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/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly
;
47 unsigned int default_hstate_idx
;
48 struct hstate hstates
[HUGE_MAX_HSTATE
];
51 static struct cma
*hugetlb_cma
[MAX_NUMNODES
];
53 static unsigned long hugetlb_cma_size __initdata
;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
61 __initdata
LIST_HEAD(huge_boot_pages
);
63 /* for command line parsing */
64 static struct hstate
* __initdata parsed_hstate
;
65 static unsigned long __initdata default_hstate_max_huge_pages
;
66 static bool __initdata parsed_valid_hugepagesz
= true;
67 static bool __initdata parsed_default_hugepagesz
;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock
);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes
;
80 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
85 static inline bool subpool_is_free(struct hugepage_subpool
*spool
)
89 if (spool
->max_hpages
!= -1)
90 return spool
->used_hpages
== 0;
91 if (spool
->min_hpages
!= -1)
92 return spool
->rsv_hpages
== spool
->min_hpages
;
97 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
99 spin_unlock(&spool
->lock
);
101 /* If no pages are used, and no other handles to the subpool
102 * remain, give up any reservations based on minimum size and
103 * free the subpool */
104 if (subpool_is_free(spool
)) {
105 if (spool
->min_hpages
!= -1)
106 hugetlb_acct_memory(spool
->hstate
,
112 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
115 struct hugepage_subpool
*spool
;
117 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
121 spin_lock_init(&spool
->lock
);
123 spool
->max_hpages
= max_hpages
;
125 spool
->min_hpages
= min_hpages
;
127 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
131 spool
->rsv_hpages
= min_hpages
;
136 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
138 spin_lock(&spool
->lock
);
139 BUG_ON(!spool
->count
);
141 unlock_or_release_subpool(spool
);
145 * Subpool accounting for allocating and reserving pages.
146 * Return -ENOMEM if there are not enough resources to satisfy the
147 * request. Otherwise, return the number of pages by which the
148 * global pools must be adjusted (upward). The returned value may
149 * only be different than the passed value (delta) in the case where
150 * a subpool minimum size must be maintained.
152 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
160 spin_lock(&spool
->lock
);
162 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
163 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
164 spool
->used_hpages
+= delta
;
171 /* minimum size accounting */
172 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
173 if (delta
> spool
->rsv_hpages
) {
175 * Asking for more reserves than those already taken on
176 * behalf of subpool. Return difference.
178 ret
= delta
- spool
->rsv_hpages
;
179 spool
->rsv_hpages
= 0;
181 ret
= 0; /* reserves already accounted for */
182 spool
->rsv_hpages
-= delta
;
187 spin_unlock(&spool
->lock
);
192 * Subpool accounting for freeing and unreserving pages.
193 * Return the number of global page reservations that must be dropped.
194 * The return value may only be different than the passed value (delta)
195 * in the case where a subpool minimum size must be maintained.
197 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
205 spin_lock(&spool
->lock
);
207 if (spool
->max_hpages
!= -1) /* maximum size accounting */
208 spool
->used_hpages
-= delta
;
210 /* minimum size accounting */
211 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
212 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
215 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
217 spool
->rsv_hpages
+= delta
;
218 if (spool
->rsv_hpages
> spool
->min_hpages
)
219 spool
->rsv_hpages
= spool
->min_hpages
;
223 * If hugetlbfs_put_super couldn't free spool due to an outstanding
224 * quota reference, free it now.
226 unlock_or_release_subpool(spool
);
231 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
233 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
236 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
238 return subpool_inode(file_inode(vma
->vm_file
));
241 /* Helper that removes a struct file_region from the resv_map cache and returns
244 static struct file_region
*
245 get_file_region_entry_from_cache(struct resv_map
*resv
, long from
, long to
)
247 struct file_region
*nrg
= NULL
;
249 VM_BUG_ON(resv
->region_cache_count
<= 0);
251 resv
->region_cache_count
--;
252 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
, link
);
253 list_del(&nrg
->link
);
261 static void copy_hugetlb_cgroup_uncharge_info(struct file_region
*nrg
,
262 struct file_region
*rg
)
264 #ifdef CONFIG_CGROUP_HUGETLB
265 nrg
->reservation_counter
= rg
->reservation_counter
;
272 /* Helper that records hugetlb_cgroup uncharge info. */
273 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup
*h_cg
,
275 struct resv_map
*resv
,
276 struct file_region
*nrg
)
278 #ifdef CONFIG_CGROUP_HUGETLB
280 nrg
->reservation_counter
=
281 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
282 nrg
->css
= &h_cg
->css
;
283 if (!resv
->pages_per_hpage
)
284 resv
->pages_per_hpage
= pages_per_huge_page(h
);
285 /* pages_per_hpage should be the same for all entries in
288 VM_BUG_ON(resv
->pages_per_hpage
!= pages_per_huge_page(h
));
290 nrg
->reservation_counter
= NULL
;
296 static bool has_same_uncharge_info(struct file_region
*rg
,
297 struct file_region
*org
)
299 #ifdef CONFIG_CGROUP_HUGETLB
301 rg
->reservation_counter
== org
->reservation_counter
&&
309 static void coalesce_file_region(struct resv_map
*resv
, struct file_region
*rg
)
311 struct file_region
*nrg
= NULL
, *prg
= NULL
;
313 prg
= list_prev_entry(rg
, link
);
314 if (&prg
->link
!= &resv
->regions
&& prg
->to
== rg
->from
&&
315 has_same_uncharge_info(prg
, rg
)) {
324 nrg
= list_next_entry(rg
, link
);
325 if (&nrg
->link
!= &resv
->regions
&& nrg
->from
== rg
->to
&&
326 has_same_uncharge_info(nrg
, rg
)) {
327 nrg
->from
= rg
->from
;
335 * Must be called with resv->lock held.
337 * Calling this with regions_needed != NULL will count the number of pages
338 * to be added but will not modify the linked list. And regions_needed will
339 * indicate the number of file_regions needed in the cache to carry out to add
340 * the regions for this range.
342 static long add_reservation_in_range(struct resv_map
*resv
, long f
, long t
,
343 struct hugetlb_cgroup
*h_cg
,
344 struct hstate
*h
, long *regions_needed
)
347 struct list_head
*head
= &resv
->regions
;
348 long last_accounted_offset
= f
;
349 struct file_region
*rg
= NULL
, *trg
= NULL
, *nrg
= NULL
;
354 /* In this loop, we essentially handle an entry for the range
355 * [last_accounted_offset, rg->from), at every iteration, with some
358 list_for_each_entry_safe(rg
, trg
, head
, link
) {
359 /* Skip irrelevant regions that start before our range. */
361 /* If this region ends after the last accounted offset,
362 * then we need to update last_accounted_offset.
364 if (rg
->to
> last_accounted_offset
)
365 last_accounted_offset
= rg
->to
;
369 /* When we find a region that starts beyond our range, we've
375 /* Add an entry for last_accounted_offset -> rg->from, and
376 * update last_accounted_offset.
378 if (rg
->from
> last_accounted_offset
) {
379 add
+= rg
->from
- last_accounted_offset
;
380 if (!regions_needed
) {
381 nrg
= get_file_region_entry_from_cache(
382 resv
, last_accounted_offset
, rg
->from
);
383 record_hugetlb_cgroup_uncharge_info(h_cg
, h
,
385 list_add(&nrg
->link
, rg
->link
.prev
);
386 coalesce_file_region(resv
, nrg
);
388 *regions_needed
+= 1;
391 last_accounted_offset
= rg
->to
;
394 /* Handle the case where our range extends beyond
395 * last_accounted_offset.
397 if (last_accounted_offset
< t
) {
398 add
+= t
- last_accounted_offset
;
399 if (!regions_needed
) {
400 nrg
= get_file_region_entry_from_cache(
401 resv
, last_accounted_offset
, t
);
402 record_hugetlb_cgroup_uncharge_info(h_cg
, h
, resv
, nrg
);
403 list_add(&nrg
->link
, rg
->link
.prev
);
404 coalesce_file_region(resv
, nrg
);
406 *regions_needed
+= 1;
413 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
415 static int allocate_file_region_entries(struct resv_map
*resv
,
417 __must_hold(&resv
->lock
)
419 struct list_head allocated_regions
;
420 int to_allocate
= 0, i
= 0;
421 struct file_region
*trg
= NULL
, *rg
= NULL
;
423 VM_BUG_ON(regions_needed
< 0);
425 INIT_LIST_HEAD(&allocated_regions
);
428 * Check for sufficient descriptors in the cache to accommodate
429 * the number of in progress add operations plus regions_needed.
431 * This is a while loop because when we drop the lock, some other call
432 * to region_add or region_del may have consumed some region_entries,
433 * so we keep looping here until we finally have enough entries for
434 * (adds_in_progress + regions_needed).
436 while (resv
->region_cache_count
<
437 (resv
->adds_in_progress
+ regions_needed
)) {
438 to_allocate
= resv
->adds_in_progress
+ regions_needed
-
439 resv
->region_cache_count
;
441 /* At this point, we should have enough entries in the cache
442 * for all the existings adds_in_progress. We should only be
443 * needing to allocate for regions_needed.
445 VM_BUG_ON(resv
->region_cache_count
< resv
->adds_in_progress
);
447 spin_unlock(&resv
->lock
);
448 for (i
= 0; i
< to_allocate
; i
++) {
449 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
452 list_add(&trg
->link
, &allocated_regions
);
455 spin_lock(&resv
->lock
);
457 list_splice(&allocated_regions
, &resv
->region_cache
);
458 resv
->region_cache_count
+= to_allocate
;
464 list_for_each_entry_safe(rg
, trg
, &allocated_regions
, link
) {
472 * Add the huge page range represented by [f, t) to the reserve
473 * map. Regions will be taken from the cache to fill in this range.
474 * Sufficient regions should exist in the cache due to the previous
475 * call to region_chg with the same range, but in some cases the cache will not
476 * have sufficient entries due to races with other code doing region_add or
477 * region_del. The extra needed entries will be allocated.
479 * regions_needed is the out value provided by a previous call to region_chg.
481 * Return the number of new huge pages added to the map. This number is greater
482 * than or equal to zero. If file_region entries needed to be allocated for
483 * this operation and we were not able to allocate, it returns -ENOMEM.
484 * region_add of regions of length 1 never allocate file_regions and cannot
485 * fail; region_chg will always allocate at least 1 entry and a region_add for
486 * 1 page will only require at most 1 entry.
488 static long region_add(struct resv_map
*resv
, long f
, long t
,
489 long in_regions_needed
, struct hstate
*h
,
490 struct hugetlb_cgroup
*h_cg
)
492 long add
= 0, actual_regions_needed
= 0;
494 spin_lock(&resv
->lock
);
497 /* Count how many regions are actually needed to execute this add. */
498 add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
499 &actual_regions_needed
);
502 * Check for sufficient descriptors in the cache to accommodate
503 * this add operation. Note that actual_regions_needed may be greater
504 * than in_regions_needed, as the resv_map may have been modified since
505 * the region_chg call. In this case, we need to make sure that we
506 * allocate extra entries, such that we have enough for all the
507 * existing adds_in_progress, plus the excess needed for this
510 if (actual_regions_needed
> in_regions_needed
&&
511 resv
->region_cache_count
<
512 resv
->adds_in_progress
+
513 (actual_regions_needed
- in_regions_needed
)) {
514 /* region_add operation of range 1 should never need to
515 * allocate file_region entries.
517 VM_BUG_ON(t
- f
<= 1);
519 if (allocate_file_region_entries(
520 resv
, actual_regions_needed
- in_regions_needed
)) {
527 add
= add_reservation_in_range(resv
, f
, t
, h_cg
, h
, NULL
);
529 resv
->adds_in_progress
-= in_regions_needed
;
531 spin_unlock(&resv
->lock
);
537 * Examine the existing reserve map and determine how many
538 * huge pages in the specified range [f, t) are NOT currently
539 * represented. This routine is called before a subsequent
540 * call to region_add that will actually modify the reserve
541 * map to add the specified range [f, t). region_chg does
542 * not change the number of huge pages represented by the
543 * map. A number of new file_region structures is added to the cache as a
544 * placeholder, for the subsequent region_add call to use. At least 1
545 * file_region structure is added.
547 * out_regions_needed is the number of regions added to the
548 * resv->adds_in_progress. This value needs to be provided to a follow up call
549 * to region_add or region_abort for proper accounting.
551 * Returns the number of huge pages that need to be added to the existing
552 * reservation map for the range [f, t). This number is greater or equal to
553 * zero. -ENOMEM is returned if a new file_region structure or cache entry
554 * is needed and can not be allocated.
556 static long region_chg(struct resv_map
*resv
, long f
, long t
,
557 long *out_regions_needed
)
561 spin_lock(&resv
->lock
);
563 /* Count how many hugepages in this range are NOT represented. */
564 chg
= add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
567 if (*out_regions_needed
== 0)
568 *out_regions_needed
= 1;
570 if (allocate_file_region_entries(resv
, *out_regions_needed
))
573 resv
->adds_in_progress
+= *out_regions_needed
;
575 spin_unlock(&resv
->lock
);
580 * Abort the in progress add operation. The adds_in_progress field
581 * of the resv_map keeps track of the operations in progress between
582 * calls to region_chg and region_add. Operations are sometimes
583 * aborted after the call to region_chg. In such cases, region_abort
584 * is called to decrement the adds_in_progress counter. regions_needed
585 * is the value returned by the region_chg call, it is used to decrement
586 * the adds_in_progress counter.
588 * NOTE: The range arguments [f, t) are not needed or used in this
589 * routine. They are kept to make reading the calling code easier as
590 * arguments will match the associated region_chg call.
592 static void region_abort(struct resv_map
*resv
, long f
, long t
,
595 spin_lock(&resv
->lock
);
596 VM_BUG_ON(!resv
->region_cache_count
);
597 resv
->adds_in_progress
-= regions_needed
;
598 spin_unlock(&resv
->lock
);
602 * Delete the specified range [f, t) from the reserve map. If the
603 * t parameter is LONG_MAX, this indicates that ALL regions after f
604 * should be deleted. Locate the regions which intersect [f, t)
605 * and either trim, delete or split the existing regions.
607 * Returns the number of huge pages deleted from the reserve map.
608 * In the normal case, the return value is zero or more. In the
609 * case where a region must be split, a new region descriptor must
610 * be allocated. If the allocation fails, -ENOMEM will be returned.
611 * NOTE: If the parameter t == LONG_MAX, then we will never split
612 * a region and possibly return -ENOMEM. Callers specifying
613 * t == LONG_MAX do not need to check for -ENOMEM error.
615 static long region_del(struct resv_map
*resv
, long f
, long t
)
617 struct list_head
*head
= &resv
->regions
;
618 struct file_region
*rg
, *trg
;
619 struct file_region
*nrg
= NULL
;
623 spin_lock(&resv
->lock
);
624 list_for_each_entry_safe(rg
, trg
, head
, link
) {
626 * Skip regions before the range to be deleted. file_region
627 * ranges are normally of the form [from, to). However, there
628 * may be a "placeholder" entry in the map which is of the form
629 * (from, to) with from == to. Check for placeholder entries
630 * at the beginning of the range to be deleted.
632 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
638 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
640 * Check for an entry in the cache before dropping
641 * lock and attempting allocation.
644 resv
->region_cache_count
> resv
->adds_in_progress
) {
645 nrg
= list_first_entry(&resv
->region_cache
,
648 list_del(&nrg
->link
);
649 resv
->region_cache_count
--;
653 spin_unlock(&resv
->lock
);
654 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
661 hugetlb_cgroup_uncharge_file_region(
664 /* New entry for end of split region */
668 copy_hugetlb_cgroup_uncharge_info(nrg
, rg
);
670 INIT_LIST_HEAD(&nrg
->link
);
672 /* Original entry is trimmed */
675 list_add(&nrg
->link
, &rg
->link
);
680 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
681 del
+= rg
->to
- rg
->from
;
682 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
689 if (f
<= rg
->from
) { /* Trim beginning of region */
690 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
695 } else { /* Trim end of region */
696 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
704 spin_unlock(&resv
->lock
);
710 * A rare out of memory error was encountered which prevented removal of
711 * the reserve map region for a page. The huge page itself was free'ed
712 * and removed from the page cache. This routine will adjust the subpool
713 * usage count, and the global reserve count if needed. By incrementing
714 * these counts, the reserve map entry which could not be deleted will
715 * appear as a "reserved" entry instead of simply dangling with incorrect
718 void hugetlb_fix_reserve_counts(struct inode
*inode
)
720 struct hugepage_subpool
*spool
= subpool_inode(inode
);
723 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
725 struct hstate
*h
= hstate_inode(inode
);
727 hugetlb_acct_memory(h
, 1);
732 * Count and return the number of huge pages in the reserve map
733 * that intersect with the range [f, t).
735 static long region_count(struct resv_map
*resv
, long f
, long t
)
737 struct list_head
*head
= &resv
->regions
;
738 struct file_region
*rg
;
741 spin_lock(&resv
->lock
);
742 /* Locate each segment we overlap with, and count that overlap. */
743 list_for_each_entry(rg
, head
, link
) {
752 seg_from
= max(rg
->from
, f
);
753 seg_to
= min(rg
->to
, t
);
755 chg
+= seg_to
- seg_from
;
757 spin_unlock(&resv
->lock
);
763 * Convert the address within this vma to the page offset within
764 * the mapping, in pagecache page units; huge pages here.
766 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
767 struct vm_area_struct
*vma
, unsigned long address
)
769 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
770 (vma
->vm_pgoff
>> huge_page_order(h
));
773 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
774 unsigned long address
)
776 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
778 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
781 * Return the size of the pages allocated when backing a VMA. In the majority
782 * cases this will be same size as used by the page table entries.
784 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
786 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
787 return vma
->vm_ops
->pagesize(vma
);
790 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
793 * Return the page size being used by the MMU to back a VMA. In the majority
794 * of cases, the page size used by the kernel matches the MMU size. On
795 * architectures where it differs, an architecture-specific 'strong'
796 * version of this symbol is required.
798 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
800 return vma_kernel_pagesize(vma
);
804 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
805 * bits of the reservation map pointer, which are always clear due to
808 #define HPAGE_RESV_OWNER (1UL << 0)
809 #define HPAGE_RESV_UNMAPPED (1UL << 1)
810 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
813 * These helpers are used to track how many pages are reserved for
814 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
815 * is guaranteed to have their future faults succeed.
817 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
818 * the reserve counters are updated with the hugetlb_lock held. It is safe
819 * to reset the VMA at fork() time as it is not in use yet and there is no
820 * chance of the global counters getting corrupted as a result of the values.
822 * The private mapping reservation is represented in a subtly different
823 * manner to a shared mapping. A shared mapping has a region map associated
824 * with the underlying file, this region map represents the backing file
825 * pages which have ever had a reservation assigned which this persists even
826 * after the page is instantiated. A private mapping has a region map
827 * associated with the original mmap which is attached to all VMAs which
828 * reference it, this region map represents those offsets which have consumed
829 * reservation ie. where pages have been instantiated.
831 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
833 return (unsigned long)vma
->vm_private_data
;
836 static void set_vma_private_data(struct vm_area_struct
*vma
,
839 vma
->vm_private_data
= (void *)value
;
843 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map
*resv_map
,
844 struct hugetlb_cgroup
*h_cg
,
847 #ifdef CONFIG_CGROUP_HUGETLB
849 resv_map
->reservation_counter
= NULL
;
850 resv_map
->pages_per_hpage
= 0;
851 resv_map
->css
= NULL
;
853 resv_map
->reservation_counter
=
854 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
855 resv_map
->pages_per_hpage
= pages_per_huge_page(h
);
856 resv_map
->css
= &h_cg
->css
;
861 struct resv_map
*resv_map_alloc(void)
863 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
864 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
866 if (!resv_map
|| !rg
) {
872 kref_init(&resv_map
->refs
);
873 spin_lock_init(&resv_map
->lock
);
874 INIT_LIST_HEAD(&resv_map
->regions
);
876 resv_map
->adds_in_progress
= 0;
878 * Initialize these to 0. On shared mappings, 0's here indicate these
879 * fields don't do cgroup accounting. On private mappings, these will be
880 * re-initialized to the proper values, to indicate that hugetlb cgroup
881 * reservations are to be un-charged from here.
883 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, NULL
, NULL
);
885 INIT_LIST_HEAD(&resv_map
->region_cache
);
886 list_add(&rg
->link
, &resv_map
->region_cache
);
887 resv_map
->region_cache_count
= 1;
892 void resv_map_release(struct kref
*ref
)
894 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
895 struct list_head
*head
= &resv_map
->region_cache
;
896 struct file_region
*rg
, *trg
;
898 /* Clear out any active regions before we release the map. */
899 region_del(resv_map
, 0, LONG_MAX
);
901 /* ... and any entries left in the cache */
902 list_for_each_entry_safe(rg
, trg
, head
, link
) {
907 VM_BUG_ON(resv_map
->adds_in_progress
);
912 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
915 * At inode evict time, i_mapping may not point to the original
916 * address space within the inode. This original address space
917 * contains the pointer to the resv_map. So, always use the
918 * address space embedded within the inode.
919 * The VERY common case is inode->mapping == &inode->i_data but,
920 * this may not be true for device special inodes.
922 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
925 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
927 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
928 if (vma
->vm_flags
& VM_MAYSHARE
) {
929 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
930 struct inode
*inode
= mapping
->host
;
932 return inode_resv_map(inode
);
935 return (struct resv_map
*)(get_vma_private_data(vma
) &
940 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
942 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
943 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
945 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
946 HPAGE_RESV_MASK
) | (unsigned long)map
);
949 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
951 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
952 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
954 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
957 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
959 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
961 return (get_vma_private_data(vma
) & flag
) != 0;
964 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
965 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
967 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
968 if (!(vma
->vm_flags
& VM_MAYSHARE
))
969 vma
->vm_private_data
= (void *)0;
972 /* Returns true if the VMA has associated reserve pages */
973 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
975 if (vma
->vm_flags
& VM_NORESERVE
) {
977 * This address is already reserved by other process(chg == 0),
978 * so, we should decrement reserved count. Without decrementing,
979 * reserve count remains after releasing inode, because this
980 * allocated page will go into page cache and is regarded as
981 * coming from reserved pool in releasing step. Currently, we
982 * don't have any other solution to deal with this situation
983 * properly, so add work-around here.
985 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
991 /* Shared mappings always use reserves */
992 if (vma
->vm_flags
& VM_MAYSHARE
) {
994 * We know VM_NORESERVE is not set. Therefore, there SHOULD
995 * be a region map for all pages. The only situation where
996 * there is no region map is if a hole was punched via
997 * fallocate. In this case, there really are no reserves to
998 * use. This situation is indicated if chg != 0.
1007 * Only the process that called mmap() has reserves for
1010 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1012 * Like the shared case above, a hole punch or truncate
1013 * could have been performed on the private mapping.
1014 * Examine the value of chg to determine if reserves
1015 * actually exist or were previously consumed.
1016 * Very Subtle - The value of chg comes from a previous
1017 * call to vma_needs_reserves(). The reserve map for
1018 * private mappings has different (opposite) semantics
1019 * than that of shared mappings. vma_needs_reserves()
1020 * has already taken this difference in semantics into
1021 * account. Therefore, the meaning of chg is the same
1022 * as in the shared case above. Code could easily be
1023 * combined, but keeping it separate draws attention to
1024 * subtle differences.
1035 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
1037 int nid
= page_to_nid(page
);
1038 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
1039 h
->free_huge_pages
++;
1040 h
->free_huge_pages_node
[nid
]++;
1041 SetHPageFreed(page
);
1044 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
1047 bool nocma
= !!(current
->flags
& PF_MEMALLOC_NOCMA
);
1049 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
) {
1050 if (nocma
&& is_migrate_cma_page(page
))
1053 if (PageHWPoison(page
))
1056 list_move(&page
->lru
, &h
->hugepage_activelist
);
1057 set_page_refcounted(page
);
1058 ClearHPageFreed(page
);
1059 h
->free_huge_pages
--;
1060 h
->free_huge_pages_node
[nid
]--;
1067 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
1070 unsigned int cpuset_mems_cookie
;
1071 struct zonelist
*zonelist
;
1074 int node
= NUMA_NO_NODE
;
1076 zonelist
= node_zonelist(nid
, gfp_mask
);
1079 cpuset_mems_cookie
= read_mems_allowed_begin();
1080 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
1083 if (!cpuset_zone_allowed(zone
, gfp_mask
))
1086 * no need to ask again on the same node. Pool is node rather than
1089 if (zone_to_nid(zone
) == node
)
1091 node
= zone_to_nid(zone
);
1093 page
= dequeue_huge_page_node_exact(h
, node
);
1097 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
1103 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
1104 struct vm_area_struct
*vma
,
1105 unsigned long address
, int avoid_reserve
,
1109 struct mempolicy
*mpol
;
1111 nodemask_t
*nodemask
;
1115 * A child process with MAP_PRIVATE mappings created by their parent
1116 * have no page reserves. This check ensures that reservations are
1117 * not "stolen". The child may still get SIGKILLed
1119 if (!vma_has_reserves(vma
, chg
) &&
1120 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1123 /* If reserves cannot be used, ensure enough pages are in the pool */
1124 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1127 gfp_mask
= htlb_alloc_mask(h
);
1128 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1129 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
1130 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
1131 SetHPageRestoreReserve(page
);
1132 h
->resv_huge_pages
--;
1135 mpol_cond_put(mpol
);
1143 * common helper functions for hstate_next_node_to_{alloc|free}.
1144 * We may have allocated or freed a huge page based on a different
1145 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1146 * be outside of *nodes_allowed. Ensure that we use an allowed
1147 * node for alloc or free.
1149 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1151 nid
= next_node_in(nid
, *nodes_allowed
);
1152 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1157 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1159 if (!node_isset(nid
, *nodes_allowed
))
1160 nid
= next_node_allowed(nid
, nodes_allowed
);
1165 * returns the previously saved node ["this node"] from which to
1166 * allocate a persistent huge page for the pool and advance the
1167 * next node from which to allocate, handling wrap at end of node
1170 static int hstate_next_node_to_alloc(struct hstate
*h
,
1171 nodemask_t
*nodes_allowed
)
1175 VM_BUG_ON(!nodes_allowed
);
1177 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1178 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1184 * helper for free_pool_huge_page() - return the previously saved
1185 * node ["this node"] from which to free a huge page. Advance the
1186 * next node id whether or not we find a free huge page to free so
1187 * that the next attempt to free addresses the next node.
1189 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1193 VM_BUG_ON(!nodes_allowed
);
1195 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1196 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1201 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1202 for (nr_nodes = nodes_weight(*mask); \
1204 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1207 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1208 for (nr_nodes = nodes_weight(*mask); \
1210 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1213 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1214 static void destroy_compound_gigantic_page(struct page
*page
,
1218 int nr_pages
= 1 << order
;
1219 struct page
*p
= page
+ 1;
1221 atomic_set(compound_mapcount_ptr(page
), 0);
1222 atomic_set(compound_pincount_ptr(page
), 0);
1224 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1225 clear_compound_head(p
);
1226 set_page_refcounted(p
);
1229 set_compound_order(page
, 0);
1230 page
[1].compound_nr
= 0;
1231 __ClearPageHead(page
);
1234 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1237 * If the page isn't allocated using the cma allocator,
1238 * cma_release() returns false.
1241 if (cma_release(hugetlb_cma
[page_to_nid(page
)], page
, 1 << order
))
1245 free_contig_range(page_to_pfn(page
), 1 << order
);
1248 #ifdef CONFIG_CONTIG_ALLOC
1249 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1250 int nid
, nodemask_t
*nodemask
)
1252 unsigned long nr_pages
= 1UL << huge_page_order(h
);
1253 if (nid
== NUMA_NO_NODE
)
1254 nid
= numa_mem_id();
1261 if (hugetlb_cma
[nid
]) {
1262 page
= cma_alloc(hugetlb_cma
[nid
], nr_pages
,
1263 huge_page_order(h
), true);
1268 if (!(gfp_mask
& __GFP_THISNODE
)) {
1269 for_each_node_mask(node
, *nodemask
) {
1270 if (node
== nid
|| !hugetlb_cma
[node
])
1273 page
= cma_alloc(hugetlb_cma
[node
], nr_pages
,
1274 huge_page_order(h
), true);
1282 return alloc_contig_pages(nr_pages
, gfp_mask
, nid
, nodemask
);
1285 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1286 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1287 #else /* !CONFIG_CONTIG_ALLOC */
1288 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1289 int nid
, nodemask_t
*nodemask
)
1293 #endif /* CONFIG_CONTIG_ALLOC */
1295 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1296 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1297 int nid
, nodemask_t
*nodemask
)
1301 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1302 static inline void destroy_compound_gigantic_page(struct page
*page
,
1303 unsigned int order
) { }
1306 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1309 struct page
*subpage
= page
;
1311 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1315 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1316 for (i
= 0; i
< pages_per_huge_page(h
);
1317 i
++, subpage
= mem_map_next(subpage
, page
, i
)) {
1318 subpage
->flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1319 1 << PG_referenced
| 1 << PG_dirty
|
1320 1 << PG_active
| 1 << PG_private
|
1323 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1324 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page
), page
);
1325 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1326 set_page_refcounted(page
);
1327 if (hstate_is_gigantic(h
)) {
1329 * Temporarily drop the hugetlb_lock, because
1330 * we might block in free_gigantic_page().
1332 spin_unlock(&hugetlb_lock
);
1333 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1334 free_gigantic_page(page
, huge_page_order(h
));
1335 spin_lock(&hugetlb_lock
);
1337 __free_pages(page
, huge_page_order(h
));
1341 struct hstate
*size_to_hstate(unsigned long size
)
1345 for_each_hstate(h
) {
1346 if (huge_page_size(h
) == size
)
1352 static void __free_huge_page(struct page
*page
)
1355 * Can't pass hstate in here because it is called from the
1356 * compound page destructor.
1358 struct hstate
*h
= page_hstate(page
);
1359 int nid
= page_to_nid(page
);
1360 struct hugepage_subpool
*spool
= hugetlb_page_subpool(page
);
1361 bool restore_reserve
;
1363 VM_BUG_ON_PAGE(page_count(page
), page
);
1364 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1366 hugetlb_set_page_subpool(page
, NULL
);
1367 page
->mapping
= NULL
;
1368 restore_reserve
= HPageRestoreReserve(page
);
1369 ClearHPageRestoreReserve(page
);
1372 * If HPageRestoreReserve was set on page, page allocation consumed a
1373 * reservation. If the page was associated with a subpool, there
1374 * would have been a page reserved in the subpool before allocation
1375 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1376 * reservation, do not call hugepage_subpool_put_pages() as this will
1377 * remove the reserved page from the subpool.
1379 if (!restore_reserve
) {
1381 * A return code of zero implies that the subpool will be
1382 * under its minimum size if the reservation is not restored
1383 * after page is free. Therefore, force restore_reserve
1386 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1387 restore_reserve
= true;
1390 spin_lock(&hugetlb_lock
);
1391 ClearHPageMigratable(page
);
1392 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1393 pages_per_huge_page(h
), page
);
1394 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
1395 pages_per_huge_page(h
), page
);
1396 if (restore_reserve
)
1397 h
->resv_huge_pages
++;
1399 if (HPageTemporary(page
)) {
1400 list_del(&page
->lru
);
1401 ClearHPageTemporary(page
);
1402 update_and_free_page(h
, page
);
1403 } else if (h
->surplus_huge_pages_node
[nid
]) {
1404 /* remove the page from active list */
1405 list_del(&page
->lru
);
1406 update_and_free_page(h
, page
);
1407 h
->surplus_huge_pages
--;
1408 h
->surplus_huge_pages_node
[nid
]--;
1410 arch_clear_hugepage_flags(page
);
1411 enqueue_huge_page(h
, page
);
1413 spin_unlock(&hugetlb_lock
);
1417 * As free_huge_page() can be called from a non-task context, we have
1418 * to defer the actual freeing in a workqueue to prevent potential
1419 * hugetlb_lock deadlock.
1421 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1422 * be freed and frees them one-by-one. As the page->mapping pointer is
1423 * going to be cleared in __free_huge_page() anyway, it is reused as the
1424 * llist_node structure of a lockless linked list of huge pages to be freed.
1426 static LLIST_HEAD(hpage_freelist
);
1428 static void free_hpage_workfn(struct work_struct
*work
)
1430 struct llist_node
*node
;
1433 node
= llist_del_all(&hpage_freelist
);
1436 page
= container_of((struct address_space
**)node
,
1437 struct page
, mapping
);
1439 __free_huge_page(page
);
1442 static DECLARE_WORK(free_hpage_work
, free_hpage_workfn
);
1444 void free_huge_page(struct page
*page
)
1447 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1451 * Only call schedule_work() if hpage_freelist is previously
1452 * empty. Otherwise, schedule_work() had been called but the
1453 * workfn hasn't retrieved the list yet.
1455 if (llist_add((struct llist_node
*)&page
->mapping
,
1457 schedule_work(&free_hpage_work
);
1461 __free_huge_page(page
);
1464 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1466 INIT_LIST_HEAD(&page
->lru
);
1467 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1468 hugetlb_set_page_subpool(page
, NULL
);
1469 set_hugetlb_cgroup(page
, NULL
);
1470 set_hugetlb_cgroup_rsvd(page
, NULL
);
1471 spin_lock(&hugetlb_lock
);
1473 h
->nr_huge_pages_node
[nid
]++;
1474 ClearHPageFreed(page
);
1475 spin_unlock(&hugetlb_lock
);
1478 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1481 int nr_pages
= 1 << order
;
1482 struct page
*p
= page
+ 1;
1484 /* we rely on prep_new_huge_page to set the destructor */
1485 set_compound_order(page
, order
);
1486 __ClearPageReserved(page
);
1487 __SetPageHead(page
);
1488 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1490 * For gigantic hugepages allocated through bootmem at
1491 * boot, it's safer to be consistent with the not-gigantic
1492 * hugepages and clear the PG_reserved bit from all tail pages
1493 * too. Otherwise drivers using get_user_pages() to access tail
1494 * pages may get the reference counting wrong if they see
1495 * PG_reserved set on a tail page (despite the head page not
1496 * having PG_reserved set). Enforcing this consistency between
1497 * head and tail pages allows drivers to optimize away a check
1498 * on the head page when they need know if put_page() is needed
1499 * after get_user_pages().
1501 __ClearPageReserved(p
);
1502 set_page_count(p
, 0);
1503 set_compound_head(p
, page
);
1505 atomic_set(compound_mapcount_ptr(page
), -1);
1506 atomic_set(compound_pincount_ptr(page
), 0);
1510 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1511 * transparent huge pages. See the PageTransHuge() documentation for more
1514 int PageHuge(struct page
*page
)
1516 if (!PageCompound(page
))
1519 page
= compound_head(page
);
1520 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1522 EXPORT_SYMBOL_GPL(PageHuge
);
1525 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1526 * normal or transparent huge pages.
1528 int PageHeadHuge(struct page
*page_head
)
1530 if (!PageHead(page_head
))
1533 return page_head
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1537 * Find and lock address space (mapping) in write mode.
1539 * Upon entry, the page is locked which means that page_mapping() is
1540 * stable. Due to locking order, we can only trylock_write. If we can
1541 * not get the lock, simply return NULL to caller.
1543 struct address_space
*hugetlb_page_mapping_lock_write(struct page
*hpage
)
1545 struct address_space
*mapping
= page_mapping(hpage
);
1550 if (i_mmap_trylock_write(mapping
))
1556 pgoff_t
__basepage_index(struct page
*page
)
1558 struct page
*page_head
= compound_head(page
);
1559 pgoff_t index
= page_index(page_head
);
1560 unsigned long compound_idx
;
1562 if (!PageHuge(page_head
))
1563 return page_index(page
);
1565 if (compound_order(page_head
) >= MAX_ORDER
)
1566 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1568 compound_idx
= page
- page_head
;
1570 return (index
<< compound_order(page_head
)) + compound_idx
;
1573 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1574 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1575 nodemask_t
*node_alloc_noretry
)
1577 int order
= huge_page_order(h
);
1579 bool alloc_try_hard
= true;
1582 * By default we always try hard to allocate the page with
1583 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1584 * a loop (to adjust global huge page counts) and previous allocation
1585 * failed, do not continue to try hard on the same node. Use the
1586 * node_alloc_noretry bitmap to manage this state information.
1588 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1589 alloc_try_hard
= false;
1590 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1592 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1593 if (nid
== NUMA_NO_NODE
)
1594 nid
= numa_mem_id();
1595 page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, nmask
);
1597 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1599 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1602 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1603 * indicates an overall state change. Clear bit so that we resume
1604 * normal 'try hard' allocations.
1606 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1607 node_clear(nid
, *node_alloc_noretry
);
1610 * If we tried hard to get a page but failed, set bit so that
1611 * subsequent attempts will not try as hard until there is an
1612 * overall state change.
1614 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1615 node_set(nid
, *node_alloc_noretry
);
1621 * Common helper to allocate a fresh hugetlb page. All specific allocators
1622 * should use this function to get new hugetlb pages
1624 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1625 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1626 nodemask_t
*node_alloc_noretry
)
1630 if (hstate_is_gigantic(h
))
1631 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1633 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1634 nid
, nmask
, node_alloc_noretry
);
1638 if (hstate_is_gigantic(h
))
1639 prep_compound_gigantic_page(page
, huge_page_order(h
));
1640 prep_new_huge_page(h
, page
, page_to_nid(page
));
1646 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1649 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1650 nodemask_t
*node_alloc_noretry
)
1654 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1656 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1657 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1658 node_alloc_noretry
);
1666 put_page(page
); /* free it into the hugepage allocator */
1672 * Free huge page from pool from next node to free.
1673 * Attempt to keep persistent huge pages more or less
1674 * balanced over allowed nodes.
1675 * Called with hugetlb_lock locked.
1677 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1683 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1685 * If we're returning unused surplus pages, only examine
1686 * nodes with surplus pages.
1688 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1689 !list_empty(&h
->hugepage_freelists
[node
])) {
1691 list_entry(h
->hugepage_freelists
[node
].next
,
1693 list_del(&page
->lru
);
1694 h
->free_huge_pages
--;
1695 h
->free_huge_pages_node
[node
]--;
1697 h
->surplus_huge_pages
--;
1698 h
->surplus_huge_pages_node
[node
]--;
1700 update_and_free_page(h
, page
);
1710 * Dissolve a given free hugepage into free buddy pages. This function does
1711 * nothing for in-use hugepages and non-hugepages.
1712 * This function returns values like below:
1714 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1715 * (allocated or reserved.)
1716 * 0: successfully dissolved free hugepages or the page is not a
1717 * hugepage (considered as already dissolved)
1719 int dissolve_free_huge_page(struct page
*page
)
1724 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1725 if (!PageHuge(page
))
1728 spin_lock(&hugetlb_lock
);
1729 if (!PageHuge(page
)) {
1734 if (!page_count(page
)) {
1735 struct page
*head
= compound_head(page
);
1736 struct hstate
*h
= page_hstate(head
);
1737 int nid
= page_to_nid(head
);
1738 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1742 * We should make sure that the page is already on the free list
1743 * when it is dissolved.
1745 if (unlikely(!HPageFreed(head
))) {
1746 spin_unlock(&hugetlb_lock
);
1750 * Theoretically, we should return -EBUSY when we
1751 * encounter this race. In fact, we have a chance
1752 * to successfully dissolve the page if we do a
1753 * retry. Because the race window is quite small.
1754 * If we seize this opportunity, it is an optimization
1755 * for increasing the success rate of dissolving page.
1761 * Move PageHWPoison flag from head page to the raw error page,
1762 * which makes any subpages rather than the error page reusable.
1764 if (PageHWPoison(head
) && page
!= head
) {
1765 SetPageHWPoison(page
);
1766 ClearPageHWPoison(head
);
1768 list_del(&head
->lru
);
1769 h
->free_huge_pages
--;
1770 h
->free_huge_pages_node
[nid
]--;
1771 h
->max_huge_pages
--;
1772 update_and_free_page(h
, head
);
1776 spin_unlock(&hugetlb_lock
);
1781 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1782 * make specified memory blocks removable from the system.
1783 * Note that this will dissolve a free gigantic hugepage completely, if any
1784 * part of it lies within the given range.
1785 * Also note that if dissolve_free_huge_page() returns with an error, all
1786 * free hugepages that were dissolved before that error are lost.
1788 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1794 if (!hugepages_supported())
1797 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1798 page
= pfn_to_page(pfn
);
1799 rc
= dissolve_free_huge_page(page
);
1808 * Allocates a fresh surplus page from the page allocator.
1810 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1811 int nid
, nodemask_t
*nmask
)
1813 struct page
*page
= NULL
;
1815 if (hstate_is_gigantic(h
))
1818 spin_lock(&hugetlb_lock
);
1819 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
1821 spin_unlock(&hugetlb_lock
);
1823 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1827 spin_lock(&hugetlb_lock
);
1829 * We could have raced with the pool size change.
1830 * Double check that and simply deallocate the new page
1831 * if we would end up overcommiting the surpluses. Abuse
1832 * temporary page to workaround the nasty free_huge_page
1835 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1836 SetHPageTemporary(page
);
1837 spin_unlock(&hugetlb_lock
);
1841 h
->surplus_huge_pages
++;
1842 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
1846 spin_unlock(&hugetlb_lock
);
1851 static struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1852 int nid
, nodemask_t
*nmask
)
1856 if (hstate_is_gigantic(h
))
1859 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1864 * We do not account these pages as surplus because they are only
1865 * temporary and will be released properly on the last reference
1867 SetHPageTemporary(page
);
1873 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1876 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1877 struct vm_area_struct
*vma
, unsigned long addr
)
1880 struct mempolicy
*mpol
;
1881 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1883 nodemask_t
*nodemask
;
1885 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
1886 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
);
1887 mpol_cond_put(mpol
);
1892 /* page migration callback function */
1893 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
1894 nodemask_t
*nmask
, gfp_t gfp_mask
)
1896 spin_lock(&hugetlb_lock
);
1897 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1900 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
1902 spin_unlock(&hugetlb_lock
);
1906 spin_unlock(&hugetlb_lock
);
1908 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
1911 /* mempolicy aware migration callback */
1912 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
1913 unsigned long address
)
1915 struct mempolicy
*mpol
;
1916 nodemask_t
*nodemask
;
1921 gfp_mask
= htlb_alloc_mask(h
);
1922 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1923 page
= alloc_huge_page_nodemask(h
, node
, nodemask
, gfp_mask
);
1924 mpol_cond_put(mpol
);
1930 * Increase the hugetlb pool such that it can accommodate a reservation
1933 static int gather_surplus_pages(struct hstate
*h
, long delta
)
1934 __must_hold(&hugetlb_lock
)
1936 struct list_head surplus_list
;
1937 struct page
*page
, *tmp
;
1940 long needed
, allocated
;
1941 bool alloc_ok
= true;
1943 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1945 h
->resv_huge_pages
+= delta
;
1950 INIT_LIST_HEAD(&surplus_list
);
1954 spin_unlock(&hugetlb_lock
);
1955 for (i
= 0; i
< needed
; i
++) {
1956 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
1957 NUMA_NO_NODE
, NULL
);
1962 list_add(&page
->lru
, &surplus_list
);
1968 * After retaking hugetlb_lock, we need to recalculate 'needed'
1969 * because either resv_huge_pages or free_huge_pages may have changed.
1971 spin_lock(&hugetlb_lock
);
1972 needed
= (h
->resv_huge_pages
+ delta
) -
1973 (h
->free_huge_pages
+ allocated
);
1978 * We were not able to allocate enough pages to
1979 * satisfy the entire reservation so we free what
1980 * we've allocated so far.
1985 * The surplus_list now contains _at_least_ the number of extra pages
1986 * needed to accommodate the reservation. Add the appropriate number
1987 * of pages to the hugetlb pool and free the extras back to the buddy
1988 * allocator. Commit the entire reservation here to prevent another
1989 * process from stealing the pages as they are added to the pool but
1990 * before they are reserved.
1992 needed
+= allocated
;
1993 h
->resv_huge_pages
+= delta
;
1996 /* Free the needed pages to the hugetlb pool */
1997 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
2003 * This page is now managed by the hugetlb allocator and has
2004 * no users -- drop the buddy allocator's reference.
2006 zeroed
= put_page_testzero(page
);
2007 VM_BUG_ON_PAGE(!zeroed
, page
);
2008 enqueue_huge_page(h
, page
);
2011 spin_unlock(&hugetlb_lock
);
2013 /* Free unnecessary surplus pages to the buddy allocator */
2014 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
2016 spin_lock(&hugetlb_lock
);
2022 * This routine has two main purposes:
2023 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2024 * in unused_resv_pages. This corresponds to the prior adjustments made
2025 * to the associated reservation map.
2026 * 2) Free any unused surplus pages that may have been allocated to satisfy
2027 * the reservation. As many as unused_resv_pages may be freed.
2029 * Called with hugetlb_lock held. However, the lock could be dropped (and
2030 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2031 * we must make sure nobody else can claim pages we are in the process of
2032 * freeing. Do this by ensuring resv_huge_page always is greater than the
2033 * number of huge pages we plan to free when dropping the lock.
2035 static void return_unused_surplus_pages(struct hstate
*h
,
2036 unsigned long unused_resv_pages
)
2038 unsigned long nr_pages
;
2040 /* Cannot return gigantic pages currently */
2041 if (hstate_is_gigantic(h
))
2045 * Part (or even all) of the reservation could have been backed
2046 * by pre-allocated pages. Only free surplus pages.
2048 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
2051 * We want to release as many surplus pages as possible, spread
2052 * evenly across all nodes with memory. Iterate across these nodes
2053 * until we can no longer free unreserved surplus pages. This occurs
2054 * when the nodes with surplus pages have no free pages.
2055 * free_pool_huge_page() will balance the freed pages across the
2056 * on-line nodes with memory and will handle the hstate accounting.
2058 * Note that we decrement resv_huge_pages as we free the pages. If
2059 * we drop the lock, resv_huge_pages will still be sufficiently large
2060 * to cover subsequent pages we may free.
2062 while (nr_pages
--) {
2063 h
->resv_huge_pages
--;
2064 unused_resv_pages
--;
2065 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
2067 cond_resched_lock(&hugetlb_lock
);
2071 /* Fully uncommit the reservation */
2072 h
->resv_huge_pages
-= unused_resv_pages
;
2077 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2078 * are used by the huge page allocation routines to manage reservations.
2080 * vma_needs_reservation is called to determine if the huge page at addr
2081 * within the vma has an associated reservation. If a reservation is
2082 * needed, the value 1 is returned. The caller is then responsible for
2083 * managing the global reservation and subpool usage counts. After
2084 * the huge page has been allocated, vma_commit_reservation is called
2085 * to add the page to the reservation map. If the page allocation fails,
2086 * the reservation must be ended instead of committed. vma_end_reservation
2087 * is called in such cases.
2089 * In the normal case, vma_commit_reservation returns the same value
2090 * as the preceding vma_needs_reservation call. The only time this
2091 * is not the case is if a reserve map was changed between calls. It
2092 * is the responsibility of the caller to notice the difference and
2093 * take appropriate action.
2095 * vma_add_reservation is used in error paths where a reservation must
2096 * be restored when a newly allocated huge page must be freed. It is
2097 * to be called after calling vma_needs_reservation to determine if a
2098 * reservation exists.
2100 enum vma_resv_mode
{
2106 static long __vma_reservation_common(struct hstate
*h
,
2107 struct vm_area_struct
*vma
, unsigned long addr
,
2108 enum vma_resv_mode mode
)
2110 struct resv_map
*resv
;
2113 long dummy_out_regions_needed
;
2115 resv
= vma_resv_map(vma
);
2119 idx
= vma_hugecache_offset(h
, vma
, addr
);
2121 case VMA_NEEDS_RESV
:
2122 ret
= region_chg(resv
, idx
, idx
+ 1, &dummy_out_regions_needed
);
2123 /* We assume that vma_reservation_* routines always operate on
2124 * 1 page, and that adding to resv map a 1 page entry can only
2125 * ever require 1 region.
2127 VM_BUG_ON(dummy_out_regions_needed
!= 1);
2129 case VMA_COMMIT_RESV
:
2130 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2131 /* region_add calls of range 1 should never fail. */
2135 region_abort(resv
, idx
, idx
+ 1, 1);
2139 if (vma
->vm_flags
& VM_MAYSHARE
) {
2140 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2141 /* region_add calls of range 1 should never fail. */
2144 region_abort(resv
, idx
, idx
+ 1, 1);
2145 ret
= region_del(resv
, idx
, idx
+ 1);
2152 if (vma
->vm_flags
& VM_MAYSHARE
)
2154 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
2156 * In most cases, reserves always exist for private mappings.
2157 * However, a file associated with mapping could have been
2158 * hole punched or truncated after reserves were consumed.
2159 * As subsequent fault on such a range will not use reserves.
2160 * Subtle - The reserve map for private mappings has the
2161 * opposite meaning than that of shared mappings. If NO
2162 * entry is in the reserve map, it means a reservation exists.
2163 * If an entry exists in the reserve map, it means the
2164 * reservation has already been consumed. As a result, the
2165 * return value of this routine is the opposite of the
2166 * value returned from reserve map manipulation routines above.
2174 return ret
< 0 ? ret
: 0;
2177 static long vma_needs_reservation(struct hstate
*h
,
2178 struct vm_area_struct
*vma
, unsigned long addr
)
2180 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2183 static long vma_commit_reservation(struct hstate
*h
,
2184 struct vm_area_struct
*vma
, unsigned long addr
)
2186 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2189 static void vma_end_reservation(struct hstate
*h
,
2190 struct vm_area_struct
*vma
, unsigned long addr
)
2192 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2195 static long vma_add_reservation(struct hstate
*h
,
2196 struct vm_area_struct
*vma
, unsigned long addr
)
2198 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2202 * This routine is called to restore a reservation on error paths. In the
2203 * specific error paths, a huge page was allocated (via alloc_huge_page)
2204 * and is about to be freed. If a reservation for the page existed,
2205 * alloc_huge_page would have consumed the reservation and set
2206 * HPageRestoreReserve in the newly allocated page. When the page is freed
2207 * via free_huge_page, the global reservation count will be incremented if
2208 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2209 * reserve map. Adjust the reserve map here to be consistent with global
2210 * reserve count adjustments to be made by free_huge_page.
2212 static void restore_reserve_on_error(struct hstate
*h
,
2213 struct vm_area_struct
*vma
, unsigned long address
,
2216 if (unlikely(HPageRestoreReserve(page
))) {
2217 long rc
= vma_needs_reservation(h
, vma
, address
);
2219 if (unlikely(rc
< 0)) {
2221 * Rare out of memory condition in reserve map
2222 * manipulation. Clear HPageRestoreReserve so that
2223 * global reserve count will not be incremented
2224 * by free_huge_page. This will make it appear
2225 * as though the reservation for this page was
2226 * consumed. This may prevent the task from
2227 * faulting in the page at a later time. This
2228 * is better than inconsistent global huge page
2229 * accounting of reserve counts.
2231 ClearHPageRestoreReserve(page
);
2233 rc
= vma_add_reservation(h
, vma
, address
);
2234 if (unlikely(rc
< 0))
2236 * See above comment about rare out of
2239 ClearHPageRestoreReserve(page
);
2241 vma_end_reservation(h
, vma
, address
);
2245 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2246 unsigned long addr
, int avoid_reserve
)
2248 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2249 struct hstate
*h
= hstate_vma(vma
);
2251 long map_chg
, map_commit
;
2254 struct hugetlb_cgroup
*h_cg
;
2255 bool deferred_reserve
;
2257 idx
= hstate_index(h
);
2259 * Examine the region/reserve map to determine if the process
2260 * has a reservation for the page to be allocated. A return
2261 * code of zero indicates a reservation exists (no change).
2263 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2265 return ERR_PTR(-ENOMEM
);
2268 * Processes that did not create the mapping will have no
2269 * reserves as indicated by the region/reserve map. Check
2270 * that the allocation will not exceed the subpool limit.
2271 * Allocations for MAP_NORESERVE mappings also need to be
2272 * checked against any subpool limit.
2274 if (map_chg
|| avoid_reserve
) {
2275 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2277 vma_end_reservation(h
, vma
, addr
);
2278 return ERR_PTR(-ENOSPC
);
2282 * Even though there was no reservation in the region/reserve
2283 * map, there could be reservations associated with the
2284 * subpool that can be used. This would be indicated if the
2285 * return value of hugepage_subpool_get_pages() is zero.
2286 * However, if avoid_reserve is specified we still avoid even
2287 * the subpool reservations.
2293 /* If this allocation is not consuming a reservation, charge it now.
2295 deferred_reserve
= map_chg
|| avoid_reserve
|| !vma_resv_map(vma
);
2296 if (deferred_reserve
) {
2297 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
2298 idx
, pages_per_huge_page(h
), &h_cg
);
2300 goto out_subpool_put
;
2303 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2305 goto out_uncharge_cgroup_reservation
;
2307 spin_lock(&hugetlb_lock
);
2309 * glb_chg is passed to indicate whether or not a page must be taken
2310 * from the global free pool (global change). gbl_chg == 0 indicates
2311 * a reservation exists for the allocation.
2313 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2315 spin_unlock(&hugetlb_lock
);
2316 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2318 goto out_uncharge_cgroup
;
2319 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2320 SetHPageRestoreReserve(page
);
2321 h
->resv_huge_pages
--;
2323 spin_lock(&hugetlb_lock
);
2324 list_add(&page
->lru
, &h
->hugepage_activelist
);
2327 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2328 /* If allocation is not consuming a reservation, also store the
2329 * hugetlb_cgroup pointer on the page.
2331 if (deferred_reserve
) {
2332 hugetlb_cgroup_commit_charge_rsvd(idx
, pages_per_huge_page(h
),
2336 spin_unlock(&hugetlb_lock
);
2338 hugetlb_set_page_subpool(page
, spool
);
2340 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2341 if (unlikely(map_chg
> map_commit
)) {
2343 * The page was added to the reservation map between
2344 * vma_needs_reservation and vma_commit_reservation.
2345 * This indicates a race with hugetlb_reserve_pages.
2346 * Adjust for the subpool count incremented above AND
2347 * in hugetlb_reserve_pages for the same page. Also,
2348 * the reservation count added in hugetlb_reserve_pages
2349 * no longer applies.
2353 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2354 hugetlb_acct_memory(h
, -rsv_adjust
);
2355 if (deferred_reserve
)
2356 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
2357 pages_per_huge_page(h
), page
);
2361 out_uncharge_cgroup
:
2362 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2363 out_uncharge_cgroup_reservation
:
2364 if (deferred_reserve
)
2365 hugetlb_cgroup_uncharge_cgroup_rsvd(idx
, pages_per_huge_page(h
),
2368 if (map_chg
|| avoid_reserve
)
2369 hugepage_subpool_put_pages(spool
, 1);
2370 vma_end_reservation(h
, vma
, addr
);
2371 return ERR_PTR(-ENOSPC
);
2374 int alloc_bootmem_huge_page(struct hstate
*h
)
2375 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2376 int __alloc_bootmem_huge_page(struct hstate
*h
)
2378 struct huge_bootmem_page
*m
;
2381 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2384 addr
= memblock_alloc_try_nid_raw(
2385 huge_page_size(h
), huge_page_size(h
),
2386 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2389 * Use the beginning of the huge page to store the
2390 * huge_bootmem_page struct (until gather_bootmem
2391 * puts them into the mem_map).
2400 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2401 /* Put them into a private list first because mem_map is not up yet */
2402 INIT_LIST_HEAD(&m
->list
);
2403 list_add(&m
->list
, &huge_boot_pages
);
2408 static void __init
prep_compound_huge_page(struct page
*page
,
2411 if (unlikely(order
> (MAX_ORDER
- 1)))
2412 prep_compound_gigantic_page(page
, order
);
2414 prep_compound_page(page
, order
);
2417 /* Put bootmem huge pages into the standard lists after mem_map is up */
2418 static void __init
gather_bootmem_prealloc(void)
2420 struct huge_bootmem_page
*m
;
2422 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2423 struct page
*page
= virt_to_page(m
);
2424 struct hstate
*h
= m
->hstate
;
2426 WARN_ON(page_count(page
) != 1);
2427 prep_compound_huge_page(page
, huge_page_order(h
));
2428 WARN_ON(PageReserved(page
));
2429 prep_new_huge_page(h
, page
, page_to_nid(page
));
2430 put_page(page
); /* free it into the hugepage allocator */
2433 * If we had gigantic hugepages allocated at boot time, we need
2434 * to restore the 'stolen' pages to totalram_pages in order to
2435 * fix confusing memory reports from free(1) and another
2436 * side-effects, like CommitLimit going negative.
2438 if (hstate_is_gigantic(h
))
2439 adjust_managed_page_count(page
, pages_per_huge_page(h
));
2444 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2447 nodemask_t
*node_alloc_noretry
;
2449 if (!hstate_is_gigantic(h
)) {
2451 * Bit mask controlling how hard we retry per-node allocations.
2452 * Ignore errors as lower level routines can deal with
2453 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2454 * time, we are likely in bigger trouble.
2456 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2459 /* allocations done at boot time */
2460 node_alloc_noretry
= NULL
;
2463 /* bit mask controlling how hard we retry per-node allocations */
2464 if (node_alloc_noretry
)
2465 nodes_clear(*node_alloc_noretry
);
2467 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2468 if (hstate_is_gigantic(h
)) {
2469 if (hugetlb_cma_size
) {
2470 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2473 if (!alloc_bootmem_huge_page(h
))
2475 } else if (!alloc_pool_huge_page(h
,
2476 &node_states
[N_MEMORY
],
2477 node_alloc_noretry
))
2481 if (i
< h
->max_huge_pages
) {
2484 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2485 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2486 h
->max_huge_pages
, buf
, i
);
2487 h
->max_huge_pages
= i
;
2490 kfree(node_alloc_noretry
);
2493 static void __init
hugetlb_init_hstates(void)
2497 for_each_hstate(h
) {
2498 if (minimum_order
> huge_page_order(h
))
2499 minimum_order
= huge_page_order(h
);
2501 /* oversize hugepages were init'ed in early boot */
2502 if (!hstate_is_gigantic(h
))
2503 hugetlb_hstate_alloc_pages(h
);
2505 VM_BUG_ON(minimum_order
== UINT_MAX
);
2508 static void __init
report_hugepages(void)
2512 for_each_hstate(h
) {
2515 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2516 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2517 buf
, h
->free_huge_pages
);
2521 #ifdef CONFIG_HIGHMEM
2522 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2523 nodemask_t
*nodes_allowed
)
2527 if (hstate_is_gigantic(h
))
2530 for_each_node_mask(i
, *nodes_allowed
) {
2531 struct page
*page
, *next
;
2532 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2533 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2534 if (count
>= h
->nr_huge_pages
)
2536 if (PageHighMem(page
))
2538 list_del(&page
->lru
);
2539 update_and_free_page(h
, page
);
2540 h
->free_huge_pages
--;
2541 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2546 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2547 nodemask_t
*nodes_allowed
)
2553 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2554 * balanced by operating on them in a round-robin fashion.
2555 * Returns 1 if an adjustment was made.
2557 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2562 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2565 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2566 if (h
->surplus_huge_pages_node
[node
])
2570 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2571 if (h
->surplus_huge_pages_node
[node
] <
2572 h
->nr_huge_pages_node
[node
])
2579 h
->surplus_huge_pages
+= delta
;
2580 h
->surplus_huge_pages_node
[node
] += delta
;
2584 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2585 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
2586 nodemask_t
*nodes_allowed
)
2588 unsigned long min_count
, ret
;
2589 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
2592 * Bit mask controlling how hard we retry per-node allocations.
2593 * If we can not allocate the bit mask, do not attempt to allocate
2594 * the requested huge pages.
2596 if (node_alloc_noretry
)
2597 nodes_clear(*node_alloc_noretry
);
2601 spin_lock(&hugetlb_lock
);
2604 * Check for a node specific request.
2605 * Changing node specific huge page count may require a corresponding
2606 * change to the global count. In any case, the passed node mask
2607 * (nodes_allowed) will restrict alloc/free to the specified node.
2609 if (nid
!= NUMA_NO_NODE
) {
2610 unsigned long old_count
= count
;
2612 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2614 * User may have specified a large count value which caused the
2615 * above calculation to overflow. In this case, they wanted
2616 * to allocate as many huge pages as possible. Set count to
2617 * largest possible value to align with their intention.
2619 if (count
< old_count
)
2624 * Gigantic pages runtime allocation depend on the capability for large
2625 * page range allocation.
2626 * If the system does not provide this feature, return an error when
2627 * the user tries to allocate gigantic pages but let the user free the
2628 * boottime allocated gigantic pages.
2630 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
2631 if (count
> persistent_huge_pages(h
)) {
2632 spin_unlock(&hugetlb_lock
);
2633 NODEMASK_FREE(node_alloc_noretry
);
2636 /* Fall through to decrease pool */
2640 * Increase the pool size
2641 * First take pages out of surplus state. Then make up the
2642 * remaining difference by allocating fresh huge pages.
2644 * We might race with alloc_surplus_huge_page() here and be unable
2645 * to convert a surplus huge page to a normal huge page. That is
2646 * not critical, though, it just means the overall size of the
2647 * pool might be one hugepage larger than it needs to be, but
2648 * within all the constraints specified by the sysctls.
2650 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2651 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2655 while (count
> persistent_huge_pages(h
)) {
2657 * If this allocation races such that we no longer need the
2658 * page, free_huge_page will handle it by freeing the page
2659 * and reducing the surplus.
2661 spin_unlock(&hugetlb_lock
);
2663 /* yield cpu to avoid soft lockup */
2666 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
2667 node_alloc_noretry
);
2668 spin_lock(&hugetlb_lock
);
2672 /* Bail for signals. Probably ctrl-c from user */
2673 if (signal_pending(current
))
2678 * Decrease the pool size
2679 * First return free pages to the buddy allocator (being careful
2680 * to keep enough around to satisfy reservations). Then place
2681 * pages into surplus state as needed so the pool will shrink
2682 * to the desired size as pages become free.
2684 * By placing pages into the surplus state independent of the
2685 * overcommit value, we are allowing the surplus pool size to
2686 * exceed overcommit. There are few sane options here. Since
2687 * alloc_surplus_huge_page() is checking the global counter,
2688 * though, we'll note that we're not allowed to exceed surplus
2689 * and won't grow the pool anywhere else. Not until one of the
2690 * sysctls are changed, or the surplus pages go out of use.
2692 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2693 min_count
= max(count
, min_count
);
2694 try_to_free_low(h
, min_count
, nodes_allowed
);
2695 while (min_count
< persistent_huge_pages(h
)) {
2696 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2698 cond_resched_lock(&hugetlb_lock
);
2700 while (count
< persistent_huge_pages(h
)) {
2701 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2705 h
->max_huge_pages
= persistent_huge_pages(h
);
2706 spin_unlock(&hugetlb_lock
);
2708 NODEMASK_FREE(node_alloc_noretry
);
2713 #define HSTATE_ATTR_RO(_name) \
2714 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2716 #define HSTATE_ATTR(_name) \
2717 static struct kobj_attribute _name##_attr = \
2718 __ATTR(_name, 0644, _name##_show, _name##_store)
2720 static struct kobject
*hugepages_kobj
;
2721 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2723 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2725 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2729 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2730 if (hstate_kobjs
[i
] == kobj
) {
2732 *nidp
= NUMA_NO_NODE
;
2736 return kobj_to_node_hstate(kobj
, nidp
);
2739 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2740 struct kobj_attribute
*attr
, char *buf
)
2743 unsigned long nr_huge_pages
;
2746 h
= kobj_to_hstate(kobj
, &nid
);
2747 if (nid
== NUMA_NO_NODE
)
2748 nr_huge_pages
= h
->nr_huge_pages
;
2750 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2752 return sysfs_emit(buf
, "%lu\n", nr_huge_pages
);
2755 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2756 struct hstate
*h
, int nid
,
2757 unsigned long count
, size_t len
)
2760 nodemask_t nodes_allowed
, *n_mask
;
2762 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
2765 if (nid
== NUMA_NO_NODE
) {
2767 * global hstate attribute
2769 if (!(obey_mempolicy
&&
2770 init_nodemask_of_mempolicy(&nodes_allowed
)))
2771 n_mask
= &node_states
[N_MEMORY
];
2773 n_mask
= &nodes_allowed
;
2776 * Node specific request. count adjustment happens in
2777 * set_max_huge_pages() after acquiring hugetlb_lock.
2779 init_nodemask_of_node(&nodes_allowed
, nid
);
2780 n_mask
= &nodes_allowed
;
2783 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
2785 return err
? err
: len
;
2788 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2789 struct kobject
*kobj
, const char *buf
,
2793 unsigned long count
;
2797 err
= kstrtoul(buf
, 10, &count
);
2801 h
= kobj_to_hstate(kobj
, &nid
);
2802 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2805 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2806 struct kobj_attribute
*attr
, char *buf
)
2808 return nr_hugepages_show_common(kobj
, attr
, buf
);
2811 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2812 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2814 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2816 HSTATE_ATTR(nr_hugepages
);
2821 * hstate attribute for optionally mempolicy-based constraint on persistent
2822 * huge page alloc/free.
2824 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2825 struct kobj_attribute
*attr
,
2828 return nr_hugepages_show_common(kobj
, attr
, buf
);
2831 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2832 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2834 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2836 HSTATE_ATTR(nr_hugepages_mempolicy
);
2840 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2841 struct kobj_attribute
*attr
, char *buf
)
2843 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2844 return sysfs_emit(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2847 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2848 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2851 unsigned long input
;
2852 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2854 if (hstate_is_gigantic(h
))
2857 err
= kstrtoul(buf
, 10, &input
);
2861 spin_lock(&hugetlb_lock
);
2862 h
->nr_overcommit_huge_pages
= input
;
2863 spin_unlock(&hugetlb_lock
);
2867 HSTATE_ATTR(nr_overcommit_hugepages
);
2869 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2870 struct kobj_attribute
*attr
, char *buf
)
2873 unsigned long free_huge_pages
;
2876 h
= kobj_to_hstate(kobj
, &nid
);
2877 if (nid
== NUMA_NO_NODE
)
2878 free_huge_pages
= h
->free_huge_pages
;
2880 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2882 return sysfs_emit(buf
, "%lu\n", free_huge_pages
);
2884 HSTATE_ATTR_RO(free_hugepages
);
2886 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2887 struct kobj_attribute
*attr
, char *buf
)
2889 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2890 return sysfs_emit(buf
, "%lu\n", h
->resv_huge_pages
);
2892 HSTATE_ATTR_RO(resv_hugepages
);
2894 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2895 struct kobj_attribute
*attr
, char *buf
)
2898 unsigned long surplus_huge_pages
;
2901 h
= kobj_to_hstate(kobj
, &nid
);
2902 if (nid
== NUMA_NO_NODE
)
2903 surplus_huge_pages
= h
->surplus_huge_pages
;
2905 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2907 return sysfs_emit(buf
, "%lu\n", surplus_huge_pages
);
2909 HSTATE_ATTR_RO(surplus_hugepages
);
2911 static struct attribute
*hstate_attrs
[] = {
2912 &nr_hugepages_attr
.attr
,
2913 &nr_overcommit_hugepages_attr
.attr
,
2914 &free_hugepages_attr
.attr
,
2915 &resv_hugepages_attr
.attr
,
2916 &surplus_hugepages_attr
.attr
,
2918 &nr_hugepages_mempolicy_attr
.attr
,
2923 static const struct attribute_group hstate_attr_group
= {
2924 .attrs
= hstate_attrs
,
2927 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2928 struct kobject
**hstate_kobjs
,
2929 const struct attribute_group
*hstate_attr_group
)
2932 int hi
= hstate_index(h
);
2934 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2935 if (!hstate_kobjs
[hi
])
2938 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2940 kobject_put(hstate_kobjs
[hi
]);
2941 hstate_kobjs
[hi
] = NULL
;
2947 static void __init
hugetlb_sysfs_init(void)
2952 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2953 if (!hugepages_kobj
)
2956 for_each_hstate(h
) {
2957 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2958 hstate_kobjs
, &hstate_attr_group
);
2960 pr_err("HugeTLB: Unable to add hstate %s", h
->name
);
2967 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2968 * with node devices in node_devices[] using a parallel array. The array
2969 * index of a node device or _hstate == node id.
2970 * This is here to avoid any static dependency of the node device driver, in
2971 * the base kernel, on the hugetlb module.
2973 struct node_hstate
{
2974 struct kobject
*hugepages_kobj
;
2975 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2977 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2980 * A subset of global hstate attributes for node devices
2982 static struct attribute
*per_node_hstate_attrs
[] = {
2983 &nr_hugepages_attr
.attr
,
2984 &free_hugepages_attr
.attr
,
2985 &surplus_hugepages_attr
.attr
,
2989 static const struct attribute_group per_node_hstate_attr_group
= {
2990 .attrs
= per_node_hstate_attrs
,
2994 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2995 * Returns node id via non-NULL nidp.
2997 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3001 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
3002 struct node_hstate
*nhs
= &node_hstates
[nid
];
3004 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
3005 if (nhs
->hstate_kobjs
[i
] == kobj
) {
3017 * Unregister hstate attributes from a single node device.
3018 * No-op if no hstate attributes attached.
3020 static void hugetlb_unregister_node(struct node
*node
)
3023 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3025 if (!nhs
->hugepages_kobj
)
3026 return; /* no hstate attributes */
3028 for_each_hstate(h
) {
3029 int idx
= hstate_index(h
);
3030 if (nhs
->hstate_kobjs
[idx
]) {
3031 kobject_put(nhs
->hstate_kobjs
[idx
]);
3032 nhs
->hstate_kobjs
[idx
] = NULL
;
3036 kobject_put(nhs
->hugepages_kobj
);
3037 nhs
->hugepages_kobj
= NULL
;
3042 * Register hstate attributes for a single node device.
3043 * No-op if attributes already registered.
3045 static void hugetlb_register_node(struct node
*node
)
3048 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3051 if (nhs
->hugepages_kobj
)
3052 return; /* already allocated */
3054 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
3056 if (!nhs
->hugepages_kobj
)
3059 for_each_hstate(h
) {
3060 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
3062 &per_node_hstate_attr_group
);
3064 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3065 h
->name
, node
->dev
.id
);
3066 hugetlb_unregister_node(node
);
3073 * hugetlb init time: register hstate attributes for all registered node
3074 * devices of nodes that have memory. All on-line nodes should have
3075 * registered their associated device by this time.
3077 static void __init
hugetlb_register_all_nodes(void)
3081 for_each_node_state(nid
, N_MEMORY
) {
3082 struct node
*node
= node_devices
[nid
];
3083 if (node
->dev
.id
== nid
)
3084 hugetlb_register_node(node
);
3088 * Let the node device driver know we're here so it can
3089 * [un]register hstate attributes on node hotplug.
3091 register_hugetlbfs_with_node(hugetlb_register_node
,
3092 hugetlb_unregister_node
);
3094 #else /* !CONFIG_NUMA */
3096 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3104 static void hugetlb_register_all_nodes(void) { }
3108 static int __init
hugetlb_init(void)
3112 BUILD_BUG_ON(sizeof_field(struct page
, private) * BITS_PER_BYTE
<
3115 if (!hugepages_supported()) {
3116 if (hugetlb_max_hstate
|| default_hstate_max_huge_pages
)
3117 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3122 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3123 * architectures depend on setup being done here.
3125 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
3126 if (!parsed_default_hugepagesz
) {
3128 * If we did not parse a default huge page size, set
3129 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3130 * number of huge pages for this default size was implicitly
3131 * specified, set that here as well.
3132 * Note that the implicit setting will overwrite an explicit
3133 * setting. A warning will be printed in this case.
3135 default_hstate_idx
= hstate_index(size_to_hstate(HPAGE_SIZE
));
3136 if (default_hstate_max_huge_pages
) {
3137 if (default_hstate
.max_huge_pages
) {
3140 string_get_size(huge_page_size(&default_hstate
),
3141 1, STRING_UNITS_2
, buf
, 32);
3142 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3143 default_hstate
.max_huge_pages
, buf
);
3144 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3145 default_hstate_max_huge_pages
);
3147 default_hstate
.max_huge_pages
=
3148 default_hstate_max_huge_pages
;
3152 hugetlb_cma_check();
3153 hugetlb_init_hstates();
3154 gather_bootmem_prealloc();
3157 hugetlb_sysfs_init();
3158 hugetlb_register_all_nodes();
3159 hugetlb_cgroup_file_init();
3162 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
3164 num_fault_mutexes
= 1;
3166 hugetlb_fault_mutex_table
=
3167 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
3169 BUG_ON(!hugetlb_fault_mutex_table
);
3171 for (i
= 0; i
< num_fault_mutexes
; i
++)
3172 mutex_init(&hugetlb_fault_mutex_table
[i
]);
3175 subsys_initcall(hugetlb_init
);
3177 /* Overwritten by architectures with more huge page sizes */
3178 bool __init
__attribute((weak
)) arch_hugetlb_valid_size(unsigned long size
)
3180 return size
== HPAGE_SIZE
;
3183 void __init
hugetlb_add_hstate(unsigned int order
)
3188 if (size_to_hstate(PAGE_SIZE
<< order
)) {
3191 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
3193 h
= &hstates
[hugetlb_max_hstate
++];
3195 h
->mask
= ~(huge_page_size(h
) - 1);
3196 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
3197 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
3198 INIT_LIST_HEAD(&h
->hugepage_activelist
);
3199 h
->next_nid_to_alloc
= first_memory_node
;
3200 h
->next_nid_to_free
= first_memory_node
;
3201 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
3202 huge_page_size(h
)/1024);
3208 * hugepages command line processing
3209 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3210 * specification. If not, ignore the hugepages value. hugepages can also
3211 * be the first huge page command line option in which case it implicitly
3212 * specifies the number of huge pages for the default size.
3214 static int __init
hugepages_setup(char *s
)
3217 static unsigned long *last_mhp
;
3219 if (!parsed_valid_hugepagesz
) {
3220 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s
);
3221 parsed_valid_hugepagesz
= true;
3226 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3227 * yet, so this hugepages= parameter goes to the "default hstate".
3228 * Otherwise, it goes with the previously parsed hugepagesz or
3229 * default_hugepagesz.
3231 else if (!hugetlb_max_hstate
)
3232 mhp
= &default_hstate_max_huge_pages
;
3234 mhp
= &parsed_hstate
->max_huge_pages
;
3236 if (mhp
== last_mhp
) {
3237 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s
);
3241 if (sscanf(s
, "%lu", mhp
) <= 0)
3245 * Global state is always initialized later in hugetlb_init.
3246 * But we need to allocate >= MAX_ORDER hstates here early to still
3247 * use the bootmem allocator.
3249 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
3250 hugetlb_hstate_alloc_pages(parsed_hstate
);
3256 __setup("hugepages=", hugepages_setup
);
3259 * hugepagesz command line processing
3260 * A specific huge page size can only be specified once with hugepagesz.
3261 * hugepagesz is followed by hugepages on the command line. The global
3262 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3263 * hugepagesz argument was valid.
3265 static int __init
hugepagesz_setup(char *s
)
3270 parsed_valid_hugepagesz
= false;
3271 size
= (unsigned long)memparse(s
, NULL
);
3273 if (!arch_hugetlb_valid_size(size
)) {
3274 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s
);
3278 h
= size_to_hstate(size
);
3281 * hstate for this size already exists. This is normally
3282 * an error, but is allowed if the existing hstate is the
3283 * default hstate. More specifically, it is only allowed if
3284 * the number of huge pages for the default hstate was not
3285 * previously specified.
3287 if (!parsed_default_hugepagesz
|| h
!= &default_hstate
||
3288 default_hstate
.max_huge_pages
) {
3289 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s
);
3294 * No need to call hugetlb_add_hstate() as hstate already
3295 * exists. But, do set parsed_hstate so that a following
3296 * hugepages= parameter will be applied to this hstate.
3299 parsed_valid_hugepagesz
= true;
3303 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3304 parsed_valid_hugepagesz
= true;
3307 __setup("hugepagesz=", hugepagesz_setup
);
3310 * default_hugepagesz command line input
3311 * Only one instance of default_hugepagesz allowed on command line.
3313 static int __init
default_hugepagesz_setup(char *s
)
3317 parsed_valid_hugepagesz
= false;
3318 if (parsed_default_hugepagesz
) {
3319 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s
);
3323 size
= (unsigned long)memparse(s
, NULL
);
3325 if (!arch_hugetlb_valid_size(size
)) {
3326 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s
);
3330 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3331 parsed_valid_hugepagesz
= true;
3332 parsed_default_hugepagesz
= true;
3333 default_hstate_idx
= hstate_index(size_to_hstate(size
));
3336 * The number of default huge pages (for this size) could have been
3337 * specified as the first hugetlb parameter: hugepages=X. If so,
3338 * then default_hstate_max_huge_pages is set. If the default huge
3339 * page size is gigantic (>= MAX_ORDER), then the pages must be
3340 * allocated here from bootmem allocator.
3342 if (default_hstate_max_huge_pages
) {
3343 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
3344 if (hstate_is_gigantic(&default_hstate
))
3345 hugetlb_hstate_alloc_pages(&default_hstate
);
3346 default_hstate_max_huge_pages
= 0;
3351 __setup("default_hugepagesz=", default_hugepagesz_setup
);
3353 static unsigned int allowed_mems_nr(struct hstate
*h
)
3356 unsigned int nr
= 0;
3357 nodemask_t
*mpol_allowed
;
3358 unsigned int *array
= h
->free_huge_pages_node
;
3359 gfp_t gfp_mask
= htlb_alloc_mask(h
);
3361 mpol_allowed
= policy_nodemask_current(gfp_mask
);
3363 for_each_node_mask(node
, cpuset_current_mems_allowed
) {
3364 if (!mpol_allowed
|| node_isset(node
, *mpol_allowed
))
3371 #ifdef CONFIG_SYSCTL
3372 static int proc_hugetlb_doulongvec_minmax(struct ctl_table
*table
, int write
,
3373 void *buffer
, size_t *length
,
3374 loff_t
*ppos
, unsigned long *out
)
3376 struct ctl_table dup_table
;
3379 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3380 * can duplicate the @table and alter the duplicate of it.
3383 dup_table
.data
= out
;
3385 return proc_doulongvec_minmax(&dup_table
, write
, buffer
, length
, ppos
);
3388 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3389 struct ctl_table
*table
, int write
,
3390 void *buffer
, size_t *length
, loff_t
*ppos
)
3392 struct hstate
*h
= &default_hstate
;
3393 unsigned long tmp
= h
->max_huge_pages
;
3396 if (!hugepages_supported())
3399 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3405 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3406 NUMA_NO_NODE
, tmp
, *length
);
3411 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3412 void *buffer
, size_t *length
, loff_t
*ppos
)
3415 return hugetlb_sysctl_handler_common(false, table
, write
,
3416 buffer
, length
, ppos
);
3420 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3421 void *buffer
, size_t *length
, loff_t
*ppos
)
3423 return hugetlb_sysctl_handler_common(true, table
, write
,
3424 buffer
, length
, ppos
);
3426 #endif /* CONFIG_NUMA */
3428 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3429 void *buffer
, size_t *length
, loff_t
*ppos
)
3431 struct hstate
*h
= &default_hstate
;
3435 if (!hugepages_supported())
3438 tmp
= h
->nr_overcommit_huge_pages
;
3440 if (write
&& hstate_is_gigantic(h
))
3443 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3449 spin_lock(&hugetlb_lock
);
3450 h
->nr_overcommit_huge_pages
= tmp
;
3451 spin_unlock(&hugetlb_lock
);
3457 #endif /* CONFIG_SYSCTL */
3459 void hugetlb_report_meminfo(struct seq_file
*m
)
3462 unsigned long total
= 0;
3464 if (!hugepages_supported())
3467 for_each_hstate(h
) {
3468 unsigned long count
= h
->nr_huge_pages
;
3470 total
+= huge_page_size(h
) * count
;
3472 if (h
== &default_hstate
)
3474 "HugePages_Total: %5lu\n"
3475 "HugePages_Free: %5lu\n"
3476 "HugePages_Rsvd: %5lu\n"
3477 "HugePages_Surp: %5lu\n"
3478 "Hugepagesize: %8lu kB\n",
3482 h
->surplus_huge_pages
,
3483 huge_page_size(h
) / SZ_1K
);
3486 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ SZ_1K
);
3489 int hugetlb_report_node_meminfo(char *buf
, int len
, int nid
)
3491 struct hstate
*h
= &default_hstate
;
3493 if (!hugepages_supported())
3496 return sysfs_emit_at(buf
, len
,
3497 "Node %d HugePages_Total: %5u\n"
3498 "Node %d HugePages_Free: %5u\n"
3499 "Node %d HugePages_Surp: %5u\n",
3500 nid
, h
->nr_huge_pages_node
[nid
],
3501 nid
, h
->free_huge_pages_node
[nid
],
3502 nid
, h
->surplus_huge_pages_node
[nid
]);
3505 void hugetlb_show_meminfo(void)
3510 if (!hugepages_supported())
3513 for_each_node_state(nid
, N_MEMORY
)
3515 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3517 h
->nr_huge_pages_node
[nid
],
3518 h
->free_huge_pages_node
[nid
],
3519 h
->surplus_huge_pages_node
[nid
],
3520 huge_page_size(h
) / SZ_1K
);
3523 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3525 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3526 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3529 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3530 unsigned long hugetlb_total_pages(void)
3533 unsigned long nr_total_pages
= 0;
3536 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3537 return nr_total_pages
;
3540 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3547 spin_lock(&hugetlb_lock
);
3549 * When cpuset is configured, it breaks the strict hugetlb page
3550 * reservation as the accounting is done on a global variable. Such
3551 * reservation is completely rubbish in the presence of cpuset because
3552 * the reservation is not checked against page availability for the
3553 * current cpuset. Application can still potentially OOM'ed by kernel
3554 * with lack of free htlb page in cpuset that the task is in.
3555 * Attempt to enforce strict accounting with cpuset is almost
3556 * impossible (or too ugly) because cpuset is too fluid that
3557 * task or memory node can be dynamically moved between cpusets.
3559 * The change of semantics for shared hugetlb mapping with cpuset is
3560 * undesirable. However, in order to preserve some of the semantics,
3561 * we fall back to check against current free page availability as
3562 * a best attempt and hopefully to minimize the impact of changing
3563 * semantics that cpuset has.
3565 * Apart from cpuset, we also have memory policy mechanism that
3566 * also determines from which node the kernel will allocate memory
3567 * in a NUMA system. So similar to cpuset, we also should consider
3568 * the memory policy of the current task. Similar to the description
3572 if (gather_surplus_pages(h
, delta
) < 0)
3575 if (delta
> allowed_mems_nr(h
)) {
3576 return_unused_surplus_pages(h
, delta
);
3583 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3586 spin_unlock(&hugetlb_lock
);
3590 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3592 struct resv_map
*resv
= vma_resv_map(vma
);
3595 * This new VMA should share its siblings reservation map if present.
3596 * The VMA will only ever have a valid reservation map pointer where
3597 * it is being copied for another still existing VMA. As that VMA
3598 * has a reference to the reservation map it cannot disappear until
3599 * after this open call completes. It is therefore safe to take a
3600 * new reference here without additional locking.
3602 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3603 kref_get(&resv
->refs
);
3606 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3608 struct hstate
*h
= hstate_vma(vma
);
3609 struct resv_map
*resv
= vma_resv_map(vma
);
3610 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3611 unsigned long reserve
, start
, end
;
3614 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3617 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3618 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3620 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3621 hugetlb_cgroup_uncharge_counter(resv
, start
, end
);
3624 * Decrement reserve counts. The global reserve count may be
3625 * adjusted if the subpool has a minimum size.
3627 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3628 hugetlb_acct_memory(h
, -gbl_reserve
);
3631 kref_put(&resv
->refs
, resv_map_release
);
3634 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
3636 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
3641 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
3643 return huge_page_size(hstate_vma(vma
));
3647 * We cannot handle pagefaults against hugetlb pages at all. They cause
3648 * handle_mm_fault() to try to instantiate regular-sized pages in the
3649 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3652 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3659 * When a new function is introduced to vm_operations_struct and added
3660 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3661 * This is because under System V memory model, mappings created via
3662 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3663 * their original vm_ops are overwritten with shm_vm_ops.
3665 const struct vm_operations_struct hugetlb_vm_ops
= {
3666 .fault
= hugetlb_vm_op_fault
,
3667 .open
= hugetlb_vm_op_open
,
3668 .close
= hugetlb_vm_op_close
,
3669 .may_split
= hugetlb_vm_op_split
,
3670 .pagesize
= hugetlb_vm_op_pagesize
,
3673 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3679 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3680 vma
->vm_page_prot
)));
3682 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3683 vma
->vm_page_prot
));
3685 entry
= pte_mkyoung(entry
);
3686 entry
= pte_mkhuge(entry
);
3687 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3692 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3693 unsigned long address
, pte_t
*ptep
)
3697 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3698 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3699 update_mmu_cache(vma
, address
, ptep
);
3702 bool is_hugetlb_entry_migration(pte_t pte
)
3706 if (huge_pte_none(pte
) || pte_present(pte
))
3708 swp
= pte_to_swp_entry(pte
);
3709 if (is_migration_entry(swp
))
3715 static bool is_hugetlb_entry_hwpoisoned(pte_t pte
)
3719 if (huge_pte_none(pte
) || pte_present(pte
))
3721 swp
= pte_to_swp_entry(pte
);
3722 if (is_hwpoison_entry(swp
))
3728 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3729 struct vm_area_struct
*vma
)
3731 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
3732 struct page
*ptepage
;
3735 struct hstate
*h
= hstate_vma(vma
);
3736 unsigned long sz
= huge_page_size(h
);
3737 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3738 struct mmu_notifier_range range
;
3741 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3744 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
3747 mmu_notifier_invalidate_range_start(&range
);
3750 * For shared mappings i_mmap_rwsem must be held to call
3751 * huge_pte_alloc, otherwise the returned ptep could go
3752 * away if part of a shared pmd and another thread calls
3755 i_mmap_lock_read(mapping
);
3758 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3759 spinlock_t
*src_ptl
, *dst_ptl
;
3760 src_pte
= huge_pte_offset(src
, addr
, sz
);
3763 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3770 * If the pagetables are shared don't copy or take references.
3771 * dst_pte == src_pte is the common case of src/dest sharing.
3773 * However, src could have 'unshared' and dst shares with
3774 * another vma. If dst_pte !none, this implies sharing.
3775 * Check here before taking page table lock, and once again
3776 * after taking the lock below.
3778 dst_entry
= huge_ptep_get(dst_pte
);
3779 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
3782 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3783 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3784 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3785 entry
= huge_ptep_get(src_pte
);
3786 dst_entry
= huge_ptep_get(dst_pte
);
3787 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
3789 * Skip if src entry none. Also, skip in the
3790 * unlikely case dst entry !none as this implies
3791 * sharing with another vma.
3794 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3795 is_hugetlb_entry_hwpoisoned(entry
))) {
3796 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3798 if (is_write_migration_entry(swp_entry
) && cow
) {
3800 * COW mappings require pages in both
3801 * parent and child to be set to read.
3803 make_migration_entry_read(&swp_entry
);
3804 entry
= swp_entry_to_pte(swp_entry
);
3805 set_huge_swap_pte_at(src
, addr
, src_pte
,
3808 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3812 * No need to notify as we are downgrading page
3813 * table protection not changing it to point
3816 * See Documentation/vm/mmu_notifier.rst
3818 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3820 entry
= huge_ptep_get(src_pte
);
3821 ptepage
= pte_page(entry
);
3823 page_dup_rmap(ptepage
, true);
3824 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3825 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3827 spin_unlock(src_ptl
);
3828 spin_unlock(dst_ptl
);
3832 mmu_notifier_invalidate_range_end(&range
);
3834 i_mmap_unlock_read(mapping
);
3839 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3840 unsigned long start
, unsigned long end
,
3841 struct page
*ref_page
)
3843 struct mm_struct
*mm
= vma
->vm_mm
;
3844 unsigned long address
;
3849 struct hstate
*h
= hstate_vma(vma
);
3850 unsigned long sz
= huge_page_size(h
);
3851 struct mmu_notifier_range range
;
3853 WARN_ON(!is_vm_hugetlb_page(vma
));
3854 BUG_ON(start
& ~huge_page_mask(h
));
3855 BUG_ON(end
& ~huge_page_mask(h
));
3858 * This is a hugetlb vma, all the pte entries should point
3861 tlb_change_page_size(tlb
, sz
);
3862 tlb_start_vma(tlb
, vma
);
3865 * If sharing possible, alert mmu notifiers of worst case.
3867 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
3869 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
3870 mmu_notifier_invalidate_range_start(&range
);
3872 for (; address
< end
; address
+= sz
) {
3873 ptep
= huge_pte_offset(mm
, address
, sz
);
3877 ptl
= huge_pte_lock(h
, mm
, ptep
);
3878 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
3881 * We just unmapped a page of PMDs by clearing a PUD.
3882 * The caller's TLB flush range should cover this area.
3887 pte
= huge_ptep_get(ptep
);
3888 if (huge_pte_none(pte
)) {
3894 * Migrating hugepage or HWPoisoned hugepage is already
3895 * unmapped and its refcount is dropped, so just clear pte here.
3897 if (unlikely(!pte_present(pte
))) {
3898 huge_pte_clear(mm
, address
, ptep
, sz
);
3903 page
= pte_page(pte
);
3905 * If a reference page is supplied, it is because a specific
3906 * page is being unmapped, not a range. Ensure the page we
3907 * are about to unmap is the actual page of interest.
3910 if (page
!= ref_page
) {
3915 * Mark the VMA as having unmapped its page so that
3916 * future faults in this VMA will fail rather than
3917 * looking like data was lost
3919 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3922 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3923 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3924 if (huge_pte_dirty(pte
))
3925 set_page_dirty(page
);
3927 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3928 page_remove_rmap(page
, true);
3931 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3933 * Bail out after unmapping reference page if supplied
3938 mmu_notifier_invalidate_range_end(&range
);
3939 tlb_end_vma(tlb
, vma
);
3942 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3943 struct vm_area_struct
*vma
, unsigned long start
,
3944 unsigned long end
, struct page
*ref_page
)
3946 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3949 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3950 * test will fail on a vma being torn down, and not grab a page table
3951 * on its way out. We're lucky that the flag has such an appropriate
3952 * name, and can in fact be safely cleared here. We could clear it
3953 * before the __unmap_hugepage_range above, but all that's necessary
3954 * is to clear it before releasing the i_mmap_rwsem. This works
3955 * because in the context this is called, the VMA is about to be
3956 * destroyed and the i_mmap_rwsem is held.
3958 vma
->vm_flags
&= ~VM_MAYSHARE
;
3961 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3962 unsigned long end
, struct page
*ref_page
)
3964 struct mmu_gather tlb
;
3966 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
3967 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3968 tlb_finish_mmu(&tlb
);
3972 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3973 * mapping it owns the reserve page for. The intention is to unmap the page
3974 * from other VMAs and let the children be SIGKILLed if they are faulting the
3977 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3978 struct page
*page
, unsigned long address
)
3980 struct hstate
*h
= hstate_vma(vma
);
3981 struct vm_area_struct
*iter_vma
;
3982 struct address_space
*mapping
;
3986 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3987 * from page cache lookup which is in HPAGE_SIZE units.
3989 address
= address
& huge_page_mask(h
);
3990 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3992 mapping
= vma
->vm_file
->f_mapping
;
3995 * Take the mapping lock for the duration of the table walk. As
3996 * this mapping should be shared between all the VMAs,
3997 * __unmap_hugepage_range() is called as the lock is already held
3999 i_mmap_lock_write(mapping
);
4000 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
4001 /* Do not unmap the current VMA */
4002 if (iter_vma
== vma
)
4006 * Shared VMAs have their own reserves and do not affect
4007 * MAP_PRIVATE accounting but it is possible that a shared
4008 * VMA is using the same page so check and skip such VMAs.
4010 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
4014 * Unmap the page from other VMAs without their own reserves.
4015 * They get marked to be SIGKILLed if they fault in these
4016 * areas. This is because a future no-page fault on this VMA
4017 * could insert a zeroed page instead of the data existing
4018 * from the time of fork. This would look like data corruption
4020 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
4021 unmap_hugepage_range(iter_vma
, address
,
4022 address
+ huge_page_size(h
), page
);
4024 i_mmap_unlock_write(mapping
);
4028 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4029 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4030 * cannot race with other handlers or page migration.
4031 * Keep the pte_same checks anyway to make transition from the mutex easier.
4033 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4034 unsigned long address
, pte_t
*ptep
,
4035 struct page
*pagecache_page
, spinlock_t
*ptl
)
4038 struct hstate
*h
= hstate_vma(vma
);
4039 struct page
*old_page
, *new_page
;
4040 int outside_reserve
= 0;
4042 unsigned long haddr
= address
& huge_page_mask(h
);
4043 struct mmu_notifier_range range
;
4045 pte
= huge_ptep_get(ptep
);
4046 old_page
= pte_page(pte
);
4049 /* If no-one else is actually using this page, avoid the copy
4050 * and just make the page writable */
4051 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
4052 page_move_anon_rmap(old_page
, vma
);
4053 set_huge_ptep_writable(vma
, haddr
, ptep
);
4058 * If the process that created a MAP_PRIVATE mapping is about to
4059 * perform a COW due to a shared page count, attempt to satisfy
4060 * the allocation without using the existing reserves. The pagecache
4061 * page is used to determine if the reserve at this address was
4062 * consumed or not. If reserves were used, a partial faulted mapping
4063 * at the time of fork() could consume its reserves on COW instead
4064 * of the full address range.
4066 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
4067 old_page
!= pagecache_page
)
4068 outside_reserve
= 1;
4073 * Drop page table lock as buddy allocator may be called. It will
4074 * be acquired again before returning to the caller, as expected.
4077 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
4079 if (IS_ERR(new_page
)) {
4081 * If a process owning a MAP_PRIVATE mapping fails to COW,
4082 * it is due to references held by a child and an insufficient
4083 * huge page pool. To guarantee the original mappers
4084 * reliability, unmap the page from child processes. The child
4085 * may get SIGKILLed if it later faults.
4087 if (outside_reserve
) {
4088 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4093 BUG_ON(huge_pte_none(pte
));
4095 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4096 * unmapping. unmapping needs to hold i_mmap_rwsem
4097 * in write mode. Dropping i_mmap_rwsem in read mode
4098 * here is OK as COW mappings do not interact with
4101 * Reacquire both after unmap operation.
4103 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4104 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4105 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4106 i_mmap_unlock_read(mapping
);
4108 unmap_ref_private(mm
, vma
, old_page
, haddr
);
4110 i_mmap_lock_read(mapping
);
4111 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4113 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4115 pte_same(huge_ptep_get(ptep
), pte
)))
4116 goto retry_avoidcopy
;
4118 * race occurs while re-acquiring page table
4119 * lock, and our job is done.
4124 ret
= vmf_error(PTR_ERR(new_page
));
4125 goto out_release_old
;
4129 * When the original hugepage is shared one, it does not have
4130 * anon_vma prepared.
4132 if (unlikely(anon_vma_prepare(vma
))) {
4134 goto out_release_all
;
4137 copy_user_huge_page(new_page
, old_page
, address
, vma
,
4138 pages_per_huge_page(h
));
4139 __SetPageUptodate(new_page
);
4141 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
4142 haddr
+ huge_page_size(h
));
4143 mmu_notifier_invalidate_range_start(&range
);
4146 * Retake the page table lock to check for racing updates
4147 * before the page tables are altered
4150 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4151 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
4152 ClearHPageRestoreReserve(new_page
);
4155 huge_ptep_clear_flush(vma
, haddr
, ptep
);
4156 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
4157 set_huge_pte_at(mm
, haddr
, ptep
,
4158 make_huge_pte(vma
, new_page
, 1));
4159 page_remove_rmap(old_page
, true);
4160 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
4161 SetHPageMigratable(new_page
);
4162 /* Make the old page be freed below */
4163 new_page
= old_page
;
4166 mmu_notifier_invalidate_range_end(&range
);
4168 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
4173 spin_lock(ptl
); /* Caller expects lock to be held */
4177 /* Return the pagecache page at a given address within a VMA */
4178 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
4179 struct vm_area_struct
*vma
, unsigned long address
)
4181 struct address_space
*mapping
;
4184 mapping
= vma
->vm_file
->f_mapping
;
4185 idx
= vma_hugecache_offset(h
, vma
, address
);
4187 return find_lock_page(mapping
, idx
);
4191 * Return whether there is a pagecache page to back given address within VMA.
4192 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4194 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
4195 struct vm_area_struct
*vma
, unsigned long address
)
4197 struct address_space
*mapping
;
4201 mapping
= vma
->vm_file
->f_mapping
;
4202 idx
= vma_hugecache_offset(h
, vma
, address
);
4204 page
= find_get_page(mapping
, idx
);
4207 return page
!= NULL
;
4210 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
4213 struct inode
*inode
= mapping
->host
;
4214 struct hstate
*h
= hstate_inode(inode
);
4215 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
4219 ClearHPageRestoreReserve(page
);
4222 * set page dirty so that it will not be removed from cache/file
4223 * by non-hugetlbfs specific code paths.
4225 set_page_dirty(page
);
4227 spin_lock(&inode
->i_lock
);
4228 inode
->i_blocks
+= blocks_per_huge_page(h
);
4229 spin_unlock(&inode
->i_lock
);
4233 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
4234 struct vm_area_struct
*vma
,
4235 struct address_space
*mapping
, pgoff_t idx
,
4236 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
4238 struct hstate
*h
= hstate_vma(vma
);
4239 vm_fault_t ret
= VM_FAULT_SIGBUS
;
4245 unsigned long haddr
= address
& huge_page_mask(h
);
4246 bool new_page
= false;
4249 * Currently, we are forced to kill the process in the event the
4250 * original mapper has unmapped pages from the child due to a failed
4251 * COW. Warn that such a situation has occurred as it may not be obvious
4253 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
4254 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4260 * We can not race with truncation due to holding i_mmap_rwsem.
4261 * i_size is modified when holding i_mmap_rwsem, so check here
4262 * once for faults beyond end of file.
4264 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4269 page
= find_lock_page(mapping
, idx
);
4272 * Check for page in userfault range
4274 if (userfaultfd_missing(vma
)) {
4276 struct vm_fault vmf
= {
4281 * Hard to debug if it ends up being
4282 * used by a callee that assumes
4283 * something about the other
4284 * uninitialized fields... same as in
4290 * hugetlb_fault_mutex and i_mmap_rwsem must be
4291 * dropped before handling userfault. Reacquire
4292 * after handling fault to make calling code simpler.
4294 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4295 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4296 i_mmap_unlock_read(mapping
);
4297 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
4298 i_mmap_lock_read(mapping
);
4299 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4303 page
= alloc_huge_page(vma
, haddr
, 0);
4306 * Returning error will result in faulting task being
4307 * sent SIGBUS. The hugetlb fault mutex prevents two
4308 * tasks from racing to fault in the same page which
4309 * could result in false unable to allocate errors.
4310 * Page migration does not take the fault mutex, but
4311 * does a clear then write of pte's under page table
4312 * lock. Page fault code could race with migration,
4313 * notice the clear pte and try to allocate a page
4314 * here. Before returning error, get ptl and make
4315 * sure there really is no pte entry.
4317 ptl
= huge_pte_lock(h
, mm
, ptep
);
4318 if (!huge_pte_none(huge_ptep_get(ptep
))) {
4324 ret
= vmf_error(PTR_ERR(page
));
4327 clear_huge_page(page
, address
, pages_per_huge_page(h
));
4328 __SetPageUptodate(page
);
4331 if (vma
->vm_flags
& VM_MAYSHARE
) {
4332 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
4341 if (unlikely(anon_vma_prepare(vma
))) {
4343 goto backout_unlocked
;
4349 * If memory error occurs between mmap() and fault, some process
4350 * don't have hwpoisoned swap entry for errored virtual address.
4351 * So we need to block hugepage fault by PG_hwpoison bit check.
4353 if (unlikely(PageHWPoison(page
))) {
4354 ret
= VM_FAULT_HWPOISON_LARGE
|
4355 VM_FAULT_SET_HINDEX(hstate_index(h
));
4356 goto backout_unlocked
;
4361 * If we are going to COW a private mapping later, we examine the
4362 * pending reservations for this page now. This will ensure that
4363 * any allocations necessary to record that reservation occur outside
4366 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4367 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4369 goto backout_unlocked
;
4371 /* Just decrements count, does not deallocate */
4372 vma_end_reservation(h
, vma
, haddr
);
4375 ptl
= huge_pte_lock(h
, mm
, ptep
);
4377 if (!huge_pte_none(huge_ptep_get(ptep
)))
4381 ClearHPageRestoreReserve(page
);
4382 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4384 page_dup_rmap(page
, true);
4385 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
4386 && (vma
->vm_flags
& VM_SHARED
)));
4387 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
4389 hugetlb_count_add(pages_per_huge_page(h
), mm
);
4390 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4391 /* Optimization, do the COW without a second fault */
4392 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
4398 * Only set HPageMigratable in newly allocated pages. Existing pages
4399 * found in the pagecache may not have HPageMigratableset if they have
4400 * been isolated for migration.
4403 SetHPageMigratable(page
);
4413 restore_reserve_on_error(h
, vma
, haddr
, page
);
4419 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4421 unsigned long key
[2];
4424 key
[0] = (unsigned long) mapping
;
4427 hash
= jhash2((u32
*)&key
, sizeof(key
)/(sizeof(u32
)), 0);
4429 return hash
& (num_fault_mutexes
- 1);
4433 * For uniprocessor systems we always use a single mutex, so just
4434 * return 0 and avoid the hashing overhead.
4436 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4442 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4443 unsigned long address
, unsigned int flags
)
4450 struct page
*page
= NULL
;
4451 struct page
*pagecache_page
= NULL
;
4452 struct hstate
*h
= hstate_vma(vma
);
4453 struct address_space
*mapping
;
4454 int need_wait_lock
= 0;
4455 unsigned long haddr
= address
& huge_page_mask(h
);
4457 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4460 * Since we hold no locks, ptep could be stale. That is
4461 * OK as we are only making decisions based on content and
4462 * not actually modifying content here.
4464 entry
= huge_ptep_get(ptep
);
4465 if (unlikely(is_hugetlb_entry_migration(entry
))) {
4466 migration_entry_wait_huge(vma
, mm
, ptep
);
4468 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
4469 return VM_FAULT_HWPOISON_LARGE
|
4470 VM_FAULT_SET_HINDEX(hstate_index(h
));
4474 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4475 * until finished with ptep. This serves two purposes:
4476 * 1) It prevents huge_pmd_unshare from being called elsewhere
4477 * and making the ptep no longer valid.
4478 * 2) It synchronizes us with i_size modifications during truncation.
4480 * ptep could have already be assigned via huge_pte_offset. That
4481 * is OK, as huge_pte_alloc will return the same value unless
4482 * something has changed.
4484 mapping
= vma
->vm_file
->f_mapping
;
4485 i_mmap_lock_read(mapping
);
4486 ptep
= huge_pte_alloc(mm
, haddr
, huge_page_size(h
));
4488 i_mmap_unlock_read(mapping
);
4489 return VM_FAULT_OOM
;
4493 * Serialize hugepage allocation and instantiation, so that we don't
4494 * get spurious allocation failures if two CPUs race to instantiate
4495 * the same page in the page cache.
4497 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4498 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4499 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4501 entry
= huge_ptep_get(ptep
);
4502 if (huge_pte_none(entry
)) {
4503 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
4510 * entry could be a migration/hwpoison entry at this point, so this
4511 * check prevents the kernel from going below assuming that we have
4512 * an active hugepage in pagecache. This goto expects the 2nd page
4513 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4514 * properly handle it.
4516 if (!pte_present(entry
))
4520 * If we are going to COW the mapping later, we examine the pending
4521 * reservations for this page now. This will ensure that any
4522 * allocations necessary to record that reservation occur outside the
4523 * spinlock. For private mappings, we also lookup the pagecache
4524 * page now as it is used to determine if a reservation has been
4527 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
4528 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4532 /* Just decrements count, does not deallocate */
4533 vma_end_reservation(h
, vma
, haddr
);
4535 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4536 pagecache_page
= hugetlbfs_pagecache_page(h
,
4540 ptl
= huge_pte_lock(h
, mm
, ptep
);
4542 /* Check for a racing update before calling hugetlb_cow */
4543 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4547 * hugetlb_cow() requires page locks of pte_page(entry) and
4548 * pagecache_page, so here we need take the former one
4549 * when page != pagecache_page or !pagecache_page.
4551 page
= pte_page(entry
);
4552 if (page
!= pagecache_page
)
4553 if (!trylock_page(page
)) {
4560 if (flags
& FAULT_FLAG_WRITE
) {
4561 if (!huge_pte_write(entry
)) {
4562 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4563 pagecache_page
, ptl
);
4566 entry
= huge_pte_mkdirty(entry
);
4568 entry
= pte_mkyoung(entry
);
4569 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
4570 flags
& FAULT_FLAG_WRITE
))
4571 update_mmu_cache(vma
, haddr
, ptep
);
4573 if (page
!= pagecache_page
)
4579 if (pagecache_page
) {
4580 unlock_page(pagecache_page
);
4581 put_page(pagecache_page
);
4584 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4585 i_mmap_unlock_read(mapping
);
4587 * Generally it's safe to hold refcount during waiting page lock. But
4588 * here we just wait to defer the next page fault to avoid busy loop and
4589 * the page is not used after unlocked before returning from the current
4590 * page fault. So we are safe from accessing freed page, even if we wait
4591 * here without taking refcount.
4594 wait_on_page_locked(page
);
4599 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4600 * modifications for huge pages.
4602 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4604 struct vm_area_struct
*dst_vma
,
4605 unsigned long dst_addr
,
4606 unsigned long src_addr
,
4607 struct page
**pagep
)
4609 struct address_space
*mapping
;
4612 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4613 struct hstate
*h
= hstate_vma(dst_vma
);
4621 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4625 ret
= copy_huge_page_from_user(page
,
4626 (const void __user
*) src_addr
,
4627 pages_per_huge_page(h
), false);
4629 /* fallback to copy_from_user outside mmap_lock */
4630 if (unlikely(ret
)) {
4633 /* don't free the page */
4642 * The memory barrier inside __SetPageUptodate makes sure that
4643 * preceding stores to the page contents become visible before
4644 * the set_pte_at() write.
4646 __SetPageUptodate(page
);
4648 mapping
= dst_vma
->vm_file
->f_mapping
;
4649 idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4652 * If shared, add to page cache
4655 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4658 goto out_release_nounlock
;
4661 * Serialization between remove_inode_hugepages() and
4662 * huge_add_to_page_cache() below happens through the
4663 * hugetlb_fault_mutex_table that here must be hold by
4666 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4668 goto out_release_nounlock
;
4671 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4675 * Recheck the i_size after holding PT lock to make sure not
4676 * to leave any page mapped (as page_mapped()) beyond the end
4677 * of the i_size (remove_inode_hugepages() is strict about
4678 * enforcing that). If we bail out here, we'll also leave a
4679 * page in the radix tree in the vm_shared case beyond the end
4680 * of the i_size, but remove_inode_hugepages() will take care
4681 * of it as soon as we drop the hugetlb_fault_mutex_table.
4683 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4686 goto out_release_unlock
;
4689 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4690 goto out_release_unlock
;
4693 page_dup_rmap(page
, true);
4695 ClearHPageRestoreReserve(page
);
4696 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4699 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4700 if (dst_vma
->vm_flags
& VM_WRITE
)
4701 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4702 _dst_pte
= pte_mkyoung(_dst_pte
);
4704 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4706 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4707 dst_vma
->vm_flags
& VM_WRITE
);
4708 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4710 /* No need to invalidate - it was non-present before */
4711 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4714 SetHPageMigratable(page
);
4724 out_release_nounlock
:
4729 static void record_subpages_vmas(struct page
*page
, struct vm_area_struct
*vma
,
4730 int refs
, struct page
**pages
,
4731 struct vm_area_struct
**vmas
)
4735 for (nr
= 0; nr
< refs
; nr
++) {
4737 pages
[nr
] = mem_map_offset(page
, nr
);
4743 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4744 struct page
**pages
, struct vm_area_struct
**vmas
,
4745 unsigned long *position
, unsigned long *nr_pages
,
4746 long i
, unsigned int flags
, int *locked
)
4748 unsigned long pfn_offset
;
4749 unsigned long vaddr
= *position
;
4750 unsigned long remainder
= *nr_pages
;
4751 struct hstate
*h
= hstate_vma(vma
);
4752 int err
= -EFAULT
, refs
;
4754 while (vaddr
< vma
->vm_end
&& remainder
) {
4756 spinlock_t
*ptl
= NULL
;
4761 * If we have a pending SIGKILL, don't keep faulting pages and
4762 * potentially allocating memory.
4764 if (fatal_signal_pending(current
)) {
4770 * Some archs (sparc64, sh*) have multiple pte_ts to
4771 * each hugepage. We have to make sure we get the
4772 * first, for the page indexing below to work.
4774 * Note that page table lock is not held when pte is null.
4776 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4779 ptl
= huge_pte_lock(h
, mm
, pte
);
4780 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4783 * When coredumping, it suits get_dump_page if we just return
4784 * an error where there's an empty slot with no huge pagecache
4785 * to back it. This way, we avoid allocating a hugepage, and
4786 * the sparse dumpfile avoids allocating disk blocks, but its
4787 * huge holes still show up with zeroes where they need to be.
4789 if (absent
&& (flags
& FOLL_DUMP
) &&
4790 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4798 * We need call hugetlb_fault for both hugepages under migration
4799 * (in which case hugetlb_fault waits for the migration,) and
4800 * hwpoisoned hugepages (in which case we need to prevent the
4801 * caller from accessing to them.) In order to do this, we use
4802 * here is_swap_pte instead of is_hugetlb_entry_migration and
4803 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4804 * both cases, and because we can't follow correct pages
4805 * directly from any kind of swap entries.
4807 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4808 ((flags
& FOLL_WRITE
) &&
4809 !huge_pte_write(huge_ptep_get(pte
)))) {
4811 unsigned int fault_flags
= 0;
4815 if (flags
& FOLL_WRITE
)
4816 fault_flags
|= FAULT_FLAG_WRITE
;
4818 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4819 FAULT_FLAG_KILLABLE
;
4820 if (flags
& FOLL_NOWAIT
)
4821 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4822 FAULT_FLAG_RETRY_NOWAIT
;
4823 if (flags
& FOLL_TRIED
) {
4825 * Note: FAULT_FLAG_ALLOW_RETRY and
4826 * FAULT_FLAG_TRIED can co-exist
4828 fault_flags
|= FAULT_FLAG_TRIED
;
4830 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4831 if (ret
& VM_FAULT_ERROR
) {
4832 err
= vm_fault_to_errno(ret
, flags
);
4836 if (ret
& VM_FAULT_RETRY
) {
4838 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
4842 * VM_FAULT_RETRY must not return an
4843 * error, it will return zero
4846 * No need to update "position" as the
4847 * caller will not check it after
4848 * *nr_pages is set to 0.
4855 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4856 page
= pte_page(huge_ptep_get(pte
));
4859 * If subpage information not requested, update counters
4860 * and skip the same_page loop below.
4862 if (!pages
&& !vmas
&& !pfn_offset
&&
4863 (vaddr
+ huge_page_size(h
) < vma
->vm_end
) &&
4864 (remainder
>= pages_per_huge_page(h
))) {
4865 vaddr
+= huge_page_size(h
);
4866 remainder
-= pages_per_huge_page(h
);
4867 i
+= pages_per_huge_page(h
);
4872 refs
= min3(pages_per_huge_page(h
) - pfn_offset
,
4873 (vma
->vm_end
- vaddr
) >> PAGE_SHIFT
, remainder
);
4876 record_subpages_vmas(mem_map_offset(page
, pfn_offset
),
4878 likely(pages
) ? pages
+ i
: NULL
,
4879 vmas
? vmas
+ i
: NULL
);
4883 * try_grab_compound_head() should always succeed here,
4884 * because: a) we hold the ptl lock, and b) we've just
4885 * checked that the huge page is present in the page
4886 * tables. If the huge page is present, then the tail
4887 * pages must also be present. The ptl prevents the
4888 * head page and tail pages from being rearranged in
4889 * any way. So this page must be available at this
4890 * point, unless the page refcount overflowed:
4892 if (WARN_ON_ONCE(!try_grab_compound_head(pages
[i
],
4902 vaddr
+= (refs
<< PAGE_SHIFT
);
4908 *nr_pages
= remainder
;
4910 * setting position is actually required only if remainder is
4911 * not zero but it's faster not to add a "if (remainder)"
4919 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4921 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4924 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4927 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4928 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4930 struct mm_struct
*mm
= vma
->vm_mm
;
4931 unsigned long start
= address
;
4934 struct hstate
*h
= hstate_vma(vma
);
4935 unsigned long pages
= 0;
4936 bool shared_pmd
= false;
4937 struct mmu_notifier_range range
;
4940 * In the case of shared PMDs, the area to flush could be beyond
4941 * start/end. Set range.start/range.end to cover the maximum possible
4942 * range if PMD sharing is possible.
4944 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
4945 0, vma
, mm
, start
, end
);
4946 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
4948 BUG_ON(address
>= end
);
4949 flush_cache_range(vma
, range
.start
, range
.end
);
4951 mmu_notifier_invalidate_range_start(&range
);
4952 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4953 for (; address
< end
; address
+= huge_page_size(h
)) {
4955 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
4958 ptl
= huge_pte_lock(h
, mm
, ptep
);
4959 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
4965 pte
= huge_ptep_get(ptep
);
4966 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4970 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4971 swp_entry_t entry
= pte_to_swp_entry(pte
);
4973 if (is_write_migration_entry(entry
)) {
4976 make_migration_entry_read(&entry
);
4977 newpte
= swp_entry_to_pte(entry
);
4978 set_huge_swap_pte_at(mm
, address
, ptep
,
4979 newpte
, huge_page_size(h
));
4985 if (!huge_pte_none(pte
)) {
4988 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
4989 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
4990 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4991 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
4997 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4998 * may have cleared our pud entry and done put_page on the page table:
4999 * once we release i_mmap_rwsem, another task can do the final put_page
5000 * and that page table be reused and filled with junk. If we actually
5001 * did unshare a page of pmds, flush the range corresponding to the pud.
5004 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
5006 flush_hugetlb_tlb_range(vma
, start
, end
);
5008 * No need to call mmu_notifier_invalidate_range() we are downgrading
5009 * page table protection not changing it to point to a new page.
5011 * See Documentation/vm/mmu_notifier.rst
5013 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
5014 mmu_notifier_invalidate_range_end(&range
);
5016 return pages
<< h
->order
;
5019 /* Return true if reservation was successful, false otherwise. */
5020 bool hugetlb_reserve_pages(struct inode
*inode
,
5022 struct vm_area_struct
*vma
,
5023 vm_flags_t vm_flags
)
5026 struct hstate
*h
= hstate_inode(inode
);
5027 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5028 struct resv_map
*resv_map
;
5029 struct hugetlb_cgroup
*h_cg
= NULL
;
5030 long gbl_reserve
, regions_needed
= 0;
5032 /* This should never happen */
5034 VM_WARN(1, "%s called with a negative range\n", __func__
);
5039 * Only apply hugepage reservation if asked. At fault time, an
5040 * attempt will be made for VM_NORESERVE to allocate a page
5041 * without using reserves
5043 if (vm_flags
& VM_NORESERVE
)
5047 * Shared mappings base their reservation on the number of pages that
5048 * are already allocated on behalf of the file. Private mappings need
5049 * to reserve the full area even if read-only as mprotect() may be
5050 * called to make the mapping read-write. Assume !vma is a shm mapping
5052 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5054 * resv_map can not be NULL as hugetlb_reserve_pages is only
5055 * called for inodes for which resv_maps were created (see
5056 * hugetlbfs_get_inode).
5058 resv_map
= inode_resv_map(inode
);
5060 chg
= region_chg(resv_map
, from
, to
, ®ions_needed
);
5063 /* Private mapping. */
5064 resv_map
= resv_map_alloc();
5070 set_vma_resv_map(vma
, resv_map
);
5071 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
5077 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h
),
5078 chg
* pages_per_huge_page(h
), &h_cg
) < 0)
5081 if (vma
&& !(vma
->vm_flags
& VM_MAYSHARE
) && h_cg
) {
5082 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5085 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, h_cg
, h
);
5089 * There must be enough pages in the subpool for the mapping. If
5090 * the subpool has a minimum size, there may be some global
5091 * reservations already in place (gbl_reserve).
5093 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
5094 if (gbl_reserve
< 0)
5095 goto out_uncharge_cgroup
;
5098 * Check enough hugepages are available for the reservation.
5099 * Hand the pages back to the subpool if there are not
5101 if (hugetlb_acct_memory(h
, gbl_reserve
) < 0)
5105 * Account for the reservations made. Shared mappings record regions
5106 * that have reservations as they are shared by multiple VMAs.
5107 * When the last VMA disappears, the region map says how much
5108 * the reservation was and the page cache tells how much of
5109 * the reservation was consumed. Private mappings are per-VMA and
5110 * only the consumed reservations are tracked. When the VMA
5111 * disappears, the original reservation is the VMA size and the
5112 * consumed reservations are stored in the map. Hence, nothing
5113 * else has to be done for private mappings here
5115 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5116 add
= region_add(resv_map
, from
, to
, regions_needed
, h
, h_cg
);
5118 if (unlikely(add
< 0)) {
5119 hugetlb_acct_memory(h
, -gbl_reserve
);
5121 } else if (unlikely(chg
> add
)) {
5123 * pages in this range were added to the reserve
5124 * map between region_chg and region_add. This
5125 * indicates a race with alloc_huge_page. Adjust
5126 * the subpool and reserve counts modified above
5127 * based on the difference.
5131 hugetlb_cgroup_uncharge_cgroup_rsvd(
5133 (chg
- add
) * pages_per_huge_page(h
), h_cg
);
5135 rsv_adjust
= hugepage_subpool_put_pages(spool
,
5137 hugetlb_acct_memory(h
, -rsv_adjust
);
5143 /* put back original number of pages, chg */
5144 (void)hugepage_subpool_put_pages(spool
, chg
);
5145 out_uncharge_cgroup
:
5146 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h
),
5147 chg
* pages_per_huge_page(h
), h_cg
);
5149 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
5150 /* Only call region_abort if the region_chg succeeded but the
5151 * region_add failed or didn't run.
5153 if (chg
>= 0 && add
< 0)
5154 region_abort(resv_map
, from
, to
, regions_needed
);
5155 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
5156 kref_put(&resv_map
->refs
, resv_map_release
);
5160 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
5163 struct hstate
*h
= hstate_inode(inode
);
5164 struct resv_map
*resv_map
= inode_resv_map(inode
);
5166 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5170 * Since this routine can be called in the evict inode path for all
5171 * hugetlbfs inodes, resv_map could be NULL.
5174 chg
= region_del(resv_map
, start
, end
);
5176 * region_del() can fail in the rare case where a region
5177 * must be split and another region descriptor can not be
5178 * allocated. If end == LONG_MAX, it will not fail.
5184 spin_lock(&inode
->i_lock
);
5185 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
5186 spin_unlock(&inode
->i_lock
);
5189 * If the subpool has a minimum size, the number of global
5190 * reservations to be released may be adjusted.
5192 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
5193 hugetlb_acct_memory(h
, -gbl_reserve
);
5198 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5199 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
5200 struct vm_area_struct
*vma
,
5201 unsigned long addr
, pgoff_t idx
)
5203 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
5205 unsigned long sbase
= saddr
& PUD_MASK
;
5206 unsigned long s_end
= sbase
+ PUD_SIZE
;
5208 /* Allow segments to share if only one is marked locked */
5209 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5210 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5213 * match the virtual addresses, permission and the alignment of the
5216 if (pmd_index(addr
) != pmd_index(saddr
) ||
5217 vm_flags
!= svm_flags
||
5218 !range_in_vma(svma
, sbase
, s_end
))
5224 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
5226 unsigned long base
= addr
& PUD_MASK
;
5227 unsigned long end
= base
+ PUD_SIZE
;
5230 * check on proper vm_flags and page table alignment
5232 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
5238 * Determine if start,end range within vma could be mapped by shared pmd.
5239 * If yes, adjust start and end to cover range associated with possible
5240 * shared pmd mappings.
5242 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5243 unsigned long *start
, unsigned long *end
)
5245 unsigned long v_start
= ALIGN(vma
->vm_start
, PUD_SIZE
),
5246 v_end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
5249 * vma need span at least one aligned PUD size and the start,end range
5250 * must at least partialy within it.
5252 if (!(vma
->vm_flags
& VM_MAYSHARE
) || !(v_end
> v_start
) ||
5253 (*end
<= v_start
) || (*start
>= v_end
))
5256 /* Extend the range to be PUD aligned for a worst case scenario */
5257 if (*start
> v_start
)
5258 *start
= ALIGN_DOWN(*start
, PUD_SIZE
);
5261 *end
= ALIGN(*end
, PUD_SIZE
);
5265 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5266 * and returns the corresponding pte. While this is not necessary for the
5267 * !shared pmd case because we can allocate the pmd later as well, it makes the
5268 * code much cleaner.
5270 * This routine must be called with i_mmap_rwsem held in at least read mode if
5271 * sharing is possible. For hugetlbfs, this prevents removal of any page
5272 * table entries associated with the address space. This is important as we
5273 * are setting up sharing based on existing page table entries (mappings).
5275 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5276 * huge_pte_alloc know that sharing is not possible and do not take
5277 * i_mmap_rwsem as a performance optimization. This is handled by the
5278 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5279 * only required for subsequent processing.
5281 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5283 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
5284 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
5285 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
5287 struct vm_area_struct
*svma
;
5288 unsigned long saddr
;
5293 if (!vma_shareable(vma
, addr
))
5294 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5296 i_mmap_assert_locked(mapping
);
5297 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
5301 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
5303 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
5304 vma_mmu_pagesize(svma
));
5306 get_page(virt_to_page(spte
));
5315 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
5316 if (pud_none(*pud
)) {
5317 pud_populate(mm
, pud
,
5318 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
5321 put_page(virt_to_page(spte
));
5325 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5330 * unmap huge page backed by shared pte.
5332 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5333 * indicated by page_count > 1, unmap is achieved by clearing pud and
5334 * decrementing the ref count. If count == 1, the pte page is not shared.
5336 * Called with page table lock held and i_mmap_rwsem held in write mode.
5338 * returns: 1 successfully unmapped a shared pte page
5339 * 0 the underlying pte page is not shared, or it is the last user
5341 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5342 unsigned long *addr
, pte_t
*ptep
)
5344 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
5345 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
5346 pud_t
*pud
= pud_offset(p4d
, *addr
);
5348 i_mmap_assert_write_locked(vma
->vm_file
->f_mapping
);
5349 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
5350 if (page_count(virt_to_page(ptep
)) == 1)
5354 put_page(virt_to_page(ptep
));
5356 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
5359 #define want_pmd_share() (1)
5360 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5361 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5366 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5367 unsigned long *addr
, pte_t
*ptep
)
5372 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5373 unsigned long *start
, unsigned long *end
)
5376 #define want_pmd_share() (0)
5377 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5379 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5380 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
5381 unsigned long addr
, unsigned long sz
)
5388 pgd
= pgd_offset(mm
, addr
);
5389 p4d
= p4d_alloc(mm
, pgd
, addr
);
5392 pud
= pud_alloc(mm
, p4d
, addr
);
5394 if (sz
== PUD_SIZE
) {
5397 BUG_ON(sz
!= PMD_SIZE
);
5398 if (want_pmd_share() && pud_none(*pud
))
5399 pte
= huge_pmd_share(mm
, addr
, pud
);
5401 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5404 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
5410 * huge_pte_offset() - Walk the page table to resolve the hugepage
5411 * entry at address @addr
5413 * Return: Pointer to page table entry (PUD or PMD) for
5414 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5415 * size @sz doesn't match the hugepage size at this level of the page
5418 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
5419 unsigned long addr
, unsigned long sz
)
5426 pgd
= pgd_offset(mm
, addr
);
5427 if (!pgd_present(*pgd
))
5429 p4d
= p4d_offset(pgd
, addr
);
5430 if (!p4d_present(*p4d
))
5433 pud
= pud_offset(p4d
, addr
);
5435 /* must be pud huge, non-present or none */
5436 return (pte_t
*)pud
;
5437 if (!pud_present(*pud
))
5439 /* must have a valid entry and size to go further */
5441 pmd
= pmd_offset(pud
, addr
);
5442 /* must be pmd huge, non-present or none */
5443 return (pte_t
*)pmd
;
5446 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5449 * These functions are overwritable if your architecture needs its own
5452 struct page
* __weak
5453 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
5456 return ERR_PTR(-EINVAL
);
5459 struct page
* __weak
5460 follow_huge_pd(struct vm_area_struct
*vma
,
5461 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
5463 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5467 struct page
* __weak
5468 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
5469 pmd_t
*pmd
, int flags
)
5471 struct page
*page
= NULL
;
5475 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5476 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
5477 (FOLL_PIN
| FOLL_GET
)))
5481 ptl
= pmd_lockptr(mm
, pmd
);
5484 * make sure that the address range covered by this pmd is not
5485 * unmapped from other threads.
5487 if (!pmd_huge(*pmd
))
5489 pte
= huge_ptep_get((pte_t
*)pmd
);
5490 if (pte_present(pte
)) {
5491 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
5493 * try_grab_page() should always succeed here, because: a) we
5494 * hold the pmd (ptl) lock, and b) we've just checked that the
5495 * huge pmd (head) page is present in the page tables. The ptl
5496 * prevents the head page and tail pages from being rearranged
5497 * in any way. So this page must be available at this point,
5498 * unless the page refcount overflowed:
5500 if (WARN_ON_ONCE(!try_grab_page(page
, flags
))) {
5505 if (is_hugetlb_entry_migration(pte
)) {
5507 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
5511 * hwpoisoned entry is treated as no_page_table in
5512 * follow_page_mask().
5520 struct page
* __weak
5521 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
5522 pud_t
*pud
, int flags
)
5524 if (flags
& (FOLL_GET
| FOLL_PIN
))
5527 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
5530 struct page
* __weak
5531 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
5533 if (flags
& (FOLL_GET
| FOLL_PIN
))
5536 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
5539 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
5543 spin_lock(&hugetlb_lock
);
5544 if (!PageHeadHuge(page
) ||
5545 !HPageMigratable(page
) ||
5546 !get_page_unless_zero(page
)) {
5550 ClearHPageMigratable(page
);
5551 list_move_tail(&page
->lru
, list
);
5553 spin_unlock(&hugetlb_lock
);
5557 void putback_active_hugepage(struct page
*page
)
5559 spin_lock(&hugetlb_lock
);
5560 SetHPageMigratable(page
);
5561 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
5562 spin_unlock(&hugetlb_lock
);
5566 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
5568 struct hstate
*h
= page_hstate(oldpage
);
5570 hugetlb_cgroup_migrate(oldpage
, newpage
);
5571 set_page_owner_migrate_reason(newpage
, reason
);
5574 * transfer temporary state of the new huge page. This is
5575 * reverse to other transitions because the newpage is going to
5576 * be final while the old one will be freed so it takes over
5577 * the temporary status.
5579 * Also note that we have to transfer the per-node surplus state
5580 * here as well otherwise the global surplus count will not match
5583 if (HPageTemporary(newpage
)) {
5584 int old_nid
= page_to_nid(oldpage
);
5585 int new_nid
= page_to_nid(newpage
);
5587 SetHPageTemporary(oldpage
);
5588 ClearHPageTemporary(newpage
);
5590 spin_lock(&hugetlb_lock
);
5591 if (h
->surplus_huge_pages_node
[old_nid
]) {
5592 h
->surplus_huge_pages_node
[old_nid
]--;
5593 h
->surplus_huge_pages_node
[new_nid
]++;
5595 spin_unlock(&hugetlb_lock
);
5600 static bool cma_reserve_called __initdata
;
5602 static int __init
cmdline_parse_hugetlb_cma(char *p
)
5604 hugetlb_cma_size
= memparse(p
, &p
);
5608 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma
);
5610 void __init
hugetlb_cma_reserve(int order
)
5612 unsigned long size
, reserved
, per_node
;
5615 cma_reserve_called
= true;
5617 if (!hugetlb_cma_size
)
5620 if (hugetlb_cma_size
< (PAGE_SIZE
<< order
)) {
5621 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5622 (PAGE_SIZE
<< order
) / SZ_1M
);
5627 * If 3 GB area is requested on a machine with 4 numa nodes,
5628 * let's allocate 1 GB on first three nodes and ignore the last one.
5630 per_node
= DIV_ROUND_UP(hugetlb_cma_size
, nr_online_nodes
);
5631 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5632 hugetlb_cma_size
/ SZ_1M
, per_node
/ SZ_1M
);
5635 for_each_node_state(nid
, N_ONLINE
) {
5637 char name
[CMA_MAX_NAME
];
5639 size
= min(per_node
, hugetlb_cma_size
- reserved
);
5640 size
= round_up(size
, PAGE_SIZE
<< order
);
5642 snprintf(name
, sizeof(name
), "hugetlb%d", nid
);
5643 res
= cma_declare_contiguous_nid(0, size
, 0, PAGE_SIZE
<< order
,
5645 &hugetlb_cma
[nid
], nid
);
5647 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5653 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5656 if (reserved
>= hugetlb_cma_size
)
5661 void __init
hugetlb_cma_check(void)
5663 if (!hugetlb_cma_size
|| cma_reserve_called
)
5666 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5669 #endif /* CONFIG_CMA */