1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
32 #include <asm/pgtable.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/node.h>
39 #include <linux/userfaultfd_k.h>
40 #include <linux/page_owner.h>
43 int hugetlb_max_hstate __read_mostly
;
44 unsigned int default_hstate_idx
;
45 struct hstate hstates
[HUGE_MAX_HSTATE
];
47 * Minimum page order among possible hugepage sizes, set to a proper value
50 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
52 __initdata
LIST_HEAD(huge_boot_pages
);
54 /* for command line parsing */
55 static struct hstate
* __initdata parsed_hstate
;
56 static unsigned long __initdata default_hstate_max_huge_pages
;
57 static unsigned long __initdata default_hstate_size
;
58 static bool __initdata parsed_valid_hugepagesz
= true;
61 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
62 * free_huge_pages, and surplus_huge_pages.
64 DEFINE_SPINLOCK(hugetlb_lock
);
67 * Serializes faults on the same logical page. This is used to
68 * prevent spurious OOMs when the hugepage pool is fully utilized.
70 static int num_fault_mutexes
;
71 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
73 /* Forward declaration */
74 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
76 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
78 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
80 spin_unlock(&spool
->lock
);
82 /* If no pages are used, and no other handles to the subpool
83 * remain, give up any reservations mased on minimum size and
86 if (spool
->min_hpages
!= -1)
87 hugetlb_acct_memory(spool
->hstate
,
93 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
96 struct hugepage_subpool
*spool
;
98 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
102 spin_lock_init(&spool
->lock
);
104 spool
->max_hpages
= max_hpages
;
106 spool
->min_hpages
= min_hpages
;
108 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
112 spool
->rsv_hpages
= min_hpages
;
117 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
119 spin_lock(&spool
->lock
);
120 BUG_ON(!spool
->count
);
122 unlock_or_release_subpool(spool
);
126 * Subpool accounting for allocating and reserving pages.
127 * Return -ENOMEM if there are not enough resources to satisfy the
128 * the request. Otherwise, return the number of pages by which the
129 * global pools must be adjusted (upward). The returned value may
130 * only be different than the passed value (delta) in the case where
131 * a subpool minimum size must be manitained.
133 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
141 spin_lock(&spool
->lock
);
143 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
144 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
145 spool
->used_hpages
+= delta
;
152 /* minimum size accounting */
153 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
154 if (delta
> spool
->rsv_hpages
) {
156 * Asking for more reserves than those already taken on
157 * behalf of subpool. Return difference.
159 ret
= delta
- spool
->rsv_hpages
;
160 spool
->rsv_hpages
= 0;
162 ret
= 0; /* reserves already accounted for */
163 spool
->rsv_hpages
-= delta
;
168 spin_unlock(&spool
->lock
);
173 * Subpool accounting for freeing and unreserving pages.
174 * Return the number of global page reservations that must be dropped.
175 * The return value may only be different than the passed value (delta)
176 * in the case where a subpool minimum size must be maintained.
178 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
186 spin_lock(&spool
->lock
);
188 if (spool
->max_hpages
!= -1) /* maximum size accounting */
189 spool
->used_hpages
-= delta
;
191 /* minimum size accounting */
192 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
193 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
196 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
198 spool
->rsv_hpages
+= delta
;
199 if (spool
->rsv_hpages
> spool
->min_hpages
)
200 spool
->rsv_hpages
= spool
->min_hpages
;
204 * If hugetlbfs_put_super couldn't free spool due to an outstanding
205 * quota reference, free it now.
207 unlock_or_release_subpool(spool
);
212 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
214 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
217 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
219 return subpool_inode(file_inode(vma
->vm_file
));
223 * Region tracking -- allows tracking of reservations and instantiated pages
224 * across the pages in a mapping.
226 * The region data structures are embedded into a resv_map and protected
227 * by a resv_map's lock. The set of regions within the resv_map represent
228 * reservations for huge pages, or huge pages that have already been
229 * instantiated within the map. The from and to elements are huge page
230 * indicies into the associated mapping. from indicates the starting index
231 * of the region. to represents the first index past the end of the region.
233 * For example, a file region structure with from == 0 and to == 4 represents
234 * four huge pages in a mapping. It is important to note that the to element
235 * represents the first element past the end of the region. This is used in
236 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
238 * Interval notation of the form [from, to) will be used to indicate that
239 * the endpoint from is inclusive and to is exclusive.
242 struct list_head link
;
248 * Add the huge page range represented by [f, t) to the reserve
249 * map. In the normal case, existing regions will be expanded
250 * to accommodate the specified range. Sufficient regions should
251 * exist for expansion due to the previous call to region_chg
252 * with the same range. However, it is possible that region_del
253 * could have been called after region_chg and modifed the map
254 * in such a way that no region exists to be expanded. In this
255 * case, pull a region descriptor from the cache associated with
256 * the map and use that for the new range.
258 * Return the number of new huge pages added to the map. This
259 * number is greater than or equal to zero.
261 static long region_add(struct resv_map
*resv
, long f
, long t
)
263 struct list_head
*head
= &resv
->regions
;
264 struct file_region
*rg
, *nrg
, *trg
;
267 spin_lock(&resv
->lock
);
268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg
, head
, link
)
274 * If no region exists which can be expanded to include the
275 * specified range, the list must have been modified by an
276 * interleving call to region_del(). Pull a region descriptor
277 * from the cache and use it for this range.
279 if (&rg
->link
== head
|| t
< rg
->from
) {
280 VM_BUG_ON(resv
->region_cache_count
<= 0);
282 resv
->region_cache_count
--;
283 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
285 list_del(&nrg
->link
);
289 list_add(&nrg
->link
, rg
->link
.prev
);
295 /* Round our left edge to the current segment if it encloses us. */
299 /* Check for and consume any regions we now overlap with. */
301 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
302 if (&rg
->link
== head
)
307 /* If this area reaches higher then extend our area to
308 * include it completely. If this is not the first area
309 * which we intend to reuse, free it. */
313 /* Decrement return value by the deleted range.
314 * Another range will span this area so that by
315 * end of routine add will be >= zero
317 add
-= (rg
->to
- rg
->from
);
323 add
+= (nrg
->from
- f
); /* Added to beginning of region */
325 add
+= t
- nrg
->to
; /* Added to end of region */
329 resv
->adds_in_progress
--;
330 spin_unlock(&resv
->lock
);
336 * Examine the existing reserve map and determine how many
337 * huge pages in the specified range [f, t) are NOT currently
338 * represented. This routine is called before a subsequent
339 * call to region_add that will actually modify the reserve
340 * map to add the specified range [f, t). region_chg does
341 * not change the number of huge pages represented by the
342 * map. However, if the existing regions in the map can not
343 * be expanded to represent the new range, a new file_region
344 * structure is added to the map as a placeholder. This is
345 * so that the subsequent region_add call will have all the
346 * regions it needs and will not fail.
348 * Upon entry, region_chg will also examine the cache of region descriptors
349 * associated with the map. If there are not enough descriptors cached, one
350 * will be allocated for the in progress add operation.
352 * Returns the number of huge pages that need to be added to the existing
353 * reservation map for the range [f, t). This number is greater or equal to
354 * zero. -ENOMEM is returned if a new file_region structure or cache entry
355 * is needed and can not be allocated.
357 static long region_chg(struct resv_map
*resv
, long f
, long t
)
359 struct list_head
*head
= &resv
->regions
;
360 struct file_region
*rg
, *nrg
= NULL
;
364 spin_lock(&resv
->lock
);
366 resv
->adds_in_progress
++;
369 * Check for sufficient descriptors in the cache to accommodate
370 * the number of in progress add operations.
372 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
373 struct file_region
*trg
;
375 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
376 /* Must drop lock to allocate a new descriptor. */
377 resv
->adds_in_progress
--;
378 spin_unlock(&resv
->lock
);
380 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
386 spin_lock(&resv
->lock
);
387 list_add(&trg
->link
, &resv
->region_cache
);
388 resv
->region_cache_count
++;
392 /* Locate the region we are before or in. */
393 list_for_each_entry(rg
, head
, link
)
397 /* If we are below the current region then a new region is required.
398 * Subtle, allocate a new region at the position but make it zero
399 * size such that we can guarantee to record the reservation. */
400 if (&rg
->link
== head
|| t
< rg
->from
) {
402 resv
->adds_in_progress
--;
403 spin_unlock(&resv
->lock
);
404 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
410 INIT_LIST_HEAD(&nrg
->link
);
414 list_add(&nrg
->link
, rg
->link
.prev
);
419 /* Round our left edge to the current segment if it encloses us. */
424 /* Check for and consume any regions we now overlap with. */
425 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
426 if (&rg
->link
== head
)
431 /* We overlap with this area, if it extends further than
432 * us then we must extend ourselves. Account for its
433 * existing reservation. */
438 chg
-= rg
->to
- rg
->from
;
442 spin_unlock(&resv
->lock
);
443 /* We already know we raced and no longer need the new region */
447 spin_unlock(&resv
->lock
);
452 * Abort the in progress add operation. The adds_in_progress field
453 * of the resv_map keeps track of the operations in progress between
454 * calls to region_chg and region_add. Operations are sometimes
455 * aborted after the call to region_chg. In such cases, region_abort
456 * is called to decrement the adds_in_progress counter.
458 * NOTE: The range arguments [f, t) are not needed or used in this
459 * routine. They are kept to make reading the calling code easier as
460 * arguments will match the associated region_chg call.
462 static void region_abort(struct resv_map
*resv
, long f
, long t
)
464 spin_lock(&resv
->lock
);
465 VM_BUG_ON(!resv
->region_cache_count
);
466 resv
->adds_in_progress
--;
467 spin_unlock(&resv
->lock
);
471 * Delete the specified range [f, t) from the reserve map. If the
472 * t parameter is LONG_MAX, this indicates that ALL regions after f
473 * should be deleted. Locate the regions which intersect [f, t)
474 * and either trim, delete or split the existing regions.
476 * Returns the number of huge pages deleted from the reserve map.
477 * In the normal case, the return value is zero or more. In the
478 * case where a region must be split, a new region descriptor must
479 * be allocated. If the allocation fails, -ENOMEM will be returned.
480 * NOTE: If the parameter t == LONG_MAX, then we will never split
481 * a region and possibly return -ENOMEM. Callers specifying
482 * t == LONG_MAX do not need to check for -ENOMEM error.
484 static long region_del(struct resv_map
*resv
, long f
, long t
)
486 struct list_head
*head
= &resv
->regions
;
487 struct file_region
*rg
, *trg
;
488 struct file_region
*nrg
= NULL
;
492 spin_lock(&resv
->lock
);
493 list_for_each_entry_safe(rg
, trg
, head
, link
) {
495 * Skip regions before the range to be deleted. file_region
496 * ranges are normally of the form [from, to). However, there
497 * may be a "placeholder" entry in the map which is of the form
498 * (from, to) with from == to. Check for placeholder entries
499 * at the beginning of the range to be deleted.
501 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
507 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
509 * Check for an entry in the cache before dropping
510 * lock and attempting allocation.
513 resv
->region_cache_count
> resv
->adds_in_progress
) {
514 nrg
= list_first_entry(&resv
->region_cache
,
517 list_del(&nrg
->link
);
518 resv
->region_cache_count
--;
522 spin_unlock(&resv
->lock
);
523 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
531 /* New entry for end of split region */
534 INIT_LIST_HEAD(&nrg
->link
);
536 /* Original entry is trimmed */
539 list_add(&nrg
->link
, &rg
->link
);
544 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
545 del
+= rg
->to
- rg
->from
;
551 if (f
<= rg
->from
) { /* Trim beginning of region */
554 } else { /* Trim end of region */
560 spin_unlock(&resv
->lock
);
566 * A rare out of memory error was encountered which prevented removal of
567 * the reserve map region for a page. The huge page itself was free'ed
568 * and removed from the page cache. This routine will adjust the subpool
569 * usage count, and the global reserve count if needed. By incrementing
570 * these counts, the reserve map entry which could not be deleted will
571 * appear as a "reserved" entry instead of simply dangling with incorrect
574 void hugetlb_fix_reserve_counts(struct inode
*inode
)
576 struct hugepage_subpool
*spool
= subpool_inode(inode
);
579 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
581 struct hstate
*h
= hstate_inode(inode
);
583 hugetlb_acct_memory(h
, 1);
588 * Count and return the number of huge pages in the reserve map
589 * that intersect with the range [f, t).
591 static long region_count(struct resv_map
*resv
, long f
, long t
)
593 struct list_head
*head
= &resv
->regions
;
594 struct file_region
*rg
;
597 spin_lock(&resv
->lock
);
598 /* Locate each segment we overlap with, and count that overlap. */
599 list_for_each_entry(rg
, head
, link
) {
608 seg_from
= max(rg
->from
, f
);
609 seg_to
= min(rg
->to
, t
);
611 chg
+= seg_to
- seg_from
;
613 spin_unlock(&resv
->lock
);
619 * Convert the address within this vma to the page offset within
620 * the mapping, in pagecache page units; huge pages here.
622 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
623 struct vm_area_struct
*vma
, unsigned long address
)
625 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
626 (vma
->vm_pgoff
>> huge_page_order(h
));
629 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
630 unsigned long address
)
632 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
634 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
637 * Return the size of the pages allocated when backing a VMA. In the majority
638 * cases this will be same size as used by the page table entries.
640 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
642 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
643 return vma
->vm_ops
->pagesize(vma
);
646 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
649 * Return the page size being used by the MMU to back a VMA. In the majority
650 * of cases, the page size used by the kernel matches the MMU size. On
651 * architectures where it differs, an architecture-specific 'strong'
652 * version of this symbol is required.
654 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
656 return vma_kernel_pagesize(vma
);
660 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
661 * bits of the reservation map pointer, which are always clear due to
664 #define HPAGE_RESV_OWNER (1UL << 0)
665 #define HPAGE_RESV_UNMAPPED (1UL << 1)
666 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
669 * These helpers are used to track how many pages are reserved for
670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671 * is guaranteed to have their future faults succeed.
673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674 * the reserve counters are updated with the hugetlb_lock held. It is safe
675 * to reset the VMA at fork() time as it is not in use yet and there is no
676 * chance of the global counters getting corrupted as a result of the values.
678 * The private mapping reservation is represented in a subtly different
679 * manner to a shared mapping. A shared mapping has a region map associated
680 * with the underlying file, this region map represents the backing file
681 * pages which have ever had a reservation assigned which this persists even
682 * after the page is instantiated. A private mapping has a region map
683 * associated with the original mmap which is attached to all VMAs which
684 * reference it, this region map represents those offsets which have consumed
685 * reservation ie. where pages have been instantiated.
687 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
689 return (unsigned long)vma
->vm_private_data
;
692 static void set_vma_private_data(struct vm_area_struct
*vma
,
695 vma
->vm_private_data
= (void *)value
;
698 struct resv_map
*resv_map_alloc(void)
700 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
701 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
703 if (!resv_map
|| !rg
) {
709 kref_init(&resv_map
->refs
);
710 spin_lock_init(&resv_map
->lock
);
711 INIT_LIST_HEAD(&resv_map
->regions
);
713 resv_map
->adds_in_progress
= 0;
715 INIT_LIST_HEAD(&resv_map
->region_cache
);
716 list_add(&rg
->link
, &resv_map
->region_cache
);
717 resv_map
->region_cache_count
= 1;
722 void resv_map_release(struct kref
*ref
)
724 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
725 struct list_head
*head
= &resv_map
->region_cache
;
726 struct file_region
*rg
, *trg
;
728 /* Clear out any active regions before we release the map. */
729 region_del(resv_map
, 0, LONG_MAX
);
731 /* ... and any entries left in the cache */
732 list_for_each_entry_safe(rg
, trg
, head
, link
) {
737 VM_BUG_ON(resv_map
->adds_in_progress
);
742 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
745 * At inode evict time, i_mapping may not point to the original
746 * address space within the inode. This original address space
747 * contains the pointer to the resv_map. So, always use the
748 * address space embedded within the inode.
749 * The VERY common case is inode->mapping == &inode->i_data but,
750 * this may not be true for device special inodes.
752 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
755 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
757 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
758 if (vma
->vm_flags
& VM_MAYSHARE
) {
759 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
760 struct inode
*inode
= mapping
->host
;
762 return inode_resv_map(inode
);
765 return (struct resv_map
*)(get_vma_private_data(vma
) &
770 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
773 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
775 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
776 HPAGE_RESV_MASK
) | (unsigned long)map
);
779 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
781 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
782 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
784 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
787 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
789 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
791 return (get_vma_private_data(vma
) & flag
) != 0;
794 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
795 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
797 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
798 if (!(vma
->vm_flags
& VM_MAYSHARE
))
799 vma
->vm_private_data
= (void *)0;
802 /* Returns true if the VMA has associated reserve pages */
803 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
805 if (vma
->vm_flags
& VM_NORESERVE
) {
807 * This address is already reserved by other process(chg == 0),
808 * so, we should decrement reserved count. Without decrementing,
809 * reserve count remains after releasing inode, because this
810 * allocated page will go into page cache and is regarded as
811 * coming from reserved pool in releasing step. Currently, we
812 * don't have any other solution to deal with this situation
813 * properly, so add work-around here.
815 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
821 /* Shared mappings always use reserves */
822 if (vma
->vm_flags
& VM_MAYSHARE
) {
824 * We know VM_NORESERVE is not set. Therefore, there SHOULD
825 * be a region map for all pages. The only situation where
826 * there is no region map is if a hole was punched via
827 * fallocate. In this case, there really are no reverves to
828 * use. This situation is indicated if chg != 0.
837 * Only the process that called mmap() has reserves for
840 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
842 * Like the shared case above, a hole punch or truncate
843 * could have been performed on the private mapping.
844 * Examine the value of chg to determine if reserves
845 * actually exist or were previously consumed.
846 * Very Subtle - The value of chg comes from a previous
847 * call to vma_needs_reserves(). The reserve map for
848 * private mappings has different (opposite) semantics
849 * than that of shared mappings. vma_needs_reserves()
850 * has already taken this difference in semantics into
851 * account. Therefore, the meaning of chg is the same
852 * as in the shared case above. Code could easily be
853 * combined, but keeping it separate draws attention to
854 * subtle differences.
865 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
867 int nid
= page_to_nid(page
);
868 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
869 h
->free_huge_pages
++;
870 h
->free_huge_pages_node
[nid
]++;
873 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
877 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
878 if (!PageHWPoison(page
))
881 * if 'non-isolated free hugepage' not found on the list,
882 * the allocation fails.
884 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
886 list_move(&page
->lru
, &h
->hugepage_activelist
);
887 set_page_refcounted(page
);
888 h
->free_huge_pages
--;
889 h
->free_huge_pages_node
[nid
]--;
893 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
896 unsigned int cpuset_mems_cookie
;
897 struct zonelist
*zonelist
;
900 int node
= NUMA_NO_NODE
;
902 zonelist
= node_zonelist(nid
, gfp_mask
);
905 cpuset_mems_cookie
= read_mems_allowed_begin();
906 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
909 if (!cpuset_zone_allowed(zone
, gfp_mask
))
912 * no need to ask again on the same node. Pool is node rather than
915 if (zone_to_nid(zone
) == node
)
917 node
= zone_to_nid(zone
);
919 page
= dequeue_huge_page_node_exact(h
, node
);
923 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
929 /* Movability of hugepages depends on migration support. */
930 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
932 if (hugepage_movable_supported(h
))
933 return GFP_HIGHUSER_MOVABLE
;
938 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
939 struct vm_area_struct
*vma
,
940 unsigned long address
, int avoid_reserve
,
944 struct mempolicy
*mpol
;
946 nodemask_t
*nodemask
;
950 * A child process with MAP_PRIVATE mappings created by their parent
951 * have no page reserves. This check ensures that reservations are
952 * not "stolen". The child may still get SIGKILLed
954 if (!vma_has_reserves(vma
, chg
) &&
955 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
958 /* If reserves cannot be used, ensure enough pages are in the pool */
959 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
962 gfp_mask
= htlb_alloc_mask(h
);
963 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
964 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
965 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
966 SetPagePrivate(page
);
967 h
->resv_huge_pages
--;
978 * common helper functions for hstate_next_node_to_{alloc|free}.
979 * We may have allocated or freed a huge page based on a different
980 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
981 * be outside of *nodes_allowed. Ensure that we use an allowed
982 * node for alloc or free.
984 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
986 nid
= next_node_in(nid
, *nodes_allowed
);
987 VM_BUG_ON(nid
>= MAX_NUMNODES
);
992 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
994 if (!node_isset(nid
, *nodes_allowed
))
995 nid
= next_node_allowed(nid
, nodes_allowed
);
1000 * returns the previously saved node ["this node"] from which to
1001 * allocate a persistent huge page for the pool and advance the
1002 * next node from which to allocate, handling wrap at end of node
1005 static int hstate_next_node_to_alloc(struct hstate
*h
,
1006 nodemask_t
*nodes_allowed
)
1010 VM_BUG_ON(!nodes_allowed
);
1012 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1013 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1019 * helper for free_pool_huge_page() - return the previously saved
1020 * node ["this node"] from which to free a huge page. Advance the
1021 * next node id whether or not we find a free huge page to free so
1022 * that the next attempt to free addresses the next node.
1024 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1028 VM_BUG_ON(!nodes_allowed
);
1030 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1031 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1036 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1037 for (nr_nodes = nodes_weight(*mask); \
1039 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1042 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1043 for (nr_nodes = nodes_weight(*mask); \
1045 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1048 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1049 static void destroy_compound_gigantic_page(struct page
*page
,
1053 int nr_pages
= 1 << order
;
1054 struct page
*p
= page
+ 1;
1056 atomic_set(compound_mapcount_ptr(page
), 0);
1057 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1058 clear_compound_head(p
);
1059 set_page_refcounted(p
);
1062 set_compound_order(page
, 0);
1063 __ClearPageHead(page
);
1066 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1068 free_contig_range(page_to_pfn(page
), 1 << order
);
1071 #ifdef CONFIG_CONTIG_ALLOC
1072 static int __alloc_gigantic_page(unsigned long start_pfn
,
1073 unsigned long nr_pages
, gfp_t gfp_mask
)
1075 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1076 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
1080 static bool pfn_range_valid_gigantic(struct zone
*z
,
1081 unsigned long start_pfn
, unsigned long nr_pages
)
1083 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1086 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1090 page
= pfn_to_page(i
);
1092 if (page_zone(page
) != z
)
1095 if (PageReserved(page
))
1098 if (page_count(page
) > 0)
1108 static bool zone_spans_last_pfn(const struct zone
*zone
,
1109 unsigned long start_pfn
, unsigned long nr_pages
)
1111 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1112 return zone_spans_pfn(zone
, last_pfn
);
1115 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1116 int nid
, nodemask_t
*nodemask
)
1118 unsigned int order
= huge_page_order(h
);
1119 unsigned long nr_pages
= 1 << order
;
1120 unsigned long ret
, pfn
, flags
;
1121 struct zonelist
*zonelist
;
1125 zonelist
= node_zonelist(nid
, gfp_mask
);
1126 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nodemask
) {
1127 spin_lock_irqsave(&zone
->lock
, flags
);
1129 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
1130 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
1131 if (pfn_range_valid_gigantic(zone
, pfn
, nr_pages
)) {
1133 * We release the zone lock here because
1134 * alloc_contig_range() will also lock the zone
1135 * at some point. If there's an allocation
1136 * spinning on this lock, it may win the race
1137 * and cause alloc_contig_range() to fail...
1139 spin_unlock_irqrestore(&zone
->lock
, flags
);
1140 ret
= __alloc_gigantic_page(pfn
, nr_pages
, gfp_mask
);
1142 return pfn_to_page(pfn
);
1143 spin_lock_irqsave(&zone
->lock
, flags
);
1148 spin_unlock_irqrestore(&zone
->lock
, flags
);
1154 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1155 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1156 #else /* !CONFIG_CONTIG_ALLOC */
1157 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1158 int nid
, nodemask_t
*nodemask
)
1162 #endif /* CONFIG_CONTIG_ALLOC */
1164 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1165 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1166 int nid
, nodemask_t
*nodemask
)
1170 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1171 static inline void destroy_compound_gigantic_page(struct page
*page
,
1172 unsigned int order
) { }
1175 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1179 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1183 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1184 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1185 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1186 1 << PG_referenced
| 1 << PG_dirty
|
1187 1 << PG_active
| 1 << PG_private
|
1190 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1191 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1192 set_page_refcounted(page
);
1193 if (hstate_is_gigantic(h
)) {
1194 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1195 free_gigantic_page(page
, huge_page_order(h
));
1197 __free_pages(page
, huge_page_order(h
));
1201 struct hstate
*size_to_hstate(unsigned long size
)
1205 for_each_hstate(h
) {
1206 if (huge_page_size(h
) == size
)
1213 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1214 * to hstate->hugepage_activelist.)
1216 * This function can be called for tail pages, but never returns true for them.
1218 bool page_huge_active(struct page
*page
)
1220 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1221 return PageHead(page
) && PagePrivate(&page
[1]);
1224 /* never called for tail page */
1225 static void set_page_huge_active(struct page
*page
)
1227 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1228 SetPagePrivate(&page
[1]);
1231 static void clear_page_huge_active(struct page
*page
)
1233 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1234 ClearPagePrivate(&page
[1]);
1238 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1241 static inline bool PageHugeTemporary(struct page
*page
)
1243 if (!PageHuge(page
))
1246 return (unsigned long)page
[2].mapping
== -1U;
1249 static inline void SetPageHugeTemporary(struct page
*page
)
1251 page
[2].mapping
= (void *)-1U;
1254 static inline void ClearPageHugeTemporary(struct page
*page
)
1256 page
[2].mapping
= NULL
;
1259 void free_huge_page(struct page
*page
)
1262 * Can't pass hstate in here because it is called from the
1263 * compound page destructor.
1265 struct hstate
*h
= page_hstate(page
);
1266 int nid
= page_to_nid(page
);
1267 struct hugepage_subpool
*spool
=
1268 (struct hugepage_subpool
*)page_private(page
);
1269 bool restore_reserve
;
1271 VM_BUG_ON_PAGE(page_count(page
), page
);
1272 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1274 set_page_private(page
, 0);
1275 page
->mapping
= NULL
;
1276 restore_reserve
= PagePrivate(page
);
1277 ClearPagePrivate(page
);
1280 * If PagePrivate() was set on page, page allocation consumed a
1281 * reservation. If the page was associated with a subpool, there
1282 * would have been a page reserved in the subpool before allocation
1283 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1284 * reservtion, do not call hugepage_subpool_put_pages() as this will
1285 * remove the reserved page from the subpool.
1287 if (!restore_reserve
) {
1289 * A return code of zero implies that the subpool will be
1290 * under its minimum size if the reservation is not restored
1291 * after page is free. Therefore, force restore_reserve
1294 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1295 restore_reserve
= true;
1298 spin_lock(&hugetlb_lock
);
1299 clear_page_huge_active(page
);
1300 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1301 pages_per_huge_page(h
), page
);
1302 if (restore_reserve
)
1303 h
->resv_huge_pages
++;
1305 if (PageHugeTemporary(page
)) {
1306 list_del(&page
->lru
);
1307 ClearPageHugeTemporary(page
);
1308 update_and_free_page(h
, page
);
1309 } else if (h
->surplus_huge_pages_node
[nid
]) {
1310 /* remove the page from active list */
1311 list_del(&page
->lru
);
1312 update_and_free_page(h
, page
);
1313 h
->surplus_huge_pages
--;
1314 h
->surplus_huge_pages_node
[nid
]--;
1316 arch_clear_hugepage_flags(page
);
1317 enqueue_huge_page(h
, page
);
1319 spin_unlock(&hugetlb_lock
);
1322 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1324 INIT_LIST_HEAD(&page
->lru
);
1325 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1326 spin_lock(&hugetlb_lock
);
1327 set_hugetlb_cgroup(page
, NULL
);
1329 h
->nr_huge_pages_node
[nid
]++;
1330 spin_unlock(&hugetlb_lock
);
1333 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1336 int nr_pages
= 1 << order
;
1337 struct page
*p
= page
+ 1;
1339 /* we rely on prep_new_huge_page to set the destructor */
1340 set_compound_order(page
, order
);
1341 __ClearPageReserved(page
);
1342 __SetPageHead(page
);
1343 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1345 * For gigantic hugepages allocated through bootmem at
1346 * boot, it's safer to be consistent with the not-gigantic
1347 * hugepages and clear the PG_reserved bit from all tail pages
1348 * too. Otherwse drivers using get_user_pages() to access tail
1349 * pages may get the reference counting wrong if they see
1350 * PG_reserved set on a tail page (despite the head page not
1351 * having PG_reserved set). Enforcing this consistency between
1352 * head and tail pages allows drivers to optimize away a check
1353 * on the head page when they need know if put_page() is needed
1354 * after get_user_pages().
1356 __ClearPageReserved(p
);
1357 set_page_count(p
, 0);
1358 set_compound_head(p
, page
);
1360 atomic_set(compound_mapcount_ptr(page
), -1);
1364 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1365 * transparent huge pages. See the PageTransHuge() documentation for more
1368 int PageHuge(struct page
*page
)
1370 if (!PageCompound(page
))
1373 page
= compound_head(page
);
1374 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1376 EXPORT_SYMBOL_GPL(PageHuge
);
1379 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1380 * normal or transparent huge pages.
1382 int PageHeadHuge(struct page
*page_head
)
1384 if (!PageHead(page_head
))
1387 return get_compound_page_dtor(page_head
) == free_huge_page
;
1390 pgoff_t
__basepage_index(struct page
*page
)
1392 struct page
*page_head
= compound_head(page
);
1393 pgoff_t index
= page_index(page_head
);
1394 unsigned long compound_idx
;
1396 if (!PageHuge(page_head
))
1397 return page_index(page
);
1399 if (compound_order(page_head
) >= MAX_ORDER
)
1400 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1402 compound_idx
= page
- page_head
;
1404 return (index
<< compound_order(page_head
)) + compound_idx
;
1407 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1408 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1409 nodemask_t
*node_alloc_noretry
)
1411 int order
= huge_page_order(h
);
1413 bool alloc_try_hard
= true;
1416 * By default we always try hard to allocate the page with
1417 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1418 * a loop (to adjust global huge page counts) and previous allocation
1419 * failed, do not continue to try hard on the same node. Use the
1420 * node_alloc_noretry bitmap to manage this state information.
1422 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1423 alloc_try_hard
= false;
1424 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1426 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1427 if (nid
== NUMA_NO_NODE
)
1428 nid
= numa_mem_id();
1429 page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, nmask
);
1431 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1433 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1436 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1437 * indicates an overall state change. Clear bit so that we resume
1438 * normal 'try hard' allocations.
1440 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1441 node_clear(nid
, *node_alloc_noretry
);
1444 * If we tried hard to get a page but failed, set bit so that
1445 * subsequent attempts will not try as hard until there is an
1446 * overall state change.
1448 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1449 node_set(nid
, *node_alloc_noretry
);
1455 * Common helper to allocate a fresh hugetlb page. All specific allocators
1456 * should use this function to get new hugetlb pages
1458 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1459 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1460 nodemask_t
*node_alloc_noretry
)
1464 if (hstate_is_gigantic(h
))
1465 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1467 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1468 nid
, nmask
, node_alloc_noretry
);
1472 if (hstate_is_gigantic(h
))
1473 prep_compound_gigantic_page(page
, huge_page_order(h
));
1474 prep_new_huge_page(h
, page
, page_to_nid(page
));
1480 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1483 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1484 nodemask_t
*node_alloc_noretry
)
1488 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1490 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1491 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1492 node_alloc_noretry
);
1500 put_page(page
); /* free it into the hugepage allocator */
1506 * Free huge page from pool from next node to free.
1507 * Attempt to keep persistent huge pages more or less
1508 * balanced over allowed nodes.
1509 * Called with hugetlb_lock locked.
1511 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1517 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1519 * If we're returning unused surplus pages, only examine
1520 * nodes with surplus pages.
1522 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1523 !list_empty(&h
->hugepage_freelists
[node
])) {
1525 list_entry(h
->hugepage_freelists
[node
].next
,
1527 list_del(&page
->lru
);
1528 h
->free_huge_pages
--;
1529 h
->free_huge_pages_node
[node
]--;
1531 h
->surplus_huge_pages
--;
1532 h
->surplus_huge_pages_node
[node
]--;
1534 update_and_free_page(h
, page
);
1544 * Dissolve a given free hugepage into free buddy pages. This function does
1545 * nothing for in-use hugepages and non-hugepages.
1546 * This function returns values like below:
1548 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1549 * (allocated or reserved.)
1550 * 0: successfully dissolved free hugepages or the page is not a
1551 * hugepage (considered as already dissolved)
1553 int dissolve_free_huge_page(struct page
*page
)
1557 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1558 if (!PageHuge(page
))
1561 spin_lock(&hugetlb_lock
);
1562 if (!PageHuge(page
)) {
1567 if (!page_count(page
)) {
1568 struct page
*head
= compound_head(page
);
1569 struct hstate
*h
= page_hstate(head
);
1570 int nid
= page_to_nid(head
);
1571 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1574 * Move PageHWPoison flag from head page to the raw error page,
1575 * which makes any subpages rather than the error page reusable.
1577 if (PageHWPoison(head
) && page
!= head
) {
1578 SetPageHWPoison(page
);
1579 ClearPageHWPoison(head
);
1581 list_del(&head
->lru
);
1582 h
->free_huge_pages
--;
1583 h
->free_huge_pages_node
[nid
]--;
1584 h
->max_huge_pages
--;
1585 update_and_free_page(h
, head
);
1589 spin_unlock(&hugetlb_lock
);
1594 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1595 * make specified memory blocks removable from the system.
1596 * Note that this will dissolve a free gigantic hugepage completely, if any
1597 * part of it lies within the given range.
1598 * Also note that if dissolve_free_huge_page() returns with an error, all
1599 * free hugepages that were dissolved before that error are lost.
1601 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1607 if (!hugepages_supported())
1610 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1611 page
= pfn_to_page(pfn
);
1612 rc
= dissolve_free_huge_page(page
);
1621 * Allocates a fresh surplus page from the page allocator.
1623 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1624 int nid
, nodemask_t
*nmask
)
1626 struct page
*page
= NULL
;
1628 if (hstate_is_gigantic(h
))
1631 spin_lock(&hugetlb_lock
);
1632 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
1634 spin_unlock(&hugetlb_lock
);
1636 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1640 spin_lock(&hugetlb_lock
);
1642 * We could have raced with the pool size change.
1643 * Double check that and simply deallocate the new page
1644 * if we would end up overcommiting the surpluses. Abuse
1645 * temporary page to workaround the nasty free_huge_page
1648 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1649 SetPageHugeTemporary(page
);
1650 spin_unlock(&hugetlb_lock
);
1654 h
->surplus_huge_pages
++;
1655 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
1659 spin_unlock(&hugetlb_lock
);
1664 struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1665 int nid
, nodemask_t
*nmask
)
1669 if (hstate_is_gigantic(h
))
1672 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1677 * We do not account these pages as surplus because they are only
1678 * temporary and will be released properly on the last reference
1680 SetPageHugeTemporary(page
);
1686 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1689 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1690 struct vm_area_struct
*vma
, unsigned long addr
)
1693 struct mempolicy
*mpol
;
1694 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1696 nodemask_t
*nodemask
;
1698 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
1699 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
);
1700 mpol_cond_put(mpol
);
1705 /* page migration callback function */
1706 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1708 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1709 struct page
*page
= NULL
;
1711 if (nid
!= NUMA_NO_NODE
)
1712 gfp_mask
|= __GFP_THISNODE
;
1714 spin_lock(&hugetlb_lock
);
1715 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1716 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, NULL
);
1717 spin_unlock(&hugetlb_lock
);
1720 page
= alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1725 /* page migration callback function */
1726 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
1729 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1731 spin_lock(&hugetlb_lock
);
1732 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1735 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
1737 spin_unlock(&hugetlb_lock
);
1741 spin_unlock(&hugetlb_lock
);
1743 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
1746 /* mempolicy aware migration callback */
1747 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
1748 unsigned long address
)
1750 struct mempolicy
*mpol
;
1751 nodemask_t
*nodemask
;
1756 gfp_mask
= htlb_alloc_mask(h
);
1757 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1758 page
= alloc_huge_page_nodemask(h
, node
, nodemask
);
1759 mpol_cond_put(mpol
);
1765 * Increase the hugetlb pool such that it can accommodate a reservation
1768 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1770 struct list_head surplus_list
;
1771 struct page
*page
, *tmp
;
1773 int needed
, allocated
;
1774 bool alloc_ok
= true;
1776 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1778 h
->resv_huge_pages
+= delta
;
1783 INIT_LIST_HEAD(&surplus_list
);
1787 spin_unlock(&hugetlb_lock
);
1788 for (i
= 0; i
< needed
; i
++) {
1789 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
1790 NUMA_NO_NODE
, NULL
);
1795 list_add(&page
->lru
, &surplus_list
);
1801 * After retaking hugetlb_lock, we need to recalculate 'needed'
1802 * because either resv_huge_pages or free_huge_pages may have changed.
1804 spin_lock(&hugetlb_lock
);
1805 needed
= (h
->resv_huge_pages
+ delta
) -
1806 (h
->free_huge_pages
+ allocated
);
1811 * We were not able to allocate enough pages to
1812 * satisfy the entire reservation so we free what
1813 * we've allocated so far.
1818 * The surplus_list now contains _at_least_ the number of extra pages
1819 * needed to accommodate the reservation. Add the appropriate number
1820 * of pages to the hugetlb pool and free the extras back to the buddy
1821 * allocator. Commit the entire reservation here to prevent another
1822 * process from stealing the pages as they are added to the pool but
1823 * before they are reserved.
1825 needed
+= allocated
;
1826 h
->resv_huge_pages
+= delta
;
1829 /* Free the needed pages to the hugetlb pool */
1830 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1834 * This page is now managed by the hugetlb allocator and has
1835 * no users -- drop the buddy allocator's reference.
1837 put_page_testzero(page
);
1838 VM_BUG_ON_PAGE(page_count(page
), page
);
1839 enqueue_huge_page(h
, page
);
1842 spin_unlock(&hugetlb_lock
);
1844 /* Free unnecessary surplus pages to the buddy allocator */
1845 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1847 spin_lock(&hugetlb_lock
);
1853 * This routine has two main purposes:
1854 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1855 * in unused_resv_pages. This corresponds to the prior adjustments made
1856 * to the associated reservation map.
1857 * 2) Free any unused surplus pages that may have been allocated to satisfy
1858 * the reservation. As many as unused_resv_pages may be freed.
1860 * Called with hugetlb_lock held. However, the lock could be dropped (and
1861 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1862 * we must make sure nobody else can claim pages we are in the process of
1863 * freeing. Do this by ensuring resv_huge_page always is greater than the
1864 * number of huge pages we plan to free when dropping the lock.
1866 static void return_unused_surplus_pages(struct hstate
*h
,
1867 unsigned long unused_resv_pages
)
1869 unsigned long nr_pages
;
1871 /* Cannot return gigantic pages currently */
1872 if (hstate_is_gigantic(h
))
1876 * Part (or even all) of the reservation could have been backed
1877 * by pre-allocated pages. Only free surplus pages.
1879 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1882 * We want to release as many surplus pages as possible, spread
1883 * evenly across all nodes with memory. Iterate across these nodes
1884 * until we can no longer free unreserved surplus pages. This occurs
1885 * when the nodes with surplus pages have no free pages.
1886 * free_pool_huge_page() will balance the the freed pages across the
1887 * on-line nodes with memory and will handle the hstate accounting.
1889 * Note that we decrement resv_huge_pages as we free the pages. If
1890 * we drop the lock, resv_huge_pages will still be sufficiently large
1891 * to cover subsequent pages we may free.
1893 while (nr_pages
--) {
1894 h
->resv_huge_pages
--;
1895 unused_resv_pages
--;
1896 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1898 cond_resched_lock(&hugetlb_lock
);
1902 /* Fully uncommit the reservation */
1903 h
->resv_huge_pages
-= unused_resv_pages
;
1908 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1909 * are used by the huge page allocation routines to manage reservations.
1911 * vma_needs_reservation is called to determine if the huge page at addr
1912 * within the vma has an associated reservation. If a reservation is
1913 * needed, the value 1 is returned. The caller is then responsible for
1914 * managing the global reservation and subpool usage counts. After
1915 * the huge page has been allocated, vma_commit_reservation is called
1916 * to add the page to the reservation map. If the page allocation fails,
1917 * the reservation must be ended instead of committed. vma_end_reservation
1918 * is called in such cases.
1920 * In the normal case, vma_commit_reservation returns the same value
1921 * as the preceding vma_needs_reservation call. The only time this
1922 * is not the case is if a reserve map was changed between calls. It
1923 * is the responsibility of the caller to notice the difference and
1924 * take appropriate action.
1926 * vma_add_reservation is used in error paths where a reservation must
1927 * be restored when a newly allocated huge page must be freed. It is
1928 * to be called after calling vma_needs_reservation to determine if a
1929 * reservation exists.
1931 enum vma_resv_mode
{
1937 static long __vma_reservation_common(struct hstate
*h
,
1938 struct vm_area_struct
*vma
, unsigned long addr
,
1939 enum vma_resv_mode mode
)
1941 struct resv_map
*resv
;
1945 resv
= vma_resv_map(vma
);
1949 idx
= vma_hugecache_offset(h
, vma
, addr
);
1951 case VMA_NEEDS_RESV
:
1952 ret
= region_chg(resv
, idx
, idx
+ 1);
1954 case VMA_COMMIT_RESV
:
1955 ret
= region_add(resv
, idx
, idx
+ 1);
1958 region_abort(resv
, idx
, idx
+ 1);
1962 if (vma
->vm_flags
& VM_MAYSHARE
)
1963 ret
= region_add(resv
, idx
, idx
+ 1);
1965 region_abort(resv
, idx
, idx
+ 1);
1966 ret
= region_del(resv
, idx
, idx
+ 1);
1973 if (vma
->vm_flags
& VM_MAYSHARE
)
1975 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1977 * In most cases, reserves always exist for private mappings.
1978 * However, a file associated with mapping could have been
1979 * hole punched or truncated after reserves were consumed.
1980 * As subsequent fault on such a range will not use reserves.
1981 * Subtle - The reserve map for private mappings has the
1982 * opposite meaning than that of shared mappings. If NO
1983 * entry is in the reserve map, it means a reservation exists.
1984 * If an entry exists in the reserve map, it means the
1985 * reservation has already been consumed. As a result, the
1986 * return value of this routine is the opposite of the
1987 * value returned from reserve map manipulation routines above.
1995 return ret
< 0 ? ret
: 0;
1998 static long vma_needs_reservation(struct hstate
*h
,
1999 struct vm_area_struct
*vma
, unsigned long addr
)
2001 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2004 static long vma_commit_reservation(struct hstate
*h
,
2005 struct vm_area_struct
*vma
, unsigned long addr
)
2007 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2010 static void vma_end_reservation(struct hstate
*h
,
2011 struct vm_area_struct
*vma
, unsigned long addr
)
2013 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2016 static long vma_add_reservation(struct hstate
*h
,
2017 struct vm_area_struct
*vma
, unsigned long addr
)
2019 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2023 * This routine is called to restore a reservation on error paths. In the
2024 * specific error paths, a huge page was allocated (via alloc_huge_page)
2025 * and is about to be freed. If a reservation for the page existed,
2026 * alloc_huge_page would have consumed the reservation and set PagePrivate
2027 * in the newly allocated page. When the page is freed via free_huge_page,
2028 * the global reservation count will be incremented if PagePrivate is set.
2029 * However, free_huge_page can not adjust the reserve map. Adjust the
2030 * reserve map here to be consistent with global reserve count adjustments
2031 * to be made by free_huge_page.
2033 static void restore_reserve_on_error(struct hstate
*h
,
2034 struct vm_area_struct
*vma
, unsigned long address
,
2037 if (unlikely(PagePrivate(page
))) {
2038 long rc
= vma_needs_reservation(h
, vma
, address
);
2040 if (unlikely(rc
< 0)) {
2042 * Rare out of memory condition in reserve map
2043 * manipulation. Clear PagePrivate so that
2044 * global reserve count will not be incremented
2045 * by free_huge_page. This will make it appear
2046 * as though the reservation for this page was
2047 * consumed. This may prevent the task from
2048 * faulting in the page at a later time. This
2049 * is better than inconsistent global huge page
2050 * accounting of reserve counts.
2052 ClearPagePrivate(page
);
2054 rc
= vma_add_reservation(h
, vma
, address
);
2055 if (unlikely(rc
< 0))
2057 * See above comment about rare out of
2060 ClearPagePrivate(page
);
2062 vma_end_reservation(h
, vma
, address
);
2066 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2067 unsigned long addr
, int avoid_reserve
)
2069 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2070 struct hstate
*h
= hstate_vma(vma
);
2072 long map_chg
, map_commit
;
2075 struct hugetlb_cgroup
*h_cg
;
2077 idx
= hstate_index(h
);
2079 * Examine the region/reserve map to determine if the process
2080 * has a reservation for the page to be allocated. A return
2081 * code of zero indicates a reservation exists (no change).
2083 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2085 return ERR_PTR(-ENOMEM
);
2088 * Processes that did not create the mapping will have no
2089 * reserves as indicated by the region/reserve map. Check
2090 * that the allocation will not exceed the subpool limit.
2091 * Allocations for MAP_NORESERVE mappings also need to be
2092 * checked against any subpool limit.
2094 if (map_chg
|| avoid_reserve
) {
2095 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2097 vma_end_reservation(h
, vma
, addr
);
2098 return ERR_PTR(-ENOSPC
);
2102 * Even though there was no reservation in the region/reserve
2103 * map, there could be reservations associated with the
2104 * subpool that can be used. This would be indicated if the
2105 * return value of hugepage_subpool_get_pages() is zero.
2106 * However, if avoid_reserve is specified we still avoid even
2107 * the subpool reservations.
2113 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2115 goto out_subpool_put
;
2117 spin_lock(&hugetlb_lock
);
2119 * glb_chg is passed to indicate whether or not a page must be taken
2120 * from the global free pool (global change). gbl_chg == 0 indicates
2121 * a reservation exists for the allocation.
2123 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2125 spin_unlock(&hugetlb_lock
);
2126 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2128 goto out_uncharge_cgroup
;
2129 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2130 SetPagePrivate(page
);
2131 h
->resv_huge_pages
--;
2133 spin_lock(&hugetlb_lock
);
2134 list_move(&page
->lru
, &h
->hugepage_activelist
);
2137 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2138 spin_unlock(&hugetlb_lock
);
2140 set_page_private(page
, (unsigned long)spool
);
2142 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2143 if (unlikely(map_chg
> map_commit
)) {
2145 * The page was added to the reservation map between
2146 * vma_needs_reservation and vma_commit_reservation.
2147 * This indicates a race with hugetlb_reserve_pages.
2148 * Adjust for the subpool count incremented above AND
2149 * in hugetlb_reserve_pages for the same page. Also,
2150 * the reservation count added in hugetlb_reserve_pages
2151 * no longer applies.
2155 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2156 hugetlb_acct_memory(h
, -rsv_adjust
);
2160 out_uncharge_cgroup
:
2161 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2163 if (map_chg
|| avoid_reserve
)
2164 hugepage_subpool_put_pages(spool
, 1);
2165 vma_end_reservation(h
, vma
, addr
);
2166 return ERR_PTR(-ENOSPC
);
2169 int alloc_bootmem_huge_page(struct hstate
*h
)
2170 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2171 int __alloc_bootmem_huge_page(struct hstate
*h
)
2173 struct huge_bootmem_page
*m
;
2176 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2179 addr
= memblock_alloc_try_nid_raw(
2180 huge_page_size(h
), huge_page_size(h
),
2181 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2184 * Use the beginning of the huge page to store the
2185 * huge_bootmem_page struct (until gather_bootmem
2186 * puts them into the mem_map).
2195 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2196 /* Put them into a private list first because mem_map is not up yet */
2197 INIT_LIST_HEAD(&m
->list
);
2198 list_add(&m
->list
, &huge_boot_pages
);
2203 static void __init
prep_compound_huge_page(struct page
*page
,
2206 if (unlikely(order
> (MAX_ORDER
- 1)))
2207 prep_compound_gigantic_page(page
, order
);
2209 prep_compound_page(page
, order
);
2212 /* Put bootmem huge pages into the standard lists after mem_map is up */
2213 static void __init
gather_bootmem_prealloc(void)
2215 struct huge_bootmem_page
*m
;
2217 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2218 struct page
*page
= virt_to_page(m
);
2219 struct hstate
*h
= m
->hstate
;
2221 WARN_ON(page_count(page
) != 1);
2222 prep_compound_huge_page(page
, h
->order
);
2223 WARN_ON(PageReserved(page
));
2224 prep_new_huge_page(h
, page
, page_to_nid(page
));
2225 put_page(page
); /* free it into the hugepage allocator */
2228 * If we had gigantic hugepages allocated at boot time, we need
2229 * to restore the 'stolen' pages to totalram_pages in order to
2230 * fix confusing memory reports from free(1) and another
2231 * side-effects, like CommitLimit going negative.
2233 if (hstate_is_gigantic(h
))
2234 adjust_managed_page_count(page
, 1 << h
->order
);
2239 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2242 nodemask_t
*node_alloc_noretry
;
2244 if (!hstate_is_gigantic(h
)) {
2246 * Bit mask controlling how hard we retry per-node allocations.
2247 * Ignore errors as lower level routines can deal with
2248 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2249 * time, we are likely in bigger trouble.
2251 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2254 /* allocations done at boot time */
2255 node_alloc_noretry
= NULL
;
2258 /* bit mask controlling how hard we retry per-node allocations */
2259 if (node_alloc_noretry
)
2260 nodes_clear(*node_alloc_noretry
);
2262 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2263 if (hstate_is_gigantic(h
)) {
2264 if (!alloc_bootmem_huge_page(h
))
2266 } else if (!alloc_pool_huge_page(h
,
2267 &node_states
[N_MEMORY
],
2268 node_alloc_noretry
))
2272 if (i
< h
->max_huge_pages
) {
2275 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2276 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2277 h
->max_huge_pages
, buf
, i
);
2278 h
->max_huge_pages
= i
;
2281 kfree(node_alloc_noretry
);
2284 static void __init
hugetlb_init_hstates(void)
2288 for_each_hstate(h
) {
2289 if (minimum_order
> huge_page_order(h
))
2290 minimum_order
= huge_page_order(h
);
2292 /* oversize hugepages were init'ed in early boot */
2293 if (!hstate_is_gigantic(h
))
2294 hugetlb_hstate_alloc_pages(h
);
2296 VM_BUG_ON(minimum_order
== UINT_MAX
);
2299 static void __init
report_hugepages(void)
2303 for_each_hstate(h
) {
2306 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2307 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2308 buf
, h
->free_huge_pages
);
2312 #ifdef CONFIG_HIGHMEM
2313 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2314 nodemask_t
*nodes_allowed
)
2318 if (hstate_is_gigantic(h
))
2321 for_each_node_mask(i
, *nodes_allowed
) {
2322 struct page
*page
, *next
;
2323 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2324 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2325 if (count
>= h
->nr_huge_pages
)
2327 if (PageHighMem(page
))
2329 list_del(&page
->lru
);
2330 update_and_free_page(h
, page
);
2331 h
->free_huge_pages
--;
2332 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2337 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2338 nodemask_t
*nodes_allowed
)
2344 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2345 * balanced by operating on them in a round-robin fashion.
2346 * Returns 1 if an adjustment was made.
2348 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2353 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2356 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2357 if (h
->surplus_huge_pages_node
[node
])
2361 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2362 if (h
->surplus_huge_pages_node
[node
] <
2363 h
->nr_huge_pages_node
[node
])
2370 h
->surplus_huge_pages
+= delta
;
2371 h
->surplus_huge_pages_node
[node
] += delta
;
2375 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2376 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
2377 nodemask_t
*nodes_allowed
)
2379 unsigned long min_count
, ret
;
2380 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
2383 * Bit mask controlling how hard we retry per-node allocations.
2384 * If we can not allocate the bit mask, do not attempt to allocate
2385 * the requested huge pages.
2387 if (node_alloc_noretry
)
2388 nodes_clear(*node_alloc_noretry
);
2392 spin_lock(&hugetlb_lock
);
2395 * Check for a node specific request.
2396 * Changing node specific huge page count may require a corresponding
2397 * change to the global count. In any case, the passed node mask
2398 * (nodes_allowed) will restrict alloc/free to the specified node.
2400 if (nid
!= NUMA_NO_NODE
) {
2401 unsigned long old_count
= count
;
2403 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2405 * User may have specified a large count value which caused the
2406 * above calculation to overflow. In this case, they wanted
2407 * to allocate as many huge pages as possible. Set count to
2408 * largest possible value to align with their intention.
2410 if (count
< old_count
)
2415 * Gigantic pages runtime allocation depend on the capability for large
2416 * page range allocation.
2417 * If the system does not provide this feature, return an error when
2418 * the user tries to allocate gigantic pages but let the user free the
2419 * boottime allocated gigantic pages.
2421 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
2422 if (count
> persistent_huge_pages(h
)) {
2423 spin_unlock(&hugetlb_lock
);
2424 NODEMASK_FREE(node_alloc_noretry
);
2427 /* Fall through to decrease pool */
2431 * Increase the pool size
2432 * First take pages out of surplus state. Then make up the
2433 * remaining difference by allocating fresh huge pages.
2435 * We might race with alloc_surplus_huge_page() here and be unable
2436 * to convert a surplus huge page to a normal huge page. That is
2437 * not critical, though, it just means the overall size of the
2438 * pool might be one hugepage larger than it needs to be, but
2439 * within all the constraints specified by the sysctls.
2441 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2442 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2446 while (count
> persistent_huge_pages(h
)) {
2448 * If this allocation races such that we no longer need the
2449 * page, free_huge_page will handle it by freeing the page
2450 * and reducing the surplus.
2452 spin_unlock(&hugetlb_lock
);
2454 /* yield cpu to avoid soft lockup */
2457 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
2458 node_alloc_noretry
);
2459 spin_lock(&hugetlb_lock
);
2463 /* Bail for signals. Probably ctrl-c from user */
2464 if (signal_pending(current
))
2469 * Decrease the pool size
2470 * First return free pages to the buddy allocator (being careful
2471 * to keep enough around to satisfy reservations). Then place
2472 * pages into surplus state as needed so the pool will shrink
2473 * to the desired size as pages become free.
2475 * By placing pages into the surplus state independent of the
2476 * overcommit value, we are allowing the surplus pool size to
2477 * exceed overcommit. There are few sane options here. Since
2478 * alloc_surplus_huge_page() is checking the global counter,
2479 * though, we'll note that we're not allowed to exceed surplus
2480 * and won't grow the pool anywhere else. Not until one of the
2481 * sysctls are changed, or the surplus pages go out of use.
2483 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2484 min_count
= max(count
, min_count
);
2485 try_to_free_low(h
, min_count
, nodes_allowed
);
2486 while (min_count
< persistent_huge_pages(h
)) {
2487 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2489 cond_resched_lock(&hugetlb_lock
);
2491 while (count
< persistent_huge_pages(h
)) {
2492 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2496 h
->max_huge_pages
= persistent_huge_pages(h
);
2497 spin_unlock(&hugetlb_lock
);
2499 NODEMASK_FREE(node_alloc_noretry
);
2504 #define HSTATE_ATTR_RO(_name) \
2505 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2507 #define HSTATE_ATTR(_name) \
2508 static struct kobj_attribute _name##_attr = \
2509 __ATTR(_name, 0644, _name##_show, _name##_store)
2511 static struct kobject
*hugepages_kobj
;
2512 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2514 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2516 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2520 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2521 if (hstate_kobjs
[i
] == kobj
) {
2523 *nidp
= NUMA_NO_NODE
;
2527 return kobj_to_node_hstate(kobj
, nidp
);
2530 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2531 struct kobj_attribute
*attr
, char *buf
)
2534 unsigned long nr_huge_pages
;
2537 h
= kobj_to_hstate(kobj
, &nid
);
2538 if (nid
== NUMA_NO_NODE
)
2539 nr_huge_pages
= h
->nr_huge_pages
;
2541 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2543 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2546 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2547 struct hstate
*h
, int nid
,
2548 unsigned long count
, size_t len
)
2551 nodemask_t nodes_allowed
, *n_mask
;
2553 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
2556 if (nid
== NUMA_NO_NODE
) {
2558 * global hstate attribute
2560 if (!(obey_mempolicy
&&
2561 init_nodemask_of_mempolicy(&nodes_allowed
)))
2562 n_mask
= &node_states
[N_MEMORY
];
2564 n_mask
= &nodes_allowed
;
2567 * Node specific request. count adjustment happens in
2568 * set_max_huge_pages() after acquiring hugetlb_lock.
2570 init_nodemask_of_node(&nodes_allowed
, nid
);
2571 n_mask
= &nodes_allowed
;
2574 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
2576 return err
? err
: len
;
2579 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2580 struct kobject
*kobj
, const char *buf
,
2584 unsigned long count
;
2588 err
= kstrtoul(buf
, 10, &count
);
2592 h
= kobj_to_hstate(kobj
, &nid
);
2593 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2596 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2597 struct kobj_attribute
*attr
, char *buf
)
2599 return nr_hugepages_show_common(kobj
, attr
, buf
);
2602 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2603 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2605 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2607 HSTATE_ATTR(nr_hugepages
);
2612 * hstate attribute for optionally mempolicy-based constraint on persistent
2613 * huge page alloc/free.
2615 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2616 struct kobj_attribute
*attr
, char *buf
)
2618 return nr_hugepages_show_common(kobj
, attr
, buf
);
2621 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2622 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2624 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2626 HSTATE_ATTR(nr_hugepages_mempolicy
);
2630 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2631 struct kobj_attribute
*attr
, char *buf
)
2633 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2634 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2637 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2638 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2641 unsigned long input
;
2642 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2644 if (hstate_is_gigantic(h
))
2647 err
= kstrtoul(buf
, 10, &input
);
2651 spin_lock(&hugetlb_lock
);
2652 h
->nr_overcommit_huge_pages
= input
;
2653 spin_unlock(&hugetlb_lock
);
2657 HSTATE_ATTR(nr_overcommit_hugepages
);
2659 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2660 struct kobj_attribute
*attr
, char *buf
)
2663 unsigned long free_huge_pages
;
2666 h
= kobj_to_hstate(kobj
, &nid
);
2667 if (nid
== NUMA_NO_NODE
)
2668 free_huge_pages
= h
->free_huge_pages
;
2670 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2672 return sprintf(buf
, "%lu\n", free_huge_pages
);
2674 HSTATE_ATTR_RO(free_hugepages
);
2676 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2677 struct kobj_attribute
*attr
, char *buf
)
2679 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2680 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2682 HSTATE_ATTR_RO(resv_hugepages
);
2684 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2685 struct kobj_attribute
*attr
, char *buf
)
2688 unsigned long surplus_huge_pages
;
2691 h
= kobj_to_hstate(kobj
, &nid
);
2692 if (nid
== NUMA_NO_NODE
)
2693 surplus_huge_pages
= h
->surplus_huge_pages
;
2695 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2697 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2699 HSTATE_ATTR_RO(surplus_hugepages
);
2701 static struct attribute
*hstate_attrs
[] = {
2702 &nr_hugepages_attr
.attr
,
2703 &nr_overcommit_hugepages_attr
.attr
,
2704 &free_hugepages_attr
.attr
,
2705 &resv_hugepages_attr
.attr
,
2706 &surplus_hugepages_attr
.attr
,
2708 &nr_hugepages_mempolicy_attr
.attr
,
2713 static const struct attribute_group hstate_attr_group
= {
2714 .attrs
= hstate_attrs
,
2717 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2718 struct kobject
**hstate_kobjs
,
2719 const struct attribute_group
*hstate_attr_group
)
2722 int hi
= hstate_index(h
);
2724 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2725 if (!hstate_kobjs
[hi
])
2728 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2730 kobject_put(hstate_kobjs
[hi
]);
2735 static void __init
hugetlb_sysfs_init(void)
2740 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2741 if (!hugepages_kobj
)
2744 for_each_hstate(h
) {
2745 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2746 hstate_kobjs
, &hstate_attr_group
);
2748 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2755 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2756 * with node devices in node_devices[] using a parallel array. The array
2757 * index of a node device or _hstate == node id.
2758 * This is here to avoid any static dependency of the node device driver, in
2759 * the base kernel, on the hugetlb module.
2761 struct node_hstate
{
2762 struct kobject
*hugepages_kobj
;
2763 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2765 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2768 * A subset of global hstate attributes for node devices
2770 static struct attribute
*per_node_hstate_attrs
[] = {
2771 &nr_hugepages_attr
.attr
,
2772 &free_hugepages_attr
.attr
,
2773 &surplus_hugepages_attr
.attr
,
2777 static const struct attribute_group per_node_hstate_attr_group
= {
2778 .attrs
= per_node_hstate_attrs
,
2782 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2783 * Returns node id via non-NULL nidp.
2785 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2789 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2790 struct node_hstate
*nhs
= &node_hstates
[nid
];
2792 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2793 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2805 * Unregister hstate attributes from a single node device.
2806 * No-op if no hstate attributes attached.
2808 static void hugetlb_unregister_node(struct node
*node
)
2811 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2813 if (!nhs
->hugepages_kobj
)
2814 return; /* no hstate attributes */
2816 for_each_hstate(h
) {
2817 int idx
= hstate_index(h
);
2818 if (nhs
->hstate_kobjs
[idx
]) {
2819 kobject_put(nhs
->hstate_kobjs
[idx
]);
2820 nhs
->hstate_kobjs
[idx
] = NULL
;
2824 kobject_put(nhs
->hugepages_kobj
);
2825 nhs
->hugepages_kobj
= NULL
;
2830 * Register hstate attributes for a single node device.
2831 * No-op if attributes already registered.
2833 static void hugetlb_register_node(struct node
*node
)
2836 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2839 if (nhs
->hugepages_kobj
)
2840 return; /* already allocated */
2842 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2844 if (!nhs
->hugepages_kobj
)
2847 for_each_hstate(h
) {
2848 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2850 &per_node_hstate_attr_group
);
2852 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2853 h
->name
, node
->dev
.id
);
2854 hugetlb_unregister_node(node
);
2861 * hugetlb init time: register hstate attributes for all registered node
2862 * devices of nodes that have memory. All on-line nodes should have
2863 * registered their associated device by this time.
2865 static void __init
hugetlb_register_all_nodes(void)
2869 for_each_node_state(nid
, N_MEMORY
) {
2870 struct node
*node
= node_devices
[nid
];
2871 if (node
->dev
.id
== nid
)
2872 hugetlb_register_node(node
);
2876 * Let the node device driver know we're here so it can
2877 * [un]register hstate attributes on node hotplug.
2879 register_hugetlbfs_with_node(hugetlb_register_node
,
2880 hugetlb_unregister_node
);
2882 #else /* !CONFIG_NUMA */
2884 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2892 static void hugetlb_register_all_nodes(void) { }
2896 static int __init
hugetlb_init(void)
2900 if (!hugepages_supported())
2903 if (!size_to_hstate(default_hstate_size
)) {
2904 if (default_hstate_size
!= 0) {
2905 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2906 default_hstate_size
, HPAGE_SIZE
);
2909 default_hstate_size
= HPAGE_SIZE
;
2910 if (!size_to_hstate(default_hstate_size
))
2911 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2913 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2914 if (default_hstate_max_huge_pages
) {
2915 if (!default_hstate
.max_huge_pages
)
2916 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2919 hugetlb_init_hstates();
2920 gather_bootmem_prealloc();
2923 hugetlb_sysfs_init();
2924 hugetlb_register_all_nodes();
2925 hugetlb_cgroup_file_init();
2928 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2930 num_fault_mutexes
= 1;
2932 hugetlb_fault_mutex_table
=
2933 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
2935 BUG_ON(!hugetlb_fault_mutex_table
);
2937 for (i
= 0; i
< num_fault_mutexes
; i
++)
2938 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2941 subsys_initcall(hugetlb_init
);
2943 /* Should be called on processing a hugepagesz=... option */
2944 void __init
hugetlb_bad_size(void)
2946 parsed_valid_hugepagesz
= false;
2949 void __init
hugetlb_add_hstate(unsigned int order
)
2954 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2955 pr_warn("hugepagesz= specified twice, ignoring\n");
2958 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2960 h
= &hstates
[hugetlb_max_hstate
++];
2962 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2963 h
->nr_huge_pages
= 0;
2964 h
->free_huge_pages
= 0;
2965 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2966 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2967 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2968 h
->next_nid_to_alloc
= first_memory_node
;
2969 h
->next_nid_to_free
= first_memory_node
;
2970 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2971 huge_page_size(h
)/1024);
2976 static int __init
hugetlb_nrpages_setup(char *s
)
2979 static unsigned long *last_mhp
;
2981 if (!parsed_valid_hugepagesz
) {
2982 pr_warn("hugepages = %s preceded by "
2983 "an unsupported hugepagesz, ignoring\n", s
);
2984 parsed_valid_hugepagesz
= true;
2988 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2989 * so this hugepages= parameter goes to the "default hstate".
2991 else if (!hugetlb_max_hstate
)
2992 mhp
= &default_hstate_max_huge_pages
;
2994 mhp
= &parsed_hstate
->max_huge_pages
;
2996 if (mhp
== last_mhp
) {
2997 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3001 if (sscanf(s
, "%lu", mhp
) <= 0)
3005 * Global state is always initialized later in hugetlb_init.
3006 * But we need to allocate >= MAX_ORDER hstates here early to still
3007 * use the bootmem allocator.
3009 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
3010 hugetlb_hstate_alloc_pages(parsed_hstate
);
3016 __setup("hugepages=", hugetlb_nrpages_setup
);
3018 static int __init
hugetlb_default_setup(char *s
)
3020 default_hstate_size
= memparse(s
, &s
);
3023 __setup("default_hugepagesz=", hugetlb_default_setup
);
3025 static unsigned int cpuset_mems_nr(unsigned int *array
)
3028 unsigned int nr
= 0;
3030 for_each_node_mask(node
, cpuset_current_mems_allowed
)
3036 #ifdef CONFIG_SYSCTL
3037 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3038 struct ctl_table
*table
, int write
,
3039 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3041 struct hstate
*h
= &default_hstate
;
3042 unsigned long tmp
= h
->max_huge_pages
;
3045 if (!hugepages_supported())
3049 table
->maxlen
= sizeof(unsigned long);
3050 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3055 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3056 NUMA_NO_NODE
, tmp
, *length
);
3061 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3062 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3065 return hugetlb_sysctl_handler_common(false, table
, write
,
3066 buffer
, length
, ppos
);
3070 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3071 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3073 return hugetlb_sysctl_handler_common(true, table
, write
,
3074 buffer
, length
, ppos
);
3076 #endif /* CONFIG_NUMA */
3078 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3079 void __user
*buffer
,
3080 size_t *length
, loff_t
*ppos
)
3082 struct hstate
*h
= &default_hstate
;
3086 if (!hugepages_supported())
3089 tmp
= h
->nr_overcommit_huge_pages
;
3091 if (write
&& hstate_is_gigantic(h
))
3095 table
->maxlen
= sizeof(unsigned long);
3096 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3101 spin_lock(&hugetlb_lock
);
3102 h
->nr_overcommit_huge_pages
= tmp
;
3103 spin_unlock(&hugetlb_lock
);
3109 #endif /* CONFIG_SYSCTL */
3111 void hugetlb_report_meminfo(struct seq_file
*m
)
3114 unsigned long total
= 0;
3116 if (!hugepages_supported())
3119 for_each_hstate(h
) {
3120 unsigned long count
= h
->nr_huge_pages
;
3122 total
+= (PAGE_SIZE
<< huge_page_order(h
)) * count
;
3124 if (h
== &default_hstate
)
3126 "HugePages_Total: %5lu\n"
3127 "HugePages_Free: %5lu\n"
3128 "HugePages_Rsvd: %5lu\n"
3129 "HugePages_Surp: %5lu\n"
3130 "Hugepagesize: %8lu kB\n",
3134 h
->surplus_huge_pages
,
3135 (PAGE_SIZE
<< huge_page_order(h
)) / 1024);
3138 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ 1024);
3141 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3143 struct hstate
*h
= &default_hstate
;
3144 if (!hugepages_supported())
3147 "Node %d HugePages_Total: %5u\n"
3148 "Node %d HugePages_Free: %5u\n"
3149 "Node %d HugePages_Surp: %5u\n",
3150 nid
, h
->nr_huge_pages_node
[nid
],
3151 nid
, h
->free_huge_pages_node
[nid
],
3152 nid
, h
->surplus_huge_pages_node
[nid
]);
3155 void hugetlb_show_meminfo(void)
3160 if (!hugepages_supported())
3163 for_each_node_state(nid
, N_MEMORY
)
3165 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3167 h
->nr_huge_pages_node
[nid
],
3168 h
->free_huge_pages_node
[nid
],
3169 h
->surplus_huge_pages_node
[nid
],
3170 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3173 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3175 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3176 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3179 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3180 unsigned long hugetlb_total_pages(void)
3183 unsigned long nr_total_pages
= 0;
3186 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3187 return nr_total_pages
;
3190 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3194 spin_lock(&hugetlb_lock
);
3196 * When cpuset is configured, it breaks the strict hugetlb page
3197 * reservation as the accounting is done on a global variable. Such
3198 * reservation is completely rubbish in the presence of cpuset because
3199 * the reservation is not checked against page availability for the
3200 * current cpuset. Application can still potentially OOM'ed by kernel
3201 * with lack of free htlb page in cpuset that the task is in.
3202 * Attempt to enforce strict accounting with cpuset is almost
3203 * impossible (or too ugly) because cpuset is too fluid that
3204 * task or memory node can be dynamically moved between cpusets.
3206 * The change of semantics for shared hugetlb mapping with cpuset is
3207 * undesirable. However, in order to preserve some of the semantics,
3208 * we fall back to check against current free page availability as
3209 * a best attempt and hopefully to minimize the impact of changing
3210 * semantics that cpuset has.
3213 if (gather_surplus_pages(h
, delta
) < 0)
3216 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3217 return_unused_surplus_pages(h
, delta
);
3224 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3227 spin_unlock(&hugetlb_lock
);
3231 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3233 struct resv_map
*resv
= vma_resv_map(vma
);
3236 * This new VMA should share its siblings reservation map if present.
3237 * The VMA will only ever have a valid reservation map pointer where
3238 * it is being copied for another still existing VMA. As that VMA
3239 * has a reference to the reservation map it cannot disappear until
3240 * after this open call completes. It is therefore safe to take a
3241 * new reference here without additional locking.
3243 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3244 kref_get(&resv
->refs
);
3247 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3249 struct hstate
*h
= hstate_vma(vma
);
3250 struct resv_map
*resv
= vma_resv_map(vma
);
3251 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3252 unsigned long reserve
, start
, end
;
3255 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3258 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3259 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3261 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3263 kref_put(&resv
->refs
, resv_map_release
);
3267 * Decrement reserve counts. The global reserve count may be
3268 * adjusted if the subpool has a minimum size.
3270 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3271 hugetlb_acct_memory(h
, -gbl_reserve
);
3275 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
3277 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
3282 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
3284 struct hstate
*hstate
= hstate_vma(vma
);
3286 return 1UL << huge_page_shift(hstate
);
3290 * We cannot handle pagefaults against hugetlb pages at all. They cause
3291 * handle_mm_fault() to try to instantiate regular-sized pages in the
3292 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3295 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3302 * When a new function is introduced to vm_operations_struct and added
3303 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3304 * This is because under System V memory model, mappings created via
3305 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3306 * their original vm_ops are overwritten with shm_vm_ops.
3308 const struct vm_operations_struct hugetlb_vm_ops
= {
3309 .fault
= hugetlb_vm_op_fault
,
3310 .open
= hugetlb_vm_op_open
,
3311 .close
= hugetlb_vm_op_close
,
3312 .split
= hugetlb_vm_op_split
,
3313 .pagesize
= hugetlb_vm_op_pagesize
,
3316 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3322 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3323 vma
->vm_page_prot
)));
3325 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3326 vma
->vm_page_prot
));
3328 entry
= pte_mkyoung(entry
);
3329 entry
= pte_mkhuge(entry
);
3330 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3335 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3336 unsigned long address
, pte_t
*ptep
)
3340 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3341 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3342 update_mmu_cache(vma
, address
, ptep
);
3345 bool is_hugetlb_entry_migration(pte_t pte
)
3349 if (huge_pte_none(pte
) || pte_present(pte
))
3351 swp
= pte_to_swp_entry(pte
);
3352 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3358 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3362 if (huge_pte_none(pte
) || pte_present(pte
))
3364 swp
= pte_to_swp_entry(pte
);
3365 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3371 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3372 struct vm_area_struct
*vma
)
3374 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
3375 struct page
*ptepage
;
3378 struct hstate
*h
= hstate_vma(vma
);
3379 unsigned long sz
= huge_page_size(h
);
3380 struct mmu_notifier_range range
;
3383 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3386 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
3389 mmu_notifier_invalidate_range_start(&range
);
3392 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3393 spinlock_t
*src_ptl
, *dst_ptl
;
3394 src_pte
= huge_pte_offset(src
, addr
, sz
);
3397 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3404 * If the pagetables are shared don't copy or take references.
3405 * dst_pte == src_pte is the common case of src/dest sharing.
3407 * However, src could have 'unshared' and dst shares with
3408 * another vma. If dst_pte !none, this implies sharing.
3409 * Check here before taking page table lock, and once again
3410 * after taking the lock below.
3412 dst_entry
= huge_ptep_get(dst_pte
);
3413 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
3416 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3417 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3418 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3419 entry
= huge_ptep_get(src_pte
);
3420 dst_entry
= huge_ptep_get(dst_pte
);
3421 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
3423 * Skip if src entry none. Also, skip in the
3424 * unlikely case dst entry !none as this implies
3425 * sharing with another vma.
3428 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3429 is_hugetlb_entry_hwpoisoned(entry
))) {
3430 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3432 if (is_write_migration_entry(swp_entry
) && cow
) {
3434 * COW mappings require pages in both
3435 * parent and child to be set to read.
3437 make_migration_entry_read(&swp_entry
);
3438 entry
= swp_entry_to_pte(swp_entry
);
3439 set_huge_swap_pte_at(src
, addr
, src_pte
,
3442 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3446 * No need to notify as we are downgrading page
3447 * table protection not changing it to point
3450 * See Documentation/vm/mmu_notifier.rst
3452 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3454 entry
= huge_ptep_get(src_pte
);
3455 ptepage
= pte_page(entry
);
3457 page_dup_rmap(ptepage
, true);
3458 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3459 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3461 spin_unlock(src_ptl
);
3462 spin_unlock(dst_ptl
);
3466 mmu_notifier_invalidate_range_end(&range
);
3471 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3472 unsigned long start
, unsigned long end
,
3473 struct page
*ref_page
)
3475 struct mm_struct
*mm
= vma
->vm_mm
;
3476 unsigned long address
;
3481 struct hstate
*h
= hstate_vma(vma
);
3482 unsigned long sz
= huge_page_size(h
);
3483 struct mmu_notifier_range range
;
3485 WARN_ON(!is_vm_hugetlb_page(vma
));
3486 BUG_ON(start
& ~huge_page_mask(h
));
3487 BUG_ON(end
& ~huge_page_mask(h
));
3490 * This is a hugetlb vma, all the pte entries should point
3493 tlb_change_page_size(tlb
, sz
);
3494 tlb_start_vma(tlb
, vma
);
3497 * If sharing possible, alert mmu notifiers of worst case.
3499 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
3501 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
3502 mmu_notifier_invalidate_range_start(&range
);
3504 for (; address
< end
; address
+= sz
) {
3505 ptep
= huge_pte_offset(mm
, address
, sz
);
3509 ptl
= huge_pte_lock(h
, mm
, ptep
);
3510 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3513 * We just unmapped a page of PMDs by clearing a PUD.
3514 * The caller's TLB flush range should cover this area.
3519 pte
= huge_ptep_get(ptep
);
3520 if (huge_pte_none(pte
)) {
3526 * Migrating hugepage or HWPoisoned hugepage is already
3527 * unmapped and its refcount is dropped, so just clear pte here.
3529 if (unlikely(!pte_present(pte
))) {
3530 huge_pte_clear(mm
, address
, ptep
, sz
);
3535 page
= pte_page(pte
);
3537 * If a reference page is supplied, it is because a specific
3538 * page is being unmapped, not a range. Ensure the page we
3539 * are about to unmap is the actual page of interest.
3542 if (page
!= ref_page
) {
3547 * Mark the VMA as having unmapped its page so that
3548 * future faults in this VMA will fail rather than
3549 * looking like data was lost
3551 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3554 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3555 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3556 if (huge_pte_dirty(pte
))
3557 set_page_dirty(page
);
3559 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3560 page_remove_rmap(page
, true);
3563 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3565 * Bail out after unmapping reference page if supplied
3570 mmu_notifier_invalidate_range_end(&range
);
3571 tlb_end_vma(tlb
, vma
);
3574 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3575 struct vm_area_struct
*vma
, unsigned long start
,
3576 unsigned long end
, struct page
*ref_page
)
3578 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3581 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3582 * test will fail on a vma being torn down, and not grab a page table
3583 * on its way out. We're lucky that the flag has such an appropriate
3584 * name, and can in fact be safely cleared here. We could clear it
3585 * before the __unmap_hugepage_range above, but all that's necessary
3586 * is to clear it before releasing the i_mmap_rwsem. This works
3587 * because in the context this is called, the VMA is about to be
3588 * destroyed and the i_mmap_rwsem is held.
3590 vma
->vm_flags
&= ~VM_MAYSHARE
;
3593 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3594 unsigned long end
, struct page
*ref_page
)
3596 struct mm_struct
*mm
;
3597 struct mmu_gather tlb
;
3598 unsigned long tlb_start
= start
;
3599 unsigned long tlb_end
= end
;
3602 * If shared PMDs were possibly used within this vma range, adjust
3603 * start/end for worst case tlb flushing.
3604 * Note that we can not be sure if PMDs are shared until we try to
3605 * unmap pages. However, we want to make sure TLB flushing covers
3606 * the largest possible range.
3608 adjust_range_if_pmd_sharing_possible(vma
, &tlb_start
, &tlb_end
);
3612 tlb_gather_mmu(&tlb
, mm
, tlb_start
, tlb_end
);
3613 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3614 tlb_finish_mmu(&tlb
, tlb_start
, tlb_end
);
3618 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3619 * mappping it owns the reserve page for. The intention is to unmap the page
3620 * from other VMAs and let the children be SIGKILLed if they are faulting the
3623 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3624 struct page
*page
, unsigned long address
)
3626 struct hstate
*h
= hstate_vma(vma
);
3627 struct vm_area_struct
*iter_vma
;
3628 struct address_space
*mapping
;
3632 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3633 * from page cache lookup which is in HPAGE_SIZE units.
3635 address
= address
& huge_page_mask(h
);
3636 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3638 mapping
= vma
->vm_file
->f_mapping
;
3641 * Take the mapping lock for the duration of the table walk. As
3642 * this mapping should be shared between all the VMAs,
3643 * __unmap_hugepage_range() is called as the lock is already held
3645 i_mmap_lock_write(mapping
);
3646 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3647 /* Do not unmap the current VMA */
3648 if (iter_vma
== vma
)
3652 * Shared VMAs have their own reserves and do not affect
3653 * MAP_PRIVATE accounting but it is possible that a shared
3654 * VMA is using the same page so check and skip such VMAs.
3656 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3660 * Unmap the page from other VMAs without their own reserves.
3661 * They get marked to be SIGKILLed if they fault in these
3662 * areas. This is because a future no-page fault on this VMA
3663 * could insert a zeroed page instead of the data existing
3664 * from the time of fork. This would look like data corruption
3666 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3667 unmap_hugepage_range(iter_vma
, address
,
3668 address
+ huge_page_size(h
), page
);
3670 i_mmap_unlock_write(mapping
);
3674 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3675 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3676 * cannot race with other handlers or page migration.
3677 * Keep the pte_same checks anyway to make transition from the mutex easier.
3679 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3680 unsigned long address
, pte_t
*ptep
,
3681 struct page
*pagecache_page
, spinlock_t
*ptl
)
3684 struct hstate
*h
= hstate_vma(vma
);
3685 struct page
*old_page
, *new_page
;
3686 int outside_reserve
= 0;
3688 unsigned long haddr
= address
& huge_page_mask(h
);
3689 struct mmu_notifier_range range
;
3691 pte
= huge_ptep_get(ptep
);
3692 old_page
= pte_page(pte
);
3695 /* If no-one else is actually using this page, avoid the copy
3696 * and just make the page writable */
3697 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3698 page_move_anon_rmap(old_page
, vma
);
3699 set_huge_ptep_writable(vma
, haddr
, ptep
);
3704 * If the process that created a MAP_PRIVATE mapping is about to
3705 * perform a COW due to a shared page count, attempt to satisfy
3706 * the allocation without using the existing reserves. The pagecache
3707 * page is used to determine if the reserve at this address was
3708 * consumed or not. If reserves were used, a partial faulted mapping
3709 * at the time of fork() could consume its reserves on COW instead
3710 * of the full address range.
3712 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3713 old_page
!= pagecache_page
)
3714 outside_reserve
= 1;
3719 * Drop page table lock as buddy allocator may be called. It will
3720 * be acquired again before returning to the caller, as expected.
3723 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
3725 if (IS_ERR(new_page
)) {
3727 * If a process owning a MAP_PRIVATE mapping fails to COW,
3728 * it is due to references held by a child and an insufficient
3729 * huge page pool. To guarantee the original mappers
3730 * reliability, unmap the page from child processes. The child
3731 * may get SIGKILLed if it later faults.
3733 if (outside_reserve
) {
3735 BUG_ON(huge_pte_none(pte
));
3736 unmap_ref_private(mm
, vma
, old_page
, haddr
);
3737 BUG_ON(huge_pte_none(pte
));
3739 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
3741 pte_same(huge_ptep_get(ptep
), pte
)))
3742 goto retry_avoidcopy
;
3744 * race occurs while re-acquiring page table
3745 * lock, and our job is done.
3750 ret
= vmf_error(PTR_ERR(new_page
));
3751 goto out_release_old
;
3755 * When the original hugepage is shared one, it does not have
3756 * anon_vma prepared.
3758 if (unlikely(anon_vma_prepare(vma
))) {
3760 goto out_release_all
;
3763 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3764 pages_per_huge_page(h
));
3765 __SetPageUptodate(new_page
);
3767 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
3768 haddr
+ huge_page_size(h
));
3769 mmu_notifier_invalidate_range_start(&range
);
3772 * Retake the page table lock to check for racing updates
3773 * before the page tables are altered
3776 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
3777 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3778 ClearPagePrivate(new_page
);
3781 huge_ptep_clear_flush(vma
, haddr
, ptep
);
3782 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
3783 set_huge_pte_at(mm
, haddr
, ptep
,
3784 make_huge_pte(vma
, new_page
, 1));
3785 page_remove_rmap(old_page
, true);
3786 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
3787 set_page_huge_active(new_page
);
3788 /* Make the old page be freed below */
3789 new_page
= old_page
;
3792 mmu_notifier_invalidate_range_end(&range
);
3794 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
3799 spin_lock(ptl
); /* Caller expects lock to be held */
3803 /* Return the pagecache page at a given address within a VMA */
3804 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3805 struct vm_area_struct
*vma
, unsigned long address
)
3807 struct address_space
*mapping
;
3810 mapping
= vma
->vm_file
->f_mapping
;
3811 idx
= vma_hugecache_offset(h
, vma
, address
);
3813 return find_lock_page(mapping
, idx
);
3817 * Return whether there is a pagecache page to back given address within VMA.
3818 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3820 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3821 struct vm_area_struct
*vma
, unsigned long address
)
3823 struct address_space
*mapping
;
3827 mapping
= vma
->vm_file
->f_mapping
;
3828 idx
= vma_hugecache_offset(h
, vma
, address
);
3830 page
= find_get_page(mapping
, idx
);
3833 return page
!= NULL
;
3836 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3839 struct inode
*inode
= mapping
->host
;
3840 struct hstate
*h
= hstate_inode(inode
);
3841 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3845 ClearPagePrivate(page
);
3848 * set page dirty so that it will not be removed from cache/file
3849 * by non-hugetlbfs specific code paths.
3851 set_page_dirty(page
);
3853 spin_lock(&inode
->i_lock
);
3854 inode
->i_blocks
+= blocks_per_huge_page(h
);
3855 spin_unlock(&inode
->i_lock
);
3859 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
3860 struct vm_area_struct
*vma
,
3861 struct address_space
*mapping
, pgoff_t idx
,
3862 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3864 struct hstate
*h
= hstate_vma(vma
);
3865 vm_fault_t ret
= VM_FAULT_SIGBUS
;
3871 unsigned long haddr
= address
& huge_page_mask(h
);
3872 bool new_page
= false;
3875 * Currently, we are forced to kill the process in the event the
3876 * original mapper has unmapped pages from the child due to a failed
3877 * COW. Warn that such a situation has occurred as it may not be obvious
3879 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3880 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3886 * Use page lock to guard against racing truncation
3887 * before we get page_table_lock.
3890 page
= find_lock_page(mapping
, idx
);
3892 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3897 * Check for page in userfault range
3899 if (userfaultfd_missing(vma
)) {
3901 struct vm_fault vmf
= {
3906 * Hard to debug if it ends up being
3907 * used by a callee that assumes
3908 * something about the other
3909 * uninitialized fields... same as in
3915 * hugetlb_fault_mutex must be dropped before
3916 * handling userfault. Reacquire after handling
3917 * fault to make calling code simpler.
3919 hash
= hugetlb_fault_mutex_hash(h
, mapping
, idx
, haddr
);
3920 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3921 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
3922 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3926 page
= alloc_huge_page(vma
, haddr
, 0);
3929 * Returning error will result in faulting task being
3930 * sent SIGBUS. The hugetlb fault mutex prevents two
3931 * tasks from racing to fault in the same page which
3932 * could result in false unable to allocate errors.
3933 * Page migration does not take the fault mutex, but
3934 * does a clear then write of pte's under page table
3935 * lock. Page fault code could race with migration,
3936 * notice the clear pte and try to allocate a page
3937 * here. Before returning error, get ptl and make
3938 * sure there really is no pte entry.
3940 ptl
= huge_pte_lock(h
, mm
, ptep
);
3941 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3947 ret
= vmf_error(PTR_ERR(page
));
3950 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3951 __SetPageUptodate(page
);
3954 if (vma
->vm_flags
& VM_MAYSHARE
) {
3955 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3964 if (unlikely(anon_vma_prepare(vma
))) {
3966 goto backout_unlocked
;
3972 * If memory error occurs between mmap() and fault, some process
3973 * don't have hwpoisoned swap entry for errored virtual address.
3974 * So we need to block hugepage fault by PG_hwpoison bit check.
3976 if (unlikely(PageHWPoison(page
))) {
3977 ret
= VM_FAULT_HWPOISON
|
3978 VM_FAULT_SET_HINDEX(hstate_index(h
));
3979 goto backout_unlocked
;
3984 * If we are going to COW a private mapping later, we examine the
3985 * pending reservations for this page now. This will ensure that
3986 * any allocations necessary to record that reservation occur outside
3989 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3990 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
3992 goto backout_unlocked
;
3994 /* Just decrements count, does not deallocate */
3995 vma_end_reservation(h
, vma
, haddr
);
3998 ptl
= huge_pte_lock(h
, mm
, ptep
);
3999 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4004 if (!huge_pte_none(huge_ptep_get(ptep
)))
4008 ClearPagePrivate(page
);
4009 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4011 page_dup_rmap(page
, true);
4012 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
4013 && (vma
->vm_flags
& VM_SHARED
)));
4014 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
4016 hugetlb_count_add(pages_per_huge_page(h
), mm
);
4017 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4018 /* Optimization, do the COW without a second fault */
4019 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
4025 * Only make newly allocated pages active. Existing pages found
4026 * in the pagecache could be !page_huge_active() if they have been
4027 * isolated for migration.
4030 set_page_huge_active(page
);
4040 restore_reserve_on_error(h
, vma
, haddr
, page
);
4046 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct address_space
*mapping
,
4047 pgoff_t idx
, unsigned long address
)
4049 unsigned long key
[2];
4052 key
[0] = (unsigned long) mapping
;
4055 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
4057 return hash
& (num_fault_mutexes
- 1);
4061 * For uniprocesor systems we always use a single mutex, so just
4062 * return 0 and avoid the hashing overhead.
4064 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct address_space
*mapping
,
4065 pgoff_t idx
, unsigned long address
)
4071 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4072 unsigned long address
, unsigned int flags
)
4079 struct page
*page
= NULL
;
4080 struct page
*pagecache_page
= NULL
;
4081 struct hstate
*h
= hstate_vma(vma
);
4082 struct address_space
*mapping
;
4083 int need_wait_lock
= 0;
4084 unsigned long haddr
= address
& huge_page_mask(h
);
4086 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4088 entry
= huge_ptep_get(ptep
);
4089 if (unlikely(is_hugetlb_entry_migration(entry
))) {
4090 migration_entry_wait_huge(vma
, mm
, ptep
);
4092 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
4093 return VM_FAULT_HWPOISON_LARGE
|
4094 VM_FAULT_SET_HINDEX(hstate_index(h
));
4096 ptep
= huge_pte_alloc(mm
, haddr
, huge_page_size(h
));
4098 return VM_FAULT_OOM
;
4101 mapping
= vma
->vm_file
->f_mapping
;
4102 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4105 * Serialize hugepage allocation and instantiation, so that we don't
4106 * get spurious allocation failures if two CPUs race to instantiate
4107 * the same page in the page cache.
4109 hash
= hugetlb_fault_mutex_hash(h
, mapping
, idx
, haddr
);
4110 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4112 entry
= huge_ptep_get(ptep
);
4113 if (huge_pte_none(entry
)) {
4114 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
4121 * entry could be a migration/hwpoison entry at this point, so this
4122 * check prevents the kernel from going below assuming that we have
4123 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4124 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4127 if (!pte_present(entry
))
4131 * If we are going to COW the mapping later, we examine the pending
4132 * reservations for this page now. This will ensure that any
4133 * allocations necessary to record that reservation occur outside the
4134 * spinlock. For private mappings, we also lookup the pagecache
4135 * page now as it is used to determine if a reservation has been
4138 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
4139 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4143 /* Just decrements count, does not deallocate */
4144 vma_end_reservation(h
, vma
, haddr
);
4146 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4147 pagecache_page
= hugetlbfs_pagecache_page(h
,
4151 ptl
= huge_pte_lock(h
, mm
, ptep
);
4153 /* Check for a racing update before calling hugetlb_cow */
4154 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4158 * hugetlb_cow() requires page locks of pte_page(entry) and
4159 * pagecache_page, so here we need take the former one
4160 * when page != pagecache_page or !pagecache_page.
4162 page
= pte_page(entry
);
4163 if (page
!= pagecache_page
)
4164 if (!trylock_page(page
)) {
4171 if (flags
& FAULT_FLAG_WRITE
) {
4172 if (!huge_pte_write(entry
)) {
4173 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4174 pagecache_page
, ptl
);
4177 entry
= huge_pte_mkdirty(entry
);
4179 entry
= pte_mkyoung(entry
);
4180 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
4181 flags
& FAULT_FLAG_WRITE
))
4182 update_mmu_cache(vma
, haddr
, ptep
);
4184 if (page
!= pagecache_page
)
4190 if (pagecache_page
) {
4191 unlock_page(pagecache_page
);
4192 put_page(pagecache_page
);
4195 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4197 * Generally it's safe to hold refcount during waiting page lock. But
4198 * here we just wait to defer the next page fault to avoid busy loop and
4199 * the page is not used after unlocked before returning from the current
4200 * page fault. So we are safe from accessing freed page, even if we wait
4201 * here without taking refcount.
4204 wait_on_page_locked(page
);
4209 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4210 * modifications for huge pages.
4212 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4214 struct vm_area_struct
*dst_vma
,
4215 unsigned long dst_addr
,
4216 unsigned long src_addr
,
4217 struct page
**pagep
)
4219 struct address_space
*mapping
;
4222 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4223 struct hstate
*h
= hstate_vma(dst_vma
);
4231 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4235 ret
= copy_huge_page_from_user(page
,
4236 (const void __user
*) src_addr
,
4237 pages_per_huge_page(h
), false);
4239 /* fallback to copy_from_user outside mmap_sem */
4240 if (unlikely(ret
)) {
4243 /* don't free the page */
4252 * The memory barrier inside __SetPageUptodate makes sure that
4253 * preceding stores to the page contents become visible before
4254 * the set_pte_at() write.
4256 __SetPageUptodate(page
);
4258 mapping
= dst_vma
->vm_file
->f_mapping
;
4259 idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4262 * If shared, add to page cache
4265 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4268 goto out_release_nounlock
;
4271 * Serialization between remove_inode_hugepages() and
4272 * huge_add_to_page_cache() below happens through the
4273 * hugetlb_fault_mutex_table that here must be hold by
4276 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4278 goto out_release_nounlock
;
4281 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4285 * Recheck the i_size after holding PT lock to make sure not
4286 * to leave any page mapped (as page_mapped()) beyond the end
4287 * of the i_size (remove_inode_hugepages() is strict about
4288 * enforcing that). If we bail out here, we'll also leave a
4289 * page in the radix tree in the vm_shared case beyond the end
4290 * of the i_size, but remove_inode_hugepages() will take care
4291 * of it as soon as we drop the hugetlb_fault_mutex_table.
4293 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4296 goto out_release_unlock
;
4299 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4300 goto out_release_unlock
;
4303 page_dup_rmap(page
, true);
4305 ClearPagePrivate(page
);
4306 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4309 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4310 if (dst_vma
->vm_flags
& VM_WRITE
)
4311 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4312 _dst_pte
= pte_mkyoung(_dst_pte
);
4314 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4316 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4317 dst_vma
->vm_flags
& VM_WRITE
);
4318 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4320 /* No need to invalidate - it was non-present before */
4321 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4324 set_page_huge_active(page
);
4334 out_release_nounlock
:
4339 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4340 struct page
**pages
, struct vm_area_struct
**vmas
,
4341 unsigned long *position
, unsigned long *nr_pages
,
4342 long i
, unsigned int flags
, int *nonblocking
)
4344 unsigned long pfn_offset
;
4345 unsigned long vaddr
= *position
;
4346 unsigned long remainder
= *nr_pages
;
4347 struct hstate
*h
= hstate_vma(vma
);
4350 while (vaddr
< vma
->vm_end
&& remainder
) {
4352 spinlock_t
*ptl
= NULL
;
4357 * If we have a pending SIGKILL, don't keep faulting pages and
4358 * potentially allocating memory.
4360 if (fatal_signal_pending(current
)) {
4366 * Some archs (sparc64, sh*) have multiple pte_ts to
4367 * each hugepage. We have to make sure we get the
4368 * first, for the page indexing below to work.
4370 * Note that page table lock is not held when pte is null.
4372 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4375 ptl
= huge_pte_lock(h
, mm
, pte
);
4376 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4379 * When coredumping, it suits get_dump_page if we just return
4380 * an error where there's an empty slot with no huge pagecache
4381 * to back it. This way, we avoid allocating a hugepage, and
4382 * the sparse dumpfile avoids allocating disk blocks, but its
4383 * huge holes still show up with zeroes where they need to be.
4385 if (absent
&& (flags
& FOLL_DUMP
) &&
4386 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4394 * We need call hugetlb_fault for both hugepages under migration
4395 * (in which case hugetlb_fault waits for the migration,) and
4396 * hwpoisoned hugepages (in which case we need to prevent the
4397 * caller from accessing to them.) In order to do this, we use
4398 * here is_swap_pte instead of is_hugetlb_entry_migration and
4399 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4400 * both cases, and because we can't follow correct pages
4401 * directly from any kind of swap entries.
4403 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4404 ((flags
& FOLL_WRITE
) &&
4405 !huge_pte_write(huge_ptep_get(pte
)))) {
4407 unsigned int fault_flags
= 0;
4411 if (flags
& FOLL_WRITE
)
4412 fault_flags
|= FAULT_FLAG_WRITE
;
4414 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4415 if (flags
& FOLL_NOWAIT
)
4416 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4417 FAULT_FLAG_RETRY_NOWAIT
;
4418 if (flags
& FOLL_TRIED
) {
4419 VM_WARN_ON_ONCE(fault_flags
&
4420 FAULT_FLAG_ALLOW_RETRY
);
4421 fault_flags
|= FAULT_FLAG_TRIED
;
4423 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4424 if (ret
& VM_FAULT_ERROR
) {
4425 err
= vm_fault_to_errno(ret
, flags
);
4429 if (ret
& VM_FAULT_RETRY
) {
4431 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
4435 * VM_FAULT_RETRY must not return an
4436 * error, it will return zero
4439 * No need to update "position" as the
4440 * caller will not check it after
4441 * *nr_pages is set to 0.
4448 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4449 page
= pte_page(huge_ptep_get(pte
));
4452 * Instead of doing 'try_get_page()' below in the same_page
4453 * loop, just check the count once here.
4455 if (unlikely(page_count(page
) <= 0)) {
4465 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4476 if (vaddr
< vma
->vm_end
&& remainder
&&
4477 pfn_offset
< pages_per_huge_page(h
)) {
4479 * We use pfn_offset to avoid touching the pageframes
4480 * of this compound page.
4486 *nr_pages
= remainder
;
4488 * setting position is actually required only if remainder is
4489 * not zero but it's faster not to add a "if (remainder)"
4497 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4499 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4502 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4505 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4506 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4508 struct mm_struct
*mm
= vma
->vm_mm
;
4509 unsigned long start
= address
;
4512 struct hstate
*h
= hstate_vma(vma
);
4513 unsigned long pages
= 0;
4514 bool shared_pmd
= false;
4515 struct mmu_notifier_range range
;
4518 * In the case of shared PMDs, the area to flush could be beyond
4519 * start/end. Set range.start/range.end to cover the maximum possible
4520 * range if PMD sharing is possible.
4522 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
4523 0, vma
, mm
, start
, end
);
4524 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
4526 BUG_ON(address
>= end
);
4527 flush_cache_range(vma
, range
.start
, range
.end
);
4529 mmu_notifier_invalidate_range_start(&range
);
4530 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4531 for (; address
< end
; address
+= huge_page_size(h
)) {
4533 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
4536 ptl
= huge_pte_lock(h
, mm
, ptep
);
4537 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4543 pte
= huge_ptep_get(ptep
);
4544 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4548 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4549 swp_entry_t entry
= pte_to_swp_entry(pte
);
4551 if (is_write_migration_entry(entry
)) {
4554 make_migration_entry_read(&entry
);
4555 newpte
= swp_entry_to_pte(entry
);
4556 set_huge_swap_pte_at(mm
, address
, ptep
,
4557 newpte
, huge_page_size(h
));
4563 if (!huge_pte_none(pte
)) {
4566 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
4567 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
4568 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4569 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
4575 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4576 * may have cleared our pud entry and done put_page on the page table:
4577 * once we release i_mmap_rwsem, another task can do the final put_page
4578 * and that page table be reused and filled with junk. If we actually
4579 * did unshare a page of pmds, flush the range corresponding to the pud.
4582 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
4584 flush_hugetlb_tlb_range(vma
, start
, end
);
4586 * No need to call mmu_notifier_invalidate_range() we are downgrading
4587 * page table protection not changing it to point to a new page.
4589 * See Documentation/vm/mmu_notifier.rst
4591 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4592 mmu_notifier_invalidate_range_end(&range
);
4594 return pages
<< h
->order
;
4597 int hugetlb_reserve_pages(struct inode
*inode
,
4599 struct vm_area_struct
*vma
,
4600 vm_flags_t vm_flags
)
4603 struct hstate
*h
= hstate_inode(inode
);
4604 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4605 struct resv_map
*resv_map
;
4608 /* This should never happen */
4610 VM_WARN(1, "%s called with a negative range\n", __func__
);
4615 * Only apply hugepage reservation if asked. At fault time, an
4616 * attempt will be made for VM_NORESERVE to allocate a page
4617 * without using reserves
4619 if (vm_flags
& VM_NORESERVE
)
4623 * Shared mappings base their reservation on the number of pages that
4624 * are already allocated on behalf of the file. Private mappings need
4625 * to reserve the full area even if read-only as mprotect() may be
4626 * called to make the mapping read-write. Assume !vma is a shm mapping
4628 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4630 * resv_map can not be NULL as hugetlb_reserve_pages is only
4631 * called for inodes for which resv_maps were created (see
4632 * hugetlbfs_get_inode).
4634 resv_map
= inode_resv_map(inode
);
4636 chg
= region_chg(resv_map
, from
, to
);
4639 resv_map
= resv_map_alloc();
4645 set_vma_resv_map(vma
, resv_map
);
4646 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4655 * There must be enough pages in the subpool for the mapping. If
4656 * the subpool has a minimum size, there may be some global
4657 * reservations already in place (gbl_reserve).
4659 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4660 if (gbl_reserve
< 0) {
4666 * Check enough hugepages are available for the reservation.
4667 * Hand the pages back to the subpool if there are not
4669 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4671 /* put back original number of pages, chg */
4672 (void)hugepage_subpool_put_pages(spool
, chg
);
4677 * Account for the reservations made. Shared mappings record regions
4678 * that have reservations as they are shared by multiple VMAs.
4679 * When the last VMA disappears, the region map says how much
4680 * the reservation was and the page cache tells how much of
4681 * the reservation was consumed. Private mappings are per-VMA and
4682 * only the consumed reservations are tracked. When the VMA
4683 * disappears, the original reservation is the VMA size and the
4684 * consumed reservations are stored in the map. Hence, nothing
4685 * else has to be done for private mappings here
4687 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4688 long add
= region_add(resv_map
, from
, to
);
4690 if (unlikely(chg
> add
)) {
4692 * pages in this range were added to the reserve
4693 * map between region_chg and region_add. This
4694 * indicates a race with alloc_huge_page. Adjust
4695 * the subpool and reserve counts modified above
4696 * based on the difference.
4700 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4702 hugetlb_acct_memory(h
, -rsv_adjust
);
4707 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4708 /* Don't call region_abort if region_chg failed */
4710 region_abort(resv_map
, from
, to
);
4711 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4712 kref_put(&resv_map
->refs
, resv_map_release
);
4716 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4719 struct hstate
*h
= hstate_inode(inode
);
4720 struct resv_map
*resv_map
= inode_resv_map(inode
);
4722 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4726 * Since this routine can be called in the evict inode path for all
4727 * hugetlbfs inodes, resv_map could be NULL.
4730 chg
= region_del(resv_map
, start
, end
);
4732 * region_del() can fail in the rare case where a region
4733 * must be split and another region descriptor can not be
4734 * allocated. If end == LONG_MAX, it will not fail.
4740 spin_lock(&inode
->i_lock
);
4741 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4742 spin_unlock(&inode
->i_lock
);
4745 * If the subpool has a minimum size, the number of global
4746 * reservations to be released may be adjusted.
4748 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4749 hugetlb_acct_memory(h
, -gbl_reserve
);
4754 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4755 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4756 struct vm_area_struct
*vma
,
4757 unsigned long addr
, pgoff_t idx
)
4759 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4761 unsigned long sbase
= saddr
& PUD_MASK
;
4762 unsigned long s_end
= sbase
+ PUD_SIZE
;
4764 /* Allow segments to share if only one is marked locked */
4765 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4766 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4769 * match the virtual addresses, permission and the alignment of the
4772 if (pmd_index(addr
) != pmd_index(saddr
) ||
4773 vm_flags
!= svm_flags
||
4774 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4780 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4782 unsigned long base
= addr
& PUD_MASK
;
4783 unsigned long end
= base
+ PUD_SIZE
;
4786 * check on proper vm_flags and page table alignment
4788 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
4794 * Determine if start,end range within vma could be mapped by shared pmd.
4795 * If yes, adjust start and end to cover range associated with possible
4796 * shared pmd mappings.
4798 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
4799 unsigned long *start
, unsigned long *end
)
4801 unsigned long check_addr
= *start
;
4803 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4806 for (check_addr
= *start
; check_addr
< *end
; check_addr
+= PUD_SIZE
) {
4807 unsigned long a_start
= check_addr
& PUD_MASK
;
4808 unsigned long a_end
= a_start
+ PUD_SIZE
;
4811 * If sharing is possible, adjust start/end if necessary.
4813 if (range_in_vma(vma
, a_start
, a_end
)) {
4814 if (a_start
< *start
)
4823 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4824 * and returns the corresponding pte. While this is not necessary for the
4825 * !shared pmd case because we can allocate the pmd later as well, it makes the
4826 * code much cleaner. pmd allocation is essential for the shared case because
4827 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4828 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4829 * bad pmd for sharing.
4831 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4833 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4834 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4835 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4837 struct vm_area_struct
*svma
;
4838 unsigned long saddr
;
4843 if (!vma_shareable(vma
, addr
))
4844 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4846 i_mmap_lock_write(mapping
);
4847 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4851 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4853 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
4854 vma_mmu_pagesize(svma
));
4856 get_page(virt_to_page(spte
));
4865 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4866 if (pud_none(*pud
)) {
4867 pud_populate(mm
, pud
,
4868 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4871 put_page(virt_to_page(spte
));
4875 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4876 i_mmap_unlock_write(mapping
);
4881 * unmap huge page backed by shared pte.
4883 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4884 * indicated by page_count > 1, unmap is achieved by clearing pud and
4885 * decrementing the ref count. If count == 1, the pte page is not shared.
4887 * called with page table lock held.
4889 * returns: 1 successfully unmapped a shared pte page
4890 * 0 the underlying pte page is not shared, or it is the last user
4892 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4894 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4895 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
4896 pud_t
*pud
= pud_offset(p4d
, *addr
);
4898 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4899 if (page_count(virt_to_page(ptep
)) == 1)
4903 put_page(virt_to_page(ptep
));
4905 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4908 #define want_pmd_share() (1)
4909 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4910 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4915 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4920 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
4921 unsigned long *start
, unsigned long *end
)
4924 #define want_pmd_share() (0)
4925 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4927 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4928 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4929 unsigned long addr
, unsigned long sz
)
4936 pgd
= pgd_offset(mm
, addr
);
4937 p4d
= p4d_alloc(mm
, pgd
, addr
);
4940 pud
= pud_alloc(mm
, p4d
, addr
);
4942 if (sz
== PUD_SIZE
) {
4945 BUG_ON(sz
!= PMD_SIZE
);
4946 if (want_pmd_share() && pud_none(*pud
))
4947 pte
= huge_pmd_share(mm
, addr
, pud
);
4949 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4952 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
4958 * huge_pte_offset() - Walk the page table to resolve the hugepage
4959 * entry at address @addr
4961 * Return: Pointer to page table or swap entry (PUD or PMD) for
4962 * address @addr, or NULL if a p*d_none() entry is encountered and the
4963 * size @sz doesn't match the hugepage size at this level of the page
4966 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
4967 unsigned long addr
, unsigned long sz
)
4974 pgd
= pgd_offset(mm
, addr
);
4975 if (!pgd_present(*pgd
))
4977 p4d
= p4d_offset(pgd
, addr
);
4978 if (!p4d_present(*p4d
))
4981 pud
= pud_offset(p4d
, addr
);
4982 if (sz
!= PUD_SIZE
&& pud_none(*pud
))
4984 /* hugepage or swap? */
4985 if (pud_huge(*pud
) || !pud_present(*pud
))
4986 return (pte_t
*)pud
;
4988 pmd
= pmd_offset(pud
, addr
);
4989 if (sz
!= PMD_SIZE
&& pmd_none(*pmd
))
4991 /* hugepage or swap? */
4992 if (pmd_huge(*pmd
) || !pmd_present(*pmd
))
4993 return (pte_t
*)pmd
;
4998 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5001 * These functions are overwritable if your architecture needs its own
5004 struct page
* __weak
5005 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
5008 return ERR_PTR(-EINVAL
);
5011 struct page
* __weak
5012 follow_huge_pd(struct vm_area_struct
*vma
,
5013 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
5015 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5019 struct page
* __weak
5020 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
5021 pmd_t
*pmd
, int flags
)
5023 struct page
*page
= NULL
;
5027 ptl
= pmd_lockptr(mm
, pmd
);
5030 * make sure that the address range covered by this pmd is not
5031 * unmapped from other threads.
5033 if (!pmd_huge(*pmd
))
5035 pte
= huge_ptep_get((pte_t
*)pmd
);
5036 if (pte_present(pte
)) {
5037 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
5038 if (flags
& FOLL_GET
)
5041 if (is_hugetlb_entry_migration(pte
)) {
5043 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
5047 * hwpoisoned entry is treated as no_page_table in
5048 * follow_page_mask().
5056 struct page
* __weak
5057 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
5058 pud_t
*pud
, int flags
)
5060 if (flags
& FOLL_GET
)
5063 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
5066 struct page
* __weak
5067 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
5069 if (flags
& FOLL_GET
)
5072 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
5075 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
5079 VM_BUG_ON_PAGE(!PageHead(page
), page
);
5080 spin_lock(&hugetlb_lock
);
5081 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
5085 clear_page_huge_active(page
);
5086 list_move_tail(&page
->lru
, list
);
5088 spin_unlock(&hugetlb_lock
);
5092 void putback_active_hugepage(struct page
*page
)
5094 VM_BUG_ON_PAGE(!PageHead(page
), page
);
5095 spin_lock(&hugetlb_lock
);
5096 set_page_huge_active(page
);
5097 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
5098 spin_unlock(&hugetlb_lock
);
5102 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
5104 struct hstate
*h
= page_hstate(oldpage
);
5106 hugetlb_cgroup_migrate(oldpage
, newpage
);
5107 set_page_owner_migrate_reason(newpage
, reason
);
5110 * transfer temporary state of the new huge page. This is
5111 * reverse to other transitions because the newpage is going to
5112 * be final while the old one will be freed so it takes over
5113 * the temporary status.
5115 * Also note that we have to transfer the per-node surplus state
5116 * here as well otherwise the global surplus count will not match
5119 if (PageHugeTemporary(newpage
)) {
5120 int old_nid
= page_to_nid(oldpage
);
5121 int new_nid
= page_to_nid(newpage
);
5123 SetPageHugeTemporary(oldpage
);
5124 ClearPageHugeTemporary(newpage
);
5126 spin_lock(&hugetlb_lock
);
5127 if (h
->surplus_huge_pages_node
[old_nid
]) {
5128 h
->surplus_huge_pages_node
[old_nid
]--;
5129 h
->surplus_huge_pages_node
[new_nid
]++;
5131 spin_unlock(&hugetlb_lock
);