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
);
632 * Return the size of the pages allocated when backing a VMA. In the majority
633 * cases this will be same size as used by the page table entries.
635 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
637 struct hstate
*hstate
;
639 if (!is_vm_hugetlb_page(vma
))
642 hstate
= hstate_vma(vma
);
644 return 1UL << huge_page_shift(hstate
);
646 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
649 * Return the page size being used by the MMU to back a VMA. In the majority
650 * of cases, the page size used by the kernel matches the MMU size. On
651 * architectures where it differs, an architecture-specific version of this
652 * function is required.
654 #ifndef vma_mmu_pagesize
655 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
657 return vma_kernel_pagesize(vma
);
662 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
663 * bits of the reservation map pointer, which are always clear due to
666 #define HPAGE_RESV_OWNER (1UL << 0)
667 #define HPAGE_RESV_UNMAPPED (1UL << 1)
668 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
671 * These helpers are used to track how many pages are reserved for
672 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
673 * is guaranteed to have their future faults succeed.
675 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
676 * the reserve counters are updated with the hugetlb_lock held. It is safe
677 * to reset the VMA at fork() time as it is not in use yet and there is no
678 * chance of the global counters getting corrupted as a result of the values.
680 * The private mapping reservation is represented in a subtly different
681 * manner to a shared mapping. A shared mapping has a region map associated
682 * with the underlying file, this region map represents the backing file
683 * pages which have ever had a reservation assigned which this persists even
684 * after the page is instantiated. A private mapping has a region map
685 * associated with the original mmap which is attached to all VMAs which
686 * reference it, this region map represents those offsets which have consumed
687 * reservation ie. where pages have been instantiated.
689 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
691 return (unsigned long)vma
->vm_private_data
;
694 static void set_vma_private_data(struct vm_area_struct
*vma
,
697 vma
->vm_private_data
= (void *)value
;
700 struct resv_map
*resv_map_alloc(void)
702 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
703 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
705 if (!resv_map
|| !rg
) {
711 kref_init(&resv_map
->refs
);
712 spin_lock_init(&resv_map
->lock
);
713 INIT_LIST_HEAD(&resv_map
->regions
);
715 resv_map
->adds_in_progress
= 0;
717 INIT_LIST_HEAD(&resv_map
->region_cache
);
718 list_add(&rg
->link
, &resv_map
->region_cache
);
719 resv_map
->region_cache_count
= 1;
724 void resv_map_release(struct kref
*ref
)
726 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
727 struct list_head
*head
= &resv_map
->region_cache
;
728 struct file_region
*rg
, *trg
;
730 /* Clear out any active regions before we release the map. */
731 region_del(resv_map
, 0, LONG_MAX
);
733 /* ... and any entries left in the cache */
734 list_for_each_entry_safe(rg
, trg
, head
, link
) {
739 VM_BUG_ON(resv_map
->adds_in_progress
);
744 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
746 return inode
->i_mapping
->private_data
;
749 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
751 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
752 if (vma
->vm_flags
& VM_MAYSHARE
) {
753 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
754 struct inode
*inode
= mapping
->host
;
756 return inode_resv_map(inode
);
759 return (struct resv_map
*)(get_vma_private_data(vma
) &
764 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
766 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
767 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
769 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
770 HPAGE_RESV_MASK
) | (unsigned long)map
);
773 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
775 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
776 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
778 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
781 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
783 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
785 return (get_vma_private_data(vma
) & flag
) != 0;
788 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
789 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
791 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
792 if (!(vma
->vm_flags
& VM_MAYSHARE
))
793 vma
->vm_private_data
= (void *)0;
796 /* Returns true if the VMA has associated reserve pages */
797 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
799 if (vma
->vm_flags
& VM_NORESERVE
) {
801 * This address is already reserved by other process(chg == 0),
802 * so, we should decrement reserved count. Without decrementing,
803 * reserve count remains after releasing inode, because this
804 * allocated page will go into page cache and is regarded as
805 * coming from reserved pool in releasing step. Currently, we
806 * don't have any other solution to deal with this situation
807 * properly, so add work-around here.
809 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
815 /* Shared mappings always use reserves */
816 if (vma
->vm_flags
& VM_MAYSHARE
) {
818 * We know VM_NORESERVE is not set. Therefore, there SHOULD
819 * be a region map for all pages. The only situation where
820 * there is no region map is if a hole was punched via
821 * fallocate. In this case, there really are no reverves to
822 * use. This situation is indicated if chg != 0.
831 * Only the process that called mmap() has reserves for
834 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
840 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
842 int nid
= page_to_nid(page
);
843 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
844 h
->free_huge_pages
++;
845 h
->free_huge_pages_node
[nid
]++;
848 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
852 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
853 if (!is_migrate_isolate_page(page
))
856 * if 'non-isolated free hugepage' not found on the list,
857 * the allocation fails.
859 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
861 list_move(&page
->lru
, &h
->hugepage_activelist
);
862 set_page_refcounted(page
);
863 h
->free_huge_pages
--;
864 h
->free_huge_pages_node
[nid
]--;
868 /* Movability of hugepages depends on migration support. */
869 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
871 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
872 return GFP_HIGHUSER_MOVABLE
;
877 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
878 struct vm_area_struct
*vma
,
879 unsigned long address
, int avoid_reserve
,
882 struct page
*page
= NULL
;
883 struct mempolicy
*mpol
;
884 nodemask_t
*nodemask
;
885 struct zonelist
*zonelist
;
888 unsigned int cpuset_mems_cookie
;
891 * A child process with MAP_PRIVATE mappings created by their parent
892 * have no page reserves. This check ensures that reservations are
893 * not "stolen". The child may still get SIGKILLed
895 if (!vma_has_reserves(vma
, chg
) &&
896 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
899 /* If reserves cannot be used, ensure enough pages are in the pool */
900 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
904 cpuset_mems_cookie
= read_mems_allowed_begin();
905 zonelist
= huge_zonelist(vma
, address
,
906 htlb_alloc_mask(h
), &mpol
, &nodemask
);
908 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
909 MAX_NR_ZONES
- 1, nodemask
) {
910 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
911 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
915 if (!vma_has_reserves(vma
, chg
))
918 SetPagePrivate(page
);
919 h
->resv_huge_pages
--;
926 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
935 * common helper functions for hstate_next_node_to_{alloc|free}.
936 * We may have allocated or freed a huge page based on a different
937 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
938 * be outside of *nodes_allowed. Ensure that we use an allowed
939 * node for alloc or free.
941 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
943 nid
= next_node_in(nid
, *nodes_allowed
);
944 VM_BUG_ON(nid
>= MAX_NUMNODES
);
949 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
951 if (!node_isset(nid
, *nodes_allowed
))
952 nid
= next_node_allowed(nid
, nodes_allowed
);
957 * returns the previously saved node ["this node"] from which to
958 * allocate a persistent huge page for the pool and advance the
959 * next node from which to allocate, handling wrap at end of node
962 static int hstate_next_node_to_alloc(struct hstate
*h
,
963 nodemask_t
*nodes_allowed
)
967 VM_BUG_ON(!nodes_allowed
);
969 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
970 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
976 * helper for free_pool_huge_page() - return the previously saved
977 * node ["this node"] from which to free a huge page. Advance the
978 * next node id whether or not we find a free huge page to free so
979 * that the next attempt to free addresses the next node.
981 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
985 VM_BUG_ON(!nodes_allowed
);
987 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
988 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
993 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
994 for (nr_nodes = nodes_weight(*mask); \
996 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
999 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1000 for (nr_nodes = nodes_weight(*mask); \
1002 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1005 #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA))
1006 static void destroy_compound_gigantic_page(struct page
*page
,
1010 int nr_pages
= 1 << order
;
1011 struct page
*p
= page
+ 1;
1013 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1014 clear_compound_head(p
);
1015 set_page_refcounted(p
);
1018 set_compound_order(page
, 0);
1019 __ClearPageHead(page
);
1022 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1024 free_contig_range(page_to_pfn(page
), 1 << order
);
1027 static int __alloc_gigantic_page(unsigned long start_pfn
,
1028 unsigned long nr_pages
)
1030 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1031 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
1034 static bool pfn_range_valid_gigantic(unsigned long start_pfn
,
1035 unsigned long nr_pages
)
1037 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1040 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1044 page
= pfn_to_page(i
);
1046 if (PageReserved(page
))
1049 if (page_count(page
) > 0)
1059 static bool zone_spans_last_pfn(const struct zone
*zone
,
1060 unsigned long start_pfn
, unsigned long nr_pages
)
1062 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1063 return zone_spans_pfn(zone
, last_pfn
);
1066 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1068 unsigned long nr_pages
= 1 << order
;
1069 unsigned long ret
, pfn
, flags
;
1072 z
= NODE_DATA(nid
)->node_zones
;
1073 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1074 spin_lock_irqsave(&z
->lock
, flags
);
1076 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1077 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1078 if (pfn_range_valid_gigantic(pfn
, nr_pages
)) {
1080 * We release the zone lock here because
1081 * alloc_contig_range() will also lock the zone
1082 * at some point. If there's an allocation
1083 * spinning on this lock, it may win the race
1084 * and cause alloc_contig_range() to fail...
1086 spin_unlock_irqrestore(&z
->lock
, flags
);
1087 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1089 return pfn_to_page(pfn
);
1090 spin_lock_irqsave(&z
->lock
, flags
);
1095 spin_unlock_irqrestore(&z
->lock
, flags
);
1101 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1102 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1104 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1108 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1110 prep_compound_gigantic_page(page
, huge_page_order(h
));
1111 prep_new_huge_page(h
, page
, nid
);
1117 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1118 nodemask_t
*nodes_allowed
)
1120 struct page
*page
= NULL
;
1123 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1124 page
= alloc_fresh_gigantic_page_node(h
, node
);
1132 static inline bool gigantic_page_supported(void) { return true; }
1134 static inline bool gigantic_page_supported(void) { return false; }
1135 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1136 static inline void destroy_compound_gigantic_page(struct page
*page
,
1137 unsigned int order
) { }
1138 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1139 nodemask_t
*nodes_allowed
) { return 0; }
1142 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1146 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1150 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1151 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1152 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1153 1 << PG_referenced
| 1 << PG_dirty
|
1154 1 << PG_active
| 1 << PG_private
|
1157 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1158 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1159 set_page_refcounted(page
);
1160 if (hstate_is_gigantic(h
)) {
1161 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1162 free_gigantic_page(page
, huge_page_order(h
));
1164 __free_pages(page
, huge_page_order(h
));
1168 struct hstate
*size_to_hstate(unsigned long size
)
1172 for_each_hstate(h
) {
1173 if (huge_page_size(h
) == size
)
1180 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1181 * to hstate->hugepage_activelist.)
1183 * This function can be called for tail pages, but never returns true for them.
1185 bool page_huge_active(struct page
*page
)
1187 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1188 return PageHead(page
) && PagePrivate(&page
[1]);
1191 /* never called for tail page */
1192 static void set_page_huge_active(struct page
*page
)
1194 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1195 SetPagePrivate(&page
[1]);
1198 static void clear_page_huge_active(struct page
*page
)
1200 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1201 ClearPagePrivate(&page
[1]);
1204 void free_huge_page(struct page
*page
)
1207 * Can't pass hstate in here because it is called from the
1208 * compound page destructor.
1210 struct hstate
*h
= page_hstate(page
);
1211 int nid
= page_to_nid(page
);
1212 struct hugepage_subpool
*spool
=
1213 (struct hugepage_subpool
*)page_private(page
);
1214 bool restore_reserve
;
1216 set_page_private(page
, 0);
1217 page
->mapping
= NULL
;
1218 VM_BUG_ON_PAGE(page_count(page
), page
);
1219 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1220 restore_reserve
= PagePrivate(page
);
1221 ClearPagePrivate(page
);
1224 * A return code of zero implies that the subpool will be under its
1225 * minimum size if the reservation is not restored after page is free.
1226 * Therefore, force restore_reserve operation.
1228 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1229 restore_reserve
= true;
1231 spin_lock(&hugetlb_lock
);
1232 clear_page_huge_active(page
);
1233 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1234 pages_per_huge_page(h
), page
);
1235 if (restore_reserve
)
1236 h
->resv_huge_pages
++;
1238 if (h
->surplus_huge_pages_node
[nid
]) {
1239 /* remove the page from active list */
1240 list_del(&page
->lru
);
1241 update_and_free_page(h
, page
);
1242 h
->surplus_huge_pages
--;
1243 h
->surplus_huge_pages_node
[nid
]--;
1245 arch_clear_hugepage_flags(page
);
1246 enqueue_huge_page(h
, page
);
1248 spin_unlock(&hugetlb_lock
);
1251 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1253 INIT_LIST_HEAD(&page
->lru
);
1254 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1255 spin_lock(&hugetlb_lock
);
1256 set_hugetlb_cgroup(page
, NULL
);
1258 h
->nr_huge_pages_node
[nid
]++;
1259 spin_unlock(&hugetlb_lock
);
1260 put_page(page
); /* free it into the hugepage allocator */
1263 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1266 int nr_pages
= 1 << order
;
1267 struct page
*p
= page
+ 1;
1269 /* we rely on prep_new_huge_page to set the destructor */
1270 set_compound_order(page
, order
);
1271 __ClearPageReserved(page
);
1272 __SetPageHead(page
);
1273 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1275 * For gigantic hugepages allocated through bootmem at
1276 * boot, it's safer to be consistent with the not-gigantic
1277 * hugepages and clear the PG_reserved bit from all tail pages
1278 * too. Otherwse drivers using get_user_pages() to access tail
1279 * pages may get the reference counting wrong if they see
1280 * PG_reserved set on a tail page (despite the head page not
1281 * having PG_reserved set). Enforcing this consistency between
1282 * head and tail pages allows drivers to optimize away a check
1283 * on the head page when they need know if put_page() is needed
1284 * after get_user_pages().
1286 __ClearPageReserved(p
);
1287 set_page_count(p
, 0);
1288 set_compound_head(p
, page
);
1290 atomic_set(compound_mapcount_ptr(page
), -1);
1294 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1295 * transparent huge pages. See the PageTransHuge() documentation for more
1298 int PageHuge(struct page
*page
)
1300 if (!PageCompound(page
))
1303 page
= compound_head(page
);
1304 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1306 EXPORT_SYMBOL_GPL(PageHuge
);
1309 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1310 * normal or transparent huge pages.
1312 int PageHeadHuge(struct page
*page_head
)
1314 if (!PageHead(page_head
))
1317 return get_compound_page_dtor(page_head
) == free_huge_page
;
1320 pgoff_t
__basepage_index(struct page
*page
)
1322 struct page
*page_head
= compound_head(page
);
1323 pgoff_t index
= page_index(page_head
);
1324 unsigned long compound_idx
;
1326 if (!PageHuge(page_head
))
1327 return page_index(page
);
1329 if (compound_order(page_head
) >= MAX_ORDER
)
1330 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1332 compound_idx
= page
- page_head
;
1334 return (index
<< compound_order(page_head
)) + compound_idx
;
1337 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1341 page
= __alloc_pages_node(nid
,
1342 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1343 __GFP_REPEAT
|__GFP_NOWARN
,
1344 huge_page_order(h
));
1346 prep_new_huge_page(h
, page
, nid
);
1352 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1358 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1359 page
= alloc_fresh_huge_page_node(h
, node
);
1367 count_vm_event(HTLB_BUDDY_PGALLOC
);
1369 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1375 * Free huge page from pool from next node to free.
1376 * Attempt to keep persistent huge pages more or less
1377 * balanced over allowed nodes.
1378 * Called with hugetlb_lock locked.
1380 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1386 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1388 * If we're returning unused surplus pages, only examine
1389 * nodes with surplus pages.
1391 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1392 !list_empty(&h
->hugepage_freelists
[node
])) {
1394 list_entry(h
->hugepage_freelists
[node
].next
,
1396 list_del(&page
->lru
);
1397 h
->free_huge_pages
--;
1398 h
->free_huge_pages_node
[node
]--;
1400 h
->surplus_huge_pages
--;
1401 h
->surplus_huge_pages_node
[node
]--;
1403 update_and_free_page(h
, page
);
1413 * Dissolve a given free hugepage into free buddy pages. This function does
1414 * nothing for in-use (including surplus) hugepages.
1416 static void dissolve_free_huge_page(struct page
*page
)
1418 spin_lock(&hugetlb_lock
);
1419 if (PageHuge(page
) && !page_count(page
)) {
1420 struct hstate
*h
= page_hstate(page
);
1421 int nid
= page_to_nid(page
);
1422 list_del(&page
->lru
);
1423 h
->free_huge_pages
--;
1424 h
->free_huge_pages_node
[nid
]--;
1425 update_and_free_page(h
, page
);
1427 spin_unlock(&hugetlb_lock
);
1431 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1432 * make specified memory blocks removable from the system.
1433 * Note that start_pfn should aligned with (minimum) hugepage size.
1435 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1439 if (!hugepages_supported())
1442 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << minimum_order
));
1443 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
)
1444 dissolve_free_huge_page(pfn_to_page(pfn
));
1448 * There are 3 ways this can get called:
1449 * 1. With vma+addr: we use the VMA's memory policy
1450 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1451 * page from any node, and let the buddy allocator itself figure
1453 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1454 * strictly from 'nid'
1456 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1457 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1459 int order
= huge_page_order(h
);
1460 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1461 unsigned int cpuset_mems_cookie
;
1464 * We need a VMA to get a memory policy. If we do not
1465 * have one, we use the 'nid' argument.
1467 * The mempolicy stuff below has some non-inlined bits
1468 * and calls ->vm_ops. That makes it hard to optimize at
1469 * compile-time, even when NUMA is off and it does
1470 * nothing. This helps the compiler optimize it out.
1472 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1474 * If a specific node is requested, make sure to
1475 * get memory from there, but only when a node
1476 * is explicitly specified.
1478 if (nid
!= NUMA_NO_NODE
)
1479 gfp
|= __GFP_THISNODE
;
1481 * Make sure to call something that can handle
1484 return alloc_pages_node(nid
, gfp
, order
);
1488 * OK, so we have a VMA. Fetch the mempolicy and try to
1489 * allocate a huge page with it. We will only reach this
1490 * when CONFIG_NUMA=y.
1494 struct mempolicy
*mpol
;
1495 struct zonelist
*zl
;
1496 nodemask_t
*nodemask
;
1498 cpuset_mems_cookie
= read_mems_allowed_begin();
1499 zl
= huge_zonelist(vma
, addr
, gfp
, &mpol
, &nodemask
);
1500 mpol_cond_put(mpol
);
1501 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nodemask
);
1504 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1510 * There are two ways to allocate a huge page:
1511 * 1. When you have a VMA and an address (like a fault)
1512 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1514 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1515 * this case which signifies that the allocation should be done with
1516 * respect for the VMA's memory policy.
1518 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1519 * implies that memory policies will not be taken in to account.
1521 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1522 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1527 if (hstate_is_gigantic(h
))
1531 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1532 * This makes sure the caller is picking _one_ of the modes with which
1533 * we can call this function, not both.
1535 if (vma
|| (addr
!= -1)) {
1536 VM_WARN_ON_ONCE(addr
== -1);
1537 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1540 * Assume we will successfully allocate the surplus page to
1541 * prevent racing processes from causing the surplus to exceed
1544 * This however introduces a different race, where a process B
1545 * tries to grow the static hugepage pool while alloc_pages() is
1546 * called by process A. B will only examine the per-node
1547 * counters in determining if surplus huge pages can be
1548 * converted to normal huge pages in adjust_pool_surplus(). A
1549 * won't be able to increment the per-node counter, until the
1550 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1551 * no more huge pages can be converted from surplus to normal
1552 * state (and doesn't try to convert again). Thus, we have a
1553 * case where a surplus huge page exists, the pool is grown, and
1554 * the surplus huge page still exists after, even though it
1555 * should just have been converted to a normal huge page. This
1556 * does not leak memory, though, as the hugepage will be freed
1557 * once it is out of use. It also does not allow the counters to
1558 * go out of whack in adjust_pool_surplus() as we don't modify
1559 * the node values until we've gotten the hugepage and only the
1560 * per-node value is checked there.
1562 spin_lock(&hugetlb_lock
);
1563 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1564 spin_unlock(&hugetlb_lock
);
1568 h
->surplus_huge_pages
++;
1570 spin_unlock(&hugetlb_lock
);
1572 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1574 spin_lock(&hugetlb_lock
);
1576 INIT_LIST_HEAD(&page
->lru
);
1577 r_nid
= page_to_nid(page
);
1578 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1579 set_hugetlb_cgroup(page
, NULL
);
1581 * We incremented the global counters already
1583 h
->nr_huge_pages_node
[r_nid
]++;
1584 h
->surplus_huge_pages_node
[r_nid
]++;
1585 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1588 h
->surplus_huge_pages
--;
1589 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1591 spin_unlock(&hugetlb_lock
);
1597 * Allocate a huge page from 'nid'. Note, 'nid' may be
1598 * NUMA_NO_NODE, which means that it may be allocated
1602 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1604 unsigned long addr
= -1;
1606 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1610 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1613 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1614 struct vm_area_struct
*vma
, unsigned long addr
)
1616 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1620 * This allocation function is useful in the context where vma is irrelevant.
1621 * E.g. soft-offlining uses this function because it only cares physical
1622 * address of error page.
1624 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1626 struct page
*page
= NULL
;
1628 spin_lock(&hugetlb_lock
);
1629 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1630 page
= dequeue_huge_page_node(h
, nid
);
1631 spin_unlock(&hugetlb_lock
);
1634 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1640 * Increase the hugetlb pool such that it can accommodate a reservation
1643 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1645 struct list_head surplus_list
;
1646 struct page
*page
, *tmp
;
1648 int needed
, allocated
;
1649 bool alloc_ok
= true;
1651 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1653 h
->resv_huge_pages
+= delta
;
1658 INIT_LIST_HEAD(&surplus_list
);
1662 spin_unlock(&hugetlb_lock
);
1663 for (i
= 0; i
< needed
; i
++) {
1664 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1669 list_add(&page
->lru
, &surplus_list
);
1674 * After retaking hugetlb_lock, we need to recalculate 'needed'
1675 * because either resv_huge_pages or free_huge_pages may have changed.
1677 spin_lock(&hugetlb_lock
);
1678 needed
= (h
->resv_huge_pages
+ delta
) -
1679 (h
->free_huge_pages
+ allocated
);
1684 * We were not able to allocate enough pages to
1685 * satisfy the entire reservation so we free what
1686 * we've allocated so far.
1691 * The surplus_list now contains _at_least_ the number of extra pages
1692 * needed to accommodate the reservation. Add the appropriate number
1693 * of pages to the hugetlb pool and free the extras back to the buddy
1694 * allocator. Commit the entire reservation here to prevent another
1695 * process from stealing the pages as they are added to the pool but
1696 * before they are reserved.
1698 needed
+= allocated
;
1699 h
->resv_huge_pages
+= delta
;
1702 /* Free the needed pages to the hugetlb pool */
1703 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1707 * This page is now managed by the hugetlb allocator and has
1708 * no users -- drop the buddy allocator's reference.
1710 put_page_testzero(page
);
1711 VM_BUG_ON_PAGE(page_count(page
), page
);
1712 enqueue_huge_page(h
, page
);
1715 spin_unlock(&hugetlb_lock
);
1717 /* Free unnecessary surplus pages to the buddy allocator */
1718 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1720 spin_lock(&hugetlb_lock
);
1726 * When releasing a hugetlb pool reservation, any surplus pages that were
1727 * allocated to satisfy the reservation must be explicitly freed if they were
1729 * Called with hugetlb_lock held.
1731 static void return_unused_surplus_pages(struct hstate
*h
,
1732 unsigned long unused_resv_pages
)
1734 unsigned long nr_pages
;
1736 /* Uncommit the reservation */
1737 h
->resv_huge_pages
-= unused_resv_pages
;
1739 /* Cannot return gigantic pages currently */
1740 if (hstate_is_gigantic(h
))
1743 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1746 * We want to release as many surplus pages as possible, spread
1747 * evenly across all nodes with memory. Iterate across these nodes
1748 * until we can no longer free unreserved surplus pages. This occurs
1749 * when the nodes with surplus pages have no free pages.
1750 * free_pool_huge_page() will balance the the freed pages across the
1751 * on-line nodes with memory and will handle the hstate accounting.
1753 while (nr_pages
--) {
1754 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1756 cond_resched_lock(&hugetlb_lock
);
1762 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1763 * are used by the huge page allocation routines to manage reservations.
1765 * vma_needs_reservation is called to determine if the huge page at addr
1766 * within the vma has an associated reservation. If a reservation is
1767 * needed, the value 1 is returned. The caller is then responsible for
1768 * managing the global reservation and subpool usage counts. After
1769 * the huge page has been allocated, vma_commit_reservation is called
1770 * to add the page to the reservation map. If the page allocation fails,
1771 * the reservation must be ended instead of committed. vma_end_reservation
1772 * is called in such cases.
1774 * In the normal case, vma_commit_reservation returns the same value
1775 * as the preceding vma_needs_reservation call. The only time this
1776 * is not the case is if a reserve map was changed between calls. It
1777 * is the responsibility of the caller to notice the difference and
1778 * take appropriate action.
1780 enum vma_resv_mode
{
1785 static long __vma_reservation_common(struct hstate
*h
,
1786 struct vm_area_struct
*vma
, unsigned long addr
,
1787 enum vma_resv_mode mode
)
1789 struct resv_map
*resv
;
1793 resv
= vma_resv_map(vma
);
1797 idx
= vma_hugecache_offset(h
, vma
, addr
);
1799 case VMA_NEEDS_RESV
:
1800 ret
= region_chg(resv
, idx
, idx
+ 1);
1802 case VMA_COMMIT_RESV
:
1803 ret
= region_add(resv
, idx
, idx
+ 1);
1806 region_abort(resv
, idx
, idx
+ 1);
1813 if (vma
->vm_flags
& VM_MAYSHARE
)
1816 return ret
< 0 ? ret
: 0;
1819 static long vma_needs_reservation(struct hstate
*h
,
1820 struct vm_area_struct
*vma
, unsigned long addr
)
1822 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1825 static long vma_commit_reservation(struct hstate
*h
,
1826 struct vm_area_struct
*vma
, unsigned long addr
)
1828 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1831 static void vma_end_reservation(struct hstate
*h
,
1832 struct vm_area_struct
*vma
, unsigned long addr
)
1834 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1837 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1838 unsigned long addr
, int avoid_reserve
)
1840 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1841 struct hstate
*h
= hstate_vma(vma
);
1843 long map_chg
, map_commit
;
1846 struct hugetlb_cgroup
*h_cg
;
1848 idx
= hstate_index(h
);
1850 * Examine the region/reserve map to determine if the process
1851 * has a reservation for the page to be allocated. A return
1852 * code of zero indicates a reservation exists (no change).
1854 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
1856 return ERR_PTR(-ENOMEM
);
1859 * Processes that did not create the mapping will have no
1860 * reserves as indicated by the region/reserve map. Check
1861 * that the allocation will not exceed the subpool limit.
1862 * Allocations for MAP_NORESERVE mappings also need to be
1863 * checked against any subpool limit.
1865 if (map_chg
|| avoid_reserve
) {
1866 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
1868 vma_end_reservation(h
, vma
, addr
);
1869 return ERR_PTR(-ENOSPC
);
1873 * Even though there was no reservation in the region/reserve
1874 * map, there could be reservations associated with the
1875 * subpool that can be used. This would be indicated if the
1876 * return value of hugepage_subpool_get_pages() is zero.
1877 * However, if avoid_reserve is specified we still avoid even
1878 * the subpool reservations.
1884 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1886 goto out_subpool_put
;
1888 spin_lock(&hugetlb_lock
);
1890 * glb_chg is passed to indicate whether or not a page must be taken
1891 * from the global free pool (global change). gbl_chg == 0 indicates
1892 * a reservation exists for the allocation.
1894 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
1896 spin_unlock(&hugetlb_lock
);
1897 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
1899 goto out_uncharge_cgroup
;
1900 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
1901 SetPagePrivate(page
);
1902 h
->resv_huge_pages
--;
1904 spin_lock(&hugetlb_lock
);
1905 list_move(&page
->lru
, &h
->hugepage_activelist
);
1908 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1909 spin_unlock(&hugetlb_lock
);
1911 set_page_private(page
, (unsigned long)spool
);
1913 map_commit
= vma_commit_reservation(h
, vma
, addr
);
1914 if (unlikely(map_chg
> map_commit
)) {
1916 * The page was added to the reservation map between
1917 * vma_needs_reservation and vma_commit_reservation.
1918 * This indicates a race with hugetlb_reserve_pages.
1919 * Adjust for the subpool count incremented above AND
1920 * in hugetlb_reserve_pages for the same page. Also,
1921 * the reservation count added in hugetlb_reserve_pages
1922 * no longer applies.
1926 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
1927 hugetlb_acct_memory(h
, -rsv_adjust
);
1931 out_uncharge_cgroup
:
1932 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1934 if (map_chg
|| avoid_reserve
)
1935 hugepage_subpool_put_pages(spool
, 1);
1936 vma_end_reservation(h
, vma
, addr
);
1937 return ERR_PTR(-ENOSPC
);
1941 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1942 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1943 * where no ERR_VALUE is expected to be returned.
1945 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1946 unsigned long addr
, int avoid_reserve
)
1948 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1954 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1956 struct huge_bootmem_page
*m
;
1959 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1962 addr
= memblock_virt_alloc_try_nid_nopanic(
1963 huge_page_size(h
), huge_page_size(h
),
1964 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1967 * Use the beginning of the huge page to store the
1968 * huge_bootmem_page struct (until gather_bootmem
1969 * puts them into the mem_map).
1978 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
1979 /* Put them into a private list first because mem_map is not up yet */
1980 list_add(&m
->list
, &huge_boot_pages
);
1985 static void __init
prep_compound_huge_page(struct page
*page
,
1988 if (unlikely(order
> (MAX_ORDER
- 1)))
1989 prep_compound_gigantic_page(page
, order
);
1991 prep_compound_page(page
, order
);
1994 /* Put bootmem huge pages into the standard lists after mem_map is up */
1995 static void __init
gather_bootmem_prealloc(void)
1997 struct huge_bootmem_page
*m
;
1999 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2000 struct hstate
*h
= m
->hstate
;
2003 #ifdef CONFIG_HIGHMEM
2004 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2005 memblock_free_late(__pa(m
),
2006 sizeof(struct huge_bootmem_page
));
2008 page
= virt_to_page(m
);
2010 WARN_ON(page_count(page
) != 1);
2011 prep_compound_huge_page(page
, h
->order
);
2012 WARN_ON(PageReserved(page
));
2013 prep_new_huge_page(h
, page
, page_to_nid(page
));
2015 * If we had gigantic hugepages allocated at boot time, we need
2016 * to restore the 'stolen' pages to totalram_pages in order to
2017 * fix confusing memory reports from free(1) and another
2018 * side-effects, like CommitLimit going negative.
2020 if (hstate_is_gigantic(h
))
2021 adjust_managed_page_count(page
, 1 << h
->order
);
2025 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2029 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2030 if (hstate_is_gigantic(h
)) {
2031 if (!alloc_bootmem_huge_page(h
))
2033 } else if (!alloc_fresh_huge_page(h
,
2034 &node_states
[N_MEMORY
]))
2037 h
->max_huge_pages
= i
;
2040 static void __init
hugetlb_init_hstates(void)
2044 for_each_hstate(h
) {
2045 if (minimum_order
> huge_page_order(h
))
2046 minimum_order
= huge_page_order(h
);
2048 /* oversize hugepages were init'ed in early boot */
2049 if (!hstate_is_gigantic(h
))
2050 hugetlb_hstate_alloc_pages(h
);
2052 VM_BUG_ON(minimum_order
== UINT_MAX
);
2055 static char * __init
memfmt(char *buf
, unsigned long n
)
2057 if (n
>= (1UL << 30))
2058 sprintf(buf
, "%lu GB", n
>> 30);
2059 else if (n
>= (1UL << 20))
2060 sprintf(buf
, "%lu MB", n
>> 20);
2062 sprintf(buf
, "%lu KB", n
>> 10);
2066 static void __init
report_hugepages(void)
2070 for_each_hstate(h
) {
2072 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2073 memfmt(buf
, huge_page_size(h
)),
2074 h
->free_huge_pages
);
2078 #ifdef CONFIG_HIGHMEM
2079 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2080 nodemask_t
*nodes_allowed
)
2084 if (hstate_is_gigantic(h
))
2087 for_each_node_mask(i
, *nodes_allowed
) {
2088 struct page
*page
, *next
;
2089 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2090 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2091 if (count
>= h
->nr_huge_pages
)
2093 if (PageHighMem(page
))
2095 list_del(&page
->lru
);
2096 update_and_free_page(h
, page
);
2097 h
->free_huge_pages
--;
2098 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2103 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2104 nodemask_t
*nodes_allowed
)
2110 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2111 * balanced by operating on them in a round-robin fashion.
2112 * Returns 1 if an adjustment was made.
2114 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2119 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2122 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2123 if (h
->surplus_huge_pages_node
[node
])
2127 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2128 if (h
->surplus_huge_pages_node
[node
] <
2129 h
->nr_huge_pages_node
[node
])
2136 h
->surplus_huge_pages
+= delta
;
2137 h
->surplus_huge_pages_node
[node
] += delta
;
2141 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2142 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2143 nodemask_t
*nodes_allowed
)
2145 unsigned long min_count
, ret
;
2147 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2148 return h
->max_huge_pages
;
2151 * Increase the pool size
2152 * First take pages out of surplus state. Then make up the
2153 * remaining difference by allocating fresh huge pages.
2155 * We might race with __alloc_buddy_huge_page() here and be unable
2156 * to convert a surplus huge page to a normal huge page. That is
2157 * not critical, though, it just means the overall size of the
2158 * pool might be one hugepage larger than it needs to be, but
2159 * within all the constraints specified by the sysctls.
2161 spin_lock(&hugetlb_lock
);
2162 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2163 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2167 while (count
> persistent_huge_pages(h
)) {
2169 * If this allocation races such that we no longer need the
2170 * page, free_huge_page will handle it by freeing the page
2171 * and reducing the surplus.
2173 spin_unlock(&hugetlb_lock
);
2174 if (hstate_is_gigantic(h
))
2175 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2177 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2178 spin_lock(&hugetlb_lock
);
2182 /* Bail for signals. Probably ctrl-c from user */
2183 if (signal_pending(current
))
2188 * Decrease the pool size
2189 * First return free pages to the buddy allocator (being careful
2190 * to keep enough around to satisfy reservations). Then place
2191 * pages into surplus state as needed so the pool will shrink
2192 * to the desired size as pages become free.
2194 * By placing pages into the surplus state independent of the
2195 * overcommit value, we are allowing the surplus pool size to
2196 * exceed overcommit. There are few sane options here. Since
2197 * __alloc_buddy_huge_page() is checking the global counter,
2198 * though, we'll note that we're not allowed to exceed surplus
2199 * and won't grow the pool anywhere else. Not until one of the
2200 * sysctls are changed, or the surplus pages go out of use.
2202 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2203 min_count
= max(count
, min_count
);
2204 try_to_free_low(h
, min_count
, nodes_allowed
);
2205 while (min_count
< persistent_huge_pages(h
)) {
2206 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2208 cond_resched_lock(&hugetlb_lock
);
2210 while (count
< persistent_huge_pages(h
)) {
2211 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2215 ret
= persistent_huge_pages(h
);
2216 spin_unlock(&hugetlb_lock
);
2220 #define HSTATE_ATTR_RO(_name) \
2221 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2223 #define HSTATE_ATTR(_name) \
2224 static struct kobj_attribute _name##_attr = \
2225 __ATTR(_name, 0644, _name##_show, _name##_store)
2227 static struct kobject
*hugepages_kobj
;
2228 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2230 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2232 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2236 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2237 if (hstate_kobjs
[i
] == kobj
) {
2239 *nidp
= NUMA_NO_NODE
;
2243 return kobj_to_node_hstate(kobj
, nidp
);
2246 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2247 struct kobj_attribute
*attr
, char *buf
)
2250 unsigned long nr_huge_pages
;
2253 h
= kobj_to_hstate(kobj
, &nid
);
2254 if (nid
== NUMA_NO_NODE
)
2255 nr_huge_pages
= h
->nr_huge_pages
;
2257 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2259 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2262 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2263 struct hstate
*h
, int nid
,
2264 unsigned long count
, size_t len
)
2267 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2269 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2274 if (nid
== NUMA_NO_NODE
) {
2276 * global hstate attribute
2278 if (!(obey_mempolicy
&&
2279 init_nodemask_of_mempolicy(nodes_allowed
))) {
2280 NODEMASK_FREE(nodes_allowed
);
2281 nodes_allowed
= &node_states
[N_MEMORY
];
2283 } else if (nodes_allowed
) {
2285 * per node hstate attribute: adjust count to global,
2286 * but restrict alloc/free to the specified node.
2288 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2289 init_nodemask_of_node(nodes_allowed
, nid
);
2291 nodes_allowed
= &node_states
[N_MEMORY
];
2293 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2295 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2296 NODEMASK_FREE(nodes_allowed
);
2300 NODEMASK_FREE(nodes_allowed
);
2304 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2305 struct kobject
*kobj
, const char *buf
,
2309 unsigned long count
;
2313 err
= kstrtoul(buf
, 10, &count
);
2317 h
= kobj_to_hstate(kobj
, &nid
);
2318 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2321 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2322 struct kobj_attribute
*attr
, char *buf
)
2324 return nr_hugepages_show_common(kobj
, attr
, buf
);
2327 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2328 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2330 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2332 HSTATE_ATTR(nr_hugepages
);
2337 * hstate attribute for optionally mempolicy-based constraint on persistent
2338 * huge page alloc/free.
2340 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2341 struct kobj_attribute
*attr
, char *buf
)
2343 return nr_hugepages_show_common(kobj
, attr
, buf
);
2346 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2347 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2349 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2351 HSTATE_ATTR(nr_hugepages_mempolicy
);
2355 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2356 struct kobj_attribute
*attr
, char *buf
)
2358 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2359 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2362 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2363 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2366 unsigned long input
;
2367 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2369 if (hstate_is_gigantic(h
))
2372 err
= kstrtoul(buf
, 10, &input
);
2376 spin_lock(&hugetlb_lock
);
2377 h
->nr_overcommit_huge_pages
= input
;
2378 spin_unlock(&hugetlb_lock
);
2382 HSTATE_ATTR(nr_overcommit_hugepages
);
2384 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2385 struct kobj_attribute
*attr
, char *buf
)
2388 unsigned long free_huge_pages
;
2391 h
= kobj_to_hstate(kobj
, &nid
);
2392 if (nid
== NUMA_NO_NODE
)
2393 free_huge_pages
= h
->free_huge_pages
;
2395 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2397 return sprintf(buf
, "%lu\n", free_huge_pages
);
2399 HSTATE_ATTR_RO(free_hugepages
);
2401 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2402 struct kobj_attribute
*attr
, char *buf
)
2404 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2405 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2407 HSTATE_ATTR_RO(resv_hugepages
);
2409 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2410 struct kobj_attribute
*attr
, char *buf
)
2413 unsigned long surplus_huge_pages
;
2416 h
= kobj_to_hstate(kobj
, &nid
);
2417 if (nid
== NUMA_NO_NODE
)
2418 surplus_huge_pages
= h
->surplus_huge_pages
;
2420 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2422 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2424 HSTATE_ATTR_RO(surplus_hugepages
);
2426 static struct attribute
*hstate_attrs
[] = {
2427 &nr_hugepages_attr
.attr
,
2428 &nr_overcommit_hugepages_attr
.attr
,
2429 &free_hugepages_attr
.attr
,
2430 &resv_hugepages_attr
.attr
,
2431 &surplus_hugepages_attr
.attr
,
2433 &nr_hugepages_mempolicy_attr
.attr
,
2438 static struct attribute_group hstate_attr_group
= {
2439 .attrs
= hstate_attrs
,
2442 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2443 struct kobject
**hstate_kobjs
,
2444 struct attribute_group
*hstate_attr_group
)
2447 int hi
= hstate_index(h
);
2449 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2450 if (!hstate_kobjs
[hi
])
2453 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2455 kobject_put(hstate_kobjs
[hi
]);
2460 static void __init
hugetlb_sysfs_init(void)
2465 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2466 if (!hugepages_kobj
)
2469 for_each_hstate(h
) {
2470 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2471 hstate_kobjs
, &hstate_attr_group
);
2473 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2480 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2481 * with node devices in node_devices[] using a parallel array. The array
2482 * index of a node device or _hstate == node id.
2483 * This is here to avoid any static dependency of the node device driver, in
2484 * the base kernel, on the hugetlb module.
2486 struct node_hstate
{
2487 struct kobject
*hugepages_kobj
;
2488 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2490 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2493 * A subset of global hstate attributes for node devices
2495 static struct attribute
*per_node_hstate_attrs
[] = {
2496 &nr_hugepages_attr
.attr
,
2497 &free_hugepages_attr
.attr
,
2498 &surplus_hugepages_attr
.attr
,
2502 static struct attribute_group per_node_hstate_attr_group
= {
2503 .attrs
= per_node_hstate_attrs
,
2507 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2508 * Returns node id via non-NULL nidp.
2510 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2514 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2515 struct node_hstate
*nhs
= &node_hstates
[nid
];
2517 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2518 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2530 * Unregister hstate attributes from a single node device.
2531 * No-op if no hstate attributes attached.
2533 static void hugetlb_unregister_node(struct node
*node
)
2536 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2538 if (!nhs
->hugepages_kobj
)
2539 return; /* no hstate attributes */
2541 for_each_hstate(h
) {
2542 int idx
= hstate_index(h
);
2543 if (nhs
->hstate_kobjs
[idx
]) {
2544 kobject_put(nhs
->hstate_kobjs
[idx
]);
2545 nhs
->hstate_kobjs
[idx
] = NULL
;
2549 kobject_put(nhs
->hugepages_kobj
);
2550 nhs
->hugepages_kobj
= NULL
;
2555 * Register hstate attributes for a single node device.
2556 * No-op if attributes already registered.
2558 static void hugetlb_register_node(struct node
*node
)
2561 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2564 if (nhs
->hugepages_kobj
)
2565 return; /* already allocated */
2567 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2569 if (!nhs
->hugepages_kobj
)
2572 for_each_hstate(h
) {
2573 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2575 &per_node_hstate_attr_group
);
2577 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2578 h
->name
, node
->dev
.id
);
2579 hugetlb_unregister_node(node
);
2586 * hugetlb init time: register hstate attributes for all registered node
2587 * devices of nodes that have memory. All on-line nodes should have
2588 * registered their associated device by this time.
2590 static void __init
hugetlb_register_all_nodes(void)
2594 for_each_node_state(nid
, N_MEMORY
) {
2595 struct node
*node
= node_devices
[nid
];
2596 if (node
->dev
.id
== nid
)
2597 hugetlb_register_node(node
);
2601 * Let the node device driver know we're here so it can
2602 * [un]register hstate attributes on node hotplug.
2604 register_hugetlbfs_with_node(hugetlb_register_node
,
2605 hugetlb_unregister_node
);
2607 #else /* !CONFIG_NUMA */
2609 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2617 static void hugetlb_register_all_nodes(void) { }
2621 static int __init
hugetlb_init(void)
2625 if (!hugepages_supported())
2628 if (!size_to_hstate(default_hstate_size
)) {
2629 default_hstate_size
= HPAGE_SIZE
;
2630 if (!size_to_hstate(default_hstate_size
))
2631 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2633 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2634 if (default_hstate_max_huge_pages
) {
2635 if (!default_hstate
.max_huge_pages
)
2636 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2639 hugetlb_init_hstates();
2640 gather_bootmem_prealloc();
2643 hugetlb_sysfs_init();
2644 hugetlb_register_all_nodes();
2645 hugetlb_cgroup_file_init();
2648 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2650 num_fault_mutexes
= 1;
2652 hugetlb_fault_mutex_table
=
2653 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2654 BUG_ON(!hugetlb_fault_mutex_table
);
2656 for (i
= 0; i
< num_fault_mutexes
; i
++)
2657 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2660 subsys_initcall(hugetlb_init
);
2662 /* Should be called on processing a hugepagesz=... option */
2663 void __init
hugetlb_bad_size(void)
2665 parsed_valid_hugepagesz
= false;
2668 void __init
hugetlb_add_hstate(unsigned int order
)
2673 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2674 pr_warn("hugepagesz= specified twice, ignoring\n");
2677 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2679 h
= &hstates
[hugetlb_max_hstate
++];
2681 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2682 h
->nr_huge_pages
= 0;
2683 h
->free_huge_pages
= 0;
2684 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2685 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2686 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2687 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2688 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2689 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2690 huge_page_size(h
)/1024);
2695 static int __init
hugetlb_nrpages_setup(char *s
)
2698 static unsigned long *last_mhp
;
2700 if (!parsed_valid_hugepagesz
) {
2701 pr_warn("hugepages = %s preceded by "
2702 "an unsupported hugepagesz, ignoring\n", s
);
2703 parsed_valid_hugepagesz
= true;
2707 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2708 * so this hugepages= parameter goes to the "default hstate".
2710 else if (!hugetlb_max_hstate
)
2711 mhp
= &default_hstate_max_huge_pages
;
2713 mhp
= &parsed_hstate
->max_huge_pages
;
2715 if (mhp
== last_mhp
) {
2716 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2720 if (sscanf(s
, "%lu", mhp
) <= 0)
2724 * Global state is always initialized later in hugetlb_init.
2725 * But we need to allocate >= MAX_ORDER hstates here early to still
2726 * use the bootmem allocator.
2728 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2729 hugetlb_hstate_alloc_pages(parsed_hstate
);
2735 __setup("hugepages=", hugetlb_nrpages_setup
);
2737 static int __init
hugetlb_default_setup(char *s
)
2739 default_hstate_size
= memparse(s
, &s
);
2742 __setup("default_hugepagesz=", hugetlb_default_setup
);
2744 static unsigned int cpuset_mems_nr(unsigned int *array
)
2747 unsigned int nr
= 0;
2749 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2755 #ifdef CONFIG_SYSCTL
2756 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2757 struct ctl_table
*table
, int write
,
2758 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2760 struct hstate
*h
= &default_hstate
;
2761 unsigned long tmp
= h
->max_huge_pages
;
2764 if (!hugepages_supported())
2768 table
->maxlen
= sizeof(unsigned long);
2769 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2774 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2775 NUMA_NO_NODE
, tmp
, *length
);
2780 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2781 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2784 return hugetlb_sysctl_handler_common(false, table
, write
,
2785 buffer
, length
, ppos
);
2789 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2790 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2792 return hugetlb_sysctl_handler_common(true, table
, write
,
2793 buffer
, length
, ppos
);
2795 #endif /* CONFIG_NUMA */
2797 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2798 void __user
*buffer
,
2799 size_t *length
, loff_t
*ppos
)
2801 struct hstate
*h
= &default_hstate
;
2805 if (!hugepages_supported())
2808 tmp
= h
->nr_overcommit_huge_pages
;
2810 if (write
&& hstate_is_gigantic(h
))
2814 table
->maxlen
= sizeof(unsigned long);
2815 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2820 spin_lock(&hugetlb_lock
);
2821 h
->nr_overcommit_huge_pages
= tmp
;
2822 spin_unlock(&hugetlb_lock
);
2828 #endif /* CONFIG_SYSCTL */
2830 void hugetlb_report_meminfo(struct seq_file
*m
)
2832 struct hstate
*h
= &default_hstate
;
2833 if (!hugepages_supported())
2836 "HugePages_Total: %5lu\n"
2837 "HugePages_Free: %5lu\n"
2838 "HugePages_Rsvd: %5lu\n"
2839 "HugePages_Surp: %5lu\n"
2840 "Hugepagesize: %8lu kB\n",
2844 h
->surplus_huge_pages
,
2845 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2848 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2850 struct hstate
*h
= &default_hstate
;
2851 if (!hugepages_supported())
2854 "Node %d HugePages_Total: %5u\n"
2855 "Node %d HugePages_Free: %5u\n"
2856 "Node %d HugePages_Surp: %5u\n",
2857 nid
, h
->nr_huge_pages_node
[nid
],
2858 nid
, h
->free_huge_pages_node
[nid
],
2859 nid
, h
->surplus_huge_pages_node
[nid
]);
2862 void hugetlb_show_meminfo(void)
2867 if (!hugepages_supported())
2870 for_each_node_state(nid
, N_MEMORY
)
2872 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2874 h
->nr_huge_pages_node
[nid
],
2875 h
->free_huge_pages_node
[nid
],
2876 h
->surplus_huge_pages_node
[nid
],
2877 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2880 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
2882 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
2883 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
2886 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2887 unsigned long hugetlb_total_pages(void)
2890 unsigned long nr_total_pages
= 0;
2893 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2894 return nr_total_pages
;
2897 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2901 spin_lock(&hugetlb_lock
);
2903 * When cpuset is configured, it breaks the strict hugetlb page
2904 * reservation as the accounting is done on a global variable. Such
2905 * reservation is completely rubbish in the presence of cpuset because
2906 * the reservation is not checked against page availability for the
2907 * current cpuset. Application can still potentially OOM'ed by kernel
2908 * with lack of free htlb page in cpuset that the task is in.
2909 * Attempt to enforce strict accounting with cpuset is almost
2910 * impossible (or too ugly) because cpuset is too fluid that
2911 * task or memory node can be dynamically moved between cpusets.
2913 * The change of semantics for shared hugetlb mapping with cpuset is
2914 * undesirable. However, in order to preserve some of the semantics,
2915 * we fall back to check against current free page availability as
2916 * a best attempt and hopefully to minimize the impact of changing
2917 * semantics that cpuset has.
2920 if (gather_surplus_pages(h
, delta
) < 0)
2923 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2924 return_unused_surplus_pages(h
, delta
);
2931 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2934 spin_unlock(&hugetlb_lock
);
2938 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2940 struct resv_map
*resv
= vma_resv_map(vma
);
2943 * This new VMA should share its siblings reservation map if present.
2944 * The VMA will only ever have a valid reservation map pointer where
2945 * it is being copied for another still existing VMA. As that VMA
2946 * has a reference to the reservation map it cannot disappear until
2947 * after this open call completes. It is therefore safe to take a
2948 * new reference here without additional locking.
2950 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2951 kref_get(&resv
->refs
);
2954 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2956 struct hstate
*h
= hstate_vma(vma
);
2957 struct resv_map
*resv
= vma_resv_map(vma
);
2958 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2959 unsigned long reserve
, start
, end
;
2962 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2965 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2966 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2968 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2970 kref_put(&resv
->refs
, resv_map_release
);
2974 * Decrement reserve counts. The global reserve count may be
2975 * adjusted if the subpool has a minimum size.
2977 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
2978 hugetlb_acct_memory(h
, -gbl_reserve
);
2983 * We cannot handle pagefaults against hugetlb pages at all. They cause
2984 * handle_mm_fault() to try to instantiate regular-sized pages in the
2985 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2988 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2994 const struct vm_operations_struct hugetlb_vm_ops
= {
2995 .fault
= hugetlb_vm_op_fault
,
2996 .open
= hugetlb_vm_op_open
,
2997 .close
= hugetlb_vm_op_close
,
3000 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3006 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3007 vma
->vm_page_prot
)));
3009 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3010 vma
->vm_page_prot
));
3012 entry
= pte_mkyoung(entry
);
3013 entry
= pte_mkhuge(entry
);
3014 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3019 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3020 unsigned long address
, pte_t
*ptep
)
3024 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3025 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3026 update_mmu_cache(vma
, address
, ptep
);
3029 static int is_hugetlb_entry_migration(pte_t pte
)
3033 if (huge_pte_none(pte
) || pte_present(pte
))
3035 swp
= pte_to_swp_entry(pte
);
3036 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3042 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3046 if (huge_pte_none(pte
) || pte_present(pte
))
3048 swp
= pte_to_swp_entry(pte
);
3049 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3055 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3056 struct vm_area_struct
*vma
)
3058 pte_t
*src_pte
, *dst_pte
, entry
;
3059 struct page
*ptepage
;
3062 struct hstate
*h
= hstate_vma(vma
);
3063 unsigned long sz
= huge_page_size(h
);
3064 unsigned long mmun_start
; /* For mmu_notifiers */
3065 unsigned long mmun_end
; /* For mmu_notifiers */
3068 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3070 mmun_start
= vma
->vm_start
;
3071 mmun_end
= vma
->vm_end
;
3073 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3075 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3076 spinlock_t
*src_ptl
, *dst_ptl
;
3077 src_pte
= huge_pte_offset(src
, addr
);
3080 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3086 /* If the pagetables are shared don't copy or take references */
3087 if (dst_pte
== src_pte
)
3090 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3091 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3092 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3093 entry
= huge_ptep_get(src_pte
);
3094 if (huge_pte_none(entry
)) { /* skip none entry */
3096 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3097 is_hugetlb_entry_hwpoisoned(entry
))) {
3098 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3100 if (is_write_migration_entry(swp_entry
) && cow
) {
3102 * COW mappings require pages in both
3103 * parent and child to be set to read.
3105 make_migration_entry_read(&swp_entry
);
3106 entry
= swp_entry_to_pte(swp_entry
);
3107 set_huge_pte_at(src
, addr
, src_pte
, entry
);
3109 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3112 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3113 mmu_notifier_invalidate_range(src
, mmun_start
,
3116 entry
= huge_ptep_get(src_pte
);
3117 ptepage
= pte_page(entry
);
3119 page_dup_rmap(ptepage
, true);
3120 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3121 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3123 spin_unlock(src_ptl
);
3124 spin_unlock(dst_ptl
);
3128 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3133 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3134 unsigned long start
, unsigned long end
,
3135 struct page
*ref_page
)
3137 int force_flush
= 0;
3138 struct mm_struct
*mm
= vma
->vm_mm
;
3139 unsigned long address
;
3144 struct hstate
*h
= hstate_vma(vma
);
3145 unsigned long sz
= huge_page_size(h
);
3146 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3147 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3149 WARN_ON(!is_vm_hugetlb_page(vma
));
3150 BUG_ON(start
& ~huge_page_mask(h
));
3151 BUG_ON(end
& ~huge_page_mask(h
));
3153 tlb_start_vma(tlb
, vma
);
3154 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3157 for (; address
< end
; address
+= sz
) {
3158 ptep
= huge_pte_offset(mm
, address
);
3162 ptl
= huge_pte_lock(h
, mm
, ptep
);
3163 if (huge_pmd_unshare(mm
, &address
, ptep
))
3166 pte
= huge_ptep_get(ptep
);
3167 if (huge_pte_none(pte
))
3171 * Migrating hugepage or HWPoisoned hugepage is already
3172 * unmapped and its refcount is dropped, so just clear pte here.
3174 if (unlikely(!pte_present(pte
))) {
3175 huge_pte_clear(mm
, address
, ptep
);
3179 page
= pte_page(pte
);
3181 * If a reference page is supplied, it is because a specific
3182 * page is being unmapped, not a range. Ensure the page we
3183 * are about to unmap is the actual page of interest.
3186 if (page
!= ref_page
)
3190 * Mark the VMA as having unmapped its page so that
3191 * future faults in this VMA will fail rather than
3192 * looking like data was lost
3194 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3197 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3198 tlb_remove_tlb_entry(tlb
, ptep
, address
);
3199 if (huge_pte_dirty(pte
))
3200 set_page_dirty(page
);
3202 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3203 page_remove_rmap(page
, true);
3204 force_flush
= !__tlb_remove_page(tlb
, page
);
3210 /* Bail out after unmapping reference page if supplied */
3219 * mmu_gather ran out of room to batch pages, we break out of
3220 * the PTE lock to avoid doing the potential expensive TLB invalidate
3221 * and page-free while holding it.
3226 if (address
< end
&& !ref_page
)
3229 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3230 tlb_end_vma(tlb
, vma
);
3233 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3234 struct vm_area_struct
*vma
, unsigned long start
,
3235 unsigned long end
, struct page
*ref_page
)
3237 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3240 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3241 * test will fail on a vma being torn down, and not grab a page table
3242 * on its way out. We're lucky that the flag has such an appropriate
3243 * name, and can in fact be safely cleared here. We could clear it
3244 * before the __unmap_hugepage_range above, but all that's necessary
3245 * is to clear it before releasing the i_mmap_rwsem. This works
3246 * because in the context this is called, the VMA is about to be
3247 * destroyed and the i_mmap_rwsem is held.
3249 vma
->vm_flags
&= ~VM_MAYSHARE
;
3252 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3253 unsigned long end
, struct page
*ref_page
)
3255 struct mm_struct
*mm
;
3256 struct mmu_gather tlb
;
3260 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3261 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3262 tlb_finish_mmu(&tlb
, start
, end
);
3266 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3267 * mappping it owns the reserve page for. The intention is to unmap the page
3268 * from other VMAs and let the children be SIGKILLed if they are faulting the
3271 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3272 struct page
*page
, unsigned long address
)
3274 struct hstate
*h
= hstate_vma(vma
);
3275 struct vm_area_struct
*iter_vma
;
3276 struct address_space
*mapping
;
3280 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3281 * from page cache lookup which is in HPAGE_SIZE units.
3283 address
= address
& huge_page_mask(h
);
3284 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3286 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
3289 * Take the mapping lock for the duration of the table walk. As
3290 * this mapping should be shared between all the VMAs,
3291 * __unmap_hugepage_range() is called as the lock is already held
3293 i_mmap_lock_write(mapping
);
3294 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3295 /* Do not unmap the current VMA */
3296 if (iter_vma
== vma
)
3300 * Shared VMAs have their own reserves and do not affect
3301 * MAP_PRIVATE accounting but it is possible that a shared
3302 * VMA is using the same page so check and skip such VMAs.
3304 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3308 * Unmap the page from other VMAs without their own reserves.
3309 * They get marked to be SIGKILLed if they fault in these
3310 * areas. This is because a future no-page fault on this VMA
3311 * could insert a zeroed page instead of the data existing
3312 * from the time of fork. This would look like data corruption
3314 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3315 unmap_hugepage_range(iter_vma
, address
,
3316 address
+ huge_page_size(h
), page
);
3318 i_mmap_unlock_write(mapping
);
3322 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3323 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3324 * cannot race with other handlers or page migration.
3325 * Keep the pte_same checks anyway to make transition from the mutex easier.
3327 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3328 unsigned long address
, pte_t
*ptep
, pte_t pte
,
3329 struct page
*pagecache_page
, spinlock_t
*ptl
)
3331 struct hstate
*h
= hstate_vma(vma
);
3332 struct page
*old_page
, *new_page
;
3333 int ret
= 0, outside_reserve
= 0;
3334 unsigned long mmun_start
; /* For mmu_notifiers */
3335 unsigned long mmun_end
; /* For mmu_notifiers */
3337 old_page
= pte_page(pte
);
3340 /* If no-one else is actually using this page, avoid the copy
3341 * and just make the page writable */
3342 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3343 page_move_anon_rmap(old_page
, vma
, address
);
3344 set_huge_ptep_writable(vma
, address
, ptep
);
3349 * If the process that created a MAP_PRIVATE mapping is about to
3350 * perform a COW due to a shared page count, attempt to satisfy
3351 * the allocation without using the existing reserves. The pagecache
3352 * page is used to determine if the reserve at this address was
3353 * consumed or not. If reserves were used, a partial faulted mapping
3354 * at the time of fork() could consume its reserves on COW instead
3355 * of the full address range.
3357 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3358 old_page
!= pagecache_page
)
3359 outside_reserve
= 1;
3364 * Drop page table lock as buddy allocator may be called. It will
3365 * be acquired again before returning to the caller, as expected.
3368 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3370 if (IS_ERR(new_page
)) {
3372 * If a process owning a MAP_PRIVATE mapping fails to COW,
3373 * it is due to references held by a child and an insufficient
3374 * huge page pool. To guarantee the original mappers
3375 * reliability, unmap the page from child processes. The child
3376 * may get SIGKILLed if it later faults.
3378 if (outside_reserve
) {
3380 BUG_ON(huge_pte_none(pte
));
3381 unmap_ref_private(mm
, vma
, old_page
, address
);
3382 BUG_ON(huge_pte_none(pte
));
3384 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3386 pte_same(huge_ptep_get(ptep
), pte
)))
3387 goto retry_avoidcopy
;
3389 * race occurs while re-acquiring page table
3390 * lock, and our job is done.
3395 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3396 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3397 goto out_release_old
;
3401 * When the original hugepage is shared one, it does not have
3402 * anon_vma prepared.
3404 if (unlikely(anon_vma_prepare(vma
))) {
3406 goto out_release_all
;
3409 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3410 pages_per_huge_page(h
));
3411 __SetPageUptodate(new_page
);
3412 set_page_huge_active(new_page
);
3414 mmun_start
= address
& huge_page_mask(h
);
3415 mmun_end
= mmun_start
+ huge_page_size(h
);
3416 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3419 * Retake the page table lock to check for racing updates
3420 * before the page tables are altered
3423 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3424 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3425 ClearPagePrivate(new_page
);
3428 huge_ptep_clear_flush(vma
, address
, ptep
);
3429 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3430 set_huge_pte_at(mm
, address
, ptep
,
3431 make_huge_pte(vma
, new_page
, 1));
3432 page_remove_rmap(old_page
, true);
3433 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3434 /* Make the old page be freed below */
3435 new_page
= old_page
;
3438 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3444 spin_lock(ptl
); /* Caller expects lock to be held */
3448 /* Return the pagecache page at a given address within a VMA */
3449 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3450 struct vm_area_struct
*vma
, unsigned long address
)
3452 struct address_space
*mapping
;
3455 mapping
= vma
->vm_file
->f_mapping
;
3456 idx
= vma_hugecache_offset(h
, vma
, address
);
3458 return find_lock_page(mapping
, idx
);
3462 * Return whether there is a pagecache page to back given address within VMA.
3463 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3465 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3466 struct vm_area_struct
*vma
, unsigned long address
)
3468 struct address_space
*mapping
;
3472 mapping
= vma
->vm_file
->f_mapping
;
3473 idx
= vma_hugecache_offset(h
, vma
, address
);
3475 page
= find_get_page(mapping
, idx
);
3478 return page
!= NULL
;
3481 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3484 struct inode
*inode
= mapping
->host
;
3485 struct hstate
*h
= hstate_inode(inode
);
3486 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3490 ClearPagePrivate(page
);
3492 spin_lock(&inode
->i_lock
);
3493 inode
->i_blocks
+= blocks_per_huge_page(h
);
3494 spin_unlock(&inode
->i_lock
);
3498 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3499 struct address_space
*mapping
, pgoff_t idx
,
3500 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3502 struct hstate
*h
= hstate_vma(vma
);
3503 int ret
= VM_FAULT_SIGBUS
;
3511 * Currently, we are forced to kill the process in the event the
3512 * original mapper has unmapped pages from the child due to a failed
3513 * COW. Warn that such a situation has occurred as it may not be obvious
3515 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3516 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3522 * Use page lock to guard against racing truncation
3523 * before we get page_table_lock.
3526 page
= find_lock_page(mapping
, idx
);
3528 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3531 page
= alloc_huge_page(vma
, address
, 0);
3533 ret
= PTR_ERR(page
);
3537 ret
= VM_FAULT_SIGBUS
;
3540 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3541 __SetPageUptodate(page
);
3542 set_page_huge_active(page
);
3544 if (vma
->vm_flags
& VM_MAYSHARE
) {
3545 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3554 if (unlikely(anon_vma_prepare(vma
))) {
3556 goto backout_unlocked
;
3562 * If memory error occurs between mmap() and fault, some process
3563 * don't have hwpoisoned swap entry for errored virtual address.
3564 * So we need to block hugepage fault by PG_hwpoison bit check.
3566 if (unlikely(PageHWPoison(page
))) {
3567 ret
= VM_FAULT_HWPOISON
|
3568 VM_FAULT_SET_HINDEX(hstate_index(h
));
3569 goto backout_unlocked
;
3574 * If we are going to COW a private mapping later, we examine the
3575 * pending reservations for this page now. This will ensure that
3576 * any allocations necessary to record that reservation occur outside
3579 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3580 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3582 goto backout_unlocked
;
3584 /* Just decrements count, does not deallocate */
3585 vma_end_reservation(h
, vma
, address
);
3588 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3590 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3595 if (!huge_pte_none(huge_ptep_get(ptep
)))
3599 ClearPagePrivate(page
);
3600 hugepage_add_new_anon_rmap(page
, vma
, address
);
3602 page_dup_rmap(page
, true);
3603 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3604 && (vma
->vm_flags
& VM_SHARED
)));
3605 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3607 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3608 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3609 /* Optimization, do the COW without a second fault */
3610 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3627 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3628 struct vm_area_struct
*vma
,
3629 struct address_space
*mapping
,
3630 pgoff_t idx
, unsigned long address
)
3632 unsigned long key
[2];
3635 if (vma
->vm_flags
& VM_SHARED
) {
3636 key
[0] = (unsigned long) mapping
;
3639 key
[0] = (unsigned long) mm
;
3640 key
[1] = address
>> huge_page_shift(h
);
3643 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3645 return hash
& (num_fault_mutexes
- 1);
3649 * For uniprocesor systems we always use a single mutex, so just
3650 * return 0 and avoid the hashing overhead.
3652 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3653 struct vm_area_struct
*vma
,
3654 struct address_space
*mapping
,
3655 pgoff_t idx
, unsigned long address
)
3661 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3662 unsigned long address
, unsigned int flags
)
3669 struct page
*page
= NULL
;
3670 struct page
*pagecache_page
= NULL
;
3671 struct hstate
*h
= hstate_vma(vma
);
3672 struct address_space
*mapping
;
3673 int need_wait_lock
= 0;
3675 address
&= huge_page_mask(h
);
3677 ptep
= huge_pte_offset(mm
, address
);
3679 entry
= huge_ptep_get(ptep
);
3680 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3681 migration_entry_wait_huge(vma
, mm
, ptep
);
3683 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3684 return VM_FAULT_HWPOISON_LARGE
|
3685 VM_FAULT_SET_HINDEX(hstate_index(h
));
3687 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3689 return VM_FAULT_OOM
;
3692 mapping
= vma
->vm_file
->f_mapping
;
3693 idx
= vma_hugecache_offset(h
, vma
, address
);
3696 * Serialize hugepage allocation and instantiation, so that we don't
3697 * get spurious allocation failures if two CPUs race to instantiate
3698 * the same page in the page cache.
3700 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3701 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3703 entry
= huge_ptep_get(ptep
);
3704 if (huge_pte_none(entry
)) {
3705 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3712 * entry could be a migration/hwpoison entry at this point, so this
3713 * check prevents the kernel from going below assuming that we have
3714 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3715 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3718 if (!pte_present(entry
))
3722 * If we are going to COW the mapping later, we examine the pending
3723 * reservations for this page now. This will ensure that any
3724 * allocations necessary to record that reservation occur outside the
3725 * spinlock. For private mappings, we also lookup the pagecache
3726 * page now as it is used to determine if a reservation has been
3729 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3730 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3734 /* Just decrements count, does not deallocate */
3735 vma_end_reservation(h
, vma
, address
);
3737 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3738 pagecache_page
= hugetlbfs_pagecache_page(h
,
3742 ptl
= huge_pte_lock(h
, mm
, ptep
);
3744 /* Check for a racing update before calling hugetlb_cow */
3745 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3749 * hugetlb_cow() requires page locks of pte_page(entry) and
3750 * pagecache_page, so here we need take the former one
3751 * when page != pagecache_page or !pagecache_page.
3753 page
= pte_page(entry
);
3754 if (page
!= pagecache_page
)
3755 if (!trylock_page(page
)) {
3762 if (flags
& FAULT_FLAG_WRITE
) {
3763 if (!huge_pte_write(entry
)) {
3764 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3765 pagecache_page
, ptl
);
3768 entry
= huge_pte_mkdirty(entry
);
3770 entry
= pte_mkyoung(entry
);
3771 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3772 flags
& FAULT_FLAG_WRITE
))
3773 update_mmu_cache(vma
, address
, ptep
);
3775 if (page
!= pagecache_page
)
3781 if (pagecache_page
) {
3782 unlock_page(pagecache_page
);
3783 put_page(pagecache_page
);
3786 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3788 * Generally it's safe to hold refcount during waiting page lock. But
3789 * here we just wait to defer the next page fault to avoid busy loop and
3790 * the page is not used after unlocked before returning from the current
3791 * page fault. So we are safe from accessing freed page, even if we wait
3792 * here without taking refcount.
3795 wait_on_page_locked(page
);
3799 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3800 struct page
**pages
, struct vm_area_struct
**vmas
,
3801 unsigned long *position
, unsigned long *nr_pages
,
3802 long i
, unsigned int flags
)
3804 unsigned long pfn_offset
;
3805 unsigned long vaddr
= *position
;
3806 unsigned long remainder
= *nr_pages
;
3807 struct hstate
*h
= hstate_vma(vma
);
3809 while (vaddr
< vma
->vm_end
&& remainder
) {
3811 spinlock_t
*ptl
= NULL
;
3816 * If we have a pending SIGKILL, don't keep faulting pages and
3817 * potentially allocating memory.
3819 if (unlikely(fatal_signal_pending(current
))) {
3825 * Some archs (sparc64, sh*) have multiple pte_ts to
3826 * each hugepage. We have to make sure we get the
3827 * first, for the page indexing below to work.
3829 * Note that page table lock is not held when pte is null.
3831 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3833 ptl
= huge_pte_lock(h
, mm
, pte
);
3834 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3837 * When coredumping, it suits get_dump_page if we just return
3838 * an error where there's an empty slot with no huge pagecache
3839 * to back it. This way, we avoid allocating a hugepage, and
3840 * the sparse dumpfile avoids allocating disk blocks, but its
3841 * huge holes still show up with zeroes where they need to be.
3843 if (absent
&& (flags
& FOLL_DUMP
) &&
3844 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3852 * We need call hugetlb_fault for both hugepages under migration
3853 * (in which case hugetlb_fault waits for the migration,) and
3854 * hwpoisoned hugepages (in which case we need to prevent the
3855 * caller from accessing to them.) In order to do this, we use
3856 * here is_swap_pte instead of is_hugetlb_entry_migration and
3857 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3858 * both cases, and because we can't follow correct pages
3859 * directly from any kind of swap entries.
3861 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3862 ((flags
& FOLL_WRITE
) &&
3863 !huge_pte_write(huge_ptep_get(pte
)))) {
3868 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3869 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3870 if (!(ret
& VM_FAULT_ERROR
))
3877 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3878 page
= pte_page(huge_ptep_get(pte
));
3881 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3892 if (vaddr
< vma
->vm_end
&& remainder
&&
3893 pfn_offset
< pages_per_huge_page(h
)) {
3895 * We use pfn_offset to avoid touching the pageframes
3896 * of this compound page.
3902 *nr_pages
= remainder
;
3905 return i
? i
: -EFAULT
;
3908 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3909 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3911 struct mm_struct
*mm
= vma
->vm_mm
;
3912 unsigned long start
= address
;
3915 struct hstate
*h
= hstate_vma(vma
);
3916 unsigned long pages
= 0;
3918 BUG_ON(address
>= end
);
3919 flush_cache_range(vma
, address
, end
);
3921 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3922 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
3923 for (; address
< end
; address
+= huge_page_size(h
)) {
3925 ptep
= huge_pte_offset(mm
, address
);
3928 ptl
= huge_pte_lock(h
, mm
, ptep
);
3929 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3934 pte
= huge_ptep_get(ptep
);
3935 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
3939 if (unlikely(is_hugetlb_entry_migration(pte
))) {
3940 swp_entry_t entry
= pte_to_swp_entry(pte
);
3942 if (is_write_migration_entry(entry
)) {
3945 make_migration_entry_read(&entry
);
3946 newpte
= swp_entry_to_pte(entry
);
3947 set_huge_pte_at(mm
, address
, ptep
, newpte
);
3953 if (!huge_pte_none(pte
)) {
3954 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3955 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3956 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3957 set_huge_pte_at(mm
, address
, ptep
, pte
);
3963 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3964 * may have cleared our pud entry and done put_page on the page table:
3965 * once we release i_mmap_rwsem, another task can do the final put_page
3966 * and that page table be reused and filled with junk.
3968 flush_tlb_range(vma
, start
, end
);
3969 mmu_notifier_invalidate_range(mm
, start
, end
);
3970 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
3971 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3973 return pages
<< h
->order
;
3976 int hugetlb_reserve_pages(struct inode
*inode
,
3978 struct vm_area_struct
*vma
,
3979 vm_flags_t vm_flags
)
3982 struct hstate
*h
= hstate_inode(inode
);
3983 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3984 struct resv_map
*resv_map
;
3988 * Only apply hugepage reservation if asked. At fault time, an
3989 * attempt will be made for VM_NORESERVE to allocate a page
3990 * without using reserves
3992 if (vm_flags
& VM_NORESERVE
)
3996 * Shared mappings base their reservation on the number of pages that
3997 * are already allocated on behalf of the file. Private mappings need
3998 * to reserve the full area even if read-only as mprotect() may be
3999 * called to make the mapping read-write. Assume !vma is a shm mapping
4001 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4002 resv_map
= inode_resv_map(inode
);
4004 chg
= region_chg(resv_map
, from
, to
);
4007 resv_map
= resv_map_alloc();
4013 set_vma_resv_map(vma
, resv_map
);
4014 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4023 * There must be enough pages in the subpool for the mapping. If
4024 * the subpool has a minimum size, there may be some global
4025 * reservations already in place (gbl_reserve).
4027 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4028 if (gbl_reserve
< 0) {
4034 * Check enough hugepages are available for the reservation.
4035 * Hand the pages back to the subpool if there are not
4037 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4039 /* put back original number of pages, chg */
4040 (void)hugepage_subpool_put_pages(spool
, chg
);
4045 * Account for the reservations made. Shared mappings record regions
4046 * that have reservations as they are shared by multiple VMAs.
4047 * When the last VMA disappears, the region map says how much
4048 * the reservation was and the page cache tells how much of
4049 * the reservation was consumed. Private mappings are per-VMA and
4050 * only the consumed reservations are tracked. When the VMA
4051 * disappears, the original reservation is the VMA size and the
4052 * consumed reservations are stored in the map. Hence, nothing
4053 * else has to be done for private mappings here
4055 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4056 long add
= region_add(resv_map
, from
, to
);
4058 if (unlikely(chg
> add
)) {
4060 * pages in this range were added to the reserve
4061 * map between region_chg and region_add. This
4062 * indicates a race with alloc_huge_page. Adjust
4063 * the subpool and reserve counts modified above
4064 * based on the difference.
4068 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4070 hugetlb_acct_memory(h
, -rsv_adjust
);
4075 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4076 region_abort(resv_map
, from
, to
);
4077 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4078 kref_put(&resv_map
->refs
, resv_map_release
);
4082 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4085 struct hstate
*h
= hstate_inode(inode
);
4086 struct resv_map
*resv_map
= inode_resv_map(inode
);
4088 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4092 chg
= region_del(resv_map
, start
, end
);
4094 * region_del() can fail in the rare case where a region
4095 * must be split and another region descriptor can not be
4096 * allocated. If end == LONG_MAX, it will not fail.
4102 spin_lock(&inode
->i_lock
);
4103 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4104 spin_unlock(&inode
->i_lock
);
4107 * If the subpool has a minimum size, the number of global
4108 * reservations to be released may be adjusted.
4110 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4111 hugetlb_acct_memory(h
, -gbl_reserve
);
4116 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4117 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4118 struct vm_area_struct
*vma
,
4119 unsigned long addr
, pgoff_t idx
)
4121 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4123 unsigned long sbase
= saddr
& PUD_MASK
;
4124 unsigned long s_end
= sbase
+ PUD_SIZE
;
4126 /* Allow segments to share if only one is marked locked */
4127 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4128 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4131 * match the virtual addresses, permission and the alignment of the
4134 if (pmd_index(addr
) != pmd_index(saddr
) ||
4135 vm_flags
!= svm_flags
||
4136 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4142 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4144 unsigned long base
= addr
& PUD_MASK
;
4145 unsigned long end
= base
+ PUD_SIZE
;
4148 * check on proper vm_flags and page table alignment
4150 if (vma
->vm_flags
& VM_MAYSHARE
&&
4151 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4157 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4158 * and returns the corresponding pte. While this is not necessary for the
4159 * !shared pmd case because we can allocate the pmd later as well, it makes the
4160 * code much cleaner. pmd allocation is essential for the shared case because
4161 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4162 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4163 * bad pmd for sharing.
4165 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4167 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4168 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4169 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4171 struct vm_area_struct
*svma
;
4172 unsigned long saddr
;
4177 if (!vma_shareable(vma
, addr
))
4178 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4180 i_mmap_lock_write(mapping
);
4181 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4185 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4187 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
4190 get_page(virt_to_page(spte
));
4199 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
4201 if (pud_none(*pud
)) {
4202 pud_populate(mm
, pud
,
4203 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4205 put_page(virt_to_page(spte
));
4210 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4211 i_mmap_unlock_write(mapping
);
4216 * unmap huge page backed by shared pte.
4218 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4219 * indicated by page_count > 1, unmap is achieved by clearing pud and
4220 * decrementing the ref count. If count == 1, the pte page is not shared.
4222 * called with page table lock held.
4224 * returns: 1 successfully unmapped a shared pte page
4225 * 0 the underlying pte page is not shared, or it is the last user
4227 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4229 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4230 pud_t
*pud
= pud_offset(pgd
, *addr
);
4232 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4233 if (page_count(virt_to_page(ptep
)) == 1)
4237 put_page(virt_to_page(ptep
));
4239 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4242 #define want_pmd_share() (1)
4243 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4244 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4249 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4253 #define want_pmd_share() (0)
4254 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4256 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4257 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4258 unsigned long addr
, unsigned long sz
)
4264 pgd
= pgd_offset(mm
, addr
);
4265 pud
= pud_alloc(mm
, pgd
, addr
);
4267 if (sz
== PUD_SIZE
) {
4270 BUG_ON(sz
!= PMD_SIZE
);
4271 if (want_pmd_share() && pud_none(*pud
))
4272 pte
= huge_pmd_share(mm
, addr
, pud
);
4274 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4277 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
4282 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
4288 pgd
= pgd_offset(mm
, addr
);
4289 if (pgd_present(*pgd
)) {
4290 pud
= pud_offset(pgd
, addr
);
4291 if (pud_present(*pud
)) {
4293 return (pte_t
*)pud
;
4294 pmd
= pmd_offset(pud
, addr
);
4297 return (pte_t
*) pmd
;
4300 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4303 * These functions are overwritable if your architecture needs its own
4306 struct page
* __weak
4307 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4310 return ERR_PTR(-EINVAL
);
4313 struct page
* __weak
4314 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4315 pmd_t
*pmd
, int flags
)
4317 struct page
*page
= NULL
;
4320 ptl
= pmd_lockptr(mm
, pmd
);
4323 * make sure that the address range covered by this pmd is not
4324 * unmapped from other threads.
4326 if (!pmd_huge(*pmd
))
4328 if (pmd_present(*pmd
)) {
4329 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4330 if (flags
& FOLL_GET
)
4333 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
4335 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4339 * hwpoisoned entry is treated as no_page_table in
4340 * follow_page_mask().
4348 struct page
* __weak
4349 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4350 pud_t
*pud
, int flags
)
4352 if (flags
& FOLL_GET
)
4355 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4358 #ifdef CONFIG_MEMORY_FAILURE
4361 * This function is called from memory failure code.
4362 * Assume the caller holds page lock of the head page.
4364 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4366 struct hstate
*h
= page_hstate(hpage
);
4367 int nid
= page_to_nid(hpage
);
4370 spin_lock(&hugetlb_lock
);
4372 * Just checking !page_huge_active is not enough, because that could be
4373 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4375 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4377 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4378 * but dangling hpage->lru can trigger list-debug warnings
4379 * (this happens when we call unpoison_memory() on it),
4380 * so let it point to itself with list_del_init().
4382 list_del_init(&hpage
->lru
);
4383 set_page_refcounted(hpage
);
4384 h
->free_huge_pages
--;
4385 h
->free_huge_pages_node
[nid
]--;
4388 spin_unlock(&hugetlb_lock
);
4393 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4397 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4398 spin_lock(&hugetlb_lock
);
4399 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4403 clear_page_huge_active(page
);
4404 list_move_tail(&page
->lru
, list
);
4406 spin_unlock(&hugetlb_lock
);
4410 void putback_active_hugepage(struct page
*page
)
4412 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4413 spin_lock(&hugetlb_lock
);
4414 set_page_huge_active(page
);
4415 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4416 spin_unlock(&hugetlb_lock
);