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>
35 #include <linux/userfaultfd_k.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
49 __initdata
LIST_HEAD(huge_boot_pages
);
51 /* for command line parsing */
52 static struct hstate
* __initdata parsed_hstate
;
53 static unsigned long __initdata default_hstate_max_huge_pages
;
54 static unsigned long __initdata default_hstate_size
;
55 static bool __initdata parsed_valid_hugepagesz
= true;
58 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
59 * free_huge_pages, and surplus_huge_pages.
61 DEFINE_SPINLOCK(hugetlb_lock
);
64 * Serializes faults on the same logical page. This is used to
65 * prevent spurious OOMs when the hugepage pool is fully utilized.
67 static int num_fault_mutexes
;
68 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
70 /* Forward declaration */
71 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
73 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
75 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
77 spin_unlock(&spool
->lock
);
79 /* If no pages are used, and no other handles to the subpool
80 * remain, give up any reservations mased on minimum size and
83 if (spool
->min_hpages
!= -1)
84 hugetlb_acct_memory(spool
->hstate
,
90 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
93 struct hugepage_subpool
*spool
;
95 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
99 spin_lock_init(&spool
->lock
);
101 spool
->max_hpages
= max_hpages
;
103 spool
->min_hpages
= min_hpages
;
105 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
109 spool
->rsv_hpages
= min_hpages
;
114 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
116 spin_lock(&spool
->lock
);
117 BUG_ON(!spool
->count
);
119 unlock_or_release_subpool(spool
);
123 * Subpool accounting for allocating and reserving pages.
124 * Return -ENOMEM if there are not enough resources to satisfy the
125 * the request. Otherwise, return the number of pages by which the
126 * global pools must be adjusted (upward). The returned value may
127 * only be different than the passed value (delta) in the case where
128 * a subpool minimum size must be manitained.
130 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
138 spin_lock(&spool
->lock
);
140 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
141 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
142 spool
->used_hpages
+= delta
;
149 /* minimum size accounting */
150 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
151 if (delta
> spool
->rsv_hpages
) {
153 * Asking for more reserves than those already taken on
154 * behalf of subpool. Return difference.
156 ret
= delta
- spool
->rsv_hpages
;
157 spool
->rsv_hpages
= 0;
159 ret
= 0; /* reserves already accounted for */
160 spool
->rsv_hpages
-= delta
;
165 spin_unlock(&spool
->lock
);
170 * Subpool accounting for freeing and unreserving pages.
171 * Return the number of global page reservations that must be dropped.
172 * The return value may only be different than the passed value (delta)
173 * in the case where a subpool minimum size must be maintained.
175 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
183 spin_lock(&spool
->lock
);
185 if (spool
->max_hpages
!= -1) /* maximum size accounting */
186 spool
->used_hpages
-= delta
;
188 /* minimum size accounting */
189 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
190 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
193 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
195 spool
->rsv_hpages
+= delta
;
196 if (spool
->rsv_hpages
> spool
->min_hpages
)
197 spool
->rsv_hpages
= spool
->min_hpages
;
201 * If hugetlbfs_put_super couldn't free spool due to an outstanding
202 * quota reference, free it now.
204 unlock_or_release_subpool(spool
);
209 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
211 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
214 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
216 return subpool_inode(file_inode(vma
->vm_file
));
220 * Region tracking -- allows tracking of reservations and instantiated pages
221 * across the pages in a mapping.
223 * The region data structures are embedded into a resv_map and protected
224 * by a resv_map's lock. The set of regions within the resv_map represent
225 * reservations for huge pages, or huge pages that have already been
226 * instantiated within the map. The from and to elements are huge page
227 * indicies into the associated mapping. from indicates the starting index
228 * of the region. to represents the first index past the end of the region.
230 * For example, a file region structure with from == 0 and to == 4 represents
231 * four huge pages in a mapping. It is important to note that the to element
232 * represents the first element past the end of the region. This is used in
233 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235 * Interval notation of the form [from, to) will be used to indicate that
236 * the endpoint from is inclusive and to is exclusive.
239 struct list_head link
;
245 * Add the huge page range represented by [f, t) to the reserve
246 * map. In the normal case, existing regions will be expanded
247 * to accommodate the specified range. Sufficient regions should
248 * exist for expansion due to the previous call to region_chg
249 * with the same range. However, it is possible that region_del
250 * could have been called after region_chg and modifed the map
251 * in such a way that no region exists to be expanded. In this
252 * case, pull a region descriptor from the cache associated with
253 * the map and use that for the new range.
255 * Return the number of new huge pages added to the map. This
256 * number is greater than or equal to zero.
258 static long region_add(struct resv_map
*resv
, long f
, long t
)
260 struct list_head
*head
= &resv
->regions
;
261 struct file_region
*rg
, *nrg
, *trg
;
264 spin_lock(&resv
->lock
);
265 /* Locate the region we are either in or before. */
266 list_for_each_entry(rg
, head
, link
)
271 * If no region exists which can be expanded to include the
272 * specified range, the list must have been modified by an
273 * interleving call to region_del(). Pull a region descriptor
274 * from the cache and use it for this range.
276 if (&rg
->link
== head
|| t
< rg
->from
) {
277 VM_BUG_ON(resv
->region_cache_count
<= 0);
279 resv
->region_cache_count
--;
280 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
282 list_del(&nrg
->link
);
286 list_add(&nrg
->link
, rg
->link
.prev
);
292 /* Round our left edge to the current segment if it encloses us. */
296 /* Check for and consume any regions we now overlap with. */
298 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
299 if (&rg
->link
== head
)
304 /* If this area reaches higher then extend our area to
305 * include it completely. If this is not the first area
306 * which we intend to reuse, free it. */
310 /* Decrement return value by the deleted range.
311 * Another range will span this area so that by
312 * end of routine add will be >= zero
314 add
-= (rg
->to
- rg
->from
);
320 add
+= (nrg
->from
- f
); /* Added to beginning of region */
322 add
+= t
- nrg
->to
; /* Added to end of region */
326 resv
->adds_in_progress
--;
327 spin_unlock(&resv
->lock
);
333 * Examine the existing reserve map and determine how many
334 * huge pages in the specified range [f, t) are NOT currently
335 * represented. This routine is called before a subsequent
336 * call to region_add that will actually modify the reserve
337 * map to add the specified range [f, t). region_chg does
338 * not change the number of huge pages represented by the
339 * map. However, if the existing regions in the map can not
340 * be expanded to represent the new range, a new file_region
341 * structure is added to the map as a placeholder. This is
342 * so that the subsequent region_add call will have all the
343 * regions it needs and will not fail.
345 * Upon entry, region_chg will also examine the cache of region descriptors
346 * associated with the map. If there are not enough descriptors cached, one
347 * will be allocated for the in progress add operation.
349 * Returns the number of huge pages that need to be added to the existing
350 * reservation map for the range [f, t). This number is greater or equal to
351 * zero. -ENOMEM is returned if a new file_region structure or cache entry
352 * is needed and can not be allocated.
354 static long region_chg(struct resv_map
*resv
, long f
, long t
)
356 struct list_head
*head
= &resv
->regions
;
357 struct file_region
*rg
, *nrg
= NULL
;
361 spin_lock(&resv
->lock
);
363 resv
->adds_in_progress
++;
366 * Check for sufficient descriptors in the cache to accommodate
367 * the number of in progress add operations.
369 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
370 struct file_region
*trg
;
372 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
373 /* Must drop lock to allocate a new descriptor. */
374 resv
->adds_in_progress
--;
375 spin_unlock(&resv
->lock
);
377 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
383 spin_lock(&resv
->lock
);
384 list_add(&trg
->link
, &resv
->region_cache
);
385 resv
->region_cache_count
++;
389 /* Locate the region we are before or in. */
390 list_for_each_entry(rg
, head
, link
)
394 /* If we are below the current region then a new region is required.
395 * Subtle, allocate a new region at the position but make it zero
396 * size such that we can guarantee to record the reservation. */
397 if (&rg
->link
== head
|| t
< rg
->from
) {
399 resv
->adds_in_progress
--;
400 spin_unlock(&resv
->lock
);
401 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
407 INIT_LIST_HEAD(&nrg
->link
);
411 list_add(&nrg
->link
, rg
->link
.prev
);
416 /* Round our left edge to the current segment if it encloses us. */
421 /* Check for and consume any regions we now overlap with. */
422 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
423 if (&rg
->link
== head
)
428 /* We overlap with this area, if it extends further than
429 * us then we must extend ourselves. Account for its
430 * existing reservation. */
435 chg
-= rg
->to
- rg
->from
;
439 spin_unlock(&resv
->lock
);
440 /* We already know we raced and no longer need the new region */
444 spin_unlock(&resv
->lock
);
449 * Abort the in progress add operation. The adds_in_progress field
450 * of the resv_map keeps track of the operations in progress between
451 * calls to region_chg and region_add. Operations are sometimes
452 * aborted after the call to region_chg. In such cases, region_abort
453 * is called to decrement the adds_in_progress counter.
455 * NOTE: The range arguments [f, t) are not needed or used in this
456 * routine. They are kept to make reading the calling code easier as
457 * arguments will match the associated region_chg call.
459 static void region_abort(struct resv_map
*resv
, long f
, long t
)
461 spin_lock(&resv
->lock
);
462 VM_BUG_ON(!resv
->region_cache_count
);
463 resv
->adds_in_progress
--;
464 spin_unlock(&resv
->lock
);
468 * Delete the specified range [f, t) from the reserve map. If the
469 * t parameter is LONG_MAX, this indicates that ALL regions after f
470 * should be deleted. Locate the regions which intersect [f, t)
471 * and either trim, delete or split the existing regions.
473 * Returns the number of huge pages deleted from the reserve map.
474 * In the normal case, the return value is zero or more. In the
475 * case where a region must be split, a new region descriptor must
476 * be allocated. If the allocation fails, -ENOMEM will be returned.
477 * NOTE: If the parameter t == LONG_MAX, then we will never split
478 * a region and possibly return -ENOMEM. Callers specifying
479 * t == LONG_MAX do not need to check for -ENOMEM error.
481 static long region_del(struct resv_map
*resv
, long f
, long t
)
483 struct list_head
*head
= &resv
->regions
;
484 struct file_region
*rg
, *trg
;
485 struct file_region
*nrg
= NULL
;
489 spin_lock(&resv
->lock
);
490 list_for_each_entry_safe(rg
, trg
, head
, link
) {
492 * Skip regions before the range to be deleted. file_region
493 * ranges are normally of the form [from, to). However, there
494 * may be a "placeholder" entry in the map which is of the form
495 * (from, to) with from == to. Check for placeholder entries
496 * at the beginning of the range to be deleted.
498 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
504 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
506 * Check for an entry in the cache before dropping
507 * lock and attempting allocation.
510 resv
->region_cache_count
> resv
->adds_in_progress
) {
511 nrg
= list_first_entry(&resv
->region_cache
,
514 list_del(&nrg
->link
);
515 resv
->region_cache_count
--;
519 spin_unlock(&resv
->lock
);
520 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
528 /* New entry for end of split region */
531 INIT_LIST_HEAD(&nrg
->link
);
533 /* Original entry is trimmed */
536 list_add(&nrg
->link
, &rg
->link
);
541 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
542 del
+= rg
->to
- rg
->from
;
548 if (f
<= rg
->from
) { /* Trim beginning of region */
551 } else { /* Trim end of region */
557 spin_unlock(&resv
->lock
);
563 * A rare out of memory error was encountered which prevented removal of
564 * the reserve map region for a page. The huge page itself was free'ed
565 * and removed from the page cache. This routine will adjust the subpool
566 * usage count, and the global reserve count if needed. By incrementing
567 * these counts, the reserve map entry which could not be deleted will
568 * appear as a "reserved" entry instead of simply dangling with incorrect
571 void hugetlb_fix_reserve_counts(struct inode
*inode
)
573 struct hugepage_subpool
*spool
= subpool_inode(inode
);
576 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
578 struct hstate
*h
= hstate_inode(inode
);
580 hugetlb_acct_memory(h
, 1);
585 * Count and return the number of huge pages in the reserve map
586 * that intersect with the range [f, t).
588 static long region_count(struct resv_map
*resv
, long f
, long t
)
590 struct list_head
*head
= &resv
->regions
;
591 struct file_region
*rg
;
594 spin_lock(&resv
->lock
);
595 /* Locate each segment we overlap with, and count that overlap. */
596 list_for_each_entry(rg
, head
, link
) {
605 seg_from
= max(rg
->from
, f
);
606 seg_to
= min(rg
->to
, t
);
608 chg
+= seg_to
- seg_from
;
610 spin_unlock(&resv
->lock
);
616 * Convert the address within this vma to the page offset within
617 * the mapping, in pagecache page units; huge pages here.
619 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
620 struct vm_area_struct
*vma
, unsigned long address
)
622 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
623 (vma
->vm_pgoff
>> huge_page_order(h
));
626 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
627 unsigned long address
)
629 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
631 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
634 * Return the size of the pages allocated when backing a VMA. In the majority
635 * cases this will be same size as used by the page table entries.
637 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
639 struct hstate
*hstate
;
641 if (!is_vm_hugetlb_page(vma
))
644 hstate
= hstate_vma(vma
);
646 return 1UL << huge_page_shift(hstate
);
648 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
651 * Return the page size being used by the MMU to back a VMA. In the majority
652 * of cases, the page size used by the kernel matches the MMU size. On
653 * architectures where it differs, an architecture-specific version of this
654 * function is required.
656 #ifndef vma_mmu_pagesize
657 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
659 return vma_kernel_pagesize(vma
);
664 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
665 * bits of the reservation map pointer, which are always clear due to
668 #define HPAGE_RESV_OWNER (1UL << 0)
669 #define HPAGE_RESV_UNMAPPED (1UL << 1)
670 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
673 * These helpers are used to track how many pages are reserved for
674 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
675 * is guaranteed to have their future faults succeed.
677 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
678 * the reserve counters are updated with the hugetlb_lock held. It is safe
679 * to reset the VMA at fork() time as it is not in use yet and there is no
680 * chance of the global counters getting corrupted as a result of the values.
682 * The private mapping reservation is represented in a subtly different
683 * manner to a shared mapping. A shared mapping has a region map associated
684 * with the underlying file, this region map represents the backing file
685 * pages which have ever had a reservation assigned which this persists even
686 * after the page is instantiated. A private mapping has a region map
687 * associated with the original mmap which is attached to all VMAs which
688 * reference it, this region map represents those offsets which have consumed
689 * reservation ie. where pages have been instantiated.
691 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
693 return (unsigned long)vma
->vm_private_data
;
696 static void set_vma_private_data(struct vm_area_struct
*vma
,
699 vma
->vm_private_data
= (void *)value
;
702 struct resv_map
*resv_map_alloc(void)
704 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
705 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
707 if (!resv_map
|| !rg
) {
713 kref_init(&resv_map
->refs
);
714 spin_lock_init(&resv_map
->lock
);
715 INIT_LIST_HEAD(&resv_map
->regions
);
717 resv_map
->adds_in_progress
= 0;
719 INIT_LIST_HEAD(&resv_map
->region_cache
);
720 list_add(&rg
->link
, &resv_map
->region_cache
);
721 resv_map
->region_cache_count
= 1;
726 void resv_map_release(struct kref
*ref
)
728 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
729 struct list_head
*head
= &resv_map
->region_cache
;
730 struct file_region
*rg
, *trg
;
732 /* Clear out any active regions before we release the map. */
733 region_del(resv_map
, 0, LONG_MAX
);
735 /* ... and any entries left in the cache */
736 list_for_each_entry_safe(rg
, trg
, head
, link
) {
741 VM_BUG_ON(resv_map
->adds_in_progress
);
746 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
748 return inode
->i_mapping
->private_data
;
751 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
753 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
754 if (vma
->vm_flags
& VM_MAYSHARE
) {
755 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
756 struct inode
*inode
= mapping
->host
;
758 return inode_resv_map(inode
);
761 return (struct resv_map
*)(get_vma_private_data(vma
) &
766 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
768 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
769 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
771 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
772 HPAGE_RESV_MASK
) | (unsigned long)map
);
775 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
777 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
778 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
780 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
783 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
785 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
787 return (get_vma_private_data(vma
) & flag
) != 0;
790 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
791 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
793 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
794 if (!(vma
->vm_flags
& VM_MAYSHARE
))
795 vma
->vm_private_data
= (void *)0;
798 /* Returns true if the VMA has associated reserve pages */
799 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
801 if (vma
->vm_flags
& VM_NORESERVE
) {
803 * This address is already reserved by other process(chg == 0),
804 * so, we should decrement reserved count. Without decrementing,
805 * reserve count remains after releasing inode, because this
806 * allocated page will go into page cache and is regarded as
807 * coming from reserved pool in releasing step. Currently, we
808 * don't have any other solution to deal with this situation
809 * properly, so add work-around here.
811 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
817 /* Shared mappings always use reserves */
818 if (vma
->vm_flags
& VM_MAYSHARE
) {
820 * We know VM_NORESERVE is not set. Therefore, there SHOULD
821 * be a region map for all pages. The only situation where
822 * there is no region map is if a hole was punched via
823 * fallocate. In this case, there really are no reverves to
824 * use. This situation is indicated if chg != 0.
833 * Only the process that called mmap() has reserves for
836 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
838 * Like the shared case above, a hole punch or truncate
839 * could have been performed on the private mapping.
840 * Examine the value of chg to determine if reserves
841 * actually exist or were previously consumed.
842 * Very Subtle - The value of chg comes from a previous
843 * call to vma_needs_reserves(). The reserve map for
844 * private mappings has different (opposite) semantics
845 * than that of shared mappings. vma_needs_reserves()
846 * has already taken this difference in semantics into
847 * account. Therefore, the meaning of chg is the same
848 * as in the shared case above. Code could easily be
849 * combined, but keeping it separate draws attention to
850 * subtle differences.
861 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
863 int nid
= page_to_nid(page
);
864 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
865 h
->free_huge_pages
++;
866 h
->free_huge_pages_node
[nid
]++;
869 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
873 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
874 if (!is_migrate_isolate_page(page
))
877 * if 'non-isolated free hugepage' not found on the list,
878 * the allocation fails.
880 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
882 list_move(&page
->lru
, &h
->hugepage_activelist
);
883 set_page_refcounted(page
);
884 h
->free_huge_pages
--;
885 h
->free_huge_pages_node
[nid
]--;
889 /* Movability of hugepages depends on migration support. */
890 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
892 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
893 return GFP_HIGHUSER_MOVABLE
;
898 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
899 struct vm_area_struct
*vma
,
900 unsigned long address
, int avoid_reserve
,
903 struct page
*page
= NULL
;
904 struct mempolicy
*mpol
;
905 nodemask_t
*nodemask
;
906 struct zonelist
*zonelist
;
909 unsigned int cpuset_mems_cookie
;
912 * A child process with MAP_PRIVATE mappings created by their parent
913 * have no page reserves. This check ensures that reservations are
914 * not "stolen". The child may still get SIGKILLed
916 if (!vma_has_reserves(vma
, chg
) &&
917 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
920 /* If reserves cannot be used, ensure enough pages are in the pool */
921 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
925 cpuset_mems_cookie
= read_mems_allowed_begin();
926 zonelist
= huge_zonelist(vma
, address
,
927 htlb_alloc_mask(h
), &mpol
, &nodemask
);
929 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
930 MAX_NR_ZONES
- 1, nodemask
) {
931 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
932 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
936 if (!vma_has_reserves(vma
, chg
))
939 SetPagePrivate(page
);
940 h
->resv_huge_pages
--;
947 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
956 * common helper functions for hstate_next_node_to_{alloc|free}.
957 * We may have allocated or freed a huge page based on a different
958 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
959 * be outside of *nodes_allowed. Ensure that we use an allowed
960 * node for alloc or free.
962 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
964 nid
= next_node_in(nid
, *nodes_allowed
);
965 VM_BUG_ON(nid
>= MAX_NUMNODES
);
970 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
972 if (!node_isset(nid
, *nodes_allowed
))
973 nid
= next_node_allowed(nid
, nodes_allowed
);
978 * returns the previously saved node ["this node"] from which to
979 * allocate a persistent huge page for the pool and advance the
980 * next node from which to allocate, handling wrap at end of node
983 static int hstate_next_node_to_alloc(struct hstate
*h
,
984 nodemask_t
*nodes_allowed
)
988 VM_BUG_ON(!nodes_allowed
);
990 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
991 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
997 * helper for free_pool_huge_page() - return the previously saved
998 * node ["this node"] from which to free a huge page. Advance the
999 * next node id whether or not we find a free huge page to free so
1000 * that the next attempt to free addresses the next node.
1002 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1006 VM_BUG_ON(!nodes_allowed
);
1008 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1009 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1014 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1015 for (nr_nodes = nodes_weight(*mask); \
1017 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1020 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1021 for (nr_nodes = nodes_weight(*mask); \
1023 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1026 #if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \
1027 ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1028 defined(CONFIG_CMA))
1029 static void destroy_compound_gigantic_page(struct page
*page
,
1033 int nr_pages
= 1 << order
;
1034 struct page
*p
= page
+ 1;
1036 atomic_set(compound_mapcount_ptr(page
), 0);
1037 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1038 clear_compound_head(p
);
1039 set_page_refcounted(p
);
1042 set_compound_order(page
, 0);
1043 __ClearPageHead(page
);
1046 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1048 free_contig_range(page_to_pfn(page
), 1 << order
);
1051 static int __alloc_gigantic_page(unsigned long start_pfn
,
1052 unsigned long nr_pages
)
1054 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1055 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
1058 static bool pfn_range_valid_gigantic(struct zone
*z
,
1059 unsigned long start_pfn
, unsigned long nr_pages
)
1061 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1064 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1068 page
= pfn_to_page(i
);
1070 if (page_zone(page
) != z
)
1073 if (PageReserved(page
))
1076 if (page_count(page
) > 0)
1086 static bool zone_spans_last_pfn(const struct zone
*zone
,
1087 unsigned long start_pfn
, unsigned long nr_pages
)
1089 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1090 return zone_spans_pfn(zone
, last_pfn
);
1093 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1095 unsigned long nr_pages
= 1 << order
;
1096 unsigned long ret
, pfn
, flags
;
1099 z
= NODE_DATA(nid
)->node_zones
;
1100 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1101 spin_lock_irqsave(&z
->lock
, flags
);
1103 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1104 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1105 if (pfn_range_valid_gigantic(z
, pfn
, nr_pages
)) {
1107 * We release the zone lock here because
1108 * alloc_contig_range() will also lock the zone
1109 * at some point. If there's an allocation
1110 * spinning on this lock, it may win the race
1111 * and cause alloc_contig_range() to fail...
1113 spin_unlock_irqrestore(&z
->lock
, flags
);
1114 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1116 return pfn_to_page(pfn
);
1117 spin_lock_irqsave(&z
->lock
, flags
);
1122 spin_unlock_irqrestore(&z
->lock
, flags
);
1128 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1129 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1131 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1135 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1137 prep_compound_gigantic_page(page
, huge_page_order(h
));
1138 prep_new_huge_page(h
, page
, nid
);
1144 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1145 nodemask_t
*nodes_allowed
)
1147 struct page
*page
= NULL
;
1150 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1151 page
= alloc_fresh_gigantic_page_node(h
, node
);
1159 static inline bool gigantic_page_supported(void) { return true; }
1161 static inline bool gigantic_page_supported(void) { return false; }
1162 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1163 static inline void destroy_compound_gigantic_page(struct page
*page
,
1164 unsigned int order
) { }
1165 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1166 nodemask_t
*nodes_allowed
) { return 0; }
1169 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1173 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1177 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1178 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1179 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1180 1 << PG_referenced
| 1 << PG_dirty
|
1181 1 << PG_active
| 1 << PG_private
|
1184 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1185 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1186 set_page_refcounted(page
);
1187 if (hstate_is_gigantic(h
)) {
1188 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1189 free_gigantic_page(page
, huge_page_order(h
));
1191 __free_pages(page
, huge_page_order(h
));
1195 struct hstate
*size_to_hstate(unsigned long size
)
1199 for_each_hstate(h
) {
1200 if (huge_page_size(h
) == size
)
1207 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1208 * to hstate->hugepage_activelist.)
1210 * This function can be called for tail pages, but never returns true for them.
1212 bool page_huge_active(struct page
*page
)
1214 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1215 return PageHead(page
) && PagePrivate(&page
[1]);
1218 /* never called for tail page */
1219 static void set_page_huge_active(struct page
*page
)
1221 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1222 SetPagePrivate(&page
[1]);
1225 static void clear_page_huge_active(struct page
*page
)
1227 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1228 ClearPagePrivate(&page
[1]);
1231 void free_huge_page(struct page
*page
)
1234 * Can't pass hstate in here because it is called from the
1235 * compound page destructor.
1237 struct hstate
*h
= page_hstate(page
);
1238 int nid
= page_to_nid(page
);
1239 struct hugepage_subpool
*spool
=
1240 (struct hugepage_subpool
*)page_private(page
);
1241 bool restore_reserve
;
1243 set_page_private(page
, 0);
1244 page
->mapping
= NULL
;
1245 VM_BUG_ON_PAGE(page_count(page
), page
);
1246 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1247 restore_reserve
= PagePrivate(page
);
1248 ClearPagePrivate(page
);
1251 * A return code of zero implies that the subpool will be under its
1252 * minimum size if the reservation is not restored after page is free.
1253 * Therefore, force restore_reserve operation.
1255 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1256 restore_reserve
= true;
1258 spin_lock(&hugetlb_lock
);
1259 clear_page_huge_active(page
);
1260 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1261 pages_per_huge_page(h
), page
);
1262 if (restore_reserve
)
1263 h
->resv_huge_pages
++;
1265 if (h
->surplus_huge_pages_node
[nid
]) {
1266 /* remove the page from active list */
1267 list_del(&page
->lru
);
1268 update_and_free_page(h
, page
);
1269 h
->surplus_huge_pages
--;
1270 h
->surplus_huge_pages_node
[nid
]--;
1272 arch_clear_hugepage_flags(page
);
1273 enqueue_huge_page(h
, page
);
1275 spin_unlock(&hugetlb_lock
);
1278 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1280 INIT_LIST_HEAD(&page
->lru
);
1281 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1282 spin_lock(&hugetlb_lock
);
1283 set_hugetlb_cgroup(page
, NULL
);
1285 h
->nr_huge_pages_node
[nid
]++;
1286 spin_unlock(&hugetlb_lock
);
1287 put_page(page
); /* free it into the hugepage allocator */
1290 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1293 int nr_pages
= 1 << order
;
1294 struct page
*p
= page
+ 1;
1296 /* we rely on prep_new_huge_page to set the destructor */
1297 set_compound_order(page
, order
);
1298 __ClearPageReserved(page
);
1299 __SetPageHead(page
);
1300 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1302 * For gigantic hugepages allocated through bootmem at
1303 * boot, it's safer to be consistent with the not-gigantic
1304 * hugepages and clear the PG_reserved bit from all tail pages
1305 * too. Otherwse drivers using get_user_pages() to access tail
1306 * pages may get the reference counting wrong if they see
1307 * PG_reserved set on a tail page (despite the head page not
1308 * having PG_reserved set). Enforcing this consistency between
1309 * head and tail pages allows drivers to optimize away a check
1310 * on the head page when they need know if put_page() is needed
1311 * after get_user_pages().
1313 __ClearPageReserved(p
);
1314 set_page_count(p
, 0);
1315 set_compound_head(p
, page
);
1317 atomic_set(compound_mapcount_ptr(page
), -1);
1321 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1322 * transparent huge pages. See the PageTransHuge() documentation for more
1325 int PageHuge(struct page
*page
)
1327 if (!PageCompound(page
))
1330 page
= compound_head(page
);
1331 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1333 EXPORT_SYMBOL_GPL(PageHuge
);
1336 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1337 * normal or transparent huge pages.
1339 int PageHeadHuge(struct page
*page_head
)
1341 if (!PageHead(page_head
))
1344 return get_compound_page_dtor(page_head
) == free_huge_page
;
1347 pgoff_t
__basepage_index(struct page
*page
)
1349 struct page
*page_head
= compound_head(page
);
1350 pgoff_t index
= page_index(page_head
);
1351 unsigned long compound_idx
;
1353 if (!PageHuge(page_head
))
1354 return page_index(page
);
1356 if (compound_order(page_head
) >= MAX_ORDER
)
1357 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1359 compound_idx
= page
- page_head
;
1361 return (index
<< compound_order(page_head
)) + compound_idx
;
1364 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1368 page
= __alloc_pages_node(nid
,
1369 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1370 __GFP_REPEAT
|__GFP_NOWARN
,
1371 huge_page_order(h
));
1373 prep_new_huge_page(h
, page
, nid
);
1379 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1385 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1386 page
= alloc_fresh_huge_page_node(h
, node
);
1394 count_vm_event(HTLB_BUDDY_PGALLOC
);
1396 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1402 * Free huge page from pool from next node to free.
1403 * Attempt to keep persistent huge pages more or less
1404 * balanced over allowed nodes.
1405 * Called with hugetlb_lock locked.
1407 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1413 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1415 * If we're returning unused surplus pages, only examine
1416 * nodes with surplus pages.
1418 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1419 !list_empty(&h
->hugepage_freelists
[node
])) {
1421 list_entry(h
->hugepage_freelists
[node
].next
,
1423 list_del(&page
->lru
);
1424 h
->free_huge_pages
--;
1425 h
->free_huge_pages_node
[node
]--;
1427 h
->surplus_huge_pages
--;
1428 h
->surplus_huge_pages_node
[node
]--;
1430 update_and_free_page(h
, page
);
1440 * Dissolve a given free hugepage into free buddy pages. This function does
1441 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1442 * number of free hugepages would be reduced below the number of reserved
1445 static int dissolve_free_huge_page(struct page
*page
)
1449 spin_lock(&hugetlb_lock
);
1450 if (PageHuge(page
) && !page_count(page
)) {
1451 struct page
*head
= compound_head(page
);
1452 struct hstate
*h
= page_hstate(head
);
1453 int nid
= page_to_nid(head
);
1454 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0) {
1458 list_del(&head
->lru
);
1459 h
->free_huge_pages
--;
1460 h
->free_huge_pages_node
[nid
]--;
1461 h
->max_huge_pages
--;
1462 update_and_free_page(h
, head
);
1465 spin_unlock(&hugetlb_lock
);
1470 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1471 * make specified memory blocks removable from the system.
1472 * Note that this will dissolve a free gigantic hugepage completely, if any
1473 * part of it lies within the given range.
1474 * Also note that if dissolve_free_huge_page() returns with an error, all
1475 * free hugepages that were dissolved before that error are lost.
1477 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1483 if (!hugepages_supported())
1486 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1487 page
= pfn_to_page(pfn
);
1488 if (PageHuge(page
) && !page_count(page
)) {
1489 rc
= dissolve_free_huge_page(page
);
1499 * There are 3 ways this can get called:
1500 * 1. With vma+addr: we use the VMA's memory policy
1501 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1502 * page from any node, and let the buddy allocator itself figure
1504 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1505 * strictly from 'nid'
1507 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1508 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1510 int order
= huge_page_order(h
);
1511 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1512 unsigned int cpuset_mems_cookie
;
1515 * We need a VMA to get a memory policy. If we do not
1516 * have one, we use the 'nid' argument.
1518 * The mempolicy stuff below has some non-inlined bits
1519 * and calls ->vm_ops. That makes it hard to optimize at
1520 * compile-time, even when NUMA is off and it does
1521 * nothing. This helps the compiler optimize it out.
1523 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1525 * If a specific node is requested, make sure to
1526 * get memory from there, but only when a node
1527 * is explicitly specified.
1529 if (nid
!= NUMA_NO_NODE
)
1530 gfp
|= __GFP_THISNODE
;
1532 * Make sure to call something that can handle
1535 return alloc_pages_node(nid
, gfp
, order
);
1539 * OK, so we have a VMA. Fetch the mempolicy and try to
1540 * allocate a huge page with it. We will only reach this
1541 * when CONFIG_NUMA=y.
1545 struct mempolicy
*mpol
;
1546 struct zonelist
*zl
;
1547 nodemask_t
*nodemask
;
1549 cpuset_mems_cookie
= read_mems_allowed_begin();
1550 zl
= huge_zonelist(vma
, addr
, gfp
, &mpol
, &nodemask
);
1551 mpol_cond_put(mpol
);
1552 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nodemask
);
1555 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1561 * There are two ways to allocate a huge page:
1562 * 1. When you have a VMA and an address (like a fault)
1563 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1565 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1566 * this case which signifies that the allocation should be done with
1567 * respect for the VMA's memory policy.
1569 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1570 * implies that memory policies will not be taken in to account.
1572 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1573 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1578 if (hstate_is_gigantic(h
))
1582 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1583 * This makes sure the caller is picking _one_ of the modes with which
1584 * we can call this function, not both.
1586 if (vma
|| (addr
!= -1)) {
1587 VM_WARN_ON_ONCE(addr
== -1);
1588 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1591 * Assume we will successfully allocate the surplus page to
1592 * prevent racing processes from causing the surplus to exceed
1595 * This however introduces a different race, where a process B
1596 * tries to grow the static hugepage pool while alloc_pages() is
1597 * called by process A. B will only examine the per-node
1598 * counters in determining if surplus huge pages can be
1599 * converted to normal huge pages in adjust_pool_surplus(). A
1600 * won't be able to increment the per-node counter, until the
1601 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1602 * no more huge pages can be converted from surplus to normal
1603 * state (and doesn't try to convert again). Thus, we have a
1604 * case where a surplus huge page exists, the pool is grown, and
1605 * the surplus huge page still exists after, even though it
1606 * should just have been converted to a normal huge page. This
1607 * does not leak memory, though, as the hugepage will be freed
1608 * once it is out of use. It also does not allow the counters to
1609 * go out of whack in adjust_pool_surplus() as we don't modify
1610 * the node values until we've gotten the hugepage and only the
1611 * per-node value is checked there.
1613 spin_lock(&hugetlb_lock
);
1614 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1615 spin_unlock(&hugetlb_lock
);
1619 h
->surplus_huge_pages
++;
1621 spin_unlock(&hugetlb_lock
);
1623 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1625 spin_lock(&hugetlb_lock
);
1627 INIT_LIST_HEAD(&page
->lru
);
1628 r_nid
= page_to_nid(page
);
1629 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1630 set_hugetlb_cgroup(page
, NULL
);
1632 * We incremented the global counters already
1634 h
->nr_huge_pages_node
[r_nid
]++;
1635 h
->surplus_huge_pages_node
[r_nid
]++;
1636 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1639 h
->surplus_huge_pages
--;
1640 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1642 spin_unlock(&hugetlb_lock
);
1648 * Allocate a huge page from 'nid'. Note, 'nid' may be
1649 * NUMA_NO_NODE, which means that it may be allocated
1653 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1655 unsigned long addr
= -1;
1657 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1661 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1664 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1665 struct vm_area_struct
*vma
, unsigned long addr
)
1667 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1671 * This allocation function is useful in the context where vma is irrelevant.
1672 * E.g. soft-offlining uses this function because it only cares physical
1673 * address of error page.
1675 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1677 struct page
*page
= NULL
;
1679 spin_lock(&hugetlb_lock
);
1680 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1681 page
= dequeue_huge_page_node(h
, nid
);
1682 spin_unlock(&hugetlb_lock
);
1685 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1691 * Increase the hugetlb pool such that it can accommodate a reservation
1694 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1696 struct list_head surplus_list
;
1697 struct page
*page
, *tmp
;
1699 int needed
, allocated
;
1700 bool alloc_ok
= true;
1702 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1704 h
->resv_huge_pages
+= delta
;
1709 INIT_LIST_HEAD(&surplus_list
);
1713 spin_unlock(&hugetlb_lock
);
1714 for (i
= 0; i
< needed
; i
++) {
1715 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1720 list_add(&page
->lru
, &surplus_list
);
1725 * After retaking hugetlb_lock, we need to recalculate 'needed'
1726 * because either resv_huge_pages or free_huge_pages may have changed.
1728 spin_lock(&hugetlb_lock
);
1729 needed
= (h
->resv_huge_pages
+ delta
) -
1730 (h
->free_huge_pages
+ allocated
);
1735 * We were not able to allocate enough pages to
1736 * satisfy the entire reservation so we free what
1737 * we've allocated so far.
1742 * The surplus_list now contains _at_least_ the number of extra pages
1743 * needed to accommodate the reservation. Add the appropriate number
1744 * of pages to the hugetlb pool and free the extras back to the buddy
1745 * allocator. Commit the entire reservation here to prevent another
1746 * process from stealing the pages as they are added to the pool but
1747 * before they are reserved.
1749 needed
+= allocated
;
1750 h
->resv_huge_pages
+= delta
;
1753 /* Free the needed pages to the hugetlb pool */
1754 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1758 * This page is now managed by the hugetlb allocator and has
1759 * no users -- drop the buddy allocator's reference.
1761 put_page_testzero(page
);
1762 VM_BUG_ON_PAGE(page_count(page
), page
);
1763 enqueue_huge_page(h
, page
);
1766 spin_unlock(&hugetlb_lock
);
1768 /* Free unnecessary surplus pages to the buddy allocator */
1769 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1771 spin_lock(&hugetlb_lock
);
1777 * This routine has two main purposes:
1778 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1779 * in unused_resv_pages. This corresponds to the prior adjustments made
1780 * to the associated reservation map.
1781 * 2) Free any unused surplus pages that may have been allocated to satisfy
1782 * the reservation. As many as unused_resv_pages may be freed.
1784 * Called with hugetlb_lock held. However, the lock could be dropped (and
1785 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1786 * we must make sure nobody else can claim pages we are in the process of
1787 * freeing. Do this by ensuring resv_huge_page always is greater than the
1788 * number of huge pages we plan to free when dropping the lock.
1790 static void return_unused_surplus_pages(struct hstate
*h
,
1791 unsigned long unused_resv_pages
)
1793 unsigned long nr_pages
;
1795 /* Cannot return gigantic pages currently */
1796 if (hstate_is_gigantic(h
))
1800 * Part (or even all) of the reservation could have been backed
1801 * by pre-allocated pages. Only free surplus pages.
1803 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1806 * We want to release as many surplus pages as possible, spread
1807 * evenly across all nodes with memory. Iterate across these nodes
1808 * until we can no longer free unreserved surplus pages. This occurs
1809 * when the nodes with surplus pages have no free pages.
1810 * free_pool_huge_page() will balance the the freed pages across the
1811 * on-line nodes with memory and will handle the hstate accounting.
1813 * Note that we decrement resv_huge_pages as we free the pages. If
1814 * we drop the lock, resv_huge_pages will still be sufficiently large
1815 * to cover subsequent pages we may free.
1817 while (nr_pages
--) {
1818 h
->resv_huge_pages
--;
1819 unused_resv_pages
--;
1820 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1822 cond_resched_lock(&hugetlb_lock
);
1826 /* Fully uncommit the reservation */
1827 h
->resv_huge_pages
-= unused_resv_pages
;
1832 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1833 * are used by the huge page allocation routines to manage reservations.
1835 * vma_needs_reservation is called to determine if the huge page at addr
1836 * within the vma has an associated reservation. If a reservation is
1837 * needed, the value 1 is returned. The caller is then responsible for
1838 * managing the global reservation and subpool usage counts. After
1839 * the huge page has been allocated, vma_commit_reservation is called
1840 * to add the page to the reservation map. If the page allocation fails,
1841 * the reservation must be ended instead of committed. vma_end_reservation
1842 * is called in such cases.
1844 * In the normal case, vma_commit_reservation returns the same value
1845 * as the preceding vma_needs_reservation call. The only time this
1846 * is not the case is if a reserve map was changed between calls. It
1847 * is the responsibility of the caller to notice the difference and
1848 * take appropriate action.
1850 * vma_add_reservation is used in error paths where a reservation must
1851 * be restored when a newly allocated huge page must be freed. It is
1852 * to be called after calling vma_needs_reservation to determine if a
1853 * reservation exists.
1855 enum vma_resv_mode
{
1861 static long __vma_reservation_common(struct hstate
*h
,
1862 struct vm_area_struct
*vma
, unsigned long addr
,
1863 enum vma_resv_mode mode
)
1865 struct resv_map
*resv
;
1869 resv
= vma_resv_map(vma
);
1873 idx
= vma_hugecache_offset(h
, vma
, addr
);
1875 case VMA_NEEDS_RESV
:
1876 ret
= region_chg(resv
, idx
, idx
+ 1);
1878 case VMA_COMMIT_RESV
:
1879 ret
= region_add(resv
, idx
, idx
+ 1);
1882 region_abort(resv
, idx
, idx
+ 1);
1886 if (vma
->vm_flags
& VM_MAYSHARE
)
1887 ret
= region_add(resv
, idx
, idx
+ 1);
1889 region_abort(resv
, idx
, idx
+ 1);
1890 ret
= region_del(resv
, idx
, idx
+ 1);
1897 if (vma
->vm_flags
& VM_MAYSHARE
)
1899 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1901 * In most cases, reserves always exist for private mappings.
1902 * However, a file associated with mapping could have been
1903 * hole punched or truncated after reserves were consumed.
1904 * As subsequent fault on such a range will not use reserves.
1905 * Subtle - The reserve map for private mappings has the
1906 * opposite meaning than that of shared mappings. If NO
1907 * entry is in the reserve map, it means a reservation exists.
1908 * If an entry exists in the reserve map, it means the
1909 * reservation has already been consumed. As a result, the
1910 * return value of this routine is the opposite of the
1911 * value returned from reserve map manipulation routines above.
1919 return ret
< 0 ? ret
: 0;
1922 static long vma_needs_reservation(struct hstate
*h
,
1923 struct vm_area_struct
*vma
, unsigned long addr
)
1925 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1928 static long vma_commit_reservation(struct hstate
*h
,
1929 struct vm_area_struct
*vma
, unsigned long addr
)
1931 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1934 static void vma_end_reservation(struct hstate
*h
,
1935 struct vm_area_struct
*vma
, unsigned long addr
)
1937 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1940 static long vma_add_reservation(struct hstate
*h
,
1941 struct vm_area_struct
*vma
, unsigned long addr
)
1943 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
1947 * This routine is called to restore a reservation on error paths. In the
1948 * specific error paths, a huge page was allocated (via alloc_huge_page)
1949 * and is about to be freed. If a reservation for the page existed,
1950 * alloc_huge_page would have consumed the reservation and set PagePrivate
1951 * in the newly allocated page. When the page is freed via free_huge_page,
1952 * the global reservation count will be incremented if PagePrivate is set.
1953 * However, free_huge_page can not adjust the reserve map. Adjust the
1954 * reserve map here to be consistent with global reserve count adjustments
1955 * to be made by free_huge_page.
1957 static void restore_reserve_on_error(struct hstate
*h
,
1958 struct vm_area_struct
*vma
, unsigned long address
,
1961 if (unlikely(PagePrivate(page
))) {
1962 long rc
= vma_needs_reservation(h
, vma
, address
);
1964 if (unlikely(rc
< 0)) {
1966 * Rare out of memory condition in reserve map
1967 * manipulation. Clear PagePrivate so that
1968 * global reserve count will not be incremented
1969 * by free_huge_page. This will make it appear
1970 * as though the reservation for this page was
1971 * consumed. This may prevent the task from
1972 * faulting in the page at a later time. This
1973 * is better than inconsistent global huge page
1974 * accounting of reserve counts.
1976 ClearPagePrivate(page
);
1978 rc
= vma_add_reservation(h
, vma
, address
);
1979 if (unlikely(rc
< 0))
1981 * See above comment about rare out of
1984 ClearPagePrivate(page
);
1986 vma_end_reservation(h
, vma
, address
);
1990 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1991 unsigned long addr
, int avoid_reserve
)
1993 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1994 struct hstate
*h
= hstate_vma(vma
);
1996 long map_chg
, map_commit
;
1999 struct hugetlb_cgroup
*h_cg
;
2001 idx
= hstate_index(h
);
2003 * Examine the region/reserve map to determine if the process
2004 * has a reservation for the page to be allocated. A return
2005 * code of zero indicates a reservation exists (no change).
2007 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2009 return ERR_PTR(-ENOMEM
);
2012 * Processes that did not create the mapping will have no
2013 * reserves as indicated by the region/reserve map. Check
2014 * that the allocation will not exceed the subpool limit.
2015 * Allocations for MAP_NORESERVE mappings also need to be
2016 * checked against any subpool limit.
2018 if (map_chg
|| avoid_reserve
) {
2019 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2021 vma_end_reservation(h
, vma
, addr
);
2022 return ERR_PTR(-ENOSPC
);
2026 * Even though there was no reservation in the region/reserve
2027 * map, there could be reservations associated with the
2028 * subpool that can be used. This would be indicated if the
2029 * return value of hugepage_subpool_get_pages() is zero.
2030 * However, if avoid_reserve is specified we still avoid even
2031 * the subpool reservations.
2037 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2039 goto out_subpool_put
;
2041 spin_lock(&hugetlb_lock
);
2043 * glb_chg is passed to indicate whether or not a page must be taken
2044 * from the global free pool (global change). gbl_chg == 0 indicates
2045 * a reservation exists for the allocation.
2047 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2049 spin_unlock(&hugetlb_lock
);
2050 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2052 goto out_uncharge_cgroup
;
2053 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2054 SetPagePrivate(page
);
2055 h
->resv_huge_pages
--;
2057 spin_lock(&hugetlb_lock
);
2058 list_move(&page
->lru
, &h
->hugepage_activelist
);
2061 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2062 spin_unlock(&hugetlb_lock
);
2064 set_page_private(page
, (unsigned long)spool
);
2066 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2067 if (unlikely(map_chg
> map_commit
)) {
2069 * The page was added to the reservation map between
2070 * vma_needs_reservation and vma_commit_reservation.
2071 * This indicates a race with hugetlb_reserve_pages.
2072 * Adjust for the subpool count incremented above AND
2073 * in hugetlb_reserve_pages for the same page. Also,
2074 * the reservation count added in hugetlb_reserve_pages
2075 * no longer applies.
2079 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2080 hugetlb_acct_memory(h
, -rsv_adjust
);
2084 out_uncharge_cgroup
:
2085 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2087 if (map_chg
|| avoid_reserve
)
2088 hugepage_subpool_put_pages(spool
, 1);
2089 vma_end_reservation(h
, vma
, addr
);
2090 return ERR_PTR(-ENOSPC
);
2094 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2095 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2096 * where no ERR_VALUE is expected to be returned.
2098 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
2099 unsigned long addr
, int avoid_reserve
)
2101 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
2107 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
2109 struct huge_bootmem_page
*m
;
2112 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2115 addr
= memblock_virt_alloc_try_nid_nopanic(
2116 huge_page_size(h
), huge_page_size(h
),
2117 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
2120 * Use the beginning of the huge page to store the
2121 * huge_bootmem_page struct (until gather_bootmem
2122 * puts them into the mem_map).
2131 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2132 /* Put them into a private list first because mem_map is not up yet */
2133 list_add(&m
->list
, &huge_boot_pages
);
2138 static void __init
prep_compound_huge_page(struct page
*page
,
2141 if (unlikely(order
> (MAX_ORDER
- 1)))
2142 prep_compound_gigantic_page(page
, order
);
2144 prep_compound_page(page
, order
);
2147 /* Put bootmem huge pages into the standard lists after mem_map is up */
2148 static void __init
gather_bootmem_prealloc(void)
2150 struct huge_bootmem_page
*m
;
2152 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2153 struct hstate
*h
= m
->hstate
;
2156 #ifdef CONFIG_HIGHMEM
2157 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2158 memblock_free_late(__pa(m
),
2159 sizeof(struct huge_bootmem_page
));
2161 page
= virt_to_page(m
);
2163 WARN_ON(page_count(page
) != 1);
2164 prep_compound_huge_page(page
, h
->order
);
2165 WARN_ON(PageReserved(page
));
2166 prep_new_huge_page(h
, page
, page_to_nid(page
));
2168 * If we had gigantic hugepages allocated at boot time, we need
2169 * to restore the 'stolen' pages to totalram_pages in order to
2170 * fix confusing memory reports from free(1) and another
2171 * side-effects, like CommitLimit going negative.
2173 if (hstate_is_gigantic(h
))
2174 adjust_managed_page_count(page
, 1 << h
->order
);
2178 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2182 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2183 if (hstate_is_gigantic(h
)) {
2184 if (!alloc_bootmem_huge_page(h
))
2186 } else if (!alloc_fresh_huge_page(h
,
2187 &node_states
[N_MEMORY
]))
2190 h
->max_huge_pages
= i
;
2193 static void __init
hugetlb_init_hstates(void)
2197 for_each_hstate(h
) {
2198 if (minimum_order
> huge_page_order(h
))
2199 minimum_order
= huge_page_order(h
);
2201 /* oversize hugepages were init'ed in early boot */
2202 if (!hstate_is_gigantic(h
))
2203 hugetlb_hstate_alloc_pages(h
);
2205 VM_BUG_ON(minimum_order
== UINT_MAX
);
2208 static char * __init
memfmt(char *buf
, unsigned long n
)
2210 if (n
>= (1UL << 30))
2211 sprintf(buf
, "%lu GB", n
>> 30);
2212 else if (n
>= (1UL << 20))
2213 sprintf(buf
, "%lu MB", n
>> 20);
2215 sprintf(buf
, "%lu KB", n
>> 10);
2219 static void __init
report_hugepages(void)
2223 for_each_hstate(h
) {
2225 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2226 memfmt(buf
, huge_page_size(h
)),
2227 h
->free_huge_pages
);
2231 #ifdef CONFIG_HIGHMEM
2232 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2233 nodemask_t
*nodes_allowed
)
2237 if (hstate_is_gigantic(h
))
2240 for_each_node_mask(i
, *nodes_allowed
) {
2241 struct page
*page
, *next
;
2242 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2243 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2244 if (count
>= h
->nr_huge_pages
)
2246 if (PageHighMem(page
))
2248 list_del(&page
->lru
);
2249 update_and_free_page(h
, page
);
2250 h
->free_huge_pages
--;
2251 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2256 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2257 nodemask_t
*nodes_allowed
)
2263 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2264 * balanced by operating on them in a round-robin fashion.
2265 * Returns 1 if an adjustment was made.
2267 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2272 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2275 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2276 if (h
->surplus_huge_pages_node
[node
])
2280 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2281 if (h
->surplus_huge_pages_node
[node
] <
2282 h
->nr_huge_pages_node
[node
])
2289 h
->surplus_huge_pages
+= delta
;
2290 h
->surplus_huge_pages_node
[node
] += delta
;
2294 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2295 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2296 nodemask_t
*nodes_allowed
)
2298 unsigned long min_count
, ret
;
2300 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2301 return h
->max_huge_pages
;
2304 * Increase the pool size
2305 * First take pages out of surplus state. Then make up the
2306 * remaining difference by allocating fresh huge pages.
2308 * We might race with __alloc_buddy_huge_page() here and be unable
2309 * to convert a surplus huge page to a normal huge page. That is
2310 * not critical, though, it just means the overall size of the
2311 * pool might be one hugepage larger than it needs to be, but
2312 * within all the constraints specified by the sysctls.
2314 spin_lock(&hugetlb_lock
);
2315 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2316 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2320 while (count
> persistent_huge_pages(h
)) {
2322 * If this allocation races such that we no longer need the
2323 * page, free_huge_page will handle it by freeing the page
2324 * and reducing the surplus.
2326 spin_unlock(&hugetlb_lock
);
2328 /* yield cpu to avoid soft lockup */
2331 if (hstate_is_gigantic(h
))
2332 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2334 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2335 spin_lock(&hugetlb_lock
);
2339 /* Bail for signals. Probably ctrl-c from user */
2340 if (signal_pending(current
))
2345 * Decrease the pool size
2346 * First return free pages to the buddy allocator (being careful
2347 * to keep enough around to satisfy reservations). Then place
2348 * pages into surplus state as needed so the pool will shrink
2349 * to the desired size as pages become free.
2351 * By placing pages into the surplus state independent of the
2352 * overcommit value, we are allowing the surplus pool size to
2353 * exceed overcommit. There are few sane options here. Since
2354 * __alloc_buddy_huge_page() is checking the global counter,
2355 * though, we'll note that we're not allowed to exceed surplus
2356 * and won't grow the pool anywhere else. Not until one of the
2357 * sysctls are changed, or the surplus pages go out of use.
2359 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2360 min_count
= max(count
, min_count
);
2361 try_to_free_low(h
, min_count
, nodes_allowed
);
2362 while (min_count
< persistent_huge_pages(h
)) {
2363 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2365 cond_resched_lock(&hugetlb_lock
);
2367 while (count
< persistent_huge_pages(h
)) {
2368 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2372 ret
= persistent_huge_pages(h
);
2373 spin_unlock(&hugetlb_lock
);
2377 #define HSTATE_ATTR_RO(_name) \
2378 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2380 #define HSTATE_ATTR(_name) \
2381 static struct kobj_attribute _name##_attr = \
2382 __ATTR(_name, 0644, _name##_show, _name##_store)
2384 static struct kobject
*hugepages_kobj
;
2385 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2387 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2389 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2393 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2394 if (hstate_kobjs
[i
] == kobj
) {
2396 *nidp
= NUMA_NO_NODE
;
2400 return kobj_to_node_hstate(kobj
, nidp
);
2403 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2404 struct kobj_attribute
*attr
, char *buf
)
2407 unsigned long nr_huge_pages
;
2410 h
= kobj_to_hstate(kobj
, &nid
);
2411 if (nid
== NUMA_NO_NODE
)
2412 nr_huge_pages
= h
->nr_huge_pages
;
2414 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2416 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2419 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2420 struct hstate
*h
, int nid
,
2421 unsigned long count
, size_t len
)
2424 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2426 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2431 if (nid
== NUMA_NO_NODE
) {
2433 * global hstate attribute
2435 if (!(obey_mempolicy
&&
2436 init_nodemask_of_mempolicy(nodes_allowed
))) {
2437 NODEMASK_FREE(nodes_allowed
);
2438 nodes_allowed
= &node_states
[N_MEMORY
];
2440 } else if (nodes_allowed
) {
2442 * per node hstate attribute: adjust count to global,
2443 * but restrict alloc/free to the specified node.
2445 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2446 init_nodemask_of_node(nodes_allowed
, nid
);
2448 nodes_allowed
= &node_states
[N_MEMORY
];
2450 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2452 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2453 NODEMASK_FREE(nodes_allowed
);
2457 NODEMASK_FREE(nodes_allowed
);
2461 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2462 struct kobject
*kobj
, const char *buf
,
2466 unsigned long count
;
2470 err
= kstrtoul(buf
, 10, &count
);
2474 h
= kobj_to_hstate(kobj
, &nid
);
2475 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2478 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2479 struct kobj_attribute
*attr
, char *buf
)
2481 return nr_hugepages_show_common(kobj
, attr
, buf
);
2484 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2485 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2487 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2489 HSTATE_ATTR(nr_hugepages
);
2494 * hstate attribute for optionally mempolicy-based constraint on persistent
2495 * huge page alloc/free.
2497 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2498 struct kobj_attribute
*attr
, char *buf
)
2500 return nr_hugepages_show_common(kobj
, attr
, buf
);
2503 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2504 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2506 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2508 HSTATE_ATTR(nr_hugepages_mempolicy
);
2512 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2513 struct kobj_attribute
*attr
, char *buf
)
2515 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2516 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2519 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2520 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2523 unsigned long input
;
2524 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2526 if (hstate_is_gigantic(h
))
2529 err
= kstrtoul(buf
, 10, &input
);
2533 spin_lock(&hugetlb_lock
);
2534 h
->nr_overcommit_huge_pages
= input
;
2535 spin_unlock(&hugetlb_lock
);
2539 HSTATE_ATTR(nr_overcommit_hugepages
);
2541 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2542 struct kobj_attribute
*attr
, char *buf
)
2545 unsigned long free_huge_pages
;
2548 h
= kobj_to_hstate(kobj
, &nid
);
2549 if (nid
== NUMA_NO_NODE
)
2550 free_huge_pages
= h
->free_huge_pages
;
2552 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2554 return sprintf(buf
, "%lu\n", free_huge_pages
);
2556 HSTATE_ATTR_RO(free_hugepages
);
2558 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2559 struct kobj_attribute
*attr
, char *buf
)
2561 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2562 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2564 HSTATE_ATTR_RO(resv_hugepages
);
2566 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2567 struct kobj_attribute
*attr
, char *buf
)
2570 unsigned long surplus_huge_pages
;
2573 h
= kobj_to_hstate(kobj
, &nid
);
2574 if (nid
== NUMA_NO_NODE
)
2575 surplus_huge_pages
= h
->surplus_huge_pages
;
2577 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2579 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2581 HSTATE_ATTR_RO(surplus_hugepages
);
2583 static struct attribute
*hstate_attrs
[] = {
2584 &nr_hugepages_attr
.attr
,
2585 &nr_overcommit_hugepages_attr
.attr
,
2586 &free_hugepages_attr
.attr
,
2587 &resv_hugepages_attr
.attr
,
2588 &surplus_hugepages_attr
.attr
,
2590 &nr_hugepages_mempolicy_attr
.attr
,
2595 static struct attribute_group hstate_attr_group
= {
2596 .attrs
= hstate_attrs
,
2599 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2600 struct kobject
**hstate_kobjs
,
2601 struct attribute_group
*hstate_attr_group
)
2604 int hi
= hstate_index(h
);
2606 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2607 if (!hstate_kobjs
[hi
])
2610 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2612 kobject_put(hstate_kobjs
[hi
]);
2617 static void __init
hugetlb_sysfs_init(void)
2622 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2623 if (!hugepages_kobj
)
2626 for_each_hstate(h
) {
2627 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2628 hstate_kobjs
, &hstate_attr_group
);
2630 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2637 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2638 * with node devices in node_devices[] using a parallel array. The array
2639 * index of a node device or _hstate == node id.
2640 * This is here to avoid any static dependency of the node device driver, in
2641 * the base kernel, on the hugetlb module.
2643 struct node_hstate
{
2644 struct kobject
*hugepages_kobj
;
2645 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2647 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2650 * A subset of global hstate attributes for node devices
2652 static struct attribute
*per_node_hstate_attrs
[] = {
2653 &nr_hugepages_attr
.attr
,
2654 &free_hugepages_attr
.attr
,
2655 &surplus_hugepages_attr
.attr
,
2659 static struct attribute_group per_node_hstate_attr_group
= {
2660 .attrs
= per_node_hstate_attrs
,
2664 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2665 * Returns node id via non-NULL nidp.
2667 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2671 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2672 struct node_hstate
*nhs
= &node_hstates
[nid
];
2674 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2675 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2687 * Unregister hstate attributes from a single node device.
2688 * No-op if no hstate attributes attached.
2690 static void hugetlb_unregister_node(struct node
*node
)
2693 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2695 if (!nhs
->hugepages_kobj
)
2696 return; /* no hstate attributes */
2698 for_each_hstate(h
) {
2699 int idx
= hstate_index(h
);
2700 if (nhs
->hstate_kobjs
[idx
]) {
2701 kobject_put(nhs
->hstate_kobjs
[idx
]);
2702 nhs
->hstate_kobjs
[idx
] = NULL
;
2706 kobject_put(nhs
->hugepages_kobj
);
2707 nhs
->hugepages_kobj
= NULL
;
2712 * Register hstate attributes for a single node device.
2713 * No-op if attributes already registered.
2715 static void hugetlb_register_node(struct node
*node
)
2718 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2721 if (nhs
->hugepages_kobj
)
2722 return; /* already allocated */
2724 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2726 if (!nhs
->hugepages_kobj
)
2729 for_each_hstate(h
) {
2730 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2732 &per_node_hstate_attr_group
);
2734 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2735 h
->name
, node
->dev
.id
);
2736 hugetlb_unregister_node(node
);
2743 * hugetlb init time: register hstate attributes for all registered node
2744 * devices of nodes that have memory. All on-line nodes should have
2745 * registered their associated device by this time.
2747 static void __init
hugetlb_register_all_nodes(void)
2751 for_each_node_state(nid
, N_MEMORY
) {
2752 struct node
*node
= node_devices
[nid
];
2753 if (node
->dev
.id
== nid
)
2754 hugetlb_register_node(node
);
2758 * Let the node device driver know we're here so it can
2759 * [un]register hstate attributes on node hotplug.
2761 register_hugetlbfs_with_node(hugetlb_register_node
,
2762 hugetlb_unregister_node
);
2764 #else /* !CONFIG_NUMA */
2766 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2774 static void hugetlb_register_all_nodes(void) { }
2778 static int __init
hugetlb_init(void)
2782 if (!hugepages_supported())
2785 if (!size_to_hstate(default_hstate_size
)) {
2786 default_hstate_size
= HPAGE_SIZE
;
2787 if (!size_to_hstate(default_hstate_size
))
2788 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2790 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2791 if (default_hstate_max_huge_pages
) {
2792 if (!default_hstate
.max_huge_pages
)
2793 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2796 hugetlb_init_hstates();
2797 gather_bootmem_prealloc();
2800 hugetlb_sysfs_init();
2801 hugetlb_register_all_nodes();
2802 hugetlb_cgroup_file_init();
2805 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2807 num_fault_mutexes
= 1;
2809 hugetlb_fault_mutex_table
=
2810 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2811 BUG_ON(!hugetlb_fault_mutex_table
);
2813 for (i
= 0; i
< num_fault_mutexes
; i
++)
2814 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2817 subsys_initcall(hugetlb_init
);
2819 /* Should be called on processing a hugepagesz=... option */
2820 void __init
hugetlb_bad_size(void)
2822 parsed_valid_hugepagesz
= false;
2825 void __init
hugetlb_add_hstate(unsigned int order
)
2830 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2831 pr_warn("hugepagesz= specified twice, ignoring\n");
2834 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2836 h
= &hstates
[hugetlb_max_hstate
++];
2838 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2839 h
->nr_huge_pages
= 0;
2840 h
->free_huge_pages
= 0;
2841 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2842 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2843 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2844 h
->next_nid_to_alloc
= first_memory_node
;
2845 h
->next_nid_to_free
= first_memory_node
;
2846 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2847 huge_page_size(h
)/1024);
2852 static int __init
hugetlb_nrpages_setup(char *s
)
2855 static unsigned long *last_mhp
;
2857 if (!parsed_valid_hugepagesz
) {
2858 pr_warn("hugepages = %s preceded by "
2859 "an unsupported hugepagesz, ignoring\n", s
);
2860 parsed_valid_hugepagesz
= true;
2864 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2865 * so this hugepages= parameter goes to the "default hstate".
2867 else if (!hugetlb_max_hstate
)
2868 mhp
= &default_hstate_max_huge_pages
;
2870 mhp
= &parsed_hstate
->max_huge_pages
;
2872 if (mhp
== last_mhp
) {
2873 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2877 if (sscanf(s
, "%lu", mhp
) <= 0)
2881 * Global state is always initialized later in hugetlb_init.
2882 * But we need to allocate >= MAX_ORDER hstates here early to still
2883 * use the bootmem allocator.
2885 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2886 hugetlb_hstate_alloc_pages(parsed_hstate
);
2892 __setup("hugepages=", hugetlb_nrpages_setup
);
2894 static int __init
hugetlb_default_setup(char *s
)
2896 default_hstate_size
= memparse(s
, &s
);
2899 __setup("default_hugepagesz=", hugetlb_default_setup
);
2901 static unsigned int cpuset_mems_nr(unsigned int *array
)
2904 unsigned int nr
= 0;
2906 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2912 #ifdef CONFIG_SYSCTL
2913 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2914 struct ctl_table
*table
, int write
,
2915 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2917 struct hstate
*h
= &default_hstate
;
2918 unsigned long tmp
= h
->max_huge_pages
;
2921 if (!hugepages_supported())
2925 table
->maxlen
= sizeof(unsigned long);
2926 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2931 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2932 NUMA_NO_NODE
, tmp
, *length
);
2937 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2938 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2941 return hugetlb_sysctl_handler_common(false, table
, write
,
2942 buffer
, length
, ppos
);
2946 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2947 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2949 return hugetlb_sysctl_handler_common(true, table
, write
,
2950 buffer
, length
, ppos
);
2952 #endif /* CONFIG_NUMA */
2954 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2955 void __user
*buffer
,
2956 size_t *length
, loff_t
*ppos
)
2958 struct hstate
*h
= &default_hstate
;
2962 if (!hugepages_supported())
2965 tmp
= h
->nr_overcommit_huge_pages
;
2967 if (write
&& hstate_is_gigantic(h
))
2971 table
->maxlen
= sizeof(unsigned long);
2972 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2977 spin_lock(&hugetlb_lock
);
2978 h
->nr_overcommit_huge_pages
= tmp
;
2979 spin_unlock(&hugetlb_lock
);
2985 #endif /* CONFIG_SYSCTL */
2987 void hugetlb_report_meminfo(struct seq_file
*m
)
2989 struct hstate
*h
= &default_hstate
;
2990 if (!hugepages_supported())
2993 "HugePages_Total: %5lu\n"
2994 "HugePages_Free: %5lu\n"
2995 "HugePages_Rsvd: %5lu\n"
2996 "HugePages_Surp: %5lu\n"
2997 "Hugepagesize: %8lu kB\n",
3001 h
->surplus_huge_pages
,
3002 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3005 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3007 struct hstate
*h
= &default_hstate
;
3008 if (!hugepages_supported())
3011 "Node %d HugePages_Total: %5u\n"
3012 "Node %d HugePages_Free: %5u\n"
3013 "Node %d HugePages_Surp: %5u\n",
3014 nid
, h
->nr_huge_pages_node
[nid
],
3015 nid
, h
->free_huge_pages_node
[nid
],
3016 nid
, h
->surplus_huge_pages_node
[nid
]);
3019 void hugetlb_show_meminfo(void)
3024 if (!hugepages_supported())
3027 for_each_node_state(nid
, N_MEMORY
)
3029 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3031 h
->nr_huge_pages_node
[nid
],
3032 h
->free_huge_pages_node
[nid
],
3033 h
->surplus_huge_pages_node
[nid
],
3034 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3037 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3039 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3040 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3043 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3044 unsigned long hugetlb_total_pages(void)
3047 unsigned long nr_total_pages
= 0;
3050 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3051 return nr_total_pages
;
3054 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3058 spin_lock(&hugetlb_lock
);
3060 * When cpuset is configured, it breaks the strict hugetlb page
3061 * reservation as the accounting is done on a global variable. Such
3062 * reservation is completely rubbish in the presence of cpuset because
3063 * the reservation is not checked against page availability for the
3064 * current cpuset. Application can still potentially OOM'ed by kernel
3065 * with lack of free htlb page in cpuset that the task is in.
3066 * Attempt to enforce strict accounting with cpuset is almost
3067 * impossible (or too ugly) because cpuset is too fluid that
3068 * task or memory node can be dynamically moved between cpusets.
3070 * The change of semantics for shared hugetlb mapping with cpuset is
3071 * undesirable. However, in order to preserve some of the semantics,
3072 * we fall back to check against current free page availability as
3073 * a best attempt and hopefully to minimize the impact of changing
3074 * semantics that cpuset has.
3077 if (gather_surplus_pages(h
, delta
) < 0)
3080 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3081 return_unused_surplus_pages(h
, delta
);
3088 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3091 spin_unlock(&hugetlb_lock
);
3095 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3097 struct resv_map
*resv
= vma_resv_map(vma
);
3100 * This new VMA should share its siblings reservation map if present.
3101 * The VMA will only ever have a valid reservation map pointer where
3102 * it is being copied for another still existing VMA. As that VMA
3103 * has a reference to the reservation map it cannot disappear until
3104 * after this open call completes. It is therefore safe to take a
3105 * new reference here without additional locking.
3107 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3108 kref_get(&resv
->refs
);
3111 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3113 struct hstate
*h
= hstate_vma(vma
);
3114 struct resv_map
*resv
= vma_resv_map(vma
);
3115 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3116 unsigned long reserve
, start
, end
;
3119 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3122 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3123 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3125 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3127 kref_put(&resv
->refs
, resv_map_release
);
3131 * Decrement reserve counts. The global reserve count may be
3132 * adjusted if the subpool has a minimum size.
3134 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3135 hugetlb_acct_memory(h
, -gbl_reserve
);
3140 * We cannot handle pagefaults against hugetlb pages at all. They cause
3141 * handle_mm_fault() to try to instantiate regular-sized pages in the
3142 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3145 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3151 const struct vm_operations_struct hugetlb_vm_ops
= {
3152 .fault
= hugetlb_vm_op_fault
,
3153 .open
= hugetlb_vm_op_open
,
3154 .close
= hugetlb_vm_op_close
,
3157 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3163 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3164 vma
->vm_page_prot
)));
3166 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3167 vma
->vm_page_prot
));
3169 entry
= pte_mkyoung(entry
);
3170 entry
= pte_mkhuge(entry
);
3171 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3176 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3177 unsigned long address
, pte_t
*ptep
)
3181 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3182 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3183 update_mmu_cache(vma
, address
, ptep
);
3186 static int is_hugetlb_entry_migration(pte_t pte
)
3190 if (huge_pte_none(pte
) || pte_present(pte
))
3192 swp
= pte_to_swp_entry(pte
);
3193 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3199 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3203 if (huge_pte_none(pte
) || pte_present(pte
))
3205 swp
= pte_to_swp_entry(pte
);
3206 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3212 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3213 struct vm_area_struct
*vma
)
3215 pte_t
*src_pte
, *dst_pte
, entry
;
3216 struct page
*ptepage
;
3219 struct hstate
*h
= hstate_vma(vma
);
3220 unsigned long sz
= huge_page_size(h
);
3221 unsigned long mmun_start
; /* For mmu_notifiers */
3222 unsigned long mmun_end
; /* For mmu_notifiers */
3225 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3227 mmun_start
= vma
->vm_start
;
3228 mmun_end
= vma
->vm_end
;
3230 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3232 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3233 spinlock_t
*src_ptl
, *dst_ptl
;
3234 src_pte
= huge_pte_offset(src
, addr
);
3237 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3243 /* If the pagetables are shared don't copy or take references */
3244 if (dst_pte
== src_pte
)
3247 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3248 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3249 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3250 entry
= huge_ptep_get(src_pte
);
3251 if (huge_pte_none(entry
)) { /* skip none entry */
3253 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3254 is_hugetlb_entry_hwpoisoned(entry
))) {
3255 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3257 if (is_write_migration_entry(swp_entry
) && cow
) {
3259 * COW mappings require pages in both
3260 * parent and child to be set to read.
3262 make_migration_entry_read(&swp_entry
);
3263 entry
= swp_entry_to_pte(swp_entry
);
3264 set_huge_pte_at(src
, addr
, src_pte
, entry
);
3266 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3269 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3270 mmu_notifier_invalidate_range(src
, mmun_start
,
3273 entry
= huge_ptep_get(src_pte
);
3274 ptepage
= pte_page(entry
);
3276 page_dup_rmap(ptepage
, true);
3277 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3278 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3280 spin_unlock(src_ptl
);
3281 spin_unlock(dst_ptl
);
3285 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3290 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3291 unsigned long start
, unsigned long end
,
3292 struct page
*ref_page
)
3294 struct mm_struct
*mm
= vma
->vm_mm
;
3295 unsigned long address
;
3300 struct hstate
*h
= hstate_vma(vma
);
3301 unsigned long sz
= huge_page_size(h
);
3302 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3303 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3305 WARN_ON(!is_vm_hugetlb_page(vma
));
3306 BUG_ON(start
& ~huge_page_mask(h
));
3307 BUG_ON(end
& ~huge_page_mask(h
));
3310 * This is a hugetlb vma, all the pte entries should point
3313 tlb_remove_check_page_size_change(tlb
, sz
);
3314 tlb_start_vma(tlb
, vma
);
3315 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3317 for (; address
< end
; address
+= sz
) {
3318 ptep
= huge_pte_offset(mm
, address
);
3322 ptl
= huge_pte_lock(h
, mm
, ptep
);
3323 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3328 pte
= huge_ptep_get(ptep
);
3329 if (huge_pte_none(pte
)) {
3335 * Migrating hugepage or HWPoisoned hugepage is already
3336 * unmapped and its refcount is dropped, so just clear pte here.
3338 if (unlikely(!pte_present(pte
))) {
3339 huge_pte_clear(mm
, address
, ptep
);
3344 page
= pte_page(pte
);
3346 * If a reference page is supplied, it is because a specific
3347 * page is being unmapped, not a range. Ensure the page we
3348 * are about to unmap is the actual page of interest.
3351 if (page
!= ref_page
) {
3356 * Mark the VMA as having unmapped its page so that
3357 * future faults in this VMA will fail rather than
3358 * looking like data was lost
3360 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3363 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3364 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3365 if (huge_pte_dirty(pte
))
3366 set_page_dirty(page
);
3368 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3369 page_remove_rmap(page
, true);
3372 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3374 * Bail out after unmapping reference page if supplied
3379 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3380 tlb_end_vma(tlb
, vma
);
3383 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3384 struct vm_area_struct
*vma
, unsigned long start
,
3385 unsigned long end
, struct page
*ref_page
)
3387 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3390 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3391 * test will fail on a vma being torn down, and not grab a page table
3392 * on its way out. We're lucky that the flag has such an appropriate
3393 * name, and can in fact be safely cleared here. We could clear it
3394 * before the __unmap_hugepage_range above, but all that's necessary
3395 * is to clear it before releasing the i_mmap_rwsem. This works
3396 * because in the context this is called, the VMA is about to be
3397 * destroyed and the i_mmap_rwsem is held.
3399 vma
->vm_flags
&= ~VM_MAYSHARE
;
3402 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3403 unsigned long end
, struct page
*ref_page
)
3405 struct mm_struct
*mm
;
3406 struct mmu_gather tlb
;
3410 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3411 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3412 tlb_finish_mmu(&tlb
, start
, end
);
3416 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3417 * mappping it owns the reserve page for. The intention is to unmap the page
3418 * from other VMAs and let the children be SIGKILLed if they are faulting the
3421 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3422 struct page
*page
, unsigned long address
)
3424 struct hstate
*h
= hstate_vma(vma
);
3425 struct vm_area_struct
*iter_vma
;
3426 struct address_space
*mapping
;
3430 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3431 * from page cache lookup which is in HPAGE_SIZE units.
3433 address
= address
& huge_page_mask(h
);
3434 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3436 mapping
= vma
->vm_file
->f_mapping
;
3439 * Take the mapping lock for the duration of the table walk. As
3440 * this mapping should be shared between all the VMAs,
3441 * __unmap_hugepage_range() is called as the lock is already held
3443 i_mmap_lock_write(mapping
);
3444 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3445 /* Do not unmap the current VMA */
3446 if (iter_vma
== vma
)
3450 * Shared VMAs have their own reserves and do not affect
3451 * MAP_PRIVATE accounting but it is possible that a shared
3452 * VMA is using the same page so check and skip such VMAs.
3454 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3458 * Unmap the page from other VMAs without their own reserves.
3459 * They get marked to be SIGKILLed if they fault in these
3460 * areas. This is because a future no-page fault on this VMA
3461 * could insert a zeroed page instead of the data existing
3462 * from the time of fork. This would look like data corruption
3464 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3465 unmap_hugepage_range(iter_vma
, address
,
3466 address
+ huge_page_size(h
), page
);
3468 i_mmap_unlock_write(mapping
);
3472 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3473 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3474 * cannot race with other handlers or page migration.
3475 * Keep the pte_same checks anyway to make transition from the mutex easier.
3477 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3478 unsigned long address
, pte_t
*ptep
,
3479 struct page
*pagecache_page
, spinlock_t
*ptl
)
3482 struct hstate
*h
= hstate_vma(vma
);
3483 struct page
*old_page
, *new_page
;
3484 int ret
= 0, outside_reserve
= 0;
3485 unsigned long mmun_start
; /* For mmu_notifiers */
3486 unsigned long mmun_end
; /* For mmu_notifiers */
3488 pte
= huge_ptep_get(ptep
);
3489 old_page
= pte_page(pte
);
3492 /* If no-one else is actually using this page, avoid the copy
3493 * and just make the page writable */
3494 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3495 page_move_anon_rmap(old_page
, vma
);
3496 set_huge_ptep_writable(vma
, address
, ptep
);
3501 * If the process that created a MAP_PRIVATE mapping is about to
3502 * perform a COW due to a shared page count, attempt to satisfy
3503 * the allocation without using the existing reserves. The pagecache
3504 * page is used to determine if the reserve at this address was
3505 * consumed or not. If reserves were used, a partial faulted mapping
3506 * at the time of fork() could consume its reserves on COW instead
3507 * of the full address range.
3509 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3510 old_page
!= pagecache_page
)
3511 outside_reserve
= 1;
3516 * Drop page table lock as buddy allocator may be called. It will
3517 * be acquired again before returning to the caller, as expected.
3520 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3522 if (IS_ERR(new_page
)) {
3524 * If a process owning a MAP_PRIVATE mapping fails to COW,
3525 * it is due to references held by a child and an insufficient
3526 * huge page pool. To guarantee the original mappers
3527 * reliability, unmap the page from child processes. The child
3528 * may get SIGKILLed if it later faults.
3530 if (outside_reserve
) {
3532 BUG_ON(huge_pte_none(pte
));
3533 unmap_ref_private(mm
, vma
, old_page
, address
);
3534 BUG_ON(huge_pte_none(pte
));
3536 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3538 pte_same(huge_ptep_get(ptep
), pte
)))
3539 goto retry_avoidcopy
;
3541 * race occurs while re-acquiring page table
3542 * lock, and our job is done.
3547 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3548 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3549 goto out_release_old
;
3553 * When the original hugepage is shared one, it does not have
3554 * anon_vma prepared.
3556 if (unlikely(anon_vma_prepare(vma
))) {
3558 goto out_release_all
;
3561 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3562 pages_per_huge_page(h
));
3563 __SetPageUptodate(new_page
);
3564 set_page_huge_active(new_page
);
3566 mmun_start
= address
& huge_page_mask(h
);
3567 mmun_end
= mmun_start
+ huge_page_size(h
);
3568 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3571 * Retake the page table lock to check for racing updates
3572 * before the page tables are altered
3575 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3576 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3577 ClearPagePrivate(new_page
);
3580 huge_ptep_clear_flush(vma
, address
, ptep
);
3581 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3582 set_huge_pte_at(mm
, address
, ptep
,
3583 make_huge_pte(vma
, new_page
, 1));
3584 page_remove_rmap(old_page
, true);
3585 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3586 /* Make the old page be freed below */
3587 new_page
= old_page
;
3590 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3592 restore_reserve_on_error(h
, vma
, address
, new_page
);
3597 spin_lock(ptl
); /* Caller expects lock to be held */
3601 /* Return the pagecache page at a given address within a VMA */
3602 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3603 struct vm_area_struct
*vma
, unsigned long address
)
3605 struct address_space
*mapping
;
3608 mapping
= vma
->vm_file
->f_mapping
;
3609 idx
= vma_hugecache_offset(h
, vma
, address
);
3611 return find_lock_page(mapping
, idx
);
3615 * Return whether there is a pagecache page to back given address within VMA.
3616 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3618 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3619 struct vm_area_struct
*vma
, unsigned long address
)
3621 struct address_space
*mapping
;
3625 mapping
= vma
->vm_file
->f_mapping
;
3626 idx
= vma_hugecache_offset(h
, vma
, address
);
3628 page
= find_get_page(mapping
, idx
);
3631 return page
!= NULL
;
3634 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3637 struct inode
*inode
= mapping
->host
;
3638 struct hstate
*h
= hstate_inode(inode
);
3639 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3643 ClearPagePrivate(page
);
3645 spin_lock(&inode
->i_lock
);
3646 inode
->i_blocks
+= blocks_per_huge_page(h
);
3647 spin_unlock(&inode
->i_lock
);
3651 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3652 struct address_space
*mapping
, pgoff_t idx
,
3653 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3655 struct hstate
*h
= hstate_vma(vma
);
3656 int ret
= VM_FAULT_SIGBUS
;
3664 * Currently, we are forced to kill the process in the event the
3665 * original mapper has unmapped pages from the child due to a failed
3666 * COW. Warn that such a situation has occurred as it may not be obvious
3668 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3669 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3675 * Use page lock to guard against racing truncation
3676 * before we get page_table_lock.
3679 page
= find_lock_page(mapping
, idx
);
3681 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3686 * Check for page in userfault range
3688 if (userfaultfd_missing(vma
)) {
3690 struct vm_fault vmf
= {
3695 * Hard to debug if it ends up being
3696 * used by a callee that assumes
3697 * something about the other
3698 * uninitialized fields... same as in
3704 * hugetlb_fault_mutex must be dropped before
3705 * handling userfault. Reacquire after handling
3706 * fault to make calling code simpler.
3708 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
,
3710 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3711 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
3712 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3716 page
= alloc_huge_page(vma
, address
, 0);
3718 ret
= PTR_ERR(page
);
3722 ret
= VM_FAULT_SIGBUS
;
3725 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3726 __SetPageUptodate(page
);
3727 set_page_huge_active(page
);
3729 if (vma
->vm_flags
& VM_MAYSHARE
) {
3730 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3739 if (unlikely(anon_vma_prepare(vma
))) {
3741 goto backout_unlocked
;
3747 * If memory error occurs between mmap() and fault, some process
3748 * don't have hwpoisoned swap entry for errored virtual address.
3749 * So we need to block hugepage fault by PG_hwpoison bit check.
3751 if (unlikely(PageHWPoison(page
))) {
3752 ret
= VM_FAULT_HWPOISON
|
3753 VM_FAULT_SET_HINDEX(hstate_index(h
));
3754 goto backout_unlocked
;
3759 * If we are going to COW a private mapping later, we examine the
3760 * pending reservations for this page now. This will ensure that
3761 * any allocations necessary to record that reservation occur outside
3764 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3765 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3767 goto backout_unlocked
;
3769 /* Just decrements count, does not deallocate */
3770 vma_end_reservation(h
, vma
, address
);
3773 ptl
= huge_pte_lock(h
, mm
, ptep
);
3774 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3779 if (!huge_pte_none(huge_ptep_get(ptep
)))
3783 ClearPagePrivate(page
);
3784 hugepage_add_new_anon_rmap(page
, vma
, address
);
3786 page_dup_rmap(page
, true);
3787 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3788 && (vma
->vm_flags
& VM_SHARED
)));
3789 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3791 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3792 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3793 /* Optimization, do the COW without a second fault */
3794 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
3806 restore_reserve_on_error(h
, vma
, address
, page
);
3812 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3813 struct vm_area_struct
*vma
,
3814 struct address_space
*mapping
,
3815 pgoff_t idx
, unsigned long address
)
3817 unsigned long key
[2];
3820 if (vma
->vm_flags
& VM_SHARED
) {
3821 key
[0] = (unsigned long) mapping
;
3824 key
[0] = (unsigned long) mm
;
3825 key
[1] = address
>> huge_page_shift(h
);
3828 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3830 return hash
& (num_fault_mutexes
- 1);
3834 * For uniprocesor systems we always use a single mutex, so just
3835 * return 0 and avoid the hashing overhead.
3837 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3838 struct vm_area_struct
*vma
,
3839 struct address_space
*mapping
,
3840 pgoff_t idx
, unsigned long address
)
3846 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3847 unsigned long address
, unsigned int flags
)
3854 struct page
*page
= NULL
;
3855 struct page
*pagecache_page
= NULL
;
3856 struct hstate
*h
= hstate_vma(vma
);
3857 struct address_space
*mapping
;
3858 int need_wait_lock
= 0;
3860 address
&= huge_page_mask(h
);
3862 ptep
= huge_pte_offset(mm
, address
);
3864 entry
= huge_ptep_get(ptep
);
3865 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3866 migration_entry_wait_huge(vma
, mm
, ptep
);
3868 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3869 return VM_FAULT_HWPOISON_LARGE
|
3870 VM_FAULT_SET_HINDEX(hstate_index(h
));
3872 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3874 return VM_FAULT_OOM
;
3877 mapping
= vma
->vm_file
->f_mapping
;
3878 idx
= vma_hugecache_offset(h
, vma
, address
);
3881 * Serialize hugepage allocation and instantiation, so that we don't
3882 * get spurious allocation failures if two CPUs race to instantiate
3883 * the same page in the page cache.
3885 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3886 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3888 entry
= huge_ptep_get(ptep
);
3889 if (huge_pte_none(entry
)) {
3890 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3897 * entry could be a migration/hwpoison entry at this point, so this
3898 * check prevents the kernel from going below assuming that we have
3899 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3900 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3903 if (!pte_present(entry
))
3907 * If we are going to COW the mapping later, we examine the pending
3908 * reservations for this page now. This will ensure that any
3909 * allocations necessary to record that reservation occur outside the
3910 * spinlock. For private mappings, we also lookup the pagecache
3911 * page now as it is used to determine if a reservation has been
3914 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3915 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3919 /* Just decrements count, does not deallocate */
3920 vma_end_reservation(h
, vma
, address
);
3922 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3923 pagecache_page
= hugetlbfs_pagecache_page(h
,
3927 ptl
= huge_pte_lock(h
, mm
, ptep
);
3929 /* Check for a racing update before calling hugetlb_cow */
3930 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3934 * hugetlb_cow() requires page locks of pte_page(entry) and
3935 * pagecache_page, so here we need take the former one
3936 * when page != pagecache_page or !pagecache_page.
3938 page
= pte_page(entry
);
3939 if (page
!= pagecache_page
)
3940 if (!trylock_page(page
)) {
3947 if (flags
& FAULT_FLAG_WRITE
) {
3948 if (!huge_pte_write(entry
)) {
3949 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
3950 pagecache_page
, ptl
);
3953 entry
= huge_pte_mkdirty(entry
);
3955 entry
= pte_mkyoung(entry
);
3956 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3957 flags
& FAULT_FLAG_WRITE
))
3958 update_mmu_cache(vma
, address
, ptep
);
3960 if (page
!= pagecache_page
)
3966 if (pagecache_page
) {
3967 unlock_page(pagecache_page
);
3968 put_page(pagecache_page
);
3971 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3973 * Generally it's safe to hold refcount during waiting page lock. But
3974 * here we just wait to defer the next page fault to avoid busy loop and
3975 * the page is not used after unlocked before returning from the current
3976 * page fault. So we are safe from accessing freed page, even if we wait
3977 * here without taking refcount.
3980 wait_on_page_locked(page
);
3985 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
3986 * modifications for huge pages.
3988 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
3990 struct vm_area_struct
*dst_vma
,
3991 unsigned long dst_addr
,
3992 unsigned long src_addr
,
3993 struct page
**pagep
)
3995 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
3996 struct hstate
*h
= hstate_vma(dst_vma
);
4004 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4008 ret
= copy_huge_page_from_user(page
,
4009 (const void __user
*) src_addr
,
4010 pages_per_huge_page(h
), false);
4012 /* fallback to copy_from_user outside mmap_sem */
4013 if (unlikely(ret
)) {
4016 /* don't free the page */
4025 * The memory barrier inside __SetPageUptodate makes sure that
4026 * preceding stores to the page contents become visible before
4027 * the set_pte_at() write.
4029 __SetPageUptodate(page
);
4030 set_page_huge_active(page
);
4033 * If shared, add to page cache
4036 struct address_space
*mapping
= dst_vma
->vm_file
->f_mapping
;
4037 pgoff_t idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4039 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4041 goto out_release_nounlock
;
4044 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4048 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4049 goto out_release_unlock
;
4052 page_dup_rmap(page
, true);
4054 ClearPagePrivate(page
);
4055 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4058 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4059 if (dst_vma
->vm_flags
& VM_WRITE
)
4060 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4061 _dst_pte
= pte_mkyoung(_dst_pte
);
4063 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4065 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4066 dst_vma
->vm_flags
& VM_WRITE
);
4067 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4069 /* No need to invalidate - it was non-present before */
4070 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4080 out_release_nounlock
:
4087 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4088 struct page
**pages
, struct vm_area_struct
**vmas
,
4089 unsigned long *position
, unsigned long *nr_pages
,
4090 long i
, unsigned int flags
, int *nonblocking
)
4092 unsigned long pfn_offset
;
4093 unsigned long vaddr
= *position
;
4094 unsigned long remainder
= *nr_pages
;
4095 struct hstate
*h
= hstate_vma(vma
);
4097 while (vaddr
< vma
->vm_end
&& remainder
) {
4099 spinlock_t
*ptl
= NULL
;
4104 * If we have a pending SIGKILL, don't keep faulting pages and
4105 * potentially allocating memory.
4107 if (unlikely(fatal_signal_pending(current
))) {
4113 * Some archs (sparc64, sh*) have multiple pte_ts to
4114 * each hugepage. We have to make sure we get the
4115 * first, for the page indexing below to work.
4117 * Note that page table lock is not held when pte is null.
4119 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
4121 ptl
= huge_pte_lock(h
, mm
, pte
);
4122 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4125 * When coredumping, it suits get_dump_page if we just return
4126 * an error where there's an empty slot with no huge pagecache
4127 * to back it. This way, we avoid allocating a hugepage, and
4128 * the sparse dumpfile avoids allocating disk blocks, but its
4129 * huge holes still show up with zeroes where they need to be.
4131 if (absent
&& (flags
& FOLL_DUMP
) &&
4132 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4140 * We need call hugetlb_fault for both hugepages under migration
4141 * (in which case hugetlb_fault waits for the migration,) and
4142 * hwpoisoned hugepages (in which case we need to prevent the
4143 * caller from accessing to them.) In order to do this, we use
4144 * here is_swap_pte instead of is_hugetlb_entry_migration and
4145 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4146 * both cases, and because we can't follow correct pages
4147 * directly from any kind of swap entries.
4149 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4150 ((flags
& FOLL_WRITE
) &&
4151 !huge_pte_write(huge_ptep_get(pte
)))) {
4153 unsigned int fault_flags
= 0;
4157 if (flags
& FOLL_WRITE
)
4158 fault_flags
|= FAULT_FLAG_WRITE
;
4160 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4161 if (flags
& FOLL_NOWAIT
)
4162 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4163 FAULT_FLAG_RETRY_NOWAIT
;
4164 if (flags
& FOLL_TRIED
) {
4165 VM_WARN_ON_ONCE(fault_flags
&
4166 FAULT_FLAG_ALLOW_RETRY
);
4167 fault_flags
|= FAULT_FLAG_TRIED
;
4169 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4170 if (ret
& VM_FAULT_ERROR
) {
4174 if (ret
& VM_FAULT_RETRY
) {
4179 * VM_FAULT_RETRY must not return an
4180 * error, it will return zero
4183 * No need to update "position" as the
4184 * caller will not check it after
4185 * *nr_pages is set to 0.
4192 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4193 page
= pte_page(huge_ptep_get(pte
));
4196 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4207 if (vaddr
< vma
->vm_end
&& remainder
&&
4208 pfn_offset
< pages_per_huge_page(h
)) {
4210 * We use pfn_offset to avoid touching the pageframes
4211 * of this compound page.
4217 *nr_pages
= remainder
;
4219 * setting position is actually required only if remainder is
4220 * not zero but it's faster not to add a "if (remainder)"
4225 return i
? i
: -EFAULT
;
4228 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4230 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4233 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4236 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4237 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4239 struct mm_struct
*mm
= vma
->vm_mm
;
4240 unsigned long start
= address
;
4243 struct hstate
*h
= hstate_vma(vma
);
4244 unsigned long pages
= 0;
4246 BUG_ON(address
>= end
);
4247 flush_cache_range(vma
, address
, end
);
4249 mmu_notifier_invalidate_range_start(mm
, start
, end
);
4250 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4251 for (; address
< end
; address
+= huge_page_size(h
)) {
4253 ptep
= huge_pte_offset(mm
, address
);
4256 ptl
= huge_pte_lock(h
, mm
, ptep
);
4257 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4262 pte
= huge_ptep_get(ptep
);
4263 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4267 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4268 swp_entry_t entry
= pte_to_swp_entry(pte
);
4270 if (is_write_migration_entry(entry
)) {
4273 make_migration_entry_read(&entry
);
4274 newpte
= swp_entry_to_pte(entry
);
4275 set_huge_pte_at(mm
, address
, ptep
, newpte
);
4281 if (!huge_pte_none(pte
)) {
4282 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4283 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
4284 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4285 set_huge_pte_at(mm
, address
, ptep
, pte
);
4291 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4292 * may have cleared our pud entry and done put_page on the page table:
4293 * once we release i_mmap_rwsem, another task can do the final put_page
4294 * and that page table be reused and filled with junk.
4296 flush_hugetlb_tlb_range(vma
, start
, end
);
4297 mmu_notifier_invalidate_range(mm
, start
, end
);
4298 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4299 mmu_notifier_invalidate_range_end(mm
, start
, end
);
4301 return pages
<< h
->order
;
4304 int hugetlb_reserve_pages(struct inode
*inode
,
4306 struct vm_area_struct
*vma
,
4307 vm_flags_t vm_flags
)
4310 struct hstate
*h
= hstate_inode(inode
);
4311 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4312 struct resv_map
*resv_map
;
4316 * Only apply hugepage reservation if asked. At fault time, an
4317 * attempt will be made for VM_NORESERVE to allocate a page
4318 * without using reserves
4320 if (vm_flags
& VM_NORESERVE
)
4324 * Shared mappings base their reservation on the number of pages that
4325 * are already allocated on behalf of the file. Private mappings need
4326 * to reserve the full area even if read-only as mprotect() may be
4327 * called to make the mapping read-write. Assume !vma is a shm mapping
4329 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4330 resv_map
= inode_resv_map(inode
);
4332 chg
= region_chg(resv_map
, from
, to
);
4335 resv_map
= resv_map_alloc();
4341 set_vma_resv_map(vma
, resv_map
);
4342 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4351 * There must be enough pages in the subpool for the mapping. If
4352 * the subpool has a minimum size, there may be some global
4353 * reservations already in place (gbl_reserve).
4355 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4356 if (gbl_reserve
< 0) {
4362 * Check enough hugepages are available for the reservation.
4363 * Hand the pages back to the subpool if there are not
4365 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4367 /* put back original number of pages, chg */
4368 (void)hugepage_subpool_put_pages(spool
, chg
);
4373 * Account for the reservations made. Shared mappings record regions
4374 * that have reservations as they are shared by multiple VMAs.
4375 * When the last VMA disappears, the region map says how much
4376 * the reservation was and the page cache tells how much of
4377 * the reservation was consumed. Private mappings are per-VMA and
4378 * only the consumed reservations are tracked. When the VMA
4379 * disappears, the original reservation is the VMA size and the
4380 * consumed reservations are stored in the map. Hence, nothing
4381 * else has to be done for private mappings here
4383 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4384 long add
= region_add(resv_map
, from
, to
);
4386 if (unlikely(chg
> add
)) {
4388 * pages in this range were added to the reserve
4389 * map between region_chg and region_add. This
4390 * indicates a race with alloc_huge_page. Adjust
4391 * the subpool and reserve counts modified above
4392 * based on the difference.
4396 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4398 hugetlb_acct_memory(h
, -rsv_adjust
);
4403 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4404 region_abort(resv_map
, from
, to
);
4405 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4406 kref_put(&resv_map
->refs
, resv_map_release
);
4410 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4413 struct hstate
*h
= hstate_inode(inode
);
4414 struct resv_map
*resv_map
= inode_resv_map(inode
);
4416 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4420 chg
= region_del(resv_map
, start
, end
);
4422 * region_del() can fail in the rare case where a region
4423 * must be split and another region descriptor can not be
4424 * allocated. If end == LONG_MAX, it will not fail.
4430 spin_lock(&inode
->i_lock
);
4431 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4432 spin_unlock(&inode
->i_lock
);
4435 * If the subpool has a minimum size, the number of global
4436 * reservations to be released may be adjusted.
4438 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4439 hugetlb_acct_memory(h
, -gbl_reserve
);
4444 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4445 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4446 struct vm_area_struct
*vma
,
4447 unsigned long addr
, pgoff_t idx
)
4449 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4451 unsigned long sbase
= saddr
& PUD_MASK
;
4452 unsigned long s_end
= sbase
+ PUD_SIZE
;
4454 /* Allow segments to share if only one is marked locked */
4455 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4456 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4459 * match the virtual addresses, permission and the alignment of the
4462 if (pmd_index(addr
) != pmd_index(saddr
) ||
4463 vm_flags
!= svm_flags
||
4464 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4470 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4472 unsigned long base
= addr
& PUD_MASK
;
4473 unsigned long end
= base
+ PUD_SIZE
;
4476 * check on proper vm_flags and page table alignment
4478 if (vma
->vm_flags
& VM_MAYSHARE
&&
4479 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4485 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4486 * and returns the corresponding pte. While this is not necessary for the
4487 * !shared pmd case because we can allocate the pmd later as well, it makes the
4488 * code much cleaner. pmd allocation is essential for the shared case because
4489 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4490 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4491 * bad pmd for sharing.
4493 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4495 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4496 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4497 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4499 struct vm_area_struct
*svma
;
4500 unsigned long saddr
;
4505 if (!vma_shareable(vma
, addr
))
4506 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4508 i_mmap_lock_write(mapping
);
4509 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4513 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4515 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
4517 get_page(virt_to_page(spte
));
4526 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4527 if (pud_none(*pud
)) {
4528 pud_populate(mm
, pud
,
4529 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4532 put_page(virt_to_page(spte
));
4536 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4537 i_mmap_unlock_write(mapping
);
4542 * unmap huge page backed by shared pte.
4544 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4545 * indicated by page_count > 1, unmap is achieved by clearing pud and
4546 * decrementing the ref count. If count == 1, the pte page is not shared.
4548 * called with page table lock held.
4550 * returns: 1 successfully unmapped a shared pte page
4551 * 0 the underlying pte page is not shared, or it is the last user
4553 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4555 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4556 pud_t
*pud
= pud_offset(pgd
, *addr
);
4558 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4559 if (page_count(virt_to_page(ptep
)) == 1)
4563 put_page(virt_to_page(ptep
));
4565 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4568 #define want_pmd_share() (1)
4569 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4570 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4575 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4579 #define want_pmd_share() (0)
4580 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4582 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4583 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4584 unsigned long addr
, unsigned long sz
)
4590 pgd
= pgd_offset(mm
, addr
);
4591 pud
= pud_alloc(mm
, pgd
, addr
);
4593 if (sz
== PUD_SIZE
) {
4596 BUG_ON(sz
!= PMD_SIZE
);
4597 if (want_pmd_share() && pud_none(*pud
))
4598 pte
= huge_pmd_share(mm
, addr
, pud
);
4600 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4603 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
4608 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
4614 pgd
= pgd_offset(mm
, addr
);
4615 if (pgd_present(*pgd
)) {
4616 pud
= pud_offset(pgd
, addr
);
4617 if (pud_present(*pud
)) {
4619 return (pte_t
*)pud
;
4620 pmd
= pmd_offset(pud
, addr
);
4623 return (pte_t
*) pmd
;
4626 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4629 * These functions are overwritable if your architecture needs its own
4632 struct page
* __weak
4633 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4636 return ERR_PTR(-EINVAL
);
4639 struct page
* __weak
4640 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4641 pmd_t
*pmd
, int flags
)
4643 struct page
*page
= NULL
;
4646 ptl
= pmd_lockptr(mm
, pmd
);
4649 * make sure that the address range covered by this pmd is not
4650 * unmapped from other threads.
4652 if (!pmd_huge(*pmd
))
4654 if (pmd_present(*pmd
)) {
4655 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4656 if (flags
& FOLL_GET
)
4659 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
4661 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4665 * hwpoisoned entry is treated as no_page_table in
4666 * follow_page_mask().
4674 struct page
* __weak
4675 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4676 pud_t
*pud
, int flags
)
4678 if (flags
& FOLL_GET
)
4681 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4684 #ifdef CONFIG_MEMORY_FAILURE
4687 * This function is called from memory failure code.
4689 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4691 struct hstate
*h
= page_hstate(hpage
);
4692 int nid
= page_to_nid(hpage
);
4695 spin_lock(&hugetlb_lock
);
4697 * Just checking !page_huge_active is not enough, because that could be
4698 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4700 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4702 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4703 * but dangling hpage->lru can trigger list-debug warnings
4704 * (this happens when we call unpoison_memory() on it),
4705 * so let it point to itself with list_del_init().
4707 list_del_init(&hpage
->lru
);
4708 set_page_refcounted(hpage
);
4709 h
->free_huge_pages
--;
4710 h
->free_huge_pages_node
[nid
]--;
4713 spin_unlock(&hugetlb_lock
);
4718 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4722 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4723 spin_lock(&hugetlb_lock
);
4724 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4728 clear_page_huge_active(page
);
4729 list_move_tail(&page
->lru
, list
);
4731 spin_unlock(&hugetlb_lock
);
4735 void putback_active_hugepage(struct page
*page
)
4737 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4738 spin_lock(&hugetlb_lock
);
4739 set_page_huge_active(page
);
4740 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4741 spin_unlock(&hugetlb_lock
);