap entry for errored virtual address.
3597 * So we need to block hugepage fault by PG_hwpoison bit check.
3599 if (unlikely(PageHWPoison(page
))) {
3600 ret
= VM_FAULT_HWPOISON
|
3601 VM_FAULT_SET_HINDEX(hstate_index(h
));
3602 goto backout_unlocked
;
3607 * If we are going to COW a private mapping later, we examine the
3608 * pending reservations for this page now. This will ensure that
3609 * any allocations necessary to record that reservation occur outside
3612 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3613 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3615 goto backout_unlocked
;
3617 /* Just decrements count, does not deallocate */
3618 vma_end_reservation(h
, vma
, address
);
3621 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3623 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3628 if (!huge_pte_none(huge_ptep_get(ptep
)))
3632 ClearPagePrivate(page
);
3633 hugepage_add_new_anon_rmap(page
, vma
, address
);
3635 page_dup_rmap(page
, true);
3636 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3637 && (vma
->vm_flags
& VM_SHARED
)));
3638 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3640 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3641 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3642 /* Optimization, do the COW without a second fault */
3643 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3660 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3661 struct vm_area_struct
*vma
,
3662 struct address_space
*mapping
,
3663 pgoff_t idx
, unsigned long address
)
3665 unsigned long key
[2];
3668 if (vma
->vm_flags
& VM_SHARED
) {
3669 key
[0] = (unsigned long) mapping
;
3672 key
[0] = (unsigned long) mm
;
3673 key
[1] = address
>> huge_page_shift(h
);
3676 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3678 return hash
& (num_fault_mutexes
- 1);
3682 * For uniprocesor systems we always use a single mutex, so just
3683 * return 0 and avoid the hashing overhead.
3685 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3686 struct vm_area_struct
*vma
,
3687 struct address_space
*mapping
,
3688 pgoff_t idx
, unsigned long address
)
3694 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3695 unsigned long address
, unsigned int flags
)
3702 struct page
*page
= NULL
;
3703 struct page
*pagecache_page
= NULL
;
3704 struct hstate
*h
= hstate_vma(vma
);
3705 struct address_space
*mapping
;
3706 int need_wait_lock
= 0;
3708 address
&= huge_page_mask(h
);
3710 ptep
= huge_pte_offset(mm
, address
);
3712 entry
= huge_ptep_get(ptep
);
3713 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3714 migration_entry_wait_huge(vma
, mm
, ptep
);
3716 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3717 return VM_FAULT_HWPOISON_LARGE
|
3718 VM_FAULT_SET_HINDEX(hstate_index(h
));
3720 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3722 return VM_FAULT_OOM
;
3725 mapping
= vma
->vm_file
->f_mapping
;
3726 idx
= vma_hugecache_offset(h
, vma
, address
);
3729 * Serialize hugepage allocation and instantiation, so that we don't
3730 * get spurious allocation failures if two CPUs race to instantiate
3731 * the same page in the page cache.
3733 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3734 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3736 entry
= huge_ptep_get(ptep
);
3737 if (huge_pte_none(entry
)) {
3738 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3745 * entry could be a migration/hwpoison entry at this point, so this
3746 * check prevents the kernel from going below assuming that we have
3747 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3748 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3751 if (!pte_present(entry
))
3755 * If we are going to COW the mapping later, we examine the pending
3756 * reservations for this page now. This will ensure that any
3757 * allocations necessary to record that reservation occur outside the
3758 * spinlock. For private mappings, we also lookup the pagecache
3759 * page now as it is used to determine if a reservation has been
3762 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3763 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3767 /* Just decrements count, does not deallocate */
3768 vma_end_reservation(h
, vma
, address
);
3770 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3771 pagecache_page
= hugetlbfs_pagecache_page(h
,
3775 ptl
= huge_pte_lock(h
, mm
, ptep
);
3777 /* Check for a racing update before calling hugetlb_cow */
3778 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3782 * hugetlb_cow() requires page locks of pte_page(entry) and
3783 * pagecache_page, so here we need take the former one
3784 * when page != pagecache_page or !pagecache_page.
3786 page
= pte_page(entry
);
3787 if (page
!= pagecache_page
)
3788 if (!trylock_page(page
)) {
3795 if (flags
& FAULT_FLAG_WRITE
) {
3796 if (!huge_pte_write(entry
)) {
3797 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3798 pagecache_page
, ptl
);
3801 entry
= huge_pte_mkdirty(entry
);
3803 entry
= pte_mkyoung(entry
);
3804 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3805 flags
& FAULT_FLAG_WRITE
))
3806 update_mmu_cache(vma
, address
, ptep
);
3808 if (page
!= pagecache_page
)
3814 if (pagecache_page
) {
3815 unlock_page(pagecache_page
);
3816 put_page(pagecache_page
);
3819 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3821 * Generally it's safe to hold refcount during waiting page lock. But
3822 * here we just wait to defer the next page fault to avoid busy loop and
3823 * the page is not used after unlocked before returning from the current
3824 * page fault. So we are safe from accessing freed page, even if we wait
3825 * here without taking refcount.
3828 wait_on_page_locked(page
);
3832 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3833 struct page
**pages
, struct vm_area_struct
**vmas
,
3834 unsigned long *position
, unsigned long *nr_pages
,
3835 long i
, unsigned int flags
)
3837 unsigned long pfn_offset
;
3838 unsigned long vaddr
= *position
;
3839 unsigned long remainder
= *nr_pages
;
3840 struct hstate
*h
= hstate_vma(vma
);
3842 while (vaddr
< vma
->vm_end
&& remainder
) {
3844 spinlock_t
*ptl
= NULL
;
3849 * If we have a pending SIGKILL, don't keep faulting pages and
3850 * potentially allocating memory.
3852 if (unlikely(fatal_signal_pending(current
))) {
3858 * Some archs (sparc64, sh*) have multiple pte_ts to
3859 * each hugepage. We have to make sure we get the
3860 * first, for the page indexing below to work.
3862 * Note that page table lock is not held when pte is null.
3864 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3866 ptl
= huge_pte_lock(h
, mm
, pte
);
3867 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3870 * When coredumping, it suits get_dump_page if we just return
3871 * an error where there's an empty slot with no huge pagecache
3872 * to back it. This way, we avoid allocating a hugepage, and
3873 * the sparse dumpfile avoids allocating disk blocks, but its
3874 * huge holes still show up with zeroes where they need to be.
3876 if (absent
&& (flags
& FOLL_DUMP
) &&
3877 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3885 * We need call hugetlb_fault for both hugepages under migration
3886 * (in which case hugetlb_fault waits for the migration,) and
3887 * hwpoisoned hugepages (in which case we need to prevent the
3888 * caller from accessing to them.) In order to do this, we use
3889 * here is_swap_pte instead of is_hugetlb_entry_migration and
3890 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3891 * both cases, and because we can't follow correct pages
3892 * directly from any kind of swap entries.
3894 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3895 ((flags
& FOLL_WRITE
) &&
3896 !huge_pte_write(huge_ptep_get(pte
)))) {
3901 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3902 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3903 if (!(ret
& VM_FAULT_ERROR
))
3910 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3911 page
= pte_page(huge_ptep_get(pte
));
3914 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3925 if (vaddr
< vma
->vm_end
&& remainder
&&
3926 pfn_offset
< pages_per_huge_page(h
)) {
3928 * We use pfn_offset to avoid touching the pageframes
3929 * of this compound page.
3935 *nr_pages
= remainder
;
3938 return i
? i
: -EFAULT
;
3941 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3942 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3944 struct mm_struct
*mm
= vma
->vm_mm
;
3945 unsigned long start
= address
;
3948 struct hstate
*h
= hstate_vma(vma
);
3949 unsigned long pages
= 0;
3951 BUG_ON(address
>= end
);
3952 flush_cache_range(vma
, address
, end
);
3954 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3955 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
3956 for (; address
< end
; address
+= huge_page_size(h
)) {
3958 ptep
= huge_pte_offset(mm
, address
);
3961 ptl
= huge_pte_lock(h
, mm
, ptep
);
3962 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3967 pte
= huge_ptep_get(ptep
);
3968 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
3972 if (unlikely(is_hugetlb_entry_migration(pte
))) {
3973 swp_entry_t entry
= pte_to_swp_entry(pte
);
3975 if (is_write_migration_entry(entry
)) {
3978 make_migration_entry_read(&entry
);
3979 newpte
= swp_entry_to_pte(entry
);
3980 set_huge_pte_at(mm
, address
, ptep
, newpte
);
3986 if (!huge_pte_none(pte
)) {
3987 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3988 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3989 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3990 set_huge_pte_at(mm
, address
, ptep
, pte
);
3996 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3997 * may have cleared our pud entry and done put_page on the page table:
3998 * once we release i_mmap_rwsem, another task can do the final put_page
3999 * and that page table be reused and filled with junk.
4001 flush_tlb_range(vma
, start
, end
);
4002 mmu_notifier_invalidate_range(mm
, start
, end
);
4003 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4004 mmu_notifier_invalidate_range_end(mm
, start
, end
);
4006 return pages
<< h
->order
;
4009 int hugetlb_reserve_pages(struct inode
*inode
,
4011 struct vm_area_struct
*vma
,
4012 vm_flags_t vm_flags
)
4015 struct hstate
*h
= hstate_inode(inode
);
4016 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4017 struct resv_map
*resv_map
;
4021 * Only apply hugepage reservation if asked. At fault time, an
4022 * attempt will be made for VM_NORESERVE to allocate a page
4023 * without using reserves
4025 if (vm_flags
& VM_NORESERVE
)
4029 * Shared mappings base their reservation on the number of pages that
4030 * are already allocated on behalf of the file. Private mappings need
4031 * to reserve the full area even if read-only as mprotect() may be
4032 * called to make the mapping read-write. Assume !vma is a shm mapping
4034 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4035 resv_map
= inode_resv_map(inode
);
4037 chg
= region_chg(resv_map
, from
, to
);
4040 resv_map
= resv_map_alloc();
4046 set_vma_resv_map(vma
, resv_map
);
4047 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4056 * There must be enough pages in the subpool for the mapping. If
4057 * the subpool has a minimum size, there may be some global
4058 * reservations already in place (gbl_reserve).
4060 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4061 if (gbl_reserve
< 0) {
4067 * Check enough hugepages are available for the reservation.
4068 * Hand the pages back to the subpool if there are not
4070 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4072 /* put back original number of pages, chg */
4073 (void)hugepage_subpool_put_pages(spool
, chg
);
4078 * Account for the reservations made. Shared mappings record regions
4079 * that have reservations as they are shared by multiple VMAs.
4080 * When the last VMA disappears, the region map says how much
4081 * the reservation was and the page cache tells how much of
4082 * the reservation was consumed. Private mappings are per-VMA and
4083 * only the consumed reservations are tracked. When the VMA
4084 * disappears, the original reservation is the VMA size and the
4085 * consumed reservations are stored in the map. Hence, nothing
4086 * else has to be done for private mappings here
4088 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4089 long add
= region_add(resv_map
, from
, to
);
4091 if (unlikely(chg
> add
)) {
4093 * pages in this range were added to the reserve
4094 * map between region_chg and region_add. This
4095 * indicates a race with alloc_huge_page. Adjust
4096 * the subpool and reserve counts modified above
4097 * based on the difference.
4101 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4103 hugetlb_acct_memory(h
, -rsv_adjust
);
4108 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4109 region_abort(resv_map
, from
, to
);
4110 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4111 kref_put(&resv_map
->refs
, resv_map_release
);
4115 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4118 struct hstate
*h
= hstate_inode(inode
);
4119 struct resv_map
*resv_map
= inode_resv_map(inode
);
4121 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4125 chg
= region_del(resv_map
, start
, end
);
4127 * region_del() can fail in the rare case where a region
4128 * must be split and another region descriptor can not be
4129 * allocated. If end == LONG_MAX, it will not fail.
4135 spin_lock(&inode
->i_lock
);
4136 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4137 spin_unlock(&inode
->i_lock
);
4140 * If the subpool has a minimum size, the number of global
4141 * reservations to be released may be adjusted.
4143 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4144 hugetlb_acct_memory(h
, -gbl_reserve
);
4149 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4150 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4151 struct vm_area_struct
*vma
,
4152 unsigned long addr
, pgoff_t idx
)
4154 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4156 unsigned long sbase
= saddr
& PUD_MASK
;
4157 unsigned long s_end
= sbase
+ PUD_SIZE
;
4159 /* Allow segments to share if only one is marked locked */
4160 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4161 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4164 * match the virtual addresses, permission and the alignment of the
4167 if (pmd_index(addr
) != pmd_index(saddr
) ||
4168 vm_flags
!= svm_flags
||
4169 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4175 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4177 unsigned long base
= addr
& PUD_MASK
;
4178 unsigned long end
= base
+ PUD_SIZE
;
4181 * check on proper vm_flags and page table alignment
4183 if (vma
->vm_flags
& VM_MAYSHARE
&&
4184 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4190 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4191 * and returns the corresponding pte. While this is not necessary for the
4192 * !shared pmd case because we can allocate the pmd later as well, it makes the
4193 * code much cleaner. pmd allocation is essential for the shared case because
4194 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4195 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4196 * bad pmd for sharing.
4198 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4200 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4201 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4202 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4204 struct vm_area_struct
*svma
;
4205 unsigned long saddr
;
4210 if (!vma_shareable(vma
, addr
))
4211 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4213 i_mmap_lock_write(mapping
);
4214 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4218 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4220 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
4222 get_page(virt_to_page(spte
));
4231 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
4233 if (pud_none(*pud
)) {
4234 pud_populate(mm
, pud
,
4235 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4238 put_page(virt_to_page(spte
));
4242 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4243 i_mmap_unlock_write(mapping
);
4248 * unmap huge page backed by shared pte.
4250 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4251 * indicated by page_count > 1, unmap is achieved by clearing pud and
4252 * decrementing the ref count. If count == 1, the pte page is not shared.
4254 * called with page table lock held.
4256 * returns: 1 successfully unmapped a shared pte page
4257 * 0 the underlying pte page is not shared, or it is the last user
4259 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4261 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4262 pud_t
*pud
= pud_offset(pgd
, *addr
);
4264 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4265 if (page_count(virt_to_page(ptep
)) == 1)
4269 put_page(virt_to_page(ptep
));
4271 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4274 #define want_pmd_share() (1)
4275 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4276 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4281 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4285 #define want_pmd_share() (0)
4286 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4288 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4289 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4290 unsigned long addr
, unsigned long sz
)
4296 pgd
= pgd_offset(mm
, addr
);
4297 pud
= pud_alloc(mm
, pgd
, addr
);
4299 if (sz
== PUD_SIZE
) {
4302 BUG_ON(sz
!= PMD_SIZE
);
4303 if (want_pmd_share() && pud_none(*pud
))
4304 pte
= huge_pmd_share(mm
, addr
, pud
);
4306 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4309 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
4314 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
4320 pgd
= pgd_offset(mm
, addr
);
4321 if (pgd_present(*pgd
)) {
4322 pud
= pud_offset(pgd
, addr
);
4323 if (pud_present(*pud
)) {
4325 return (pte_t
*)pud
;
4326 pmd
= pmd_offset(pud
, addr
);
4329 return (pte_t
*) pmd
;
4332 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4335 * These functions are overwritable if your architecture needs its own
4338 struct page
* __weak
4339 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4342 return ERR_PTR(-EINVAL
);
4345 struct page
* __weak
4346 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4347 pmd_t
*pmd
, int flags
)
4349 struct page
*page
= NULL
;
4352 ptl
= pmd_lockptr(mm
, pmd
);
4355 * make sure that the address range covered by this pmd is not
4356 * unmapped from other threads.
4358 if (!pmd_huge(*pmd
))
4360 if (pmd_present(*pmd
)) {
4361 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4362 if (flags
& FOLL_GET
)
4365 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
4367 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4371 * hwpoisoned entry is treated as no_page_table in
4372 * follow_page_mask().
4380 struct page
* __weak
4381 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4382 pud_t
*pud
, int flags
)
4384 if (flags
& FOLL_GET
)
4387 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4390 #ifdef CONFIG_MEMORY_FAILURE
4393 * This function is called from memory failure code.
4394 * Assume the caller holds page lock of the head page.
4396 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4398 struct hstate
*h
= page_hstate(hpage
);
4399 int nid
= page_to_nid(hpage
);
4402 spin_lock(&hugetlb_lock
);
4404 * Just checking !page_huge_active is not enough, because that could be
4405 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4407 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4409 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4410 * but dangling hpage->lru can trigger list-debug warnings
4411 * (this happens when we call unpoison_memory() on it),
4412 * so let it point to itself with list_del_init().
4414 list_del_init(&hpage
->lru
);
4415 set_page_refcounted(hpage
);
4416 h
->free_huge_pages
--;
4417 h
->free_huge_pages_node
[nid
]--;
4420 spin_unlock(&hugetlb_lock
);
4425 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4429 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4430 spin_lock(&hugetlb_lock
);
4431 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4435 clear_page_huge_active(page
);
4436 list_move_tail(&page
->lru
, list
);
4438 spin_unlock(&hugetlb_lock
);
4442 void putback_active_hugepage(struct page
*page
)
4444 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4445 spin_lock(&hugetlb_lock
);
4446 set_page_huge_active(page
);
4447 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4448 spin_unlock(&hugetlb_lock
);