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/sched/signal.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.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_exact(struct hstate
*h
, int nid
)
873 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
874 if (!PageHWPoison(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 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
894 if (nid
!= NUMA_NO_NODE
)
895 return dequeue_huge_page_node_exact(h
, nid
);
897 for_each_online_node(node
) {
898 page
= dequeue_huge_page_node_exact(h
, node
);
905 /* Movability of hugepages depends on migration support. */
906 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
908 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
909 return GFP_HIGHUSER_MOVABLE
;
914 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
915 struct vm_area_struct
*vma
,
916 unsigned long address
, int avoid_reserve
,
919 struct page
*page
= NULL
;
920 struct mempolicy
*mpol
;
921 nodemask_t
*nodemask
;
924 struct zonelist
*zonelist
;
927 unsigned int cpuset_mems_cookie
;
930 * A child process with MAP_PRIVATE mappings created by their parent
931 * have no page reserves. This check ensures that reservations are
932 * not "stolen". The child may still get SIGKILLed
934 if (!vma_has_reserves(vma
, chg
) &&
935 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
938 /* If reserves cannot be used, ensure enough pages are in the pool */
939 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
943 cpuset_mems_cookie
= read_mems_allowed_begin();
944 gfp_mask
= htlb_alloc_mask(h
);
945 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
946 zonelist
= node_zonelist(nid
, gfp_mask
);
948 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
949 MAX_NR_ZONES
- 1, nodemask
) {
950 if (cpuset_zone_allowed(zone
, gfp_mask
)) {
951 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
955 if (!vma_has_reserves(vma
, chg
))
958 SetPagePrivate(page
);
959 h
->resv_huge_pages
--;
966 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
975 * common helper functions for hstate_next_node_to_{alloc|free}.
976 * We may have allocated or freed a huge page based on a different
977 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
978 * be outside of *nodes_allowed. Ensure that we use an allowed
979 * node for alloc or free.
981 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
983 nid
= next_node_in(nid
, *nodes_allowed
);
984 VM_BUG_ON(nid
>= MAX_NUMNODES
);
989 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
991 if (!node_isset(nid
, *nodes_allowed
))
992 nid
= next_node_allowed(nid
, nodes_allowed
);
997 * returns the previously saved node ["this node"] from which to
998 * allocate a persistent huge page for the pool and advance the
999 * next node from which to allocate, handling wrap at end of node
1002 static int hstate_next_node_to_alloc(struct hstate
*h
,
1003 nodemask_t
*nodes_allowed
)
1007 VM_BUG_ON(!nodes_allowed
);
1009 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1010 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1016 * helper for free_pool_huge_page() - return the previously saved
1017 * node ["this node"] from which to free a huge page. Advance the
1018 * next node id whether or not we find a free huge page to free so
1019 * that the next attempt to free addresses the next node.
1021 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1025 VM_BUG_ON(!nodes_allowed
);
1027 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1028 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1033 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1036 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1039 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1040 for (nr_nodes = nodes_weight(*mask); \
1042 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1045 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1046 static void destroy_compound_gigantic_page(struct page
*page
,
1050 int nr_pages
= 1 << order
;
1051 struct page
*p
= page
+ 1;
1053 atomic_set(compound_mapcount_ptr(page
), 0);
1054 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1055 clear_compound_head(p
);
1056 set_page_refcounted(p
);
1059 set_compound_order(page
, 0);
1060 __ClearPageHead(page
);
1063 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1065 free_contig_range(page_to_pfn(page
), 1 << order
);
1068 static int __alloc_gigantic_page(unsigned long start_pfn
,
1069 unsigned long nr_pages
)
1071 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1072 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
1076 static bool pfn_range_valid_gigantic(struct zone
*z
,
1077 unsigned long start_pfn
, unsigned long nr_pages
)
1079 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1082 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1086 page
= pfn_to_page(i
);
1088 if (page_zone(page
) != z
)
1091 if (PageReserved(page
))
1094 if (page_count(page
) > 0)
1104 static bool zone_spans_last_pfn(const struct zone
*zone
,
1105 unsigned long start_pfn
, unsigned long nr_pages
)
1107 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1108 return zone_spans_pfn(zone
, last_pfn
);
1111 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1113 unsigned long nr_pages
= 1 << order
;
1114 unsigned long ret
, pfn
, flags
;
1117 z
= NODE_DATA(nid
)->node_zones
;
1118 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1119 spin_lock_irqsave(&z
->lock
, flags
);
1121 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1122 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1123 if (pfn_range_valid_gigantic(z
, pfn
, nr_pages
)) {
1125 * We release the zone lock here because
1126 * alloc_contig_range() will also lock the zone
1127 * at some point. If there's an allocation
1128 * spinning on this lock, it may win the race
1129 * and cause alloc_contig_range() to fail...
1131 spin_unlock_irqrestore(&z
->lock
, flags
);
1132 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1134 return pfn_to_page(pfn
);
1135 spin_lock_irqsave(&z
->lock
, flags
);
1140 spin_unlock_irqrestore(&z
->lock
, flags
);
1146 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1147 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1149 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1153 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1155 prep_compound_gigantic_page(page
, huge_page_order(h
));
1156 prep_new_huge_page(h
, page
, nid
);
1162 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1163 nodemask_t
*nodes_allowed
)
1165 struct page
*page
= NULL
;
1168 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1169 page
= alloc_fresh_gigantic_page_node(h
, node
);
1177 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1178 static inline bool gigantic_page_supported(void) { return false; }
1179 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1180 static inline void destroy_compound_gigantic_page(struct page
*page
,
1181 unsigned int order
) { }
1182 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1183 nodemask_t
*nodes_allowed
) { return 0; }
1186 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1190 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1194 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1195 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1196 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1197 1 << PG_referenced
| 1 << PG_dirty
|
1198 1 << PG_active
| 1 << PG_private
|
1201 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1202 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1203 set_page_refcounted(page
);
1204 if (hstate_is_gigantic(h
)) {
1205 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1206 free_gigantic_page(page
, huge_page_order(h
));
1208 __free_pages(page
, huge_page_order(h
));
1212 struct hstate
*size_to_hstate(unsigned long size
)
1216 for_each_hstate(h
) {
1217 if (huge_page_size(h
) == size
)
1224 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1225 * to hstate->hugepage_activelist.)
1227 * This function can be called for tail pages, but never returns true for them.
1229 bool page_huge_active(struct page
*page
)
1231 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1232 return PageHead(page
) && PagePrivate(&page
[1]);
1235 /* never called for tail page */
1236 static void set_page_huge_active(struct page
*page
)
1238 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1239 SetPagePrivate(&page
[1]);
1242 static void clear_page_huge_active(struct page
*page
)
1244 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1245 ClearPagePrivate(&page
[1]);
1248 void free_huge_page(struct page
*page
)
1251 * Can't pass hstate in here because it is called from the
1252 * compound page destructor.
1254 struct hstate
*h
= page_hstate(page
);
1255 int nid
= page_to_nid(page
);
1256 struct hugepage_subpool
*spool
=
1257 (struct hugepage_subpool
*)page_private(page
);
1258 bool restore_reserve
;
1260 set_page_private(page
, 0);
1261 page
->mapping
= NULL
;
1262 VM_BUG_ON_PAGE(page_count(page
), page
);
1263 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1264 restore_reserve
= PagePrivate(page
);
1265 ClearPagePrivate(page
);
1268 * A return code of zero implies that the subpool will be under its
1269 * minimum size if the reservation is not restored after page is free.
1270 * Therefore, force restore_reserve operation.
1272 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1273 restore_reserve
= true;
1275 spin_lock(&hugetlb_lock
);
1276 clear_page_huge_active(page
);
1277 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1278 pages_per_huge_page(h
), page
);
1279 if (restore_reserve
)
1280 h
->resv_huge_pages
++;
1282 if (h
->surplus_huge_pages_node
[nid
]) {
1283 /* remove the page from active list */
1284 list_del(&page
->lru
);
1285 update_and_free_page(h
, page
);
1286 h
->surplus_huge_pages
--;
1287 h
->surplus_huge_pages_node
[nid
]--;
1289 arch_clear_hugepage_flags(page
);
1290 enqueue_huge_page(h
, page
);
1292 spin_unlock(&hugetlb_lock
);
1295 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1297 INIT_LIST_HEAD(&page
->lru
);
1298 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1299 spin_lock(&hugetlb_lock
);
1300 set_hugetlb_cgroup(page
, NULL
);
1302 h
->nr_huge_pages_node
[nid
]++;
1303 spin_unlock(&hugetlb_lock
);
1304 put_page(page
); /* free it into the hugepage allocator */
1307 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1310 int nr_pages
= 1 << order
;
1311 struct page
*p
= page
+ 1;
1313 /* we rely on prep_new_huge_page to set the destructor */
1314 set_compound_order(page
, order
);
1315 __ClearPageReserved(page
);
1316 __SetPageHead(page
);
1317 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1319 * For gigantic hugepages allocated through bootmem at
1320 * boot, it's safer to be consistent with the not-gigantic
1321 * hugepages and clear the PG_reserved bit from all tail pages
1322 * too. Otherwse drivers using get_user_pages() to access tail
1323 * pages may get the reference counting wrong if they see
1324 * PG_reserved set on a tail page (despite the head page not
1325 * having PG_reserved set). Enforcing this consistency between
1326 * head and tail pages allows drivers to optimize away a check
1327 * on the head page when they need know if put_page() is needed
1328 * after get_user_pages().
1330 __ClearPageReserved(p
);
1331 set_page_count(p
, 0);
1332 set_compound_head(p
, page
);
1334 atomic_set(compound_mapcount_ptr(page
), -1);
1338 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1339 * transparent huge pages. See the PageTransHuge() documentation for more
1342 int PageHuge(struct page
*page
)
1344 if (!PageCompound(page
))
1347 page
= compound_head(page
);
1348 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1350 EXPORT_SYMBOL_GPL(PageHuge
);
1353 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1354 * normal or transparent huge pages.
1356 int PageHeadHuge(struct page
*page_head
)
1358 if (!PageHead(page_head
))
1361 return get_compound_page_dtor(page_head
) == free_huge_page
;
1364 pgoff_t
__basepage_index(struct page
*page
)
1366 struct page
*page_head
= compound_head(page
);
1367 pgoff_t index
= page_index(page_head
);
1368 unsigned long compound_idx
;
1370 if (!PageHuge(page_head
))
1371 return page_index(page
);
1373 if (compound_order(page_head
) >= MAX_ORDER
)
1374 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1376 compound_idx
= page
- page_head
;
1378 return (index
<< compound_order(page_head
)) + compound_idx
;
1381 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1385 page
= __alloc_pages_node(nid
,
1386 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1387 __GFP_REPEAT
|__GFP_NOWARN
,
1388 huge_page_order(h
));
1390 prep_new_huge_page(h
, page
, nid
);
1396 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1402 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1403 page
= alloc_fresh_huge_page_node(h
, node
);
1411 count_vm_event(HTLB_BUDDY_PGALLOC
);
1413 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1419 * Free huge page from pool from next node to free.
1420 * Attempt to keep persistent huge pages more or less
1421 * balanced over allowed nodes.
1422 * Called with hugetlb_lock locked.
1424 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1430 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1432 * If we're returning unused surplus pages, only examine
1433 * nodes with surplus pages.
1435 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1436 !list_empty(&h
->hugepage_freelists
[node
])) {
1438 list_entry(h
->hugepage_freelists
[node
].next
,
1440 list_del(&page
->lru
);
1441 h
->free_huge_pages
--;
1442 h
->free_huge_pages_node
[node
]--;
1444 h
->surplus_huge_pages
--;
1445 h
->surplus_huge_pages_node
[node
]--;
1447 update_and_free_page(h
, page
);
1457 * Dissolve a given free hugepage into free buddy pages. This function does
1458 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1459 * number of free hugepages would be reduced below the number of reserved
1462 static int dissolve_free_huge_page(struct page
*page
)
1466 spin_lock(&hugetlb_lock
);
1467 if (PageHuge(page
) && !page_count(page
)) {
1468 struct page
*head
= compound_head(page
);
1469 struct hstate
*h
= page_hstate(head
);
1470 int nid
= page_to_nid(head
);
1471 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0) {
1475 list_del(&head
->lru
);
1476 h
->free_huge_pages
--;
1477 h
->free_huge_pages_node
[nid
]--;
1478 h
->max_huge_pages
--;
1479 update_and_free_page(h
, head
);
1482 spin_unlock(&hugetlb_lock
);
1487 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1488 * make specified memory blocks removable from the system.
1489 * Note that this will dissolve a free gigantic hugepage completely, if any
1490 * part of it lies within the given range.
1491 * Also note that if dissolve_free_huge_page() returns with an error, all
1492 * free hugepages that were dissolved before that error are lost.
1494 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1500 if (!hugepages_supported())
1503 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1504 page
= pfn_to_page(pfn
);
1505 if (PageHuge(page
) && !page_count(page
)) {
1506 rc
= dissolve_free_huge_page(page
);
1516 * There are 3 ways this can get called:
1517 * 1. With vma+addr: we use the VMA's memory policy
1518 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1519 * page from any node, and let the buddy allocator itself figure
1521 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1522 * strictly from 'nid'
1524 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1525 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1527 int order
= huge_page_order(h
);
1528 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1529 unsigned int cpuset_mems_cookie
;
1532 * We need a VMA to get a memory policy. If we do not
1533 * have one, we use the 'nid' argument.
1535 * The mempolicy stuff below has some non-inlined bits
1536 * and calls ->vm_ops. That makes it hard to optimize at
1537 * compile-time, even when NUMA is off and it does
1538 * nothing. This helps the compiler optimize it out.
1540 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1542 * If a specific node is requested, make sure to
1543 * get memory from there, but only when a node
1544 * is explicitly specified.
1546 if (nid
!= NUMA_NO_NODE
)
1547 gfp
|= __GFP_THISNODE
;
1549 * Make sure to call something that can handle
1552 return alloc_pages_node(nid
, gfp
, order
);
1556 * OK, so we have a VMA. Fetch the mempolicy and try to
1557 * allocate a huge page with it. We will only reach this
1558 * when CONFIG_NUMA=y.
1562 struct mempolicy
*mpol
;
1564 nodemask_t
*nodemask
;
1566 cpuset_mems_cookie
= read_mems_allowed_begin();
1567 nid
= huge_node(vma
, addr
, gfp
, &mpol
, &nodemask
);
1568 mpol_cond_put(mpol
);
1569 page
= __alloc_pages_nodemask(gfp
, order
, nid
, nodemask
);
1572 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1578 * There are two ways to allocate a huge page:
1579 * 1. When you have a VMA and an address (like a fault)
1580 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1582 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1583 * this case which signifies that the allocation should be done with
1584 * respect for the VMA's memory policy.
1586 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1587 * implies that memory policies will not be taken in to account.
1589 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1590 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1595 if (hstate_is_gigantic(h
))
1599 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1600 * This makes sure the caller is picking _one_ of the modes with which
1601 * we can call this function, not both.
1603 if (vma
|| (addr
!= -1)) {
1604 VM_WARN_ON_ONCE(addr
== -1);
1605 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1608 * Assume we will successfully allocate the surplus page to
1609 * prevent racing processes from causing the surplus to exceed
1612 * This however introduces a different race, where a process B
1613 * tries to grow the static hugepage pool while alloc_pages() is
1614 * called by process A. B will only examine the per-node
1615 * counters in determining if surplus huge pages can be
1616 * converted to normal huge pages in adjust_pool_surplus(). A
1617 * won't be able to increment the per-node counter, until the
1618 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1619 * no more huge pages can be converted from surplus to normal
1620 * state (and doesn't try to convert again). Thus, we have a
1621 * case where a surplus huge page exists, the pool is grown, and
1622 * the surplus huge page still exists after, even though it
1623 * should just have been converted to a normal huge page. This
1624 * does not leak memory, though, as the hugepage will be freed
1625 * once it is out of use. It also does not allow the counters to
1626 * go out of whack in adjust_pool_surplus() as we don't modify
1627 * the node values until we've gotten the hugepage and only the
1628 * per-node value is checked there.
1630 spin_lock(&hugetlb_lock
);
1631 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1632 spin_unlock(&hugetlb_lock
);
1636 h
->surplus_huge_pages
++;
1638 spin_unlock(&hugetlb_lock
);
1640 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1642 spin_lock(&hugetlb_lock
);
1644 INIT_LIST_HEAD(&page
->lru
);
1645 r_nid
= page_to_nid(page
);
1646 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1647 set_hugetlb_cgroup(page
, NULL
);
1649 * We incremented the global counters already
1651 h
->nr_huge_pages_node
[r_nid
]++;
1652 h
->surplus_huge_pages_node
[r_nid
]++;
1653 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1656 h
->surplus_huge_pages
--;
1657 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1659 spin_unlock(&hugetlb_lock
);
1665 * Allocate a huge page from 'nid'. Note, 'nid' may be
1666 * NUMA_NO_NODE, which means that it may be allocated
1670 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1672 unsigned long addr
= -1;
1674 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1678 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1681 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1682 struct vm_area_struct
*vma
, unsigned long addr
)
1684 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1688 * This allocation function is useful in the context where vma is irrelevant.
1689 * E.g. soft-offlining uses this function because it only cares physical
1690 * address of error page.
1692 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1694 struct page
*page
= NULL
;
1696 spin_lock(&hugetlb_lock
);
1697 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1698 page
= dequeue_huge_page_node(h
, nid
);
1699 spin_unlock(&hugetlb_lock
);
1702 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1708 * Increase the hugetlb pool such that it can accommodate a reservation
1711 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1713 struct list_head surplus_list
;
1714 struct page
*page
, *tmp
;
1716 int needed
, allocated
;
1717 bool alloc_ok
= true;
1719 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1721 h
->resv_huge_pages
+= delta
;
1726 INIT_LIST_HEAD(&surplus_list
);
1730 spin_unlock(&hugetlb_lock
);
1731 for (i
= 0; i
< needed
; i
++) {
1732 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1737 list_add(&page
->lru
, &surplus_list
);
1742 * After retaking hugetlb_lock, we need to recalculate 'needed'
1743 * because either resv_huge_pages or free_huge_pages may have changed.
1745 spin_lock(&hugetlb_lock
);
1746 needed
= (h
->resv_huge_pages
+ delta
) -
1747 (h
->free_huge_pages
+ allocated
);
1752 * We were not able to allocate enough pages to
1753 * satisfy the entire reservation so we free what
1754 * we've allocated so far.
1759 * The surplus_list now contains _at_least_ the number of extra pages
1760 * needed to accommodate the reservation. Add the appropriate number
1761 * of pages to the hugetlb pool and free the extras back to the buddy
1762 * allocator. Commit the entire reservation here to prevent another
1763 * process from stealing the pages as they are added to the pool but
1764 * before they are reserved.
1766 needed
+= allocated
;
1767 h
->resv_huge_pages
+= delta
;
1770 /* Free the needed pages to the hugetlb pool */
1771 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1775 * This page is now managed by the hugetlb allocator and has
1776 * no users -- drop the buddy allocator's reference.
1778 put_page_testzero(page
);
1779 VM_BUG_ON_PAGE(page_count(page
), page
);
1780 enqueue_huge_page(h
, page
);
1783 spin_unlock(&hugetlb_lock
);
1785 /* Free unnecessary surplus pages to the buddy allocator */
1786 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1788 spin_lock(&hugetlb_lock
);
1794 * This routine has two main purposes:
1795 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1796 * in unused_resv_pages. This corresponds to the prior adjustments made
1797 * to the associated reservation map.
1798 * 2) Free any unused surplus pages that may have been allocated to satisfy
1799 * the reservation. As many as unused_resv_pages may be freed.
1801 * Called with hugetlb_lock held. However, the lock could be dropped (and
1802 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1803 * we must make sure nobody else can claim pages we are in the process of
1804 * freeing. Do this by ensuring resv_huge_page always is greater than the
1805 * number of huge pages we plan to free when dropping the lock.
1807 static void return_unused_surplus_pages(struct hstate
*h
,
1808 unsigned long unused_resv_pages
)
1810 unsigned long nr_pages
;
1812 /* Cannot return gigantic pages currently */
1813 if (hstate_is_gigantic(h
))
1817 * Part (or even all) of the reservation could have been backed
1818 * by pre-allocated pages. Only free surplus pages.
1820 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1823 * We want to release as many surplus pages as possible, spread
1824 * evenly across all nodes with memory. Iterate across these nodes
1825 * until we can no longer free unreserved surplus pages. This occurs
1826 * when the nodes with surplus pages have no free pages.
1827 * free_pool_huge_page() will balance the the freed pages across the
1828 * on-line nodes with memory and will handle the hstate accounting.
1830 * Note that we decrement resv_huge_pages as we free the pages. If
1831 * we drop the lock, resv_huge_pages will still be sufficiently large
1832 * to cover subsequent pages we may free.
1834 while (nr_pages
--) {
1835 h
->resv_huge_pages
--;
1836 unused_resv_pages
--;
1837 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1839 cond_resched_lock(&hugetlb_lock
);
1843 /* Fully uncommit the reservation */
1844 h
->resv_huge_pages
-= unused_resv_pages
;
1849 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1850 * are used by the huge page allocation routines to manage reservations.
1852 * vma_needs_reservation is called to determine if the huge page at addr
1853 * within the vma has an associated reservation. If a reservation is
1854 * needed, the value 1 is returned. The caller is then responsible for
1855 * managing the global reservation and subpool usage counts. After
1856 * the huge page has been allocated, vma_commit_reservation is called
1857 * to add the page to the reservation map. If the page allocation fails,
1858 * the reservation must be ended instead of committed. vma_end_reservation
1859 * is called in such cases.
1861 * In the normal case, vma_commit_reservation returns the same value
1862 * as the preceding vma_needs_reservation call. The only time this
1863 * is not the case is if a reserve map was changed between calls. It
1864 * is the responsibility of the caller to notice the difference and
1865 * take appropriate action.
1867 * vma_add_reservation is used in error paths where a reservation must
1868 * be restored when a newly allocated huge page must be freed. It is
1869 * to be called after calling vma_needs_reservation to determine if a
1870 * reservation exists.
1872 enum vma_resv_mode
{
1878 static long __vma_reservation_common(struct hstate
*h
,
1879 struct vm_area_struct
*vma
, unsigned long addr
,
1880 enum vma_resv_mode mode
)
1882 struct resv_map
*resv
;
1886 resv
= vma_resv_map(vma
);
1890 idx
= vma_hugecache_offset(h
, vma
, addr
);
1892 case VMA_NEEDS_RESV
:
1893 ret
= region_chg(resv
, idx
, idx
+ 1);
1895 case VMA_COMMIT_RESV
:
1896 ret
= region_add(resv
, idx
, idx
+ 1);
1899 region_abort(resv
, idx
, idx
+ 1);
1903 if (vma
->vm_flags
& VM_MAYSHARE
)
1904 ret
= region_add(resv
, idx
, idx
+ 1);
1906 region_abort(resv
, idx
, idx
+ 1);
1907 ret
= region_del(resv
, idx
, idx
+ 1);
1914 if (vma
->vm_flags
& VM_MAYSHARE
)
1916 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1918 * In most cases, reserves always exist for private mappings.
1919 * However, a file associated with mapping could have been
1920 * hole punched or truncated after reserves were consumed.
1921 * As subsequent fault on such a range will not use reserves.
1922 * Subtle - The reserve map for private mappings has the
1923 * opposite meaning than that of shared mappings. If NO
1924 * entry is in the reserve map, it means a reservation exists.
1925 * If an entry exists in the reserve map, it means the
1926 * reservation has already been consumed. As a result, the
1927 * return value of this routine is the opposite of the
1928 * value returned from reserve map manipulation routines above.
1936 return ret
< 0 ? ret
: 0;
1939 static long vma_needs_reservation(struct hstate
*h
,
1940 struct vm_area_struct
*vma
, unsigned long addr
)
1942 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1945 static long vma_commit_reservation(struct hstate
*h
,
1946 struct vm_area_struct
*vma
, unsigned long addr
)
1948 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1951 static void vma_end_reservation(struct hstate
*h
,
1952 struct vm_area_struct
*vma
, unsigned long addr
)
1954 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1957 static long vma_add_reservation(struct hstate
*h
,
1958 struct vm_area_struct
*vma
, unsigned long addr
)
1960 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
1964 * This routine is called to restore a reservation on error paths. In the
1965 * specific error paths, a huge page was allocated (via alloc_huge_page)
1966 * and is about to be freed. If a reservation for the page existed,
1967 * alloc_huge_page would have consumed the reservation and set PagePrivate
1968 * in the newly allocated page. When the page is freed via free_huge_page,
1969 * the global reservation count will be incremented if PagePrivate is set.
1970 * However, free_huge_page can not adjust the reserve map. Adjust the
1971 * reserve map here to be consistent with global reserve count adjustments
1972 * to be made by free_huge_page.
1974 static void restore_reserve_on_error(struct hstate
*h
,
1975 struct vm_area_struct
*vma
, unsigned long address
,
1978 if (unlikely(PagePrivate(page
))) {
1979 long rc
= vma_needs_reservation(h
, vma
, address
);
1981 if (unlikely(rc
< 0)) {
1983 * Rare out of memory condition in reserve map
1984 * manipulation. Clear PagePrivate so that
1985 * global reserve count will not be incremented
1986 * by free_huge_page. This will make it appear
1987 * as though the reservation for this page was
1988 * consumed. This may prevent the task from
1989 * faulting in the page at a later time. This
1990 * is better than inconsistent global huge page
1991 * accounting of reserve counts.
1993 ClearPagePrivate(page
);
1995 rc
= vma_add_reservation(h
, vma
, address
);
1996 if (unlikely(rc
< 0))
1998 * See above comment about rare out of
2001 ClearPagePrivate(page
);
2003 vma_end_reservation(h
, vma
, address
);
2007 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2008 unsigned long addr
, int avoid_reserve
)
2010 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2011 struct hstate
*h
= hstate_vma(vma
);
2013 long map_chg
, map_commit
;
2016 struct hugetlb_cgroup
*h_cg
;
2018 idx
= hstate_index(h
);
2020 * Examine the region/reserve map to determine if the process
2021 * has a reservation for the page to be allocated. A return
2022 * code of zero indicates a reservation exists (no change).
2024 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2026 return ERR_PTR(-ENOMEM
);
2029 * Processes that did not create the mapping will have no
2030 * reserves as indicated by the region/reserve map. Check
2031 * that the allocation will not exceed the subpool limit.
2032 * Allocations for MAP_NORESERVE mappings also need to be
2033 * checked against any subpool limit.
2035 if (map_chg
|| avoid_reserve
) {
2036 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2038 vma_end_reservation(h
, vma
, addr
);
2039 return ERR_PTR(-ENOSPC
);
2043 * Even though there was no reservation in the region/reserve
2044 * map, there could be reservations associated with the
2045 * subpool that can be used. This would be indicated if the
2046 * return value of hugepage_subpool_get_pages() is zero.
2047 * However, if avoid_reserve is specified we still avoid even
2048 * the subpool reservations.
2054 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2056 goto out_subpool_put
;
2058 spin_lock(&hugetlb_lock
);
2060 * glb_chg is passed to indicate whether or not a page must be taken
2061 * from the global free pool (global change). gbl_chg == 0 indicates
2062 * a reservation exists for the allocation.
2064 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2066 spin_unlock(&hugetlb_lock
);
2067 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2069 goto out_uncharge_cgroup
;
2070 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2071 SetPagePrivate(page
);
2072 h
->resv_huge_pages
--;
2074 spin_lock(&hugetlb_lock
);
2075 list_move(&page
->lru
, &h
->hugepage_activelist
);
2078 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2079 spin_unlock(&hugetlb_lock
);
2081 set_page_private(page
, (unsigned long)spool
);
2083 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2084 if (unlikely(map_chg
> map_commit
)) {
2086 * The page was added to the reservation map between
2087 * vma_needs_reservation and vma_commit_reservation.
2088 * This indicates a race with hugetlb_reserve_pages.
2089 * Adjust for the subpool count incremented above AND
2090 * in hugetlb_reserve_pages for the same page. Also,
2091 * the reservation count added in hugetlb_reserve_pages
2092 * no longer applies.
2096 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2097 hugetlb_acct_memory(h
, -rsv_adjust
);
2101 out_uncharge_cgroup
:
2102 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2104 if (map_chg
|| avoid_reserve
)
2105 hugepage_subpool_put_pages(spool
, 1);
2106 vma_end_reservation(h
, vma
, addr
);
2107 return ERR_PTR(-ENOSPC
);
2111 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2112 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2113 * where no ERR_VALUE is expected to be returned.
2115 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
2116 unsigned long addr
, int avoid_reserve
)
2118 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
2124 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
2126 struct huge_bootmem_page
*m
;
2129 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2132 addr
= memblock_virt_alloc_try_nid_nopanic(
2133 huge_page_size(h
), huge_page_size(h
),
2134 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
2137 * Use the beginning of the huge page to store the
2138 * huge_bootmem_page struct (until gather_bootmem
2139 * puts them into the mem_map).
2148 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2149 /* Put them into a private list first because mem_map is not up yet */
2150 list_add(&m
->list
, &huge_boot_pages
);
2155 static void __init
prep_compound_huge_page(struct page
*page
,
2158 if (unlikely(order
> (MAX_ORDER
- 1)))
2159 prep_compound_gigantic_page(page
, order
);
2161 prep_compound_page(page
, order
);
2164 /* Put bootmem huge pages into the standard lists after mem_map is up */
2165 static void __init
gather_bootmem_prealloc(void)
2167 struct huge_bootmem_page
*m
;
2169 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2170 struct hstate
*h
= m
->hstate
;
2173 #ifdef CONFIG_HIGHMEM
2174 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2175 memblock_free_late(__pa(m
),
2176 sizeof(struct huge_bootmem_page
));
2178 page
= virt_to_page(m
);
2180 WARN_ON(page_count(page
) != 1);
2181 prep_compound_huge_page(page
, h
->order
);
2182 WARN_ON(PageReserved(page
));
2183 prep_new_huge_page(h
, page
, page_to_nid(page
));
2185 * If we had gigantic hugepages allocated at boot time, we need
2186 * to restore the 'stolen' pages to totalram_pages in order to
2187 * fix confusing memory reports from free(1) and another
2188 * side-effects, like CommitLimit going negative.
2190 if (hstate_is_gigantic(h
))
2191 adjust_managed_page_count(page
, 1 << h
->order
);
2195 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2199 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2200 if (hstate_is_gigantic(h
)) {
2201 if (!alloc_bootmem_huge_page(h
))
2203 } else if (!alloc_fresh_huge_page(h
,
2204 &node_states
[N_MEMORY
]))
2207 h
->max_huge_pages
= i
;
2210 static void __init
hugetlb_init_hstates(void)
2214 for_each_hstate(h
) {
2215 if (minimum_order
> huge_page_order(h
))
2216 minimum_order
= huge_page_order(h
);
2218 /* oversize hugepages were init'ed in early boot */
2219 if (!hstate_is_gigantic(h
))
2220 hugetlb_hstate_alloc_pages(h
);
2222 VM_BUG_ON(minimum_order
== UINT_MAX
);
2225 static char * __init
memfmt(char *buf
, unsigned long n
)
2227 if (n
>= (1UL << 30))
2228 sprintf(buf
, "%lu GB", n
>> 30);
2229 else if (n
>= (1UL << 20))
2230 sprintf(buf
, "%lu MB", n
>> 20);
2232 sprintf(buf
, "%lu KB", n
>> 10);
2236 static void __init
report_hugepages(void)
2240 for_each_hstate(h
) {
2242 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2243 memfmt(buf
, huge_page_size(h
)),
2244 h
->free_huge_pages
);
2248 #ifdef CONFIG_HIGHMEM
2249 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2250 nodemask_t
*nodes_allowed
)
2254 if (hstate_is_gigantic(h
))
2257 for_each_node_mask(i
, *nodes_allowed
) {
2258 struct page
*page
, *next
;
2259 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2260 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2261 if (count
>= h
->nr_huge_pages
)
2263 if (PageHighMem(page
))
2265 list_del(&page
->lru
);
2266 update_and_free_page(h
, page
);
2267 h
->free_huge_pages
--;
2268 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2273 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2274 nodemask_t
*nodes_allowed
)
2280 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2281 * balanced by operating on them in a round-robin fashion.
2282 * Returns 1 if an adjustment was made.
2284 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2289 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2292 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2293 if (h
->surplus_huge_pages_node
[node
])
2297 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2298 if (h
->surplus_huge_pages_node
[node
] <
2299 h
->nr_huge_pages_node
[node
])
2306 h
->surplus_huge_pages
+= delta
;
2307 h
->surplus_huge_pages_node
[node
] += delta
;
2311 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2312 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2313 nodemask_t
*nodes_allowed
)
2315 unsigned long min_count
, ret
;
2317 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2318 return h
->max_huge_pages
;
2321 * Increase the pool size
2322 * First take pages out of surplus state. Then make up the
2323 * remaining difference by allocating fresh huge pages.
2325 * We might race with __alloc_buddy_huge_page() here and be unable
2326 * to convert a surplus huge page to a normal huge page. That is
2327 * not critical, though, it just means the overall size of the
2328 * pool might be one hugepage larger than it needs to be, but
2329 * within all the constraints specified by the sysctls.
2331 spin_lock(&hugetlb_lock
);
2332 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2333 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2337 while (count
> persistent_huge_pages(h
)) {
2339 * If this allocation races such that we no longer need the
2340 * page, free_huge_page will handle it by freeing the page
2341 * and reducing the surplus.
2343 spin_unlock(&hugetlb_lock
);
2345 /* yield cpu to avoid soft lockup */
2348 if (hstate_is_gigantic(h
))
2349 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2351 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2352 spin_lock(&hugetlb_lock
);
2356 /* Bail for signals. Probably ctrl-c from user */
2357 if (signal_pending(current
))
2362 * Decrease the pool size
2363 * First return free pages to the buddy allocator (being careful
2364 * to keep enough around to satisfy reservations). Then place
2365 * pages into surplus state as needed so the pool will shrink
2366 * to the desired size as pages become free.
2368 * By placing pages into the surplus state independent of the
2369 * overcommit value, we are allowing the surplus pool size to
2370 * exceed overcommit. There are few sane options here. Since
2371 * __alloc_buddy_huge_page() is checking the global counter,
2372 * though, we'll note that we're not allowed to exceed surplus
2373 * and won't grow the pool anywhere else. Not until one of the
2374 * sysctls are changed, or the surplus pages go out of use.
2376 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2377 min_count
= max(count
, min_count
);
2378 try_to_free_low(h
, min_count
, nodes_allowed
);
2379 while (min_count
< persistent_huge_pages(h
)) {
2380 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2382 cond_resched_lock(&hugetlb_lock
);
2384 while (count
< persistent_huge_pages(h
)) {
2385 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2389 ret
= persistent_huge_pages(h
);
2390 spin_unlock(&hugetlb_lock
);
2394 #define HSTATE_ATTR_RO(_name) \
2395 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2397 #define HSTATE_ATTR(_name) \
2398 static struct kobj_attribute _name##_attr = \
2399 __ATTR(_name, 0644, _name##_show, _name##_store)
2401 static struct kobject
*hugepages_kobj
;
2402 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2404 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2406 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2410 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2411 if (hstate_kobjs
[i
] == kobj
) {
2413 *nidp
= NUMA_NO_NODE
;
2417 return kobj_to_node_hstate(kobj
, nidp
);
2420 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2421 struct kobj_attribute
*attr
, char *buf
)
2424 unsigned long nr_huge_pages
;
2427 h
= kobj_to_hstate(kobj
, &nid
);
2428 if (nid
== NUMA_NO_NODE
)
2429 nr_huge_pages
= h
->nr_huge_pages
;
2431 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2433 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2436 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2437 struct hstate
*h
, int nid
,
2438 unsigned long count
, size_t len
)
2441 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2443 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2448 if (nid
== NUMA_NO_NODE
) {
2450 * global hstate attribute
2452 if (!(obey_mempolicy
&&
2453 init_nodemask_of_mempolicy(nodes_allowed
))) {
2454 NODEMASK_FREE(nodes_allowed
);
2455 nodes_allowed
= &node_states
[N_MEMORY
];
2457 } else if (nodes_allowed
) {
2459 * per node hstate attribute: adjust count to global,
2460 * but restrict alloc/free to the specified node.
2462 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2463 init_nodemask_of_node(nodes_allowed
, nid
);
2465 nodes_allowed
= &node_states
[N_MEMORY
];
2467 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2469 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2470 NODEMASK_FREE(nodes_allowed
);
2474 NODEMASK_FREE(nodes_allowed
);
2478 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2479 struct kobject
*kobj
, const char *buf
,
2483 unsigned long count
;
2487 err
= kstrtoul(buf
, 10, &count
);
2491 h
= kobj_to_hstate(kobj
, &nid
);
2492 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2495 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2496 struct kobj_attribute
*attr
, char *buf
)
2498 return nr_hugepages_show_common(kobj
, attr
, buf
);
2501 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2502 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2504 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2506 HSTATE_ATTR(nr_hugepages
);
2511 * hstate attribute for optionally mempolicy-based constraint on persistent
2512 * huge page alloc/free.
2514 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2515 struct kobj_attribute
*attr
, char *buf
)
2517 return nr_hugepages_show_common(kobj
, attr
, buf
);
2520 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2521 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2523 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2525 HSTATE_ATTR(nr_hugepages_mempolicy
);
2529 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2530 struct kobj_attribute
*attr
, char *buf
)
2532 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2533 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2536 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2537 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2540 unsigned long input
;
2541 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2543 if (hstate_is_gigantic(h
))
2546 err
= kstrtoul(buf
, 10, &input
);
2550 spin_lock(&hugetlb_lock
);
2551 h
->nr_overcommit_huge_pages
= input
;
2552 spin_unlock(&hugetlb_lock
);
2556 HSTATE_ATTR(nr_overcommit_hugepages
);
2558 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2559 struct kobj_attribute
*attr
, char *buf
)
2562 unsigned long free_huge_pages
;
2565 h
= kobj_to_hstate(kobj
, &nid
);
2566 if (nid
== NUMA_NO_NODE
)
2567 free_huge_pages
= h
->free_huge_pages
;
2569 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2571 return sprintf(buf
, "%lu\n", free_huge_pages
);
2573 HSTATE_ATTR_RO(free_hugepages
);
2575 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2576 struct kobj_attribute
*attr
, char *buf
)
2578 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2579 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2581 HSTATE_ATTR_RO(resv_hugepages
);
2583 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2584 struct kobj_attribute
*attr
, char *buf
)
2587 unsigned long surplus_huge_pages
;
2590 h
= kobj_to_hstate(kobj
, &nid
);
2591 if (nid
== NUMA_NO_NODE
)
2592 surplus_huge_pages
= h
->surplus_huge_pages
;
2594 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2596 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2598 HSTATE_ATTR_RO(surplus_hugepages
);
2600 static struct attribute
*hstate_attrs
[] = {
2601 &nr_hugepages_attr
.attr
,
2602 &nr_overcommit_hugepages_attr
.attr
,
2603 &free_hugepages_attr
.attr
,
2604 &resv_hugepages_attr
.attr
,
2605 &surplus_hugepages_attr
.attr
,
2607 &nr_hugepages_mempolicy_attr
.attr
,
2612 static struct attribute_group hstate_attr_group
= {
2613 .attrs
= hstate_attrs
,
2616 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2617 struct kobject
**hstate_kobjs
,
2618 struct attribute_group
*hstate_attr_group
)
2621 int hi
= hstate_index(h
);
2623 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2624 if (!hstate_kobjs
[hi
])
2627 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2629 kobject_put(hstate_kobjs
[hi
]);
2634 static void __init
hugetlb_sysfs_init(void)
2639 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2640 if (!hugepages_kobj
)
2643 for_each_hstate(h
) {
2644 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2645 hstate_kobjs
, &hstate_attr_group
);
2647 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2654 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2655 * with node devices in node_devices[] using a parallel array. The array
2656 * index of a node device or _hstate == node id.
2657 * This is here to avoid any static dependency of the node device driver, in
2658 * the base kernel, on the hugetlb module.
2660 struct node_hstate
{
2661 struct kobject
*hugepages_kobj
;
2662 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2664 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2667 * A subset of global hstate attributes for node devices
2669 static struct attribute
*per_node_hstate_attrs
[] = {
2670 &nr_hugepages_attr
.attr
,
2671 &free_hugepages_attr
.attr
,
2672 &surplus_hugepages_attr
.attr
,
2676 static struct attribute_group per_node_hstate_attr_group
= {
2677 .attrs
= per_node_hstate_attrs
,
2681 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2682 * Returns node id via non-NULL nidp.
2684 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2688 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2689 struct node_hstate
*nhs
= &node_hstates
[nid
];
2691 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2692 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2704 * Unregister hstate attributes from a single node device.
2705 * No-op if no hstate attributes attached.
2707 static void hugetlb_unregister_node(struct node
*node
)
2710 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2712 if (!nhs
->hugepages_kobj
)
2713 return; /* no hstate attributes */
2715 for_each_hstate(h
) {
2716 int idx
= hstate_index(h
);
2717 if (nhs
->hstate_kobjs
[idx
]) {
2718 kobject_put(nhs
->hstate_kobjs
[idx
]);
2719 nhs
->hstate_kobjs
[idx
] = NULL
;
2723 kobject_put(nhs
->hugepages_kobj
);
2724 nhs
->hugepages_kobj
= NULL
;
2729 * Register hstate attributes for a single node device.
2730 * No-op if attributes already registered.
2732 static void hugetlb_register_node(struct node
*node
)
2735 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2738 if (nhs
->hugepages_kobj
)
2739 return; /* already allocated */
2741 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2743 if (!nhs
->hugepages_kobj
)
2746 for_each_hstate(h
) {
2747 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2749 &per_node_hstate_attr_group
);
2751 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2752 h
->name
, node
->dev
.id
);
2753 hugetlb_unregister_node(node
);
2760 * hugetlb init time: register hstate attributes for all registered node
2761 * devices of nodes that have memory. All on-line nodes should have
2762 * registered their associated device by this time.
2764 static void __init
hugetlb_register_all_nodes(void)
2768 for_each_node_state(nid
, N_MEMORY
) {
2769 struct node
*node
= node_devices
[nid
];
2770 if (node
->dev
.id
== nid
)
2771 hugetlb_register_node(node
);
2775 * Let the node device driver know we're here so it can
2776 * [un]register hstate attributes on node hotplug.
2778 register_hugetlbfs_with_node(hugetlb_register_node
,
2779 hugetlb_unregister_node
);
2781 #else /* !CONFIG_NUMA */
2783 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2791 static void hugetlb_register_all_nodes(void) { }
2795 static int __init
hugetlb_init(void)
2799 if (!hugepages_supported())
2802 if (!size_to_hstate(default_hstate_size
)) {
2803 default_hstate_size
= HPAGE_SIZE
;
2804 if (!size_to_hstate(default_hstate_size
))
2805 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2807 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2808 if (default_hstate_max_huge_pages
) {
2809 if (!default_hstate
.max_huge_pages
)
2810 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2813 hugetlb_init_hstates();
2814 gather_bootmem_prealloc();
2817 hugetlb_sysfs_init();
2818 hugetlb_register_all_nodes();
2819 hugetlb_cgroup_file_init();
2822 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2824 num_fault_mutexes
= 1;
2826 hugetlb_fault_mutex_table
=
2827 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2828 BUG_ON(!hugetlb_fault_mutex_table
);
2830 for (i
= 0; i
< num_fault_mutexes
; i
++)
2831 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2834 subsys_initcall(hugetlb_init
);
2836 /* Should be called on processing a hugepagesz=... option */
2837 void __init
hugetlb_bad_size(void)
2839 parsed_valid_hugepagesz
= false;
2842 void __init
hugetlb_add_hstate(unsigned int order
)
2847 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2848 pr_warn("hugepagesz= specified twice, ignoring\n");
2851 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2853 h
= &hstates
[hugetlb_max_hstate
++];
2855 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2856 h
->nr_huge_pages
= 0;
2857 h
->free_huge_pages
= 0;
2858 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2859 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2860 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2861 h
->next_nid_to_alloc
= first_memory_node
;
2862 h
->next_nid_to_free
= first_memory_node
;
2863 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2864 huge_page_size(h
)/1024);
2869 static int __init
hugetlb_nrpages_setup(char *s
)
2872 static unsigned long *last_mhp
;
2874 if (!parsed_valid_hugepagesz
) {
2875 pr_warn("hugepages = %s preceded by "
2876 "an unsupported hugepagesz, ignoring\n", s
);
2877 parsed_valid_hugepagesz
= true;
2881 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2882 * so this hugepages= parameter goes to the "default hstate".
2884 else if (!hugetlb_max_hstate
)
2885 mhp
= &default_hstate_max_huge_pages
;
2887 mhp
= &parsed_hstate
->max_huge_pages
;
2889 if (mhp
== last_mhp
) {
2890 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2894 if (sscanf(s
, "%lu", mhp
) <= 0)
2898 * Global state is always initialized later in hugetlb_init.
2899 * But we need to allocate >= MAX_ORDER hstates here early to still
2900 * use the bootmem allocator.
2902 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2903 hugetlb_hstate_alloc_pages(parsed_hstate
);
2909 __setup("hugepages=", hugetlb_nrpages_setup
);
2911 static int __init
hugetlb_default_setup(char *s
)
2913 default_hstate_size
= memparse(s
, &s
);
2916 __setup("default_hugepagesz=", hugetlb_default_setup
);
2918 static unsigned int cpuset_mems_nr(unsigned int *array
)
2921 unsigned int nr
= 0;
2923 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2929 #ifdef CONFIG_SYSCTL
2930 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2931 struct ctl_table
*table
, int write
,
2932 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2934 struct hstate
*h
= &default_hstate
;
2935 unsigned long tmp
= h
->max_huge_pages
;
2938 if (!hugepages_supported())
2942 table
->maxlen
= sizeof(unsigned long);
2943 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2948 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2949 NUMA_NO_NODE
, tmp
, *length
);
2954 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2955 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2958 return hugetlb_sysctl_handler_common(false, table
, write
,
2959 buffer
, length
, ppos
);
2963 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2964 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2966 return hugetlb_sysctl_handler_common(true, table
, write
,
2967 buffer
, length
, ppos
);
2969 #endif /* CONFIG_NUMA */
2971 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2972 void __user
*buffer
,
2973 size_t *length
, loff_t
*ppos
)
2975 struct hstate
*h
= &default_hstate
;
2979 if (!hugepages_supported())
2982 tmp
= h
->nr_overcommit_huge_pages
;
2984 if (write
&& hstate_is_gigantic(h
))
2988 table
->maxlen
= sizeof(unsigned long);
2989 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2994 spin_lock(&hugetlb_lock
);
2995 h
->nr_overcommit_huge_pages
= tmp
;
2996 spin_unlock(&hugetlb_lock
);
3002 #endif /* CONFIG_SYSCTL */
3004 void hugetlb_report_meminfo(struct seq_file
*m
)
3006 struct hstate
*h
= &default_hstate
;
3007 if (!hugepages_supported())
3010 "HugePages_Total: %5lu\n"
3011 "HugePages_Free: %5lu\n"
3012 "HugePages_Rsvd: %5lu\n"
3013 "HugePages_Surp: %5lu\n"
3014 "Hugepagesize: %8lu kB\n",
3018 h
->surplus_huge_pages
,
3019 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3022 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3024 struct hstate
*h
= &default_hstate
;
3025 if (!hugepages_supported())
3028 "Node %d HugePages_Total: %5u\n"
3029 "Node %d HugePages_Free: %5u\n"
3030 "Node %d HugePages_Surp: %5u\n",
3031 nid
, h
->nr_huge_pages_node
[nid
],
3032 nid
, h
->free_huge_pages_node
[nid
],
3033 nid
, h
->surplus_huge_pages_node
[nid
]);
3036 void hugetlb_show_meminfo(void)
3041 if (!hugepages_supported())
3044 for_each_node_state(nid
, N_MEMORY
)
3046 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3048 h
->nr_huge_pages_node
[nid
],
3049 h
->free_huge_pages_node
[nid
],
3050 h
->surplus_huge_pages_node
[nid
],
3051 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3054 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3056 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3057 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3060 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3061 unsigned long hugetlb_total_pages(void)
3064 unsigned long nr_total_pages
= 0;
3067 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3068 return nr_total_pages
;
3071 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3075 spin_lock(&hugetlb_lock
);
3077 * When cpuset is configured, it breaks the strict hugetlb page
3078 * reservation as the accounting is done on a global variable. Such
3079 * reservation is completely rubbish in the presence of cpuset because
3080 * the reservation is not checked against page availability for the
3081 * current cpuset. Application can still potentially OOM'ed by kernel
3082 * with lack of free htlb page in cpuset that the task is in.
3083 * Attempt to enforce strict accounting with cpuset is almost
3084 * impossible (or too ugly) because cpuset is too fluid that
3085 * task or memory node can be dynamically moved between cpusets.
3087 * The change of semantics for shared hugetlb mapping with cpuset is
3088 * undesirable. However, in order to preserve some of the semantics,
3089 * we fall back to check against current free page availability as
3090 * a best attempt and hopefully to minimize the impact of changing
3091 * semantics that cpuset has.
3094 if (gather_surplus_pages(h
, delta
) < 0)
3097 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3098 return_unused_surplus_pages(h
, delta
);
3105 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3108 spin_unlock(&hugetlb_lock
);
3112 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3114 struct resv_map
*resv
= vma_resv_map(vma
);
3117 * This new VMA should share its siblings reservation map if present.
3118 * The VMA will only ever have a valid reservation map pointer where
3119 * it is being copied for another still existing VMA. As that VMA
3120 * has a reference to the reservation map it cannot disappear until
3121 * after this open call completes. It is therefore safe to take a
3122 * new reference here without additional locking.
3124 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3125 kref_get(&resv
->refs
);
3128 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3130 struct hstate
*h
= hstate_vma(vma
);
3131 struct resv_map
*resv
= vma_resv_map(vma
);
3132 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3133 unsigned long reserve
, start
, end
;
3136 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3139 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3140 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3142 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3144 kref_put(&resv
->refs
, resv_map_release
);
3148 * Decrement reserve counts. The global reserve count may be
3149 * adjusted if the subpool has a minimum size.
3151 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3152 hugetlb_acct_memory(h
, -gbl_reserve
);
3157 * We cannot handle pagefaults against hugetlb pages at all. They cause
3158 * handle_mm_fault() to try to instantiate regular-sized pages in the
3159 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3162 static int hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3168 const struct vm_operations_struct hugetlb_vm_ops
= {
3169 .fault
= hugetlb_vm_op_fault
,
3170 .open
= hugetlb_vm_op_open
,
3171 .close
= hugetlb_vm_op_close
,
3174 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3180 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3181 vma
->vm_page_prot
)));
3183 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3184 vma
->vm_page_prot
));
3186 entry
= pte_mkyoung(entry
);
3187 entry
= pte_mkhuge(entry
);
3188 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3193 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3194 unsigned long address
, pte_t
*ptep
)
3198 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3199 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3200 update_mmu_cache(vma
, address
, ptep
);
3203 bool is_hugetlb_entry_migration(pte_t pte
)
3207 if (huge_pte_none(pte
) || pte_present(pte
))
3209 swp
= pte_to_swp_entry(pte
);
3210 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3216 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3220 if (huge_pte_none(pte
) || pte_present(pte
))
3222 swp
= pte_to_swp_entry(pte
);
3223 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3229 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3230 struct vm_area_struct
*vma
)
3232 pte_t
*src_pte
, *dst_pte
, entry
;
3233 struct page
*ptepage
;
3236 struct hstate
*h
= hstate_vma(vma
);
3237 unsigned long sz
= huge_page_size(h
);
3238 unsigned long mmun_start
; /* For mmu_notifiers */
3239 unsigned long mmun_end
; /* For mmu_notifiers */
3242 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3244 mmun_start
= vma
->vm_start
;
3245 mmun_end
= vma
->vm_end
;
3247 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3249 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3250 spinlock_t
*src_ptl
, *dst_ptl
;
3251 src_pte
= huge_pte_offset(src
, addr
, sz
);
3254 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3260 /* If the pagetables are shared don't copy or take references */
3261 if (dst_pte
== src_pte
)
3264 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3265 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3266 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3267 entry
= huge_ptep_get(src_pte
);
3268 if (huge_pte_none(entry
)) { /* skip none entry */
3270 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3271 is_hugetlb_entry_hwpoisoned(entry
))) {
3272 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3274 if (is_write_migration_entry(swp_entry
) && cow
) {
3276 * COW mappings require pages in both
3277 * parent and child to be set to read.
3279 make_migration_entry_read(&swp_entry
);
3280 entry
= swp_entry_to_pte(swp_entry
);
3281 set_huge_swap_pte_at(src
, addr
, src_pte
,
3284 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3287 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3288 mmu_notifier_invalidate_range(src
, mmun_start
,
3291 entry
= huge_ptep_get(src_pte
);
3292 ptepage
= pte_page(entry
);
3294 page_dup_rmap(ptepage
, true);
3295 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3296 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3298 spin_unlock(src_ptl
);
3299 spin_unlock(dst_ptl
);
3303 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3308 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3309 unsigned long start
, unsigned long end
,
3310 struct page
*ref_page
)
3312 struct mm_struct
*mm
= vma
->vm_mm
;
3313 unsigned long address
;
3318 struct hstate
*h
= hstate_vma(vma
);
3319 unsigned long sz
= huge_page_size(h
);
3320 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3321 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3323 WARN_ON(!is_vm_hugetlb_page(vma
));
3324 BUG_ON(start
& ~huge_page_mask(h
));
3325 BUG_ON(end
& ~huge_page_mask(h
));
3328 * This is a hugetlb vma, all the pte entries should point
3331 tlb_remove_check_page_size_change(tlb
, sz
);
3332 tlb_start_vma(tlb
, vma
);
3333 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3335 for (; address
< end
; address
+= sz
) {
3336 ptep
= huge_pte_offset(mm
, address
, sz
);
3340 ptl
= huge_pte_lock(h
, mm
, ptep
);
3341 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3346 pte
= huge_ptep_get(ptep
);
3347 if (huge_pte_none(pte
)) {
3353 * Migrating hugepage or HWPoisoned hugepage is already
3354 * unmapped and its refcount is dropped, so just clear pte here.
3356 if (unlikely(!pte_present(pte
))) {
3357 huge_pte_clear(mm
, address
, ptep
, sz
);
3362 page
= pte_page(pte
);
3364 * If a reference page is supplied, it is because a specific
3365 * page is being unmapped, not a range. Ensure the page we
3366 * are about to unmap is the actual page of interest.
3369 if (page
!= ref_page
) {
3374 * Mark the VMA as having unmapped its page so that
3375 * future faults in this VMA will fail rather than
3376 * looking like data was lost
3378 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3381 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3382 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3383 if (huge_pte_dirty(pte
))
3384 set_page_dirty(page
);
3386 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3387 page_remove_rmap(page
, true);
3390 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3392 * Bail out after unmapping reference page if supplied
3397 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3398 tlb_end_vma(tlb
, vma
);
3401 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3402 struct vm_area_struct
*vma
, unsigned long start
,
3403 unsigned long end
, struct page
*ref_page
)
3405 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3408 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3409 * test will fail on a vma being torn down, and not grab a page table
3410 * on its way out. We're lucky that the flag has such an appropriate
3411 * name, and can in fact be safely cleared here. We could clear it
3412 * before the __unmap_hugepage_range above, but all that's necessary
3413 * is to clear it before releasing the i_mmap_rwsem. This works
3414 * because in the context this is called, the VMA is about to be
3415 * destroyed and the i_mmap_rwsem is held.
3417 vma
->vm_flags
&= ~VM_MAYSHARE
;
3420 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3421 unsigned long end
, struct page
*ref_page
)
3423 struct mm_struct
*mm
;
3424 struct mmu_gather tlb
;
3428 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3429 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3430 tlb_finish_mmu(&tlb
, start
, end
);
3434 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3435 * mappping it owns the reserve page for. The intention is to unmap the page
3436 * from other VMAs and let the children be SIGKILLed if they are faulting the
3439 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3440 struct page
*page
, unsigned long address
)
3442 struct hstate
*h
= hstate_vma(vma
);
3443 struct vm_area_struct
*iter_vma
;
3444 struct address_space
*mapping
;
3448 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3449 * from page cache lookup which is in HPAGE_SIZE units.
3451 address
= address
& huge_page_mask(h
);
3452 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3454 mapping
= vma
->vm_file
->f_mapping
;
3457 * Take the mapping lock for the duration of the table walk. As
3458 * this mapping should be shared between all the VMAs,
3459 * __unmap_hugepage_range() is called as the lock is already held
3461 i_mmap_lock_write(mapping
);
3462 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3463 /* Do not unmap the current VMA */
3464 if (iter_vma
== vma
)
3468 * Shared VMAs have their own reserves and do not affect
3469 * MAP_PRIVATE accounting but it is possible that a shared
3470 * VMA is using the same page so check and skip such VMAs.
3472 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3476 * Unmap the page from other VMAs without their own reserves.
3477 * They get marked to be SIGKILLed if they fault in these
3478 * areas. This is because a future no-page fault on this VMA
3479 * could insert a zeroed page instead of the data existing
3480 * from the time of fork. This would look like data corruption
3482 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3483 unmap_hugepage_range(iter_vma
, address
,
3484 address
+ huge_page_size(h
), page
);
3486 i_mmap_unlock_write(mapping
);
3490 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3491 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3492 * cannot race with other handlers or page migration.
3493 * Keep the pte_same checks anyway to make transition from the mutex easier.
3495 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3496 unsigned long address
, pte_t
*ptep
,
3497 struct page
*pagecache_page
, spinlock_t
*ptl
)
3500 struct hstate
*h
= hstate_vma(vma
);
3501 struct page
*old_page
, *new_page
;
3502 int ret
= 0, outside_reserve
= 0;
3503 unsigned long mmun_start
; /* For mmu_notifiers */
3504 unsigned long mmun_end
; /* For mmu_notifiers */
3506 pte
= huge_ptep_get(ptep
);
3507 old_page
= pte_page(pte
);
3510 /* If no-one else is actually using this page, avoid the copy
3511 * and just make the page writable */
3512 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3513 page_move_anon_rmap(old_page
, vma
);
3514 set_huge_ptep_writable(vma
, address
, ptep
);
3519 * If the process that created a MAP_PRIVATE mapping is about to
3520 * perform a COW due to a shared page count, attempt to satisfy
3521 * the allocation without using the existing reserves. The pagecache
3522 * page is used to determine if the reserve at this address was
3523 * consumed or not. If reserves were used, a partial faulted mapping
3524 * at the time of fork() could consume its reserves on COW instead
3525 * of the full address range.
3527 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3528 old_page
!= pagecache_page
)
3529 outside_reserve
= 1;
3534 * Drop page table lock as buddy allocator may be called. It will
3535 * be acquired again before returning to the caller, as expected.
3538 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3540 if (IS_ERR(new_page
)) {
3542 * If a process owning a MAP_PRIVATE mapping fails to COW,
3543 * it is due to references held by a child and an insufficient
3544 * huge page pool. To guarantee the original mappers
3545 * reliability, unmap the page from child processes. The child
3546 * may get SIGKILLed if it later faults.
3548 if (outside_reserve
) {
3550 BUG_ON(huge_pte_none(pte
));
3551 unmap_ref_private(mm
, vma
, old_page
, address
);
3552 BUG_ON(huge_pte_none(pte
));
3554 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
),
3557 pte_same(huge_ptep_get(ptep
), pte
)))
3558 goto retry_avoidcopy
;
3560 * race occurs while re-acquiring page table
3561 * lock, and our job is done.
3566 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3567 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3568 goto out_release_old
;
3572 * When the original hugepage is shared one, it does not have
3573 * anon_vma prepared.
3575 if (unlikely(anon_vma_prepare(vma
))) {
3577 goto out_release_all
;
3580 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3581 pages_per_huge_page(h
));
3582 __SetPageUptodate(new_page
);
3583 set_page_huge_active(new_page
);
3585 mmun_start
= address
& huge_page_mask(h
);
3586 mmun_end
= mmun_start
+ huge_page_size(h
);
3587 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3590 * Retake the page table lock to check for racing updates
3591 * before the page tables are altered
3594 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
),
3596 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3597 ClearPagePrivate(new_page
);
3600 huge_ptep_clear_flush(vma
, address
, ptep
);
3601 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3602 set_huge_pte_at(mm
, address
, ptep
,
3603 make_huge_pte(vma
, new_page
, 1));
3604 page_remove_rmap(old_page
, true);
3605 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3606 /* Make the old page be freed below */
3607 new_page
= old_page
;
3610 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3612 restore_reserve_on_error(h
, vma
, address
, new_page
);
3617 spin_lock(ptl
); /* Caller expects lock to be held */
3621 /* Return the pagecache page at a given address within a VMA */
3622 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3623 struct vm_area_struct
*vma
, unsigned long address
)
3625 struct address_space
*mapping
;
3628 mapping
= vma
->vm_file
->f_mapping
;
3629 idx
= vma_hugecache_offset(h
, vma
, address
);
3631 return find_lock_page(mapping
, idx
);
3635 * Return whether there is a pagecache page to back given address within VMA.
3636 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3638 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3639 struct vm_area_struct
*vma
, unsigned long address
)
3641 struct address_space
*mapping
;
3645 mapping
= vma
->vm_file
->f_mapping
;
3646 idx
= vma_hugecache_offset(h
, vma
, address
);
3648 page
= find_get_page(mapping
, idx
);
3651 return page
!= NULL
;
3654 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3657 struct inode
*inode
= mapping
->host
;
3658 struct hstate
*h
= hstate_inode(inode
);
3659 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3663 ClearPagePrivate(page
);
3665 spin_lock(&inode
->i_lock
);
3666 inode
->i_blocks
+= blocks_per_huge_page(h
);
3667 spin_unlock(&inode
->i_lock
);
3671 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3672 struct address_space
*mapping
, pgoff_t idx
,
3673 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3675 struct hstate
*h
= hstate_vma(vma
);
3676 int ret
= VM_FAULT_SIGBUS
;
3684 * Currently, we are forced to kill the process in the event the
3685 * original mapper has unmapped pages from the child due to a failed
3686 * COW. Warn that such a situation has occurred as it may not be obvious
3688 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3689 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3695 * Use page lock to guard against racing truncation
3696 * before we get page_table_lock.
3699 page
= find_lock_page(mapping
, idx
);
3701 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3706 * Check for page in userfault range
3708 if (userfaultfd_missing(vma
)) {
3710 struct vm_fault vmf
= {
3715 * Hard to debug if it ends up being
3716 * used by a callee that assumes
3717 * something about the other
3718 * uninitialized fields... same as in
3724 * hugetlb_fault_mutex must be dropped before
3725 * handling userfault. Reacquire after handling
3726 * fault to make calling code simpler.
3728 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
,
3730 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3731 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
3732 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3736 page
= alloc_huge_page(vma
, address
, 0);
3738 ret
= PTR_ERR(page
);
3742 ret
= VM_FAULT_SIGBUS
;
3745 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3746 __SetPageUptodate(page
);
3747 set_page_huge_active(page
);
3749 if (vma
->vm_flags
& VM_MAYSHARE
) {
3750 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3759 if (unlikely(anon_vma_prepare(vma
))) {
3761 goto backout_unlocked
;
3767 * If memory error occurs between mmap() and fault, some process
3768 * don't have hwpoisoned swap entry for errored virtual address.
3769 * So we need to block hugepage fault by PG_hwpoison bit check.
3771 if (unlikely(PageHWPoison(page
))) {
3772 ret
= VM_FAULT_HWPOISON
|
3773 VM_FAULT_SET_HINDEX(hstate_index(h
));
3774 goto backout_unlocked
;
3779 * If we are going to COW a private mapping later, we examine the
3780 * pending reservations for this page now. This will ensure that
3781 * any allocations necessary to record that reservation occur outside
3784 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3785 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3787 goto backout_unlocked
;
3789 /* Just decrements count, does not deallocate */
3790 vma_end_reservation(h
, vma
, address
);
3793 ptl
= huge_pte_lock(h
, mm
, ptep
);
3794 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3799 if (!huge_pte_none(huge_ptep_get(ptep
)))
3803 ClearPagePrivate(page
);
3804 hugepage_add_new_anon_rmap(page
, vma
, address
);
3806 page_dup_rmap(page
, true);
3807 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3808 && (vma
->vm_flags
& VM_SHARED
)));
3809 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3811 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3812 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3813 /* Optimization, do the COW without a second fault */
3814 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
3826 restore_reserve_on_error(h
, vma
, address
, page
);
3832 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3833 struct vm_area_struct
*vma
,
3834 struct address_space
*mapping
,
3835 pgoff_t idx
, unsigned long address
)
3837 unsigned long key
[2];
3840 if (vma
->vm_flags
& VM_SHARED
) {
3841 key
[0] = (unsigned long) mapping
;
3844 key
[0] = (unsigned long) mm
;
3845 key
[1] = address
>> huge_page_shift(h
);
3848 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3850 return hash
& (num_fault_mutexes
- 1);
3854 * For uniprocesor systems we always use a single mutex, so just
3855 * return 0 and avoid the hashing overhead.
3857 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3858 struct vm_area_struct
*vma
,
3859 struct address_space
*mapping
,
3860 pgoff_t idx
, unsigned long address
)
3866 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3867 unsigned long address
, unsigned int flags
)
3874 struct page
*page
= NULL
;
3875 struct page
*pagecache_page
= NULL
;
3876 struct hstate
*h
= hstate_vma(vma
);
3877 struct address_space
*mapping
;
3878 int need_wait_lock
= 0;
3880 address
&= huge_page_mask(h
);
3882 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
3884 entry
= huge_ptep_get(ptep
);
3885 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3886 migration_entry_wait_huge(vma
, mm
, ptep
);
3888 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3889 return VM_FAULT_HWPOISON_LARGE
|
3890 VM_FAULT_SET_HINDEX(hstate_index(h
));
3892 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3894 return VM_FAULT_OOM
;
3897 mapping
= vma
->vm_file
->f_mapping
;
3898 idx
= vma_hugecache_offset(h
, vma
, address
);
3901 * Serialize hugepage allocation and instantiation, so that we don't
3902 * get spurious allocation failures if two CPUs race to instantiate
3903 * the same page in the page cache.
3905 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3906 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3908 entry
= huge_ptep_get(ptep
);
3909 if (huge_pte_none(entry
)) {
3910 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3917 * entry could be a migration/hwpoison entry at this point, so this
3918 * check prevents the kernel from going below assuming that we have
3919 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3920 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3923 if (!pte_present(entry
))
3927 * If we are going to COW the mapping later, we examine the pending
3928 * reservations for this page now. This will ensure that any
3929 * allocations necessary to record that reservation occur outside the
3930 * spinlock. For private mappings, we also lookup the pagecache
3931 * page now as it is used to determine if a reservation has been
3934 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3935 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3939 /* Just decrements count, does not deallocate */
3940 vma_end_reservation(h
, vma
, address
);
3942 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3943 pagecache_page
= hugetlbfs_pagecache_page(h
,
3947 ptl
= huge_pte_lock(h
, mm
, ptep
);
3949 /* Check for a racing update before calling hugetlb_cow */
3950 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3954 * hugetlb_cow() requires page locks of pte_page(entry) and
3955 * pagecache_page, so here we need take the former one
3956 * when page != pagecache_page or !pagecache_page.
3958 page
= pte_page(entry
);
3959 if (page
!= pagecache_page
)
3960 if (!trylock_page(page
)) {
3967 if (flags
& FAULT_FLAG_WRITE
) {
3968 if (!huge_pte_write(entry
)) {
3969 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
3970 pagecache_page
, ptl
);
3973 entry
= huge_pte_mkdirty(entry
);
3975 entry
= pte_mkyoung(entry
);
3976 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3977 flags
& FAULT_FLAG_WRITE
))
3978 update_mmu_cache(vma
, address
, ptep
);
3980 if (page
!= pagecache_page
)
3986 if (pagecache_page
) {
3987 unlock_page(pagecache_page
);
3988 put_page(pagecache_page
);
3991 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3993 * Generally it's safe to hold refcount during waiting page lock. But
3994 * here we just wait to defer the next page fault to avoid busy loop and
3995 * the page is not used after unlocked before returning from the current
3996 * page fault. So we are safe from accessing freed page, even if we wait
3997 * here without taking refcount.
4000 wait_on_page_locked(page
);
4005 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4006 * modifications for huge pages.
4008 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4010 struct vm_area_struct
*dst_vma
,
4011 unsigned long dst_addr
,
4012 unsigned long src_addr
,
4013 struct page
**pagep
)
4015 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4016 struct hstate
*h
= hstate_vma(dst_vma
);
4024 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4028 ret
= copy_huge_page_from_user(page
,
4029 (const void __user
*) src_addr
,
4030 pages_per_huge_page(h
), false);
4032 /* fallback to copy_from_user outside mmap_sem */
4033 if (unlikely(ret
)) {
4036 /* don't free the page */
4045 * The memory barrier inside __SetPageUptodate makes sure that
4046 * preceding stores to the page contents become visible before
4047 * the set_pte_at() write.
4049 __SetPageUptodate(page
);
4050 set_page_huge_active(page
);
4053 * If shared, add to page cache
4056 struct address_space
*mapping
= dst_vma
->vm_file
->f_mapping
;
4057 pgoff_t idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4059 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4061 goto out_release_nounlock
;
4064 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4068 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4069 goto out_release_unlock
;
4072 page_dup_rmap(page
, true);
4074 ClearPagePrivate(page
);
4075 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4078 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4079 if (dst_vma
->vm_flags
& VM_WRITE
)
4080 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4081 _dst_pte
= pte_mkyoung(_dst_pte
);
4083 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4085 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4086 dst_vma
->vm_flags
& VM_WRITE
);
4087 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4089 /* No need to invalidate - it was non-present before */
4090 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4100 out_release_nounlock
:
4107 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4108 struct page
**pages
, struct vm_area_struct
**vmas
,
4109 unsigned long *position
, unsigned long *nr_pages
,
4110 long i
, unsigned int flags
, int *nonblocking
)
4112 unsigned long pfn_offset
;
4113 unsigned long vaddr
= *position
;
4114 unsigned long remainder
= *nr_pages
;
4115 struct hstate
*h
= hstate_vma(vma
);
4117 while (vaddr
< vma
->vm_end
&& remainder
) {
4119 spinlock_t
*ptl
= NULL
;
4124 * If we have a pending SIGKILL, don't keep faulting pages and
4125 * potentially allocating memory.
4127 if (unlikely(fatal_signal_pending(current
))) {
4133 * Some archs (sparc64, sh*) have multiple pte_ts to
4134 * each hugepage. We have to make sure we get the
4135 * first, for the page indexing below to work.
4137 * Note that page table lock is not held when pte is null.
4139 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4142 ptl
= huge_pte_lock(h
, mm
, pte
);
4143 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4146 * When coredumping, it suits get_dump_page if we just return
4147 * an error where there's an empty slot with no huge pagecache
4148 * to back it. This way, we avoid allocating a hugepage, and
4149 * the sparse dumpfile avoids allocating disk blocks, but its
4150 * huge holes still show up with zeroes where they need to be.
4152 if (absent
&& (flags
& FOLL_DUMP
) &&
4153 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4161 * We need call hugetlb_fault for both hugepages under migration
4162 * (in which case hugetlb_fault waits for the migration,) and
4163 * hwpoisoned hugepages (in which case we need to prevent the
4164 * caller from accessing to them.) In order to do this, we use
4165 * here is_swap_pte instead of is_hugetlb_entry_migration and
4166 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4167 * both cases, and because we can't follow correct pages
4168 * directly from any kind of swap entries.
4170 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4171 ((flags
& FOLL_WRITE
) &&
4172 !huge_pte_write(huge_ptep_get(pte
)))) {
4174 unsigned int fault_flags
= 0;
4178 if (flags
& FOLL_WRITE
)
4179 fault_flags
|= FAULT_FLAG_WRITE
;
4181 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4182 if (flags
& FOLL_NOWAIT
)
4183 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4184 FAULT_FLAG_RETRY_NOWAIT
;
4185 if (flags
& FOLL_TRIED
) {
4186 VM_WARN_ON_ONCE(fault_flags
&
4187 FAULT_FLAG_ALLOW_RETRY
);
4188 fault_flags
|= FAULT_FLAG_TRIED
;
4190 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4191 if (ret
& VM_FAULT_ERROR
) {
4192 int err
= vm_fault_to_errno(ret
, flags
);
4200 if (ret
& VM_FAULT_RETRY
) {
4205 * VM_FAULT_RETRY must not return an
4206 * error, it will return zero
4209 * No need to update "position" as the
4210 * caller will not check it after
4211 * *nr_pages is set to 0.
4218 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4219 page
= pte_page(huge_ptep_get(pte
));
4222 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4233 if (vaddr
< vma
->vm_end
&& remainder
&&
4234 pfn_offset
< pages_per_huge_page(h
)) {
4236 * We use pfn_offset to avoid touching the pageframes
4237 * of this compound page.
4243 *nr_pages
= remainder
;
4245 * setting position is actually required only if remainder is
4246 * not zero but it's faster not to add a "if (remainder)"
4251 return i
? i
: -EFAULT
;
4254 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4256 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4259 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4262 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4263 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4265 struct mm_struct
*mm
= vma
->vm_mm
;
4266 unsigned long start
= address
;
4269 struct hstate
*h
= hstate_vma(vma
);
4270 unsigned long pages
= 0;
4272 BUG_ON(address
>= end
);
4273 flush_cache_range(vma
, address
, end
);
4275 mmu_notifier_invalidate_range_start(mm
, start
, end
);
4276 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4277 for (; address
< end
; address
+= huge_page_size(h
)) {
4279 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
4282 ptl
= huge_pte_lock(h
, mm
, ptep
);
4283 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4288 pte
= huge_ptep_get(ptep
);
4289 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4293 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4294 swp_entry_t entry
= pte_to_swp_entry(pte
);
4296 if (is_write_migration_entry(entry
)) {
4299 make_migration_entry_read(&entry
);
4300 newpte
= swp_entry_to_pte(entry
);
4301 set_huge_swap_pte_at(mm
, address
, ptep
,
4302 newpte
, huge_page_size(h
));
4308 if (!huge_pte_none(pte
)) {
4309 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4310 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
4311 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4312 set_huge_pte_at(mm
, address
, ptep
, pte
);
4318 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4319 * may have cleared our pud entry and done put_page on the page table:
4320 * once we release i_mmap_rwsem, another task can do the final put_page
4321 * and that page table be reused and filled with junk.
4323 flush_hugetlb_tlb_range(vma
, start
, end
);
4324 mmu_notifier_invalidate_range(mm
, start
, end
);
4325 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4326 mmu_notifier_invalidate_range_end(mm
, start
, end
);
4328 return pages
<< h
->order
;
4331 int hugetlb_reserve_pages(struct inode
*inode
,
4333 struct vm_area_struct
*vma
,
4334 vm_flags_t vm_flags
)
4337 struct hstate
*h
= hstate_inode(inode
);
4338 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4339 struct resv_map
*resv_map
;
4343 * Only apply hugepage reservation if asked. At fault time, an
4344 * attempt will be made for VM_NORESERVE to allocate a page
4345 * without using reserves
4347 if (vm_flags
& VM_NORESERVE
)
4351 * Shared mappings base their reservation on the number of pages that
4352 * are already allocated on behalf of the file. Private mappings need
4353 * to reserve the full area even if read-only as mprotect() may be
4354 * called to make the mapping read-write. Assume !vma is a shm mapping
4356 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4357 resv_map
= inode_resv_map(inode
);
4359 chg
= region_chg(resv_map
, from
, to
);
4362 resv_map
= resv_map_alloc();
4368 set_vma_resv_map(vma
, resv_map
);
4369 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4378 * There must be enough pages in the subpool for the mapping. If
4379 * the subpool has a minimum size, there may be some global
4380 * reservations already in place (gbl_reserve).
4382 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4383 if (gbl_reserve
< 0) {
4389 * Check enough hugepages are available for the reservation.
4390 * Hand the pages back to the subpool if there are not
4392 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4394 /* put back original number of pages, chg */
4395 (void)hugepage_subpool_put_pages(spool
, chg
);
4400 * Account for the reservations made. Shared mappings record regions
4401 * that have reservations as they are shared by multiple VMAs.
4402 * When the last VMA disappears, the region map says how much
4403 * the reservation was and the page cache tells how much of
4404 * the reservation was consumed. Private mappings are per-VMA and
4405 * only the consumed reservations are tracked. When the VMA
4406 * disappears, the original reservation is the VMA size and the
4407 * consumed reservations are stored in the map. Hence, nothing
4408 * else has to be done for private mappings here
4410 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4411 long add
= region_add(resv_map
, from
, to
);
4413 if (unlikely(chg
> add
)) {
4415 * pages in this range were added to the reserve
4416 * map between region_chg and region_add. This
4417 * indicates a race with alloc_huge_page. Adjust
4418 * the subpool and reserve counts modified above
4419 * based on the difference.
4423 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4425 hugetlb_acct_memory(h
, -rsv_adjust
);
4430 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4431 /* Don't call region_abort if region_chg failed */
4433 region_abort(resv_map
, from
, to
);
4434 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4435 kref_put(&resv_map
->refs
, resv_map_release
);
4439 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4442 struct hstate
*h
= hstate_inode(inode
);
4443 struct resv_map
*resv_map
= inode_resv_map(inode
);
4445 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4449 chg
= region_del(resv_map
, start
, end
);
4451 * region_del() can fail in the rare case where a region
4452 * must be split and another region descriptor can not be
4453 * allocated. If end == LONG_MAX, it will not fail.
4459 spin_lock(&inode
->i_lock
);
4460 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4461 spin_unlock(&inode
->i_lock
);
4464 * If the subpool has a minimum size, the number of global
4465 * reservations to be released may be adjusted.
4467 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4468 hugetlb_acct_memory(h
, -gbl_reserve
);
4473 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4474 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4475 struct vm_area_struct
*vma
,
4476 unsigned long addr
, pgoff_t idx
)
4478 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4480 unsigned long sbase
= saddr
& PUD_MASK
;
4481 unsigned long s_end
= sbase
+ PUD_SIZE
;
4483 /* Allow segments to share if only one is marked locked */
4484 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4485 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4488 * match the virtual addresses, permission and the alignment of the
4491 if (pmd_index(addr
) != pmd_index(saddr
) ||
4492 vm_flags
!= svm_flags
||
4493 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4499 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4501 unsigned long base
= addr
& PUD_MASK
;
4502 unsigned long end
= base
+ PUD_SIZE
;
4505 * check on proper vm_flags and page table alignment
4507 if (vma
->vm_flags
& VM_MAYSHARE
&&
4508 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4514 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4515 * and returns the corresponding pte. While this is not necessary for the
4516 * !shared pmd case because we can allocate the pmd later as well, it makes the
4517 * code much cleaner. pmd allocation is essential for the shared case because
4518 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4519 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4520 * bad pmd for sharing.
4522 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4524 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4525 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4526 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4528 struct vm_area_struct
*svma
;
4529 unsigned long saddr
;
4534 if (!vma_shareable(vma
, addr
))
4535 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4537 i_mmap_lock_write(mapping
);
4538 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4542 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4544 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
4545 vma_mmu_pagesize(svma
));
4547 get_page(virt_to_page(spte
));
4556 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4557 if (pud_none(*pud
)) {
4558 pud_populate(mm
, pud
,
4559 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4562 put_page(virt_to_page(spte
));
4566 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4567 i_mmap_unlock_write(mapping
);
4572 * unmap huge page backed by shared pte.
4574 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4575 * indicated by page_count > 1, unmap is achieved by clearing pud and
4576 * decrementing the ref count. If count == 1, the pte page is not shared.
4578 * called with page table lock held.
4580 * returns: 1 successfully unmapped a shared pte page
4581 * 0 the underlying pte page is not shared, or it is the last user
4583 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4585 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4586 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
4587 pud_t
*pud
= pud_offset(p4d
, *addr
);
4589 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4590 if (page_count(virt_to_page(ptep
)) == 1)
4594 put_page(virt_to_page(ptep
));
4596 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4599 #define want_pmd_share() (1)
4600 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4601 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4606 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4610 #define want_pmd_share() (0)
4611 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4613 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4614 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4615 unsigned long addr
, unsigned long sz
)
4622 pgd
= pgd_offset(mm
, addr
);
4623 p4d
= p4d_offset(pgd
, addr
);
4624 pud
= pud_alloc(mm
, p4d
, addr
);
4626 if (sz
== PUD_SIZE
) {
4629 BUG_ON(sz
!= PMD_SIZE
);
4630 if (want_pmd_share() && pud_none(*pud
))
4631 pte
= huge_pmd_share(mm
, addr
, pud
);
4633 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4636 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
4641 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
4642 unsigned long addr
, unsigned long sz
)
4649 pgd
= pgd_offset(mm
, addr
);
4650 if (!pgd_present(*pgd
))
4652 p4d
= p4d_offset(pgd
, addr
);
4653 if (!p4d_present(*p4d
))
4655 pud
= pud_offset(p4d
, addr
);
4656 if (!pud_present(*pud
))
4659 return (pte_t
*)pud
;
4660 pmd
= pmd_offset(pud
, addr
);
4661 return (pte_t
*) pmd
;
4664 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4667 * These functions are overwritable if your architecture needs its own
4670 struct page
* __weak
4671 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4674 return ERR_PTR(-EINVAL
);
4677 struct page
* __weak
4678 follow_huge_pd(struct vm_area_struct
*vma
,
4679 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
4681 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4685 struct page
* __weak
4686 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4687 pmd_t
*pmd
, int flags
)
4689 struct page
*page
= NULL
;
4693 ptl
= pmd_lockptr(mm
, pmd
);
4696 * make sure that the address range covered by this pmd is not
4697 * unmapped from other threads.
4699 if (!pmd_huge(*pmd
))
4701 pte
= huge_ptep_get((pte_t
*)pmd
);
4702 if (pte_present(pte
)) {
4703 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4704 if (flags
& FOLL_GET
)
4707 if (is_hugetlb_entry_migration(pte
)) {
4709 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4713 * hwpoisoned entry is treated as no_page_table in
4714 * follow_page_mask().
4722 struct page
* __weak
4723 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4724 pud_t
*pud
, int flags
)
4726 if (flags
& FOLL_GET
)
4729 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4732 struct page
* __weak
4733 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
4735 if (flags
& FOLL_GET
)
4738 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
4741 #ifdef CONFIG_MEMORY_FAILURE
4744 * This function is called from memory failure code.
4746 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4748 struct hstate
*h
= page_hstate(hpage
);
4749 int nid
= page_to_nid(hpage
);
4752 spin_lock(&hugetlb_lock
);
4754 * Just checking !page_huge_active is not enough, because that could be
4755 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4757 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4759 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4760 * but dangling hpage->lru can trigger list-debug warnings
4761 * (this happens when we call unpoison_memory() on it),
4762 * so let it point to itself with list_del_init().
4764 list_del_init(&hpage
->lru
);
4765 set_page_refcounted(hpage
);
4766 h
->free_huge_pages
--;
4767 h
->free_huge_pages_node
[nid
]--;
4770 spin_unlock(&hugetlb_lock
);
4775 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4779 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4780 spin_lock(&hugetlb_lock
);
4781 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4785 clear_page_huge_active(page
);
4786 list_move_tail(&page
->lru
, list
);
4788 spin_unlock(&hugetlb_lock
);
4792 void putback_active_hugepage(struct page
*page
)
4794 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4795 spin_lock(&hugetlb_lock
);
4796 set_page_huge_active(page
);
4797 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4798 spin_unlock(&hugetlb_lock
);