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 static inline bool PageHugeFreed(struct page
*head
)
84 return page_private(head
+ 4) == -1UL;
87 static inline void SetPageHugeFreed(struct page
*head
)
89 set_page_private(head
+ 4, -1UL);
92 static inline void ClearPageHugeFreed(struct page
*head
)
94 set_page_private(head
+ 4, 0);
97 /* Forward declaration */
98 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
100 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
102 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
104 spin_unlock(&spool
->lock
);
106 /* If no pages are used, and no other handles to the subpool
107 * remain, give up any reservations based on minimum size and
108 * free the subpool */
110 if (spool
->min_hpages
!= -1)
111 hugetlb_acct_memory(spool
->hstate
,
117 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
120 struct hugepage_subpool
*spool
;
122 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
126 spin_lock_init(&spool
->lock
);
128 spool
->max_hpages
= max_hpages
;
130 spool
->min_hpages
= min_hpages
;
132 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
136 spool
->rsv_hpages
= min_hpages
;
141 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
143 spin_lock(&spool
->lock
);
144 BUG_ON(!spool
->count
);
146 unlock_or_release_subpool(spool
);
150 * Subpool accounting for allocating and reserving pages.
151 * Return -ENOMEM if there are not enough resources to satisfy the
152 * request. Otherwise, return the number of pages by which the
153 * global pools must be adjusted (upward). The returned value may
154 * only be different than the passed value (delta) in the case where
155 * a subpool minimum size must be maintained.
157 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
165 spin_lock(&spool
->lock
);
167 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
168 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
169 spool
->used_hpages
+= delta
;
176 /* minimum size accounting */
177 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
178 if (delta
> spool
->rsv_hpages
) {
180 * Asking for more reserves than those already taken on
181 * behalf of subpool. Return difference.
183 ret
= delta
- spool
->rsv_hpages
;
184 spool
->rsv_hpages
= 0;
186 ret
= 0; /* reserves already accounted for */
187 spool
->rsv_hpages
-= delta
;
192 spin_unlock(&spool
->lock
);
197 * Subpool accounting for freeing and unreserving pages.
198 * Return the number of global page reservations that must be dropped.
199 * The return value may only be different than the passed value (delta)
200 * in the case where a subpool minimum size must be maintained.
202 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
210 spin_lock(&spool
->lock
);
212 if (spool
->max_hpages
!= -1) /* maximum size accounting */
213 spool
->used_hpages
-= delta
;
215 /* minimum size accounting */
216 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
217 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
220 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
222 spool
->rsv_hpages
+= delta
;
223 if (spool
->rsv_hpages
> spool
->min_hpages
)
224 spool
->rsv_hpages
= spool
->min_hpages
;
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
231 unlock_or_release_subpool(spool
);
236 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
238 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
241 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
243 return subpool_inode(file_inode(vma
->vm_file
));
246 /* Helper that removes a struct file_region from the resv_map cache and returns
249 static struct file_region
*
250 get_file_region_entry_from_cache(struct resv_map
*resv
, long from
, long to
)
252 struct file_region
*nrg
= NULL
;
254 VM_BUG_ON(resv
->region_cache_count
<= 0);
256 resv
->region_cache_count
--;
257 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
, link
);
258 list_del(&nrg
->link
);
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region
*nrg
,
267 struct file_region
*rg
)
269 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg
->reservation_counter
= rg
->reservation_counter
;
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup
*h_cg
,
280 struct resv_map
*resv
,
281 struct file_region
*nrg
)
283 #ifdef CONFIG_CGROUP_HUGETLB
285 nrg
->reservation_counter
=
286 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
287 nrg
->css
= &h_cg
->css
;
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
299 if (!resv
->pages_per_hpage
)
300 resv
->pages_per_hpage
= pages_per_huge_page(h
);
301 /* pages_per_hpage should be the same for all entries in
304 VM_BUG_ON(resv
->pages_per_hpage
!= pages_per_huge_page(h
));
306 nrg
->reservation_counter
= NULL
;
312 static void put_uncharge_info(struct file_region
*rg
)
314 #ifdef CONFIG_CGROUP_HUGETLB
320 static bool has_same_uncharge_info(struct file_region
*rg
,
321 struct file_region
*org
)
323 #ifdef CONFIG_CGROUP_HUGETLB
325 rg
->reservation_counter
== org
->reservation_counter
&&
333 static void coalesce_file_region(struct resv_map
*resv
, struct file_region
*rg
)
335 struct file_region
*nrg
= NULL
, *prg
= NULL
;
337 prg
= list_prev_entry(rg
, link
);
338 if (&prg
->link
!= &resv
->regions
&& prg
->to
== rg
->from
&&
339 has_same_uncharge_info(prg
, rg
)) {
343 put_uncharge_info(rg
);
349 nrg
= list_next_entry(rg
, link
);
350 if (&nrg
->link
!= &resv
->regions
&& nrg
->from
== rg
->to
&&
351 has_same_uncharge_info(nrg
, rg
)) {
352 nrg
->from
= rg
->from
;
355 put_uncharge_info(rg
);
361 * Must be called with resv->lock held.
363 * Calling this with regions_needed != NULL will count the number of pages
364 * to be added but will not modify the linked list. And regions_needed will
365 * indicate the number of file_regions needed in the cache to carry out to add
366 * the regions for this range.
368 static long add_reservation_in_range(struct resv_map
*resv
, long f
, long t
,
369 struct hugetlb_cgroup
*h_cg
,
370 struct hstate
*h
, long *regions_needed
)
373 struct list_head
*head
= &resv
->regions
;
374 long last_accounted_offset
= f
;
375 struct file_region
*rg
= NULL
, *trg
= NULL
, *nrg
= NULL
;
380 /* In this loop, we essentially handle an entry for the range
381 * [last_accounted_offset, rg->from), at every iteration, with some
384 list_for_each_entry_safe(rg
, trg
, head
, link
) {
385 /* Skip irrelevant regions that start before our range. */
387 /* If this region ends after the last accounted offset,
388 * then we need to update last_accounted_offset.
390 if (rg
->to
> last_accounted_offset
)
391 last_accounted_offset
= rg
->to
;
395 /* When we find a region that starts beyond our range, we've
401 /* Add an entry for last_accounted_offset -> rg->from, and
402 * update last_accounted_offset.
404 if (rg
->from
> last_accounted_offset
) {
405 add
+= rg
->from
- last_accounted_offset
;
406 if (!regions_needed
) {
407 nrg
= get_file_region_entry_from_cache(
408 resv
, last_accounted_offset
, rg
->from
);
409 record_hugetlb_cgroup_uncharge_info(h_cg
, h
,
411 list_add(&nrg
->link
, rg
->link
.prev
);
412 coalesce_file_region(resv
, nrg
);
414 *regions_needed
+= 1;
417 last_accounted_offset
= rg
->to
;
420 /* Handle the case where our range extends beyond
421 * last_accounted_offset.
423 if (last_accounted_offset
< t
) {
424 add
+= t
- last_accounted_offset
;
425 if (!regions_needed
) {
426 nrg
= get_file_region_entry_from_cache(
427 resv
, last_accounted_offset
, t
);
428 record_hugetlb_cgroup_uncharge_info(h_cg
, h
, resv
, nrg
);
429 list_add(&nrg
->link
, rg
->link
.prev
);
430 coalesce_file_region(resv
, nrg
);
432 *regions_needed
+= 1;
439 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
441 static int allocate_file_region_entries(struct resv_map
*resv
,
443 __must_hold(&resv
->lock
)
445 struct list_head allocated_regions
;
446 int to_allocate
= 0, i
= 0;
447 struct file_region
*trg
= NULL
, *rg
= NULL
;
449 VM_BUG_ON(regions_needed
< 0);
451 INIT_LIST_HEAD(&allocated_regions
);
454 * Check for sufficient descriptors in the cache to accommodate
455 * the number of in progress add operations plus regions_needed.
457 * This is a while loop because when we drop the lock, some other call
458 * to region_add or region_del may have consumed some region_entries,
459 * so we keep looping here until we finally have enough entries for
460 * (adds_in_progress + regions_needed).
462 while (resv
->region_cache_count
<
463 (resv
->adds_in_progress
+ regions_needed
)) {
464 to_allocate
= resv
->adds_in_progress
+ regions_needed
-
465 resv
->region_cache_count
;
467 /* At this point, we should have enough entries in the cache
468 * for all the existings adds_in_progress. We should only be
469 * needing to allocate for regions_needed.
471 VM_BUG_ON(resv
->region_cache_count
< resv
->adds_in_progress
);
473 spin_unlock(&resv
->lock
);
474 for (i
= 0; i
< to_allocate
; i
++) {
475 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
478 list_add(&trg
->link
, &allocated_regions
);
481 spin_lock(&resv
->lock
);
483 list_splice(&allocated_regions
, &resv
->region_cache
);
484 resv
->region_cache_count
+= to_allocate
;
490 list_for_each_entry_safe(rg
, trg
, &allocated_regions
, link
) {
498 * Add the huge page range represented by [f, t) to the reserve
499 * map. Regions will be taken from the cache to fill in this range.
500 * Sufficient regions should exist in the cache due to the previous
501 * call to region_chg with the same range, but in some cases the cache will not
502 * have sufficient entries due to races with other code doing region_add or
503 * region_del. The extra needed entries will be allocated.
505 * regions_needed is the out value provided by a previous call to region_chg.
507 * Return the number of new huge pages added to the map. This number is greater
508 * than or equal to zero. If file_region entries needed to be allocated for
509 * this operation and we were not able to allocate, it returns -ENOMEM.
510 * region_add of regions of length 1 never allocate file_regions and cannot
511 * fail; region_chg will always allocate at least 1 entry and a region_add for
512 * 1 page will only require at most 1 entry.
514 static long region_add(struct resv_map
*resv
, long f
, long t
,
515 long in_regions_needed
, struct hstate
*h
,
516 struct hugetlb_cgroup
*h_cg
)
518 long add
= 0, actual_regions_needed
= 0;
520 spin_lock(&resv
->lock
);
523 /* Count how many regions are actually needed to execute this add. */
524 add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
525 &actual_regions_needed
);
528 * Check for sufficient descriptors in the cache to accommodate
529 * this add operation. Note that actual_regions_needed may be greater
530 * than in_regions_needed, as the resv_map may have been modified since
531 * the region_chg call. In this case, we need to make sure that we
532 * allocate extra entries, such that we have enough for all the
533 * existing adds_in_progress, plus the excess needed for this
536 if (actual_regions_needed
> in_regions_needed
&&
537 resv
->region_cache_count
<
538 resv
->adds_in_progress
+
539 (actual_regions_needed
- in_regions_needed
)) {
540 /* region_add operation of range 1 should never need to
541 * allocate file_region entries.
543 VM_BUG_ON(t
- f
<= 1);
545 if (allocate_file_region_entries(
546 resv
, actual_regions_needed
- in_regions_needed
)) {
553 add
= add_reservation_in_range(resv
, f
, t
, h_cg
, h
, NULL
);
555 resv
->adds_in_progress
-= in_regions_needed
;
557 spin_unlock(&resv
->lock
);
563 * Examine the existing reserve map and determine how many
564 * huge pages in the specified range [f, t) are NOT currently
565 * represented. This routine is called before a subsequent
566 * call to region_add that will actually modify the reserve
567 * map to add the specified range [f, t). region_chg does
568 * not change the number of huge pages represented by the
569 * map. A number of new file_region structures is added to the cache as a
570 * placeholder, for the subsequent region_add call to use. At least 1
571 * file_region structure is added.
573 * out_regions_needed is the number of regions added to the
574 * resv->adds_in_progress. This value needs to be provided to a follow up call
575 * to region_add or region_abort for proper accounting.
577 * Returns the number of huge pages that need to be added to the existing
578 * reservation map for the range [f, t). This number is greater or equal to
579 * zero. -ENOMEM is returned if a new file_region structure or cache entry
580 * is needed and can not be allocated.
582 static long region_chg(struct resv_map
*resv
, long f
, long t
,
583 long *out_regions_needed
)
587 spin_lock(&resv
->lock
);
589 /* Count how many hugepages in this range are NOT represented. */
590 chg
= add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
593 if (*out_regions_needed
== 0)
594 *out_regions_needed
= 1;
596 if (allocate_file_region_entries(resv
, *out_regions_needed
))
599 resv
->adds_in_progress
+= *out_regions_needed
;
601 spin_unlock(&resv
->lock
);
606 * Abort the in progress add operation. The adds_in_progress field
607 * of the resv_map keeps track of the operations in progress between
608 * calls to region_chg and region_add. Operations are sometimes
609 * aborted after the call to region_chg. In such cases, region_abort
610 * is called to decrement the adds_in_progress counter. regions_needed
611 * is the value returned by the region_chg call, it is used to decrement
612 * the adds_in_progress counter.
614 * NOTE: The range arguments [f, t) are not needed or used in this
615 * routine. They are kept to make reading the calling code easier as
616 * arguments will match the associated region_chg call.
618 static void region_abort(struct resv_map
*resv
, long f
, long t
,
621 spin_lock(&resv
->lock
);
622 VM_BUG_ON(!resv
->region_cache_count
);
623 resv
->adds_in_progress
-= regions_needed
;
624 spin_unlock(&resv
->lock
);
628 * Delete the specified range [f, t) from the reserve map. If the
629 * t parameter is LONG_MAX, this indicates that ALL regions after f
630 * should be deleted. Locate the regions which intersect [f, t)
631 * and either trim, delete or split the existing regions.
633 * Returns the number of huge pages deleted from the reserve map.
634 * In the normal case, the return value is zero or more. In the
635 * case where a region must be split, a new region descriptor must
636 * be allocated. If the allocation fails, -ENOMEM will be returned.
637 * NOTE: If the parameter t == LONG_MAX, then we will never split
638 * a region and possibly return -ENOMEM. Callers specifying
639 * t == LONG_MAX do not need to check for -ENOMEM error.
641 static long region_del(struct resv_map
*resv
, long f
, long t
)
643 struct list_head
*head
= &resv
->regions
;
644 struct file_region
*rg
, *trg
;
645 struct file_region
*nrg
= NULL
;
649 spin_lock(&resv
->lock
);
650 list_for_each_entry_safe(rg
, trg
, head
, link
) {
652 * Skip regions before the range to be deleted. file_region
653 * ranges are normally of the form [from, to). However, there
654 * may be a "placeholder" entry in the map which is of the form
655 * (from, to) with from == to. Check for placeholder entries
656 * at the beginning of the range to be deleted.
658 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
664 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
666 * Check for an entry in the cache before dropping
667 * lock and attempting allocation.
670 resv
->region_cache_count
> resv
->adds_in_progress
) {
671 nrg
= list_first_entry(&resv
->region_cache
,
674 list_del(&nrg
->link
);
675 resv
->region_cache_count
--;
679 spin_unlock(&resv
->lock
);
680 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
687 hugetlb_cgroup_uncharge_file_region(
688 resv
, rg
, t
- f
, false);
690 /* New entry for end of split region */
694 copy_hugetlb_cgroup_uncharge_info(nrg
, rg
);
696 INIT_LIST_HEAD(&nrg
->link
);
698 /* Original entry is trimmed */
701 list_add(&nrg
->link
, &rg
->link
);
706 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
707 del
+= rg
->to
- rg
->from
;
708 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
709 rg
->to
- rg
->from
, true);
715 if (f
<= rg
->from
) { /* Trim beginning of region */
716 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
717 t
- rg
->from
, false);
721 } else { /* Trim end of region */
722 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
730 spin_unlock(&resv
->lock
);
736 * A rare out of memory error was encountered which prevented removal of
737 * the reserve map region for a page. The huge page itself was free'ed
738 * and removed from the page cache. This routine will adjust the subpool
739 * usage count, and the global reserve count if needed. By incrementing
740 * these counts, the reserve map entry which could not be deleted will
741 * appear as a "reserved" entry instead of simply dangling with incorrect
744 void hugetlb_fix_reserve_counts(struct inode
*inode
)
746 struct hugepage_subpool
*spool
= subpool_inode(inode
);
748 bool reserved
= false;
750 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
751 if (rsv_adjust
> 0) {
752 struct hstate
*h
= hstate_inode(inode
);
754 if (!hugetlb_acct_memory(h
, 1))
756 } else if (!rsv_adjust
) {
761 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
765 * Count and return the number of huge pages in the reserve map
766 * that intersect with the range [f, t).
768 static long region_count(struct resv_map
*resv
, long f
, long t
)
770 struct list_head
*head
= &resv
->regions
;
771 struct file_region
*rg
;
774 spin_lock(&resv
->lock
);
775 /* Locate each segment we overlap with, and count that overlap. */
776 list_for_each_entry(rg
, head
, link
) {
785 seg_from
= max(rg
->from
, f
);
786 seg_to
= min(rg
->to
, t
);
788 chg
+= seg_to
- seg_from
;
790 spin_unlock(&resv
->lock
);
796 * Convert the address within this vma to the page offset within
797 * the mapping, in pagecache page units; huge pages here.
799 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
800 struct vm_area_struct
*vma
, unsigned long address
)
802 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
803 (vma
->vm_pgoff
>> huge_page_order(h
));
806 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
807 unsigned long address
)
809 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
811 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
814 * Return the size of the pages allocated when backing a VMA. In the majority
815 * cases this will be same size as used by the page table entries.
817 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
819 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
820 return vma
->vm_ops
->pagesize(vma
);
823 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
826 * Return the page size being used by the MMU to back a VMA. In the majority
827 * of cases, the page size used by the kernel matches the MMU size. On
828 * architectures where it differs, an architecture-specific 'strong'
829 * version of this symbol is required.
831 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
833 return vma_kernel_pagesize(vma
);
837 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
838 * bits of the reservation map pointer, which are always clear due to
841 #define HPAGE_RESV_OWNER (1UL << 0)
842 #define HPAGE_RESV_UNMAPPED (1UL << 1)
843 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
846 * These helpers are used to track how many pages are reserved for
847 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
848 * is guaranteed to have their future faults succeed.
850 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
851 * the reserve counters are updated with the hugetlb_lock held. It is safe
852 * to reset the VMA at fork() time as it is not in use yet and there is no
853 * chance of the global counters getting corrupted as a result of the values.
855 * The private mapping reservation is represented in a subtly different
856 * manner to a shared mapping. A shared mapping has a region map associated
857 * with the underlying file, this region map represents the backing file
858 * pages which have ever had a reservation assigned which this persists even
859 * after the page is instantiated. A private mapping has a region map
860 * associated with the original mmap which is attached to all VMAs which
861 * reference it, this region map represents those offsets which have consumed
862 * reservation ie. where pages have been instantiated.
864 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
866 return (unsigned long)vma
->vm_private_data
;
869 static void set_vma_private_data(struct vm_area_struct
*vma
,
872 vma
->vm_private_data
= (void *)value
;
876 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map
*resv_map
,
877 struct hugetlb_cgroup
*h_cg
,
880 #ifdef CONFIG_CGROUP_HUGETLB
882 resv_map
->reservation_counter
= NULL
;
883 resv_map
->pages_per_hpage
= 0;
884 resv_map
->css
= NULL
;
886 resv_map
->reservation_counter
=
887 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
888 resv_map
->pages_per_hpage
= pages_per_huge_page(h
);
889 resv_map
->css
= &h_cg
->css
;
894 struct resv_map
*resv_map_alloc(void)
896 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
897 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
899 if (!resv_map
|| !rg
) {
905 kref_init(&resv_map
->refs
);
906 spin_lock_init(&resv_map
->lock
);
907 INIT_LIST_HEAD(&resv_map
->regions
);
909 resv_map
->adds_in_progress
= 0;
911 * Initialize these to 0. On shared mappings, 0's here indicate these
912 * fields don't do cgroup accounting. On private mappings, these will be
913 * re-initialized to the proper values, to indicate that hugetlb cgroup
914 * reservations are to be un-charged from here.
916 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, NULL
, NULL
);
918 INIT_LIST_HEAD(&resv_map
->region_cache
);
919 list_add(&rg
->link
, &resv_map
->region_cache
);
920 resv_map
->region_cache_count
= 1;
925 void resv_map_release(struct kref
*ref
)
927 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
928 struct list_head
*head
= &resv_map
->region_cache
;
929 struct file_region
*rg
, *trg
;
931 /* Clear out any active regions before we release the map. */
932 region_del(resv_map
, 0, LONG_MAX
);
934 /* ... and any entries left in the cache */
935 list_for_each_entry_safe(rg
, trg
, head
, link
) {
940 VM_BUG_ON(resv_map
->adds_in_progress
);
945 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
948 * At inode evict time, i_mapping may not point to the original
949 * address space within the inode. This original address space
950 * contains the pointer to the resv_map. So, always use the
951 * address space embedded within the inode.
952 * The VERY common case is inode->mapping == &inode->i_data but,
953 * this may not be true for device special inodes.
955 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
958 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
960 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
961 if (vma
->vm_flags
& VM_MAYSHARE
) {
962 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
963 struct inode
*inode
= mapping
->host
;
965 return inode_resv_map(inode
);
968 return (struct resv_map
*)(get_vma_private_data(vma
) &
973 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
975 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
976 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
978 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
979 HPAGE_RESV_MASK
) | (unsigned long)map
);
982 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
984 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
985 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
987 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
990 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
992 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
994 return (get_vma_private_data(vma
) & flag
) != 0;
997 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
998 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
1000 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
1001 if (!(vma
->vm_flags
& VM_MAYSHARE
))
1002 vma
->vm_private_data
= (void *)0;
1005 /* Returns true if the VMA has associated reserve pages */
1006 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
1008 if (vma
->vm_flags
& VM_NORESERVE
) {
1010 * This address is already reserved by other process(chg == 0),
1011 * so, we should decrement reserved count. Without decrementing,
1012 * reserve count remains after releasing inode, because this
1013 * allocated page will go into page cache and is regarded as
1014 * coming from reserved pool in releasing step. Currently, we
1015 * don't have any other solution to deal with this situation
1016 * properly, so add work-around here.
1018 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
1024 /* Shared mappings always use reserves */
1025 if (vma
->vm_flags
& VM_MAYSHARE
) {
1027 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1028 * be a region map for all pages. The only situation where
1029 * there is no region map is if a hole was punched via
1030 * fallocate. In this case, there really are no reserves to
1031 * use. This situation is indicated if chg != 0.
1040 * Only the process that called mmap() has reserves for
1043 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1045 * Like the shared case above, a hole punch or truncate
1046 * could have been performed on the private mapping.
1047 * Examine the value of chg to determine if reserves
1048 * actually exist or were previously consumed.
1049 * Very Subtle - The value of chg comes from a previous
1050 * call to vma_needs_reserves(). The reserve map for
1051 * private mappings has different (opposite) semantics
1052 * than that of shared mappings. vma_needs_reserves()
1053 * has already taken this difference in semantics into
1054 * account. Therefore, the meaning of chg is the same
1055 * as in the shared case above. Code could easily be
1056 * combined, but keeping it separate draws attention to
1057 * subtle differences.
1068 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
1070 int nid
= page_to_nid(page
);
1071 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
1072 h
->free_huge_pages
++;
1073 h
->free_huge_pages_node
[nid
]++;
1074 SetPageHugeFreed(page
);
1077 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
1080 bool nocma
= !!(current
->flags
& PF_MEMALLOC_NOCMA
);
1082 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
) {
1083 if (nocma
&& is_migrate_cma_page(page
))
1086 if (PageHWPoison(page
))
1089 list_move(&page
->lru
, &h
->hugepage_activelist
);
1090 set_page_refcounted(page
);
1091 ClearPageHugeFreed(page
);
1092 h
->free_huge_pages
--;
1093 h
->free_huge_pages_node
[nid
]--;
1100 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
1103 unsigned int cpuset_mems_cookie
;
1104 struct zonelist
*zonelist
;
1107 int node
= NUMA_NO_NODE
;
1109 zonelist
= node_zonelist(nid
, gfp_mask
);
1112 cpuset_mems_cookie
= read_mems_allowed_begin();
1113 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
1116 if (!cpuset_zone_allowed(zone
, gfp_mask
))
1119 * no need to ask again on the same node. Pool is node rather than
1122 if (zone_to_nid(zone
) == node
)
1124 node
= zone_to_nid(zone
);
1126 page
= dequeue_huge_page_node_exact(h
, node
);
1130 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
1136 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
1137 struct vm_area_struct
*vma
,
1138 unsigned long address
, int avoid_reserve
,
1142 struct mempolicy
*mpol
;
1144 nodemask_t
*nodemask
;
1148 * A child process with MAP_PRIVATE mappings created by their parent
1149 * have no page reserves. This check ensures that reservations are
1150 * not "stolen". The child may still get SIGKILLed
1152 if (!vma_has_reserves(vma
, chg
) &&
1153 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1156 /* If reserves cannot be used, ensure enough pages are in the pool */
1157 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1160 gfp_mask
= htlb_alloc_mask(h
);
1161 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1162 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
1163 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
1164 SetPagePrivate(page
);
1165 h
->resv_huge_pages
--;
1168 mpol_cond_put(mpol
);
1176 * common helper functions for hstate_next_node_to_{alloc|free}.
1177 * We may have allocated or freed a huge page based on a different
1178 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1179 * be outside of *nodes_allowed. Ensure that we use an allowed
1180 * node for alloc or free.
1182 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1184 nid
= next_node_in(nid
, *nodes_allowed
);
1185 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1190 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1192 if (!node_isset(nid
, *nodes_allowed
))
1193 nid
= next_node_allowed(nid
, nodes_allowed
);
1198 * returns the previously saved node ["this node"] from which to
1199 * allocate a persistent huge page for the pool and advance the
1200 * next node from which to allocate, handling wrap at end of node
1203 static int hstate_next_node_to_alloc(struct hstate
*h
,
1204 nodemask_t
*nodes_allowed
)
1208 VM_BUG_ON(!nodes_allowed
);
1210 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1211 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1217 * helper for free_pool_huge_page() - return the previously saved
1218 * node ["this node"] from which to free a huge page. Advance the
1219 * next node id whether or not we find a free huge page to free so
1220 * that the next attempt to free addresses the next node.
1222 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1226 VM_BUG_ON(!nodes_allowed
);
1228 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1229 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1234 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1235 for (nr_nodes = nodes_weight(*mask); \
1237 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1240 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1241 for (nr_nodes = nodes_weight(*mask); \
1243 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1246 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1247 static void destroy_compound_gigantic_page(struct page
*page
,
1251 int nr_pages
= 1 << order
;
1252 struct page
*p
= page
+ 1;
1254 atomic_set(compound_mapcount_ptr(page
), 0);
1255 if (hpage_pincount_available(page
))
1256 atomic_set(compound_pincount_ptr(page
), 0);
1258 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1259 clear_compound_head(p
);
1260 set_page_refcounted(p
);
1263 set_compound_order(page
, 0);
1264 page
[1].compound_nr
= 0;
1265 __ClearPageHead(page
);
1268 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1271 * If the page isn't allocated using the cma allocator,
1272 * cma_release() returns false.
1275 if (cma_release(hugetlb_cma
[page_to_nid(page
)], page
, 1 << order
))
1279 free_contig_range(page_to_pfn(page
), 1 << order
);
1282 #ifdef CONFIG_CONTIG_ALLOC
1283 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1284 int nid
, nodemask_t
*nodemask
)
1286 unsigned long nr_pages
= 1UL << huge_page_order(h
);
1287 if (nid
== NUMA_NO_NODE
)
1288 nid
= numa_mem_id();
1295 if (hugetlb_cma
[nid
]) {
1296 page
= cma_alloc(hugetlb_cma
[nid
], nr_pages
,
1297 huge_page_order(h
), true);
1302 if (!(gfp_mask
& __GFP_THISNODE
)) {
1303 for_each_node_mask(node
, *nodemask
) {
1304 if (node
== nid
|| !hugetlb_cma
[node
])
1307 page
= cma_alloc(hugetlb_cma
[node
], nr_pages
,
1308 huge_page_order(h
), true);
1316 return alloc_contig_pages(nr_pages
, gfp_mask
, nid
, nodemask
);
1319 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1320 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1321 #else /* !CONFIG_CONTIG_ALLOC */
1322 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1323 int nid
, nodemask_t
*nodemask
)
1327 #endif /* CONFIG_CONTIG_ALLOC */
1329 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1330 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1331 int nid
, nodemask_t
*nodemask
)
1335 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1336 static inline void destroy_compound_gigantic_page(struct page
*page
,
1337 unsigned int order
) { }
1340 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1343 struct page
*subpage
= page
;
1345 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1349 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1350 for (i
= 0; i
< pages_per_huge_page(h
);
1351 i
++, subpage
= mem_map_next(subpage
, page
, i
)) {
1352 subpage
->flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1353 1 << PG_referenced
| 1 << PG_dirty
|
1354 1 << PG_active
| 1 << PG_private
|
1357 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1358 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page
), page
);
1359 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1360 set_page_refcounted(page
);
1361 if (hstate_is_gigantic(h
)) {
1363 * Temporarily drop the hugetlb_lock, because
1364 * we might block in free_gigantic_page().
1366 spin_unlock(&hugetlb_lock
);
1367 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1368 free_gigantic_page(page
, huge_page_order(h
));
1369 spin_lock(&hugetlb_lock
);
1371 __free_pages(page
, huge_page_order(h
));
1375 struct hstate
*size_to_hstate(unsigned long size
)
1379 for_each_hstate(h
) {
1380 if (huge_page_size(h
) == size
)
1387 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1388 * to hstate->hugepage_activelist.)
1390 * This function can be called for tail pages, but never returns true for them.
1392 bool page_huge_active(struct page
*page
)
1394 return PageHeadHuge(page
) && PagePrivate(&page
[1]);
1397 /* never called for tail page */
1398 void set_page_huge_active(struct page
*page
)
1400 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1401 SetPagePrivate(&page
[1]);
1404 static void clear_page_huge_active(struct page
*page
)
1406 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1407 ClearPagePrivate(&page
[1]);
1411 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1414 static inline bool PageHugeTemporary(struct page
*page
)
1416 if (!PageHuge(page
))
1419 return (unsigned long)page
[2].mapping
== -1U;
1422 static inline void SetPageHugeTemporary(struct page
*page
)
1424 page
[2].mapping
= (void *)-1U;
1427 static inline void ClearPageHugeTemporary(struct page
*page
)
1429 page
[2].mapping
= NULL
;
1432 static void __free_huge_page(struct page
*page
)
1435 * Can't pass hstate in here because it is called from the
1436 * compound page destructor.
1438 struct hstate
*h
= page_hstate(page
);
1439 int nid
= page_to_nid(page
);
1440 struct hugepage_subpool
*spool
=
1441 (struct hugepage_subpool
*)page_private(page
);
1442 bool restore_reserve
;
1444 VM_BUG_ON_PAGE(page_count(page
), page
);
1445 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1447 set_page_private(page
, 0);
1448 page
->mapping
= NULL
;
1449 restore_reserve
= PagePrivate(page
);
1450 ClearPagePrivate(page
);
1453 * If PagePrivate() was set on page, page allocation consumed a
1454 * reservation. If the page was associated with a subpool, there
1455 * would have been a page reserved in the subpool before allocation
1456 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1457 * reservtion, do not call hugepage_subpool_put_pages() as this will
1458 * remove the reserved page from the subpool.
1460 if (!restore_reserve
) {
1462 * A return code of zero implies that the subpool will be
1463 * under its minimum size if the reservation is not restored
1464 * after page is free. Therefore, force restore_reserve
1467 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1468 restore_reserve
= true;
1471 spin_lock(&hugetlb_lock
);
1472 clear_page_huge_active(page
);
1473 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1474 pages_per_huge_page(h
), page
);
1475 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
1476 pages_per_huge_page(h
), page
);
1477 if (restore_reserve
)
1478 h
->resv_huge_pages
++;
1480 if (PageHugeTemporary(page
)) {
1481 list_del(&page
->lru
);
1482 ClearPageHugeTemporary(page
);
1483 update_and_free_page(h
, page
);
1484 } else if (h
->surplus_huge_pages_node
[nid
]) {
1485 /* remove the page from active list */
1486 list_del(&page
->lru
);
1487 update_and_free_page(h
, page
);
1488 h
->surplus_huge_pages
--;
1489 h
->surplus_huge_pages_node
[nid
]--;
1491 arch_clear_hugepage_flags(page
);
1492 enqueue_huge_page(h
, page
);
1494 spin_unlock(&hugetlb_lock
);
1498 * As free_huge_page() can be called from a non-task context, we have
1499 * to defer the actual freeing in a workqueue to prevent potential
1500 * hugetlb_lock deadlock.
1502 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1503 * be freed and frees them one-by-one. As the page->mapping pointer is
1504 * going to be cleared in __free_huge_page() anyway, it is reused as the
1505 * llist_node structure of a lockless linked list of huge pages to be freed.
1507 static LLIST_HEAD(hpage_freelist
);
1509 static void free_hpage_workfn(struct work_struct
*work
)
1511 struct llist_node
*node
;
1514 node
= llist_del_all(&hpage_freelist
);
1517 page
= container_of((struct address_space
**)node
,
1518 struct page
, mapping
);
1520 __free_huge_page(page
);
1523 static DECLARE_WORK(free_hpage_work
, free_hpage_workfn
);
1525 void free_huge_page(struct page
*page
)
1528 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1532 * Only call schedule_work() if hpage_freelist is previously
1533 * empty. Otherwise, schedule_work() had been called but the
1534 * workfn hasn't retrieved the list yet.
1536 if (llist_add((struct llist_node
*)&page
->mapping
,
1538 schedule_work(&free_hpage_work
);
1542 __free_huge_page(page
);
1545 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1547 INIT_LIST_HEAD(&page
->lru
);
1548 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1549 set_hugetlb_cgroup(page
, NULL
);
1550 set_hugetlb_cgroup_rsvd(page
, NULL
);
1551 spin_lock(&hugetlb_lock
);
1553 h
->nr_huge_pages_node
[nid
]++;
1554 ClearPageHugeFreed(page
);
1555 spin_unlock(&hugetlb_lock
);
1558 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1561 int nr_pages
= 1 << order
;
1562 struct page
*p
= page
+ 1;
1564 /* we rely on prep_new_huge_page to set the destructor */
1565 set_compound_order(page
, order
);
1566 __ClearPageReserved(page
);
1567 __SetPageHead(page
);
1568 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1570 * For gigantic hugepages allocated through bootmem at
1571 * boot, it's safer to be consistent with the not-gigantic
1572 * hugepages and clear the PG_reserved bit from all tail pages
1573 * too. Otherwise drivers using get_user_pages() to access tail
1574 * pages may get the reference counting wrong if they see
1575 * PG_reserved set on a tail page (despite the head page not
1576 * having PG_reserved set). Enforcing this consistency between
1577 * head and tail pages allows drivers to optimize away a check
1578 * on the head page when they need know if put_page() is needed
1579 * after get_user_pages().
1581 __ClearPageReserved(p
);
1582 set_page_count(p
, 0);
1583 set_compound_head(p
, page
);
1585 atomic_set(compound_mapcount_ptr(page
), -1);
1587 if (hpage_pincount_available(page
))
1588 atomic_set(compound_pincount_ptr(page
), 0);
1592 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1593 * transparent huge pages. See the PageTransHuge() documentation for more
1596 int PageHuge(struct page
*page
)
1598 if (!PageCompound(page
))
1601 page
= compound_head(page
);
1602 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1604 EXPORT_SYMBOL_GPL(PageHuge
);
1607 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1608 * normal or transparent huge pages.
1610 int PageHeadHuge(struct page
*page_head
)
1612 if (!PageHead(page_head
))
1615 return page_head
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1619 * Find and lock address space (mapping) in write mode.
1621 * Upon entry, the page is locked which means that page_mapping() is
1622 * stable. Due to locking order, we can only trylock_write. If we can
1623 * not get the lock, simply return NULL to caller.
1625 struct address_space
*hugetlb_page_mapping_lock_write(struct page
*hpage
)
1627 struct address_space
*mapping
= page_mapping(hpage
);
1632 if (i_mmap_trylock_write(mapping
))
1638 pgoff_t
__basepage_index(struct page
*page
)
1640 struct page
*page_head
= compound_head(page
);
1641 pgoff_t index
= page_index(page_head
);
1642 unsigned long compound_idx
;
1644 if (!PageHuge(page_head
))
1645 return page_index(page
);
1647 if (compound_order(page_head
) >= MAX_ORDER
)
1648 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1650 compound_idx
= page
- page_head
;
1652 return (index
<< compound_order(page_head
)) + compound_idx
;
1655 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1656 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1657 nodemask_t
*node_alloc_noretry
)
1659 int order
= huge_page_order(h
);
1661 bool alloc_try_hard
= true;
1664 * By default we always try hard to allocate the page with
1665 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1666 * a loop (to adjust global huge page counts) and previous allocation
1667 * failed, do not continue to try hard on the same node. Use the
1668 * node_alloc_noretry bitmap to manage this state information.
1670 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1671 alloc_try_hard
= false;
1672 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1674 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1675 if (nid
== NUMA_NO_NODE
)
1676 nid
= numa_mem_id();
1677 page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, nmask
);
1679 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1681 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1684 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1685 * indicates an overall state change. Clear bit so that we resume
1686 * normal 'try hard' allocations.
1688 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1689 node_clear(nid
, *node_alloc_noretry
);
1692 * If we tried hard to get a page but failed, set bit so that
1693 * subsequent attempts will not try as hard until there is an
1694 * overall state change.
1696 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1697 node_set(nid
, *node_alloc_noretry
);
1703 * Common helper to allocate a fresh hugetlb page. All specific allocators
1704 * should use this function to get new hugetlb pages
1706 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1707 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1708 nodemask_t
*node_alloc_noretry
)
1712 if (hstate_is_gigantic(h
))
1713 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1715 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1716 nid
, nmask
, node_alloc_noretry
);
1720 if (hstate_is_gigantic(h
))
1721 prep_compound_gigantic_page(page
, huge_page_order(h
));
1722 prep_new_huge_page(h
, page
, page_to_nid(page
));
1728 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1731 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1732 nodemask_t
*node_alloc_noretry
)
1736 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1738 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1739 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1740 node_alloc_noretry
);
1748 put_page(page
); /* free it into the hugepage allocator */
1754 * Free huge page from pool from next node to free.
1755 * Attempt to keep persistent huge pages more or less
1756 * balanced over allowed nodes.
1757 * Called with hugetlb_lock locked.
1759 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1765 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1767 * If we're returning unused surplus pages, only examine
1768 * nodes with surplus pages.
1770 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1771 !list_empty(&h
->hugepage_freelists
[node
])) {
1773 list_entry(h
->hugepage_freelists
[node
].next
,
1775 list_del(&page
->lru
);
1776 h
->free_huge_pages
--;
1777 h
->free_huge_pages_node
[node
]--;
1779 h
->surplus_huge_pages
--;
1780 h
->surplus_huge_pages_node
[node
]--;
1782 update_and_free_page(h
, page
);
1792 * Dissolve a given free hugepage into free buddy pages. This function does
1793 * nothing for in-use hugepages and non-hugepages.
1794 * This function returns values like below:
1796 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1797 * (allocated or reserved.)
1798 * 0: successfully dissolved free hugepages or the page is not a
1799 * hugepage (considered as already dissolved)
1801 int dissolve_free_huge_page(struct page
*page
)
1806 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1807 if (!PageHuge(page
))
1810 spin_lock(&hugetlb_lock
);
1811 if (!PageHuge(page
)) {
1816 if (!page_count(page
)) {
1817 struct page
*head
= compound_head(page
);
1818 struct hstate
*h
= page_hstate(head
);
1819 int nid
= page_to_nid(head
);
1820 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1824 * We should make sure that the page is already on the free list
1825 * when it is dissolved.
1827 if (unlikely(!PageHugeFreed(head
))) {
1828 spin_unlock(&hugetlb_lock
);
1832 * Theoretically, we should return -EBUSY when we
1833 * encounter this race. In fact, we have a chance
1834 * to successfully dissolve the page if we do a
1835 * retry. Because the race window is quite small.
1836 * If we seize this opportunity, it is an optimization
1837 * for increasing the success rate of dissolving page.
1843 * Move PageHWPoison flag from head page to the raw error page,
1844 * which makes any subpages rather than the error page reusable.
1846 if (PageHWPoison(head
) && page
!= head
) {
1847 SetPageHWPoison(page
);
1848 ClearPageHWPoison(head
);
1850 list_del(&head
->lru
);
1851 h
->free_huge_pages
--;
1852 h
->free_huge_pages_node
[nid
]--;
1853 h
->max_huge_pages
--;
1854 update_and_free_page(h
, head
);
1858 spin_unlock(&hugetlb_lock
);
1863 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1864 * make specified memory blocks removable from the system.
1865 * Note that this will dissolve a free gigantic hugepage completely, if any
1866 * part of it lies within the given range.
1867 * Also note that if dissolve_free_huge_page() returns with an error, all
1868 * free hugepages that were dissolved before that error are lost.
1870 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1876 if (!hugepages_supported())
1879 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1880 page
= pfn_to_page(pfn
);
1881 rc
= dissolve_free_huge_page(page
);
1890 * Allocates a fresh surplus page from the page allocator.
1892 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1893 int nid
, nodemask_t
*nmask
)
1895 struct page
*page
= NULL
;
1897 if (hstate_is_gigantic(h
))
1900 spin_lock(&hugetlb_lock
);
1901 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
1903 spin_unlock(&hugetlb_lock
);
1905 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1909 spin_lock(&hugetlb_lock
);
1911 * We could have raced with the pool size change.
1912 * Double check that and simply deallocate the new page
1913 * if we would end up overcommiting the surpluses. Abuse
1914 * temporary page to workaround the nasty free_huge_page
1917 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1918 SetPageHugeTemporary(page
);
1919 spin_unlock(&hugetlb_lock
);
1923 h
->surplus_huge_pages
++;
1924 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
1928 spin_unlock(&hugetlb_lock
);
1933 static struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1934 int nid
, nodemask_t
*nmask
)
1938 if (hstate_is_gigantic(h
))
1941 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1946 * We do not account these pages as surplus because they are only
1947 * temporary and will be released properly on the last reference
1949 SetPageHugeTemporary(page
);
1955 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1958 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1959 struct vm_area_struct
*vma
, unsigned long addr
)
1962 struct mempolicy
*mpol
;
1963 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1965 nodemask_t
*nodemask
;
1967 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
1968 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
);
1969 mpol_cond_put(mpol
);
1974 /* page migration callback function */
1975 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
1976 nodemask_t
*nmask
, gfp_t gfp_mask
)
1978 spin_lock(&hugetlb_lock
);
1979 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1982 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
1984 spin_unlock(&hugetlb_lock
);
1988 spin_unlock(&hugetlb_lock
);
1990 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
1993 /* mempolicy aware migration callback */
1994 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
1995 unsigned long address
)
1997 struct mempolicy
*mpol
;
1998 nodemask_t
*nodemask
;
2003 gfp_mask
= htlb_alloc_mask(h
);
2004 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
2005 page
= alloc_huge_page_nodemask(h
, node
, nodemask
, gfp_mask
);
2006 mpol_cond_put(mpol
);
2012 * Increase the hugetlb pool such that it can accommodate a reservation
2015 static int gather_surplus_pages(struct hstate
*h
, long delta
)
2016 __must_hold(&hugetlb_lock
)
2018 struct list_head surplus_list
;
2019 struct page
*page
, *tmp
;
2022 long needed
, allocated
;
2023 bool alloc_ok
= true;
2025 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
2027 h
->resv_huge_pages
+= delta
;
2032 INIT_LIST_HEAD(&surplus_list
);
2036 spin_unlock(&hugetlb_lock
);
2037 for (i
= 0; i
< needed
; i
++) {
2038 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
2039 NUMA_NO_NODE
, NULL
);
2044 list_add(&page
->lru
, &surplus_list
);
2050 * After retaking hugetlb_lock, we need to recalculate 'needed'
2051 * because either resv_huge_pages or free_huge_pages may have changed.
2053 spin_lock(&hugetlb_lock
);
2054 needed
= (h
->resv_huge_pages
+ delta
) -
2055 (h
->free_huge_pages
+ allocated
);
2060 * We were not able to allocate enough pages to
2061 * satisfy the entire reservation so we free what
2062 * we've allocated so far.
2067 * The surplus_list now contains _at_least_ the number of extra pages
2068 * needed to accommodate the reservation. Add the appropriate number
2069 * of pages to the hugetlb pool and free the extras back to the buddy
2070 * allocator. Commit the entire reservation here to prevent another
2071 * process from stealing the pages as they are added to the pool but
2072 * before they are reserved.
2074 needed
+= allocated
;
2075 h
->resv_huge_pages
+= delta
;
2078 /* Free the needed pages to the hugetlb pool */
2079 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
2085 * This page is now managed by the hugetlb allocator and has
2086 * no users -- drop the buddy allocator's reference.
2088 zeroed
= put_page_testzero(page
);
2089 VM_BUG_ON_PAGE(!zeroed
, page
);
2090 enqueue_huge_page(h
, page
);
2093 spin_unlock(&hugetlb_lock
);
2095 /* Free unnecessary surplus pages to the buddy allocator */
2096 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
2098 spin_lock(&hugetlb_lock
);
2104 * This routine has two main purposes:
2105 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2106 * in unused_resv_pages. This corresponds to the prior adjustments made
2107 * to the associated reservation map.
2108 * 2) Free any unused surplus pages that may have been allocated to satisfy
2109 * the reservation. As many as unused_resv_pages may be freed.
2111 * Called with hugetlb_lock held. However, the lock could be dropped (and
2112 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2113 * we must make sure nobody else can claim pages we are in the process of
2114 * freeing. Do this by ensuring resv_huge_page always is greater than the
2115 * number of huge pages we plan to free when dropping the lock.
2117 static void return_unused_surplus_pages(struct hstate
*h
,
2118 unsigned long unused_resv_pages
)
2120 unsigned long nr_pages
;
2122 /* Cannot return gigantic pages currently */
2123 if (hstate_is_gigantic(h
))
2127 * Part (or even all) of the reservation could have been backed
2128 * by pre-allocated pages. Only free surplus pages.
2130 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
2133 * We want to release as many surplus pages as possible, spread
2134 * evenly across all nodes with memory. Iterate across these nodes
2135 * until we can no longer free unreserved surplus pages. This occurs
2136 * when the nodes with surplus pages have no free pages.
2137 * free_pool_huge_page() will balance the freed pages across the
2138 * on-line nodes with memory and will handle the hstate accounting.
2140 * Note that we decrement resv_huge_pages as we free the pages. If
2141 * we drop the lock, resv_huge_pages will still be sufficiently large
2142 * to cover subsequent pages we may free.
2144 while (nr_pages
--) {
2145 h
->resv_huge_pages
--;
2146 unused_resv_pages
--;
2147 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
2149 cond_resched_lock(&hugetlb_lock
);
2153 /* Fully uncommit the reservation */
2154 h
->resv_huge_pages
-= unused_resv_pages
;
2159 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2160 * are used by the huge page allocation routines to manage reservations.
2162 * vma_needs_reservation is called to determine if the huge page at addr
2163 * within the vma has an associated reservation. If a reservation is
2164 * needed, the value 1 is returned. The caller is then responsible for
2165 * managing the global reservation and subpool usage counts. After
2166 * the huge page has been allocated, vma_commit_reservation is called
2167 * to add the page to the reservation map. If the page allocation fails,
2168 * the reservation must be ended instead of committed. vma_end_reservation
2169 * is called in such cases.
2171 * In the normal case, vma_commit_reservation returns the same value
2172 * as the preceding vma_needs_reservation call. The only time this
2173 * is not the case is if a reserve map was changed between calls. It
2174 * is the responsibility of the caller to notice the difference and
2175 * take appropriate action.
2177 * vma_add_reservation is used in error paths where a reservation must
2178 * be restored when a newly allocated huge page must be freed. It is
2179 * to be called after calling vma_needs_reservation to determine if a
2180 * reservation exists.
2182 enum vma_resv_mode
{
2188 static long __vma_reservation_common(struct hstate
*h
,
2189 struct vm_area_struct
*vma
, unsigned long addr
,
2190 enum vma_resv_mode mode
)
2192 struct resv_map
*resv
;
2195 long dummy_out_regions_needed
;
2197 resv
= vma_resv_map(vma
);
2201 idx
= vma_hugecache_offset(h
, vma
, addr
);
2203 case VMA_NEEDS_RESV
:
2204 ret
= region_chg(resv
, idx
, idx
+ 1, &dummy_out_regions_needed
);
2205 /* We assume that vma_reservation_* routines always operate on
2206 * 1 page, and that adding to resv map a 1 page entry can only
2207 * ever require 1 region.
2209 VM_BUG_ON(dummy_out_regions_needed
!= 1);
2211 case VMA_COMMIT_RESV
:
2212 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2213 /* region_add calls of range 1 should never fail. */
2217 region_abort(resv
, idx
, idx
+ 1, 1);
2221 if (vma
->vm_flags
& VM_MAYSHARE
) {
2222 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2223 /* region_add calls of range 1 should never fail. */
2226 region_abort(resv
, idx
, idx
+ 1, 1);
2227 ret
= region_del(resv
, idx
, idx
+ 1);
2234 if (vma
->vm_flags
& VM_MAYSHARE
)
2236 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
2238 * In most cases, reserves always exist for private mappings.
2239 * However, a file associated with mapping could have been
2240 * hole punched or truncated after reserves were consumed.
2241 * As subsequent fault on such a range will not use reserves.
2242 * Subtle - The reserve map for private mappings has the
2243 * opposite meaning than that of shared mappings. If NO
2244 * entry is in the reserve map, it means a reservation exists.
2245 * If an entry exists in the reserve map, it means the
2246 * reservation has already been consumed. As a result, the
2247 * return value of this routine is the opposite of the
2248 * value returned from reserve map manipulation routines above.
2256 return ret
< 0 ? ret
: 0;
2259 static long vma_needs_reservation(struct hstate
*h
,
2260 struct vm_area_struct
*vma
, unsigned long addr
)
2262 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2265 static long vma_commit_reservation(struct hstate
*h
,
2266 struct vm_area_struct
*vma
, unsigned long addr
)
2268 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2271 static void vma_end_reservation(struct hstate
*h
,
2272 struct vm_area_struct
*vma
, unsigned long addr
)
2274 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2277 static long vma_add_reservation(struct hstate
*h
,
2278 struct vm_area_struct
*vma
, unsigned long addr
)
2280 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2284 * This routine is called to restore a reservation on error paths. In the
2285 * specific error paths, a huge page was allocated (via alloc_huge_page)
2286 * and is about to be freed. If a reservation for the page existed,
2287 * alloc_huge_page would have consumed the reservation and set PagePrivate
2288 * in the newly allocated page. When the page is freed via free_huge_page,
2289 * the global reservation count will be incremented if PagePrivate is set.
2290 * However, free_huge_page can not adjust the reserve map. Adjust the
2291 * reserve map here to be consistent with global reserve count adjustments
2292 * to be made by free_huge_page.
2294 static void restore_reserve_on_error(struct hstate
*h
,
2295 struct vm_area_struct
*vma
, unsigned long address
,
2298 if (unlikely(PagePrivate(page
))) {
2299 long rc
= vma_needs_reservation(h
, vma
, address
);
2301 if (unlikely(rc
< 0)) {
2303 * Rare out of memory condition in reserve map
2304 * manipulation. Clear PagePrivate so that
2305 * global reserve count will not be incremented
2306 * by free_huge_page. This will make it appear
2307 * as though the reservation for this page was
2308 * consumed. This may prevent the task from
2309 * faulting in the page at a later time. This
2310 * is better than inconsistent global huge page
2311 * accounting of reserve counts.
2313 ClearPagePrivate(page
);
2315 rc
= vma_add_reservation(h
, vma
, address
);
2316 if (unlikely(rc
< 0))
2318 * See above comment about rare out of
2321 ClearPagePrivate(page
);
2323 vma_end_reservation(h
, vma
, address
);
2327 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2328 unsigned long addr
, int avoid_reserve
)
2330 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2331 struct hstate
*h
= hstate_vma(vma
);
2333 long map_chg
, map_commit
;
2336 struct hugetlb_cgroup
*h_cg
;
2337 bool deferred_reserve
;
2339 idx
= hstate_index(h
);
2341 * Examine the region/reserve map to determine if the process
2342 * has a reservation for the page to be allocated. A return
2343 * code of zero indicates a reservation exists (no change).
2345 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2347 return ERR_PTR(-ENOMEM
);
2350 * Processes that did not create the mapping will have no
2351 * reserves as indicated by the region/reserve map. Check
2352 * that the allocation will not exceed the subpool limit.
2353 * Allocations for MAP_NORESERVE mappings also need to be
2354 * checked against any subpool limit.
2356 if (map_chg
|| avoid_reserve
) {
2357 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2359 vma_end_reservation(h
, vma
, addr
);
2360 return ERR_PTR(-ENOSPC
);
2364 * Even though there was no reservation in the region/reserve
2365 * map, there could be reservations associated with the
2366 * subpool that can be used. This would be indicated if the
2367 * return value of hugepage_subpool_get_pages() is zero.
2368 * However, if avoid_reserve is specified we still avoid even
2369 * the subpool reservations.
2375 /* If this allocation is not consuming a reservation, charge it now.
2377 deferred_reserve
= map_chg
|| avoid_reserve
|| !vma_resv_map(vma
);
2378 if (deferred_reserve
) {
2379 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
2380 idx
, pages_per_huge_page(h
), &h_cg
);
2382 goto out_subpool_put
;
2385 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2387 goto out_uncharge_cgroup_reservation
;
2389 spin_lock(&hugetlb_lock
);
2391 * glb_chg is passed to indicate whether or not a page must be taken
2392 * from the global free pool (global change). gbl_chg == 0 indicates
2393 * a reservation exists for the allocation.
2395 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2397 spin_unlock(&hugetlb_lock
);
2398 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2400 goto out_uncharge_cgroup
;
2401 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2402 SetPagePrivate(page
);
2403 h
->resv_huge_pages
--;
2405 spin_lock(&hugetlb_lock
);
2406 list_add(&page
->lru
, &h
->hugepage_activelist
);
2409 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2410 /* If allocation is not consuming a reservation, also store the
2411 * hugetlb_cgroup pointer on the page.
2413 if (deferred_reserve
) {
2414 hugetlb_cgroup_commit_charge_rsvd(idx
, pages_per_huge_page(h
),
2418 spin_unlock(&hugetlb_lock
);
2420 set_page_private(page
, (unsigned long)spool
);
2422 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2423 if (unlikely(map_chg
> map_commit
)) {
2425 * The page was added to the reservation map between
2426 * vma_needs_reservation and vma_commit_reservation.
2427 * This indicates a race with hugetlb_reserve_pages.
2428 * Adjust for the subpool count incremented above AND
2429 * in hugetlb_reserve_pages for the same page. Also,
2430 * the reservation count added in hugetlb_reserve_pages
2431 * no longer applies.
2435 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2436 hugetlb_acct_memory(h
, -rsv_adjust
);
2437 if (deferred_reserve
)
2438 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
2439 pages_per_huge_page(h
), page
);
2443 out_uncharge_cgroup
:
2444 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2445 out_uncharge_cgroup_reservation
:
2446 if (deferred_reserve
)
2447 hugetlb_cgroup_uncharge_cgroup_rsvd(idx
, pages_per_huge_page(h
),
2450 if (map_chg
|| avoid_reserve
)
2451 hugepage_subpool_put_pages(spool
, 1);
2452 vma_end_reservation(h
, vma
, addr
);
2453 return ERR_PTR(-ENOSPC
);
2456 int alloc_bootmem_huge_page(struct hstate
*h
)
2457 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2458 int __alloc_bootmem_huge_page(struct hstate
*h
)
2460 struct huge_bootmem_page
*m
;
2463 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2466 addr
= memblock_alloc_try_nid_raw(
2467 huge_page_size(h
), huge_page_size(h
),
2468 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2471 * Use the beginning of the huge page to store the
2472 * huge_bootmem_page struct (until gather_bootmem
2473 * puts them into the mem_map).
2482 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2483 /* Put them into a private list first because mem_map is not up yet */
2484 INIT_LIST_HEAD(&m
->list
);
2485 list_add(&m
->list
, &huge_boot_pages
);
2490 static void __init
prep_compound_huge_page(struct page
*page
,
2493 if (unlikely(order
> (MAX_ORDER
- 1)))
2494 prep_compound_gigantic_page(page
, order
);
2496 prep_compound_page(page
, order
);
2499 /* Put bootmem huge pages into the standard lists after mem_map is up */
2500 static void __init
gather_bootmem_prealloc(void)
2502 struct huge_bootmem_page
*m
;
2504 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2505 struct page
*page
= virt_to_page(m
);
2506 struct hstate
*h
= m
->hstate
;
2508 WARN_ON(page_count(page
) != 1);
2509 prep_compound_huge_page(page
, h
->order
);
2510 WARN_ON(PageReserved(page
));
2511 prep_new_huge_page(h
, page
, page_to_nid(page
));
2512 put_page(page
); /* free it into the hugepage allocator */
2515 * If we had gigantic hugepages allocated at boot time, we need
2516 * to restore the 'stolen' pages to totalram_pages in order to
2517 * fix confusing memory reports from free(1) and another
2518 * side-effects, like CommitLimit going negative.
2520 if (hstate_is_gigantic(h
))
2521 adjust_managed_page_count(page
, 1 << h
->order
);
2526 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2529 nodemask_t
*node_alloc_noretry
;
2531 if (!hstate_is_gigantic(h
)) {
2533 * Bit mask controlling how hard we retry per-node allocations.
2534 * Ignore errors as lower level routines can deal with
2535 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2536 * time, we are likely in bigger trouble.
2538 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2541 /* allocations done at boot time */
2542 node_alloc_noretry
= NULL
;
2545 /* bit mask controlling how hard we retry per-node allocations */
2546 if (node_alloc_noretry
)
2547 nodes_clear(*node_alloc_noretry
);
2549 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2550 if (hstate_is_gigantic(h
)) {
2551 if (hugetlb_cma_size
) {
2552 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2555 if (!alloc_bootmem_huge_page(h
))
2557 } else if (!alloc_pool_huge_page(h
,
2558 &node_states
[N_MEMORY
],
2559 node_alloc_noretry
))
2563 if (i
< h
->max_huge_pages
) {
2566 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2567 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2568 h
->max_huge_pages
, buf
, i
);
2569 h
->max_huge_pages
= i
;
2572 kfree(node_alloc_noretry
);
2575 static void __init
hugetlb_init_hstates(void)
2579 for_each_hstate(h
) {
2580 if (minimum_order
> huge_page_order(h
))
2581 minimum_order
= huge_page_order(h
);
2583 /* oversize hugepages were init'ed in early boot */
2584 if (!hstate_is_gigantic(h
))
2585 hugetlb_hstate_alloc_pages(h
);
2587 VM_BUG_ON(minimum_order
== UINT_MAX
);
2590 static void __init
report_hugepages(void)
2594 for_each_hstate(h
) {
2597 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2598 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2599 buf
, h
->free_huge_pages
);
2603 #ifdef CONFIG_HIGHMEM
2604 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2605 nodemask_t
*nodes_allowed
)
2609 if (hstate_is_gigantic(h
))
2612 for_each_node_mask(i
, *nodes_allowed
) {
2613 struct page
*page
, *next
;
2614 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2615 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2616 if (count
>= h
->nr_huge_pages
)
2618 if (PageHighMem(page
))
2620 list_del(&page
->lru
);
2621 update_and_free_page(h
, page
);
2622 h
->free_huge_pages
--;
2623 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2628 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2629 nodemask_t
*nodes_allowed
)
2635 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2636 * balanced by operating on them in a round-robin fashion.
2637 * Returns 1 if an adjustment was made.
2639 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2644 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2647 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2648 if (h
->surplus_huge_pages_node
[node
])
2652 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2653 if (h
->surplus_huge_pages_node
[node
] <
2654 h
->nr_huge_pages_node
[node
])
2661 h
->surplus_huge_pages
+= delta
;
2662 h
->surplus_huge_pages_node
[node
] += delta
;
2666 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2667 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
2668 nodemask_t
*nodes_allowed
)
2670 unsigned long min_count
, ret
;
2671 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
2674 * Bit mask controlling how hard we retry per-node allocations.
2675 * If we can not allocate the bit mask, do not attempt to allocate
2676 * the requested huge pages.
2678 if (node_alloc_noretry
)
2679 nodes_clear(*node_alloc_noretry
);
2683 spin_lock(&hugetlb_lock
);
2686 * Check for a node specific request.
2687 * Changing node specific huge page count may require a corresponding
2688 * change to the global count. In any case, the passed node mask
2689 * (nodes_allowed) will restrict alloc/free to the specified node.
2691 if (nid
!= NUMA_NO_NODE
) {
2692 unsigned long old_count
= count
;
2694 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2696 * User may have specified a large count value which caused the
2697 * above calculation to overflow. In this case, they wanted
2698 * to allocate as many huge pages as possible. Set count to
2699 * largest possible value to align with their intention.
2701 if (count
< old_count
)
2706 * Gigantic pages runtime allocation depend on the capability for large
2707 * page range allocation.
2708 * If the system does not provide this feature, return an error when
2709 * the user tries to allocate gigantic pages but let the user free the
2710 * boottime allocated gigantic pages.
2712 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
2713 if (count
> persistent_huge_pages(h
)) {
2714 spin_unlock(&hugetlb_lock
);
2715 NODEMASK_FREE(node_alloc_noretry
);
2718 /* Fall through to decrease pool */
2722 * Increase the pool size
2723 * First take pages out of surplus state. Then make up the
2724 * remaining difference by allocating fresh huge pages.
2726 * We might race with alloc_surplus_huge_page() here and be unable
2727 * to convert a surplus huge page to a normal huge page. That is
2728 * not critical, though, it just means the overall size of the
2729 * pool might be one hugepage larger than it needs to be, but
2730 * within all the constraints specified by the sysctls.
2732 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2733 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2737 while (count
> persistent_huge_pages(h
)) {
2739 * If this allocation races such that we no longer need the
2740 * page, free_huge_page will handle it by freeing the page
2741 * and reducing the surplus.
2743 spin_unlock(&hugetlb_lock
);
2745 /* yield cpu to avoid soft lockup */
2748 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
2749 node_alloc_noretry
);
2750 spin_lock(&hugetlb_lock
);
2754 /* Bail for signals. Probably ctrl-c from user */
2755 if (signal_pending(current
))
2760 * Decrease the pool size
2761 * First return free pages to the buddy allocator (being careful
2762 * to keep enough around to satisfy reservations). Then place
2763 * pages into surplus state as needed so the pool will shrink
2764 * to the desired size as pages become free.
2766 * By placing pages into the surplus state independent of the
2767 * overcommit value, we are allowing the surplus pool size to
2768 * exceed overcommit. There are few sane options here. Since
2769 * alloc_surplus_huge_page() is checking the global counter,
2770 * though, we'll note that we're not allowed to exceed surplus
2771 * and won't grow the pool anywhere else. Not until one of the
2772 * sysctls are changed, or the surplus pages go out of use.
2774 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2775 min_count
= max(count
, min_count
);
2776 try_to_free_low(h
, min_count
, nodes_allowed
);
2777 while (min_count
< persistent_huge_pages(h
)) {
2778 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2780 cond_resched_lock(&hugetlb_lock
);
2782 while (count
< persistent_huge_pages(h
)) {
2783 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2787 h
->max_huge_pages
= persistent_huge_pages(h
);
2788 spin_unlock(&hugetlb_lock
);
2790 NODEMASK_FREE(node_alloc_noretry
);
2795 #define HSTATE_ATTR_RO(_name) \
2796 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2798 #define HSTATE_ATTR(_name) \
2799 static struct kobj_attribute _name##_attr = \
2800 __ATTR(_name, 0644, _name##_show, _name##_store)
2802 static struct kobject
*hugepages_kobj
;
2803 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2805 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2807 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2811 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2812 if (hstate_kobjs
[i
] == kobj
) {
2814 *nidp
= NUMA_NO_NODE
;
2818 return kobj_to_node_hstate(kobj
, nidp
);
2821 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2822 struct kobj_attribute
*attr
, char *buf
)
2825 unsigned long nr_huge_pages
;
2828 h
= kobj_to_hstate(kobj
, &nid
);
2829 if (nid
== NUMA_NO_NODE
)
2830 nr_huge_pages
= h
->nr_huge_pages
;
2832 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2834 return sysfs_emit(buf
, "%lu\n", nr_huge_pages
);
2837 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2838 struct hstate
*h
, int nid
,
2839 unsigned long count
, size_t len
)
2842 nodemask_t nodes_allowed
, *n_mask
;
2844 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
2847 if (nid
== NUMA_NO_NODE
) {
2849 * global hstate attribute
2851 if (!(obey_mempolicy
&&
2852 init_nodemask_of_mempolicy(&nodes_allowed
)))
2853 n_mask
= &node_states
[N_MEMORY
];
2855 n_mask
= &nodes_allowed
;
2858 * Node specific request. count adjustment happens in
2859 * set_max_huge_pages() after acquiring hugetlb_lock.
2861 init_nodemask_of_node(&nodes_allowed
, nid
);
2862 n_mask
= &nodes_allowed
;
2865 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
2867 return err
? err
: len
;
2870 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2871 struct kobject
*kobj
, const char *buf
,
2875 unsigned long count
;
2879 err
= kstrtoul(buf
, 10, &count
);
2883 h
= kobj_to_hstate(kobj
, &nid
);
2884 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2887 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2888 struct kobj_attribute
*attr
, char *buf
)
2890 return nr_hugepages_show_common(kobj
, attr
, buf
);
2893 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2894 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2896 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2898 HSTATE_ATTR(nr_hugepages
);
2903 * hstate attribute for optionally mempolicy-based constraint on persistent
2904 * huge page alloc/free.
2906 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2907 struct kobj_attribute
*attr
,
2910 return nr_hugepages_show_common(kobj
, attr
, buf
);
2913 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2914 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2916 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2918 HSTATE_ATTR(nr_hugepages_mempolicy
);
2922 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2923 struct kobj_attribute
*attr
, char *buf
)
2925 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2926 return sysfs_emit(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2929 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2930 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2933 unsigned long input
;
2934 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2936 if (hstate_is_gigantic(h
))
2939 err
= kstrtoul(buf
, 10, &input
);
2943 spin_lock(&hugetlb_lock
);
2944 h
->nr_overcommit_huge_pages
= input
;
2945 spin_unlock(&hugetlb_lock
);
2949 HSTATE_ATTR(nr_overcommit_hugepages
);
2951 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2952 struct kobj_attribute
*attr
, char *buf
)
2955 unsigned long free_huge_pages
;
2958 h
= kobj_to_hstate(kobj
, &nid
);
2959 if (nid
== NUMA_NO_NODE
)
2960 free_huge_pages
= h
->free_huge_pages
;
2962 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2964 return sysfs_emit(buf
, "%lu\n", free_huge_pages
);
2966 HSTATE_ATTR_RO(free_hugepages
);
2968 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2969 struct kobj_attribute
*attr
, char *buf
)
2971 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2972 return sysfs_emit(buf
, "%lu\n", h
->resv_huge_pages
);
2974 HSTATE_ATTR_RO(resv_hugepages
);
2976 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2977 struct kobj_attribute
*attr
, char *buf
)
2980 unsigned long surplus_huge_pages
;
2983 h
= kobj_to_hstate(kobj
, &nid
);
2984 if (nid
== NUMA_NO_NODE
)
2985 surplus_huge_pages
= h
->surplus_huge_pages
;
2987 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2989 return sysfs_emit(buf
, "%lu\n", surplus_huge_pages
);
2991 HSTATE_ATTR_RO(surplus_hugepages
);
2993 static struct attribute
*hstate_attrs
[] = {
2994 &nr_hugepages_attr
.attr
,
2995 &nr_overcommit_hugepages_attr
.attr
,
2996 &free_hugepages_attr
.attr
,
2997 &resv_hugepages_attr
.attr
,
2998 &surplus_hugepages_attr
.attr
,
3000 &nr_hugepages_mempolicy_attr
.attr
,
3005 static const struct attribute_group hstate_attr_group
= {
3006 .attrs
= hstate_attrs
,
3009 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
3010 struct kobject
**hstate_kobjs
,
3011 const struct attribute_group
*hstate_attr_group
)
3014 int hi
= hstate_index(h
);
3016 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
3017 if (!hstate_kobjs
[hi
])
3020 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
3022 kobject_put(hstate_kobjs
[hi
]);
3023 hstate_kobjs
[hi
] = NULL
;
3029 static void __init
hugetlb_sysfs_init(void)
3034 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
3035 if (!hugepages_kobj
)
3038 for_each_hstate(h
) {
3039 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
3040 hstate_kobjs
, &hstate_attr_group
);
3042 pr_err("HugeTLB: Unable to add hstate %s", h
->name
);
3049 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3050 * with node devices in node_devices[] using a parallel array. The array
3051 * index of a node device or _hstate == node id.
3052 * This is here to avoid any static dependency of the node device driver, in
3053 * the base kernel, on the hugetlb module.
3055 struct node_hstate
{
3056 struct kobject
*hugepages_kobj
;
3057 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
3059 static struct node_hstate node_hstates
[MAX_NUMNODES
];
3062 * A subset of global hstate attributes for node devices
3064 static struct attribute
*per_node_hstate_attrs
[] = {
3065 &nr_hugepages_attr
.attr
,
3066 &free_hugepages_attr
.attr
,
3067 &surplus_hugepages_attr
.attr
,
3071 static const struct attribute_group per_node_hstate_attr_group
= {
3072 .attrs
= per_node_hstate_attrs
,
3076 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3077 * Returns node id via non-NULL nidp.
3079 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3083 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
3084 struct node_hstate
*nhs
= &node_hstates
[nid
];
3086 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
3087 if (nhs
->hstate_kobjs
[i
] == kobj
) {
3099 * Unregister hstate attributes from a single node device.
3100 * No-op if no hstate attributes attached.
3102 static void hugetlb_unregister_node(struct node
*node
)
3105 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3107 if (!nhs
->hugepages_kobj
)
3108 return; /* no hstate attributes */
3110 for_each_hstate(h
) {
3111 int idx
= hstate_index(h
);
3112 if (nhs
->hstate_kobjs
[idx
]) {
3113 kobject_put(nhs
->hstate_kobjs
[idx
]);
3114 nhs
->hstate_kobjs
[idx
] = NULL
;
3118 kobject_put(nhs
->hugepages_kobj
);
3119 nhs
->hugepages_kobj
= NULL
;
3124 * Register hstate attributes for a single node device.
3125 * No-op if attributes already registered.
3127 static void hugetlb_register_node(struct node
*node
)
3130 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3133 if (nhs
->hugepages_kobj
)
3134 return; /* already allocated */
3136 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
3138 if (!nhs
->hugepages_kobj
)
3141 for_each_hstate(h
) {
3142 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
3144 &per_node_hstate_attr_group
);
3146 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3147 h
->name
, node
->dev
.id
);
3148 hugetlb_unregister_node(node
);
3155 * hugetlb init time: register hstate attributes for all registered node
3156 * devices of nodes that have memory. All on-line nodes should have
3157 * registered their associated device by this time.
3159 static void __init
hugetlb_register_all_nodes(void)
3163 for_each_node_state(nid
, N_MEMORY
) {
3164 struct node
*node
= node_devices
[nid
];
3165 if (node
->dev
.id
== nid
)
3166 hugetlb_register_node(node
);
3170 * Let the node device driver know we're here so it can
3171 * [un]register hstate attributes on node hotplug.
3173 register_hugetlbfs_with_node(hugetlb_register_node
,
3174 hugetlb_unregister_node
);
3176 #else /* !CONFIG_NUMA */
3178 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3186 static void hugetlb_register_all_nodes(void) { }
3190 static int __init
hugetlb_init(void)
3194 if (!hugepages_supported()) {
3195 if (hugetlb_max_hstate
|| default_hstate_max_huge_pages
)
3196 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3201 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3202 * architectures depend on setup being done here.
3204 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
3205 if (!parsed_default_hugepagesz
) {
3207 * If we did not parse a default huge page size, set
3208 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3209 * number of huge pages for this default size was implicitly
3210 * specified, set that here as well.
3211 * Note that the implicit setting will overwrite an explicit
3212 * setting. A warning will be printed in this case.
3214 default_hstate_idx
= hstate_index(size_to_hstate(HPAGE_SIZE
));
3215 if (default_hstate_max_huge_pages
) {
3216 if (default_hstate
.max_huge_pages
) {
3219 string_get_size(huge_page_size(&default_hstate
),
3220 1, STRING_UNITS_2
, buf
, 32);
3221 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3222 default_hstate
.max_huge_pages
, buf
);
3223 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3224 default_hstate_max_huge_pages
);
3226 default_hstate
.max_huge_pages
=
3227 default_hstate_max_huge_pages
;
3231 hugetlb_cma_check();
3232 hugetlb_init_hstates();
3233 gather_bootmem_prealloc();
3236 hugetlb_sysfs_init();
3237 hugetlb_register_all_nodes();
3238 hugetlb_cgroup_file_init();
3241 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
3243 num_fault_mutexes
= 1;
3245 hugetlb_fault_mutex_table
=
3246 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
3248 BUG_ON(!hugetlb_fault_mutex_table
);
3250 for (i
= 0; i
< num_fault_mutexes
; i
++)
3251 mutex_init(&hugetlb_fault_mutex_table
[i
]);
3254 subsys_initcall(hugetlb_init
);
3256 /* Overwritten by architectures with more huge page sizes */
3257 bool __init
__attribute((weak
)) arch_hugetlb_valid_size(unsigned long size
)
3259 return size
== HPAGE_SIZE
;
3262 void __init
hugetlb_add_hstate(unsigned int order
)
3267 if (size_to_hstate(PAGE_SIZE
<< order
)) {
3270 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
3272 h
= &hstates
[hugetlb_max_hstate
++];
3274 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
3275 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
3276 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
3277 INIT_LIST_HEAD(&h
->hugepage_activelist
);
3278 h
->next_nid_to_alloc
= first_memory_node
;
3279 h
->next_nid_to_free
= first_memory_node
;
3280 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
3281 huge_page_size(h
)/1024);
3287 * hugepages command line processing
3288 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3289 * specification. If not, ignore the hugepages value. hugepages can also
3290 * be the first huge page command line option in which case it implicitly
3291 * specifies the number of huge pages for the default size.
3293 static int __init
hugepages_setup(char *s
)
3296 static unsigned long *last_mhp
;
3298 if (!parsed_valid_hugepagesz
) {
3299 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s
);
3300 parsed_valid_hugepagesz
= true;
3305 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3306 * yet, so this hugepages= parameter goes to the "default hstate".
3307 * Otherwise, it goes with the previously parsed hugepagesz or
3308 * default_hugepagesz.
3310 else if (!hugetlb_max_hstate
)
3311 mhp
= &default_hstate_max_huge_pages
;
3313 mhp
= &parsed_hstate
->max_huge_pages
;
3315 if (mhp
== last_mhp
) {
3316 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s
);
3320 if (sscanf(s
, "%lu", mhp
) <= 0)
3324 * Global state is always initialized later in hugetlb_init.
3325 * But we need to allocate >= MAX_ORDER hstates here early to still
3326 * use the bootmem allocator.
3328 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
3329 hugetlb_hstate_alloc_pages(parsed_hstate
);
3335 __setup("hugepages=", hugepages_setup
);
3338 * hugepagesz command line processing
3339 * A specific huge page size can only be specified once with hugepagesz.
3340 * hugepagesz is followed by hugepages on the command line. The global
3341 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3342 * hugepagesz argument was valid.
3344 static int __init
hugepagesz_setup(char *s
)
3349 parsed_valid_hugepagesz
= false;
3350 size
= (unsigned long)memparse(s
, NULL
);
3352 if (!arch_hugetlb_valid_size(size
)) {
3353 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s
);
3357 h
= size_to_hstate(size
);
3360 * hstate for this size already exists. This is normally
3361 * an error, but is allowed if the existing hstate is the
3362 * default hstate. More specifically, it is only allowed if
3363 * the number of huge pages for the default hstate was not
3364 * previously specified.
3366 if (!parsed_default_hugepagesz
|| h
!= &default_hstate
||
3367 default_hstate
.max_huge_pages
) {
3368 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s
);
3373 * No need to call hugetlb_add_hstate() as hstate already
3374 * exists. But, do set parsed_hstate so that a following
3375 * hugepages= parameter will be applied to this hstate.
3378 parsed_valid_hugepagesz
= true;
3382 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3383 parsed_valid_hugepagesz
= true;
3386 __setup("hugepagesz=", hugepagesz_setup
);
3389 * default_hugepagesz command line input
3390 * Only one instance of default_hugepagesz allowed on command line.
3392 static int __init
default_hugepagesz_setup(char *s
)
3396 parsed_valid_hugepagesz
= false;
3397 if (parsed_default_hugepagesz
) {
3398 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s
);
3402 size
= (unsigned long)memparse(s
, NULL
);
3404 if (!arch_hugetlb_valid_size(size
)) {
3405 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s
);
3409 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3410 parsed_valid_hugepagesz
= true;
3411 parsed_default_hugepagesz
= true;
3412 default_hstate_idx
= hstate_index(size_to_hstate(size
));
3415 * The number of default huge pages (for this size) could have been
3416 * specified as the first hugetlb parameter: hugepages=X. If so,
3417 * then default_hstate_max_huge_pages is set. If the default huge
3418 * page size is gigantic (>= MAX_ORDER), then the pages must be
3419 * allocated here from bootmem allocator.
3421 if (default_hstate_max_huge_pages
) {
3422 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
3423 if (hstate_is_gigantic(&default_hstate
))
3424 hugetlb_hstate_alloc_pages(&default_hstate
);
3425 default_hstate_max_huge_pages
= 0;
3430 __setup("default_hugepagesz=", default_hugepagesz_setup
);
3432 static unsigned int allowed_mems_nr(struct hstate
*h
)
3435 unsigned int nr
= 0;
3436 nodemask_t
*mpol_allowed
;
3437 unsigned int *array
= h
->free_huge_pages_node
;
3438 gfp_t gfp_mask
= htlb_alloc_mask(h
);
3440 mpol_allowed
= policy_nodemask_current(gfp_mask
);
3442 for_each_node_mask(node
, cpuset_current_mems_allowed
) {
3443 if (!mpol_allowed
||
3444 (mpol_allowed
&& node_isset(node
, *mpol_allowed
)))
3451 #ifdef CONFIG_SYSCTL
3452 static int proc_hugetlb_doulongvec_minmax(struct ctl_table
*table
, int write
,
3453 void *buffer
, size_t *length
,
3454 loff_t
*ppos
, unsigned long *out
)
3456 struct ctl_table dup_table
;
3459 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3460 * can duplicate the @table and alter the duplicate of it.
3463 dup_table
.data
= out
;
3465 return proc_doulongvec_minmax(&dup_table
, write
, buffer
, length
, ppos
);
3468 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3469 struct ctl_table
*table
, int write
,
3470 void *buffer
, size_t *length
, loff_t
*ppos
)
3472 struct hstate
*h
= &default_hstate
;
3473 unsigned long tmp
= h
->max_huge_pages
;
3476 if (!hugepages_supported())
3479 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3485 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3486 NUMA_NO_NODE
, tmp
, *length
);
3491 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3492 void *buffer
, size_t *length
, loff_t
*ppos
)
3495 return hugetlb_sysctl_handler_common(false, table
, write
,
3496 buffer
, length
, ppos
);
3500 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3501 void *buffer
, size_t *length
, loff_t
*ppos
)
3503 return hugetlb_sysctl_handler_common(true, table
, write
,
3504 buffer
, length
, ppos
);
3506 #endif /* CONFIG_NUMA */
3508 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3509 void *buffer
, size_t *length
, loff_t
*ppos
)
3511 struct hstate
*h
= &default_hstate
;
3515 if (!hugepages_supported())
3518 tmp
= h
->nr_overcommit_huge_pages
;
3520 if (write
&& hstate_is_gigantic(h
))
3523 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3529 spin_lock(&hugetlb_lock
);
3530 h
->nr_overcommit_huge_pages
= tmp
;
3531 spin_unlock(&hugetlb_lock
);
3537 #endif /* CONFIG_SYSCTL */
3539 void hugetlb_report_meminfo(struct seq_file
*m
)
3542 unsigned long total
= 0;
3544 if (!hugepages_supported())
3547 for_each_hstate(h
) {
3548 unsigned long count
= h
->nr_huge_pages
;
3550 total
+= (PAGE_SIZE
<< huge_page_order(h
)) * count
;
3552 if (h
== &default_hstate
)
3554 "HugePages_Total: %5lu\n"
3555 "HugePages_Free: %5lu\n"
3556 "HugePages_Rsvd: %5lu\n"
3557 "HugePages_Surp: %5lu\n"
3558 "Hugepagesize: %8lu kB\n",
3562 h
->surplus_huge_pages
,
3563 (PAGE_SIZE
<< huge_page_order(h
)) / 1024);
3566 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ 1024);
3569 int hugetlb_report_node_meminfo(char *buf
, int len
, int nid
)
3571 struct hstate
*h
= &default_hstate
;
3573 if (!hugepages_supported())
3576 return sysfs_emit_at(buf
, len
,
3577 "Node %d HugePages_Total: %5u\n"
3578 "Node %d HugePages_Free: %5u\n"
3579 "Node %d HugePages_Surp: %5u\n",
3580 nid
, h
->nr_huge_pages_node
[nid
],
3581 nid
, h
->free_huge_pages_node
[nid
],
3582 nid
, h
->surplus_huge_pages_node
[nid
]);
3585 void hugetlb_show_meminfo(void)
3590 if (!hugepages_supported())
3593 for_each_node_state(nid
, N_MEMORY
)
3595 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3597 h
->nr_huge_pages_node
[nid
],
3598 h
->free_huge_pages_node
[nid
],
3599 h
->surplus_huge_pages_node
[nid
],
3600 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3603 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3605 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3606 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3609 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3610 unsigned long hugetlb_total_pages(void)
3613 unsigned long nr_total_pages
= 0;
3616 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3617 return nr_total_pages
;
3620 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3624 spin_lock(&hugetlb_lock
);
3626 * When cpuset is configured, it breaks the strict hugetlb page
3627 * reservation as the accounting is done on a global variable. Such
3628 * reservation is completely rubbish in the presence of cpuset because
3629 * the reservation is not checked against page availability for the
3630 * current cpuset. Application can still potentially OOM'ed by kernel
3631 * with lack of free htlb page in cpuset that the task is in.
3632 * Attempt to enforce strict accounting with cpuset is almost
3633 * impossible (or too ugly) because cpuset is too fluid that
3634 * task or memory node can be dynamically moved between cpusets.
3636 * The change of semantics for shared hugetlb mapping with cpuset is
3637 * undesirable. However, in order to preserve some of the semantics,
3638 * we fall back to check against current free page availability as
3639 * a best attempt and hopefully to minimize the impact of changing
3640 * semantics that cpuset has.
3642 * Apart from cpuset, we also have memory policy mechanism that
3643 * also determines from which node the kernel will allocate memory
3644 * in a NUMA system. So similar to cpuset, we also should consider
3645 * the memory policy of the current task. Similar to the description
3649 if (gather_surplus_pages(h
, delta
) < 0)
3652 if (delta
> allowed_mems_nr(h
)) {
3653 return_unused_surplus_pages(h
, delta
);
3660 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3663 spin_unlock(&hugetlb_lock
);
3667 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3669 struct resv_map
*resv
= vma_resv_map(vma
);
3672 * This new VMA should share its siblings reservation map if present.
3673 * The VMA will only ever have a valid reservation map pointer where
3674 * it is being copied for another still existing VMA. As that VMA
3675 * has a reference to the reservation map it cannot disappear until
3676 * after this open call completes. It is therefore safe to take a
3677 * new reference here without additional locking.
3679 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3680 kref_get(&resv
->refs
);
3683 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3685 struct hstate
*h
= hstate_vma(vma
);
3686 struct resv_map
*resv
= vma_resv_map(vma
);
3687 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3688 unsigned long reserve
, start
, end
;
3691 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3694 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3695 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3697 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3698 hugetlb_cgroup_uncharge_counter(resv
, start
, end
);
3701 * Decrement reserve counts. The global reserve count may be
3702 * adjusted if the subpool has a minimum size.
3704 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3705 hugetlb_acct_memory(h
, -gbl_reserve
);
3708 kref_put(&resv
->refs
, resv_map_release
);
3711 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
3713 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
3718 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
3720 struct hstate
*hstate
= hstate_vma(vma
);
3722 return 1UL << huge_page_shift(hstate
);
3726 * We cannot handle pagefaults against hugetlb pages at all. They cause
3727 * handle_mm_fault() to try to instantiate regular-sized pages in the
3728 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3731 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3738 * When a new function is introduced to vm_operations_struct and added
3739 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3740 * This is because under System V memory model, mappings created via
3741 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3742 * their original vm_ops are overwritten with shm_vm_ops.
3744 const struct vm_operations_struct hugetlb_vm_ops
= {
3745 .fault
= hugetlb_vm_op_fault
,
3746 .open
= hugetlb_vm_op_open
,
3747 .close
= hugetlb_vm_op_close
,
3748 .may_split
= hugetlb_vm_op_split
,
3749 .pagesize
= hugetlb_vm_op_pagesize
,
3752 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3758 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3759 vma
->vm_page_prot
)));
3761 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3762 vma
->vm_page_prot
));
3764 entry
= pte_mkyoung(entry
);
3765 entry
= pte_mkhuge(entry
);
3766 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3771 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3772 unsigned long address
, pte_t
*ptep
)
3776 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3777 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3778 update_mmu_cache(vma
, address
, ptep
);
3781 bool is_hugetlb_entry_migration(pte_t pte
)
3785 if (huge_pte_none(pte
) || pte_present(pte
))
3787 swp
= pte_to_swp_entry(pte
);
3788 if (is_migration_entry(swp
))
3794 static bool is_hugetlb_entry_hwpoisoned(pte_t pte
)
3798 if (huge_pte_none(pte
) || pte_present(pte
))
3800 swp
= pte_to_swp_entry(pte
);
3801 if (is_hwpoison_entry(swp
))
3807 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3808 struct vm_area_struct
*vma
)
3810 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
3811 struct page
*ptepage
;
3814 struct hstate
*h
= hstate_vma(vma
);
3815 unsigned long sz
= huge_page_size(h
);
3816 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3817 struct mmu_notifier_range range
;
3820 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3823 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
3826 mmu_notifier_invalidate_range_start(&range
);
3829 * For shared mappings i_mmap_rwsem must be held to call
3830 * huge_pte_alloc, otherwise the returned ptep could go
3831 * away if part of a shared pmd and another thread calls
3834 i_mmap_lock_read(mapping
);
3837 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3838 spinlock_t
*src_ptl
, *dst_ptl
;
3839 src_pte
= huge_pte_offset(src
, addr
, sz
);
3842 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3849 * If the pagetables are shared don't copy or take references.
3850 * dst_pte == src_pte is the common case of src/dest sharing.
3852 * However, src could have 'unshared' and dst shares with
3853 * another vma. If dst_pte !none, this implies sharing.
3854 * Check here before taking page table lock, and once again
3855 * after taking the lock below.
3857 dst_entry
= huge_ptep_get(dst_pte
);
3858 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
3861 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3862 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3863 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3864 entry
= huge_ptep_get(src_pte
);
3865 dst_entry
= huge_ptep_get(dst_pte
);
3866 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
3868 * Skip if src entry none. Also, skip in the
3869 * unlikely case dst entry !none as this implies
3870 * sharing with another vma.
3873 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3874 is_hugetlb_entry_hwpoisoned(entry
))) {
3875 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3877 if (is_write_migration_entry(swp_entry
) && cow
) {
3879 * COW mappings require pages in both
3880 * parent and child to be set to read.
3882 make_migration_entry_read(&swp_entry
);
3883 entry
= swp_entry_to_pte(swp_entry
);
3884 set_huge_swap_pte_at(src
, addr
, src_pte
,
3887 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3891 * No need to notify as we are downgrading page
3892 * table protection not changing it to point
3895 * See Documentation/vm/mmu_notifier.rst
3897 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3899 entry
= huge_ptep_get(src_pte
);
3900 ptepage
= pte_page(entry
);
3902 page_dup_rmap(ptepage
, true);
3903 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3904 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3906 spin_unlock(src_ptl
);
3907 spin_unlock(dst_ptl
);
3911 mmu_notifier_invalidate_range_end(&range
);
3913 i_mmap_unlock_read(mapping
);
3918 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3919 unsigned long start
, unsigned long end
,
3920 struct page
*ref_page
)
3922 struct mm_struct
*mm
= vma
->vm_mm
;
3923 unsigned long address
;
3928 struct hstate
*h
= hstate_vma(vma
);
3929 unsigned long sz
= huge_page_size(h
);
3930 struct mmu_notifier_range range
;
3932 WARN_ON(!is_vm_hugetlb_page(vma
));
3933 BUG_ON(start
& ~huge_page_mask(h
));
3934 BUG_ON(end
& ~huge_page_mask(h
));
3937 * This is a hugetlb vma, all the pte entries should point
3940 tlb_change_page_size(tlb
, sz
);
3941 tlb_start_vma(tlb
, vma
);
3944 * If sharing possible, alert mmu notifiers of worst case.
3946 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
3948 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
3949 mmu_notifier_invalidate_range_start(&range
);
3951 for (; address
< end
; address
+= sz
) {
3952 ptep
= huge_pte_offset(mm
, address
, sz
);
3956 ptl
= huge_pte_lock(h
, mm
, ptep
);
3957 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
3960 * We just unmapped a page of PMDs by clearing a PUD.
3961 * The caller's TLB flush range should cover this area.
3966 pte
= huge_ptep_get(ptep
);
3967 if (huge_pte_none(pte
)) {
3973 * Migrating hugepage or HWPoisoned hugepage is already
3974 * unmapped and its refcount is dropped, so just clear pte here.
3976 if (unlikely(!pte_present(pte
))) {
3977 huge_pte_clear(mm
, address
, ptep
, sz
);
3982 page
= pte_page(pte
);
3984 * If a reference page is supplied, it is because a specific
3985 * page is being unmapped, not a range. Ensure the page we
3986 * are about to unmap is the actual page of interest.
3989 if (page
!= ref_page
) {
3994 * Mark the VMA as having unmapped its page so that
3995 * future faults in this VMA will fail rather than
3996 * looking like data was lost
3998 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
4001 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4002 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
4003 if (huge_pte_dirty(pte
))
4004 set_page_dirty(page
);
4006 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
4007 page_remove_rmap(page
, true);
4010 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
4012 * Bail out after unmapping reference page if supplied
4017 mmu_notifier_invalidate_range_end(&range
);
4018 tlb_end_vma(tlb
, vma
);
4021 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
4022 struct vm_area_struct
*vma
, unsigned long start
,
4023 unsigned long end
, struct page
*ref_page
)
4025 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
4028 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4029 * test will fail on a vma being torn down, and not grab a page table
4030 * on its way out. We're lucky that the flag has such an appropriate
4031 * name, and can in fact be safely cleared here. We could clear it
4032 * before the __unmap_hugepage_range above, but all that's necessary
4033 * is to clear it before releasing the i_mmap_rwsem. This works
4034 * because in the context this is called, the VMA is about to be
4035 * destroyed and the i_mmap_rwsem is held.
4037 vma
->vm_flags
&= ~VM_MAYSHARE
;
4040 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
4041 unsigned long end
, struct page
*ref_page
)
4043 struct mm_struct
*mm
;
4044 struct mmu_gather tlb
;
4045 unsigned long tlb_start
= start
;
4046 unsigned long tlb_end
= end
;
4049 * If shared PMDs were possibly used within this vma range, adjust
4050 * start/end for worst case tlb flushing.
4051 * Note that we can not be sure if PMDs are shared until we try to
4052 * unmap pages. However, we want to make sure TLB flushing covers
4053 * the largest possible range.
4055 adjust_range_if_pmd_sharing_possible(vma
, &tlb_start
, &tlb_end
);
4059 tlb_gather_mmu(&tlb
, mm
, tlb_start
, tlb_end
);
4060 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
4061 tlb_finish_mmu(&tlb
, tlb_start
, tlb_end
);
4065 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4066 * mappping it owns the reserve page for. The intention is to unmap the page
4067 * from other VMAs and let the children be SIGKILLed if they are faulting the
4070 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4071 struct page
*page
, unsigned long address
)
4073 struct hstate
*h
= hstate_vma(vma
);
4074 struct vm_area_struct
*iter_vma
;
4075 struct address_space
*mapping
;
4079 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4080 * from page cache lookup which is in HPAGE_SIZE units.
4082 address
= address
& huge_page_mask(h
);
4083 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
4085 mapping
= vma
->vm_file
->f_mapping
;
4088 * Take the mapping lock for the duration of the table walk. As
4089 * this mapping should be shared between all the VMAs,
4090 * __unmap_hugepage_range() is called as the lock is already held
4092 i_mmap_lock_write(mapping
);
4093 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
4094 /* Do not unmap the current VMA */
4095 if (iter_vma
== vma
)
4099 * Shared VMAs have their own reserves and do not affect
4100 * MAP_PRIVATE accounting but it is possible that a shared
4101 * VMA is using the same page so check and skip such VMAs.
4103 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
4107 * Unmap the page from other VMAs without their own reserves.
4108 * They get marked to be SIGKILLed if they fault in these
4109 * areas. This is because a future no-page fault on this VMA
4110 * could insert a zeroed page instead of the data existing
4111 * from the time of fork. This would look like data corruption
4113 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
4114 unmap_hugepage_range(iter_vma
, address
,
4115 address
+ huge_page_size(h
), page
);
4117 i_mmap_unlock_write(mapping
);
4121 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4122 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4123 * cannot race with other handlers or page migration.
4124 * Keep the pte_same checks anyway to make transition from the mutex easier.
4126 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4127 unsigned long address
, pte_t
*ptep
,
4128 struct page
*pagecache_page
, spinlock_t
*ptl
)
4131 struct hstate
*h
= hstate_vma(vma
);
4132 struct page
*old_page
, *new_page
;
4133 int outside_reserve
= 0;
4135 unsigned long haddr
= address
& huge_page_mask(h
);
4136 struct mmu_notifier_range range
;
4138 pte
= huge_ptep_get(ptep
);
4139 old_page
= pte_page(pte
);
4142 /* If no-one else is actually using this page, avoid the copy
4143 * and just make the page writable */
4144 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
4145 page_move_anon_rmap(old_page
, vma
);
4146 set_huge_ptep_writable(vma
, haddr
, ptep
);
4151 * If the process that created a MAP_PRIVATE mapping is about to
4152 * perform a COW due to a shared page count, attempt to satisfy
4153 * the allocation without using the existing reserves. The pagecache
4154 * page is used to determine if the reserve at this address was
4155 * consumed or not. If reserves were used, a partial faulted mapping
4156 * at the time of fork() could consume its reserves on COW instead
4157 * of the full address range.
4159 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
4160 old_page
!= pagecache_page
)
4161 outside_reserve
= 1;
4166 * Drop page table lock as buddy allocator may be called. It will
4167 * be acquired again before returning to the caller, as expected.
4170 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
4172 if (IS_ERR(new_page
)) {
4174 * If a process owning a MAP_PRIVATE mapping fails to COW,
4175 * it is due to references held by a child and an insufficient
4176 * huge page pool. To guarantee the original mappers
4177 * reliability, unmap the page from child processes. The child
4178 * may get SIGKILLed if it later faults.
4180 if (outside_reserve
) {
4181 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4186 BUG_ON(huge_pte_none(pte
));
4188 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4189 * unmapping. unmapping needs to hold i_mmap_rwsem
4190 * in write mode. Dropping i_mmap_rwsem in read mode
4191 * here is OK as COW mappings do not interact with
4194 * Reacquire both after unmap operation.
4196 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4197 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4198 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4199 i_mmap_unlock_read(mapping
);
4201 unmap_ref_private(mm
, vma
, old_page
, haddr
);
4203 i_mmap_lock_read(mapping
);
4204 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4206 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4208 pte_same(huge_ptep_get(ptep
), pte
)))
4209 goto retry_avoidcopy
;
4211 * race occurs while re-acquiring page table
4212 * lock, and our job is done.
4217 ret
= vmf_error(PTR_ERR(new_page
));
4218 goto out_release_old
;
4222 * When the original hugepage is shared one, it does not have
4223 * anon_vma prepared.
4225 if (unlikely(anon_vma_prepare(vma
))) {
4227 goto out_release_all
;
4230 copy_user_huge_page(new_page
, old_page
, address
, vma
,
4231 pages_per_huge_page(h
));
4232 __SetPageUptodate(new_page
);
4234 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
4235 haddr
+ huge_page_size(h
));
4236 mmu_notifier_invalidate_range_start(&range
);
4239 * Retake the page table lock to check for racing updates
4240 * before the page tables are altered
4243 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4244 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
4245 ClearPagePrivate(new_page
);
4248 huge_ptep_clear_flush(vma
, haddr
, ptep
);
4249 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
4250 set_huge_pte_at(mm
, haddr
, ptep
,
4251 make_huge_pte(vma
, new_page
, 1));
4252 page_remove_rmap(old_page
, true);
4253 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
4254 set_page_huge_active(new_page
);
4255 /* Make the old page be freed below */
4256 new_page
= old_page
;
4259 mmu_notifier_invalidate_range_end(&range
);
4261 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
4266 spin_lock(ptl
); /* Caller expects lock to be held */
4270 /* Return the pagecache page at a given address within a VMA */
4271 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
4272 struct vm_area_struct
*vma
, unsigned long address
)
4274 struct address_space
*mapping
;
4277 mapping
= vma
->vm_file
->f_mapping
;
4278 idx
= vma_hugecache_offset(h
, vma
, address
);
4280 return find_lock_page(mapping
, idx
);
4284 * Return whether there is a pagecache page to back given address within VMA.
4285 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4287 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
4288 struct vm_area_struct
*vma
, unsigned long address
)
4290 struct address_space
*mapping
;
4294 mapping
= vma
->vm_file
->f_mapping
;
4295 idx
= vma_hugecache_offset(h
, vma
, address
);
4297 page
= find_get_page(mapping
, idx
);
4300 return page
!= NULL
;
4303 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
4306 struct inode
*inode
= mapping
->host
;
4307 struct hstate
*h
= hstate_inode(inode
);
4308 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
4312 ClearPagePrivate(page
);
4315 * set page dirty so that it will not be removed from cache/file
4316 * by non-hugetlbfs specific code paths.
4318 set_page_dirty(page
);
4320 spin_lock(&inode
->i_lock
);
4321 inode
->i_blocks
+= blocks_per_huge_page(h
);
4322 spin_unlock(&inode
->i_lock
);
4326 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
4327 struct vm_area_struct
*vma
,
4328 struct address_space
*mapping
, pgoff_t idx
,
4329 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
4331 struct hstate
*h
= hstate_vma(vma
);
4332 vm_fault_t ret
= VM_FAULT_SIGBUS
;
4338 unsigned long haddr
= address
& huge_page_mask(h
);
4339 bool new_page
= false;
4342 * Currently, we are forced to kill the process in the event the
4343 * original mapper has unmapped pages from the child due to a failed
4344 * COW. Warn that such a situation has occurred as it may not be obvious
4346 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
4347 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4353 * We can not race with truncation due to holding i_mmap_rwsem.
4354 * i_size is modified when holding i_mmap_rwsem, so check here
4355 * once for faults beyond end of file.
4357 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4362 page
= find_lock_page(mapping
, idx
);
4365 * Check for page in userfault range
4367 if (userfaultfd_missing(vma
)) {
4369 struct vm_fault vmf
= {
4374 * Hard to debug if it ends up being
4375 * used by a callee that assumes
4376 * something about the other
4377 * uninitialized fields... same as in
4383 * hugetlb_fault_mutex and i_mmap_rwsem must be
4384 * dropped before handling userfault. Reacquire
4385 * after handling fault to make calling code simpler.
4387 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4388 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4389 i_mmap_unlock_read(mapping
);
4390 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
4391 i_mmap_lock_read(mapping
);
4392 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4396 page
= alloc_huge_page(vma
, haddr
, 0);
4399 * Returning error will result in faulting task being
4400 * sent SIGBUS. The hugetlb fault mutex prevents two
4401 * tasks from racing to fault in the same page which
4402 * could result in false unable to allocate errors.
4403 * Page migration does not take the fault mutex, but
4404 * does a clear then write of pte's under page table
4405 * lock. Page fault code could race with migration,
4406 * notice the clear pte and try to allocate a page
4407 * here. Before returning error, get ptl and make
4408 * sure there really is no pte entry.
4410 ptl
= huge_pte_lock(h
, mm
, ptep
);
4411 if (!huge_pte_none(huge_ptep_get(ptep
))) {
4417 ret
= vmf_error(PTR_ERR(page
));
4420 clear_huge_page(page
, address
, pages_per_huge_page(h
));
4421 __SetPageUptodate(page
);
4424 if (vma
->vm_flags
& VM_MAYSHARE
) {
4425 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
4434 if (unlikely(anon_vma_prepare(vma
))) {
4436 goto backout_unlocked
;
4442 * If memory error occurs between mmap() and fault, some process
4443 * don't have hwpoisoned swap entry for errored virtual address.
4444 * So we need to block hugepage fault by PG_hwpoison bit check.
4446 if (unlikely(PageHWPoison(page
))) {
4447 ret
= VM_FAULT_HWPOISON_LARGE
|
4448 VM_FAULT_SET_HINDEX(hstate_index(h
));
4449 goto backout_unlocked
;
4454 * If we are going to COW a private mapping later, we examine the
4455 * pending reservations for this page now. This will ensure that
4456 * any allocations necessary to record that reservation occur outside
4459 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4460 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4462 goto backout_unlocked
;
4464 /* Just decrements count, does not deallocate */
4465 vma_end_reservation(h
, vma
, haddr
);
4468 ptl
= huge_pte_lock(h
, mm
, ptep
);
4470 if (!huge_pte_none(huge_ptep_get(ptep
)))
4474 ClearPagePrivate(page
);
4475 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4477 page_dup_rmap(page
, true);
4478 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
4479 && (vma
->vm_flags
& VM_SHARED
)));
4480 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
4482 hugetlb_count_add(pages_per_huge_page(h
), mm
);
4483 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4484 /* Optimization, do the COW without a second fault */
4485 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
4491 * Only make newly allocated pages active. Existing pages found
4492 * in the pagecache could be !page_huge_active() if they have been
4493 * isolated for migration.
4496 set_page_huge_active(page
);
4506 restore_reserve_on_error(h
, vma
, haddr
, page
);
4512 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4514 unsigned long key
[2];
4517 key
[0] = (unsigned long) mapping
;
4520 hash
= jhash2((u32
*)&key
, sizeof(key
)/(sizeof(u32
)), 0);
4522 return hash
& (num_fault_mutexes
- 1);
4526 * For uniprocesor systems we always use a single mutex, so just
4527 * return 0 and avoid the hashing overhead.
4529 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4535 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4536 unsigned long address
, unsigned int flags
)
4543 struct page
*page
= NULL
;
4544 struct page
*pagecache_page
= NULL
;
4545 struct hstate
*h
= hstate_vma(vma
);
4546 struct address_space
*mapping
;
4547 int need_wait_lock
= 0;
4548 unsigned long haddr
= address
& huge_page_mask(h
);
4550 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4553 * Since we hold no locks, ptep could be stale. That is
4554 * OK as we are only making decisions based on content and
4555 * not actually modifying content here.
4557 entry
= huge_ptep_get(ptep
);
4558 if (unlikely(is_hugetlb_entry_migration(entry
))) {
4559 migration_entry_wait_huge(vma
, mm
, ptep
);
4561 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
4562 return VM_FAULT_HWPOISON_LARGE
|
4563 VM_FAULT_SET_HINDEX(hstate_index(h
));
4567 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4568 * until finished with ptep. This serves two purposes:
4569 * 1) It prevents huge_pmd_unshare from being called elsewhere
4570 * and making the ptep no longer valid.
4571 * 2) It synchronizes us with i_size modifications during truncation.
4573 * ptep could have already be assigned via huge_pte_offset. That
4574 * is OK, as huge_pte_alloc will return the same value unless
4575 * something has changed.
4577 mapping
= vma
->vm_file
->f_mapping
;
4578 i_mmap_lock_read(mapping
);
4579 ptep
= huge_pte_alloc(mm
, haddr
, huge_page_size(h
));
4581 i_mmap_unlock_read(mapping
);
4582 return VM_FAULT_OOM
;
4586 * Serialize hugepage allocation and instantiation, so that we don't
4587 * get spurious allocation failures if two CPUs race to instantiate
4588 * the same page in the page cache.
4590 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4591 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4592 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4594 entry
= huge_ptep_get(ptep
);
4595 if (huge_pte_none(entry
)) {
4596 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
4603 * entry could be a migration/hwpoison entry at this point, so this
4604 * check prevents the kernel from going below assuming that we have
4605 * an active hugepage in pagecache. This goto expects the 2nd page
4606 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4607 * properly handle it.
4609 if (!pte_present(entry
))
4613 * If we are going to COW the mapping later, we examine the pending
4614 * reservations for this page now. This will ensure that any
4615 * allocations necessary to record that reservation occur outside the
4616 * spinlock. For private mappings, we also lookup the pagecache
4617 * page now as it is used to determine if a reservation has been
4620 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
4621 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4625 /* Just decrements count, does not deallocate */
4626 vma_end_reservation(h
, vma
, haddr
);
4628 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4629 pagecache_page
= hugetlbfs_pagecache_page(h
,
4633 ptl
= huge_pte_lock(h
, mm
, ptep
);
4635 /* Check for a racing update before calling hugetlb_cow */
4636 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4640 * hugetlb_cow() requires page locks of pte_page(entry) and
4641 * pagecache_page, so here we need take the former one
4642 * when page != pagecache_page or !pagecache_page.
4644 page
= pte_page(entry
);
4645 if (page
!= pagecache_page
)
4646 if (!trylock_page(page
)) {
4653 if (flags
& FAULT_FLAG_WRITE
) {
4654 if (!huge_pte_write(entry
)) {
4655 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4656 pagecache_page
, ptl
);
4659 entry
= huge_pte_mkdirty(entry
);
4661 entry
= pte_mkyoung(entry
);
4662 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
4663 flags
& FAULT_FLAG_WRITE
))
4664 update_mmu_cache(vma
, haddr
, ptep
);
4666 if (page
!= pagecache_page
)
4672 if (pagecache_page
) {
4673 unlock_page(pagecache_page
);
4674 put_page(pagecache_page
);
4677 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4678 i_mmap_unlock_read(mapping
);
4680 * Generally it's safe to hold refcount during waiting page lock. But
4681 * here we just wait to defer the next page fault to avoid busy loop and
4682 * the page is not used after unlocked before returning from the current
4683 * page fault. So we are safe from accessing freed page, even if we wait
4684 * here without taking refcount.
4687 wait_on_page_locked(page
);
4692 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4693 * modifications for huge pages.
4695 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4697 struct vm_area_struct
*dst_vma
,
4698 unsigned long dst_addr
,
4699 unsigned long src_addr
,
4700 struct page
**pagep
)
4702 struct address_space
*mapping
;
4705 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4706 struct hstate
*h
= hstate_vma(dst_vma
);
4713 /* If a page already exists, then it's UFFDIO_COPY for
4714 * a non-missing case. Return -EEXIST.
4717 hugetlbfs_pagecache_present(h
, dst_vma
, dst_addr
)) {
4722 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4728 ret
= copy_huge_page_from_user(page
,
4729 (const void __user
*) src_addr
,
4730 pages_per_huge_page(h
), false);
4732 /* fallback to copy_from_user outside mmap_lock */
4733 if (unlikely(ret
)) {
4736 /* don't free the page */
4745 * The memory barrier inside __SetPageUptodate makes sure that
4746 * preceding stores to the page contents become visible before
4747 * the set_pte_at() write.
4749 __SetPageUptodate(page
);
4751 mapping
= dst_vma
->vm_file
->f_mapping
;
4752 idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4755 * If shared, add to page cache
4758 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4761 goto out_release_nounlock
;
4764 * Serialization between remove_inode_hugepages() and
4765 * huge_add_to_page_cache() below happens through the
4766 * hugetlb_fault_mutex_table that here must be hold by
4769 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4771 goto out_release_nounlock
;
4774 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4778 * Recheck the i_size after holding PT lock to make sure not
4779 * to leave any page mapped (as page_mapped()) beyond the end
4780 * of the i_size (remove_inode_hugepages() is strict about
4781 * enforcing that). If we bail out here, we'll also leave a
4782 * page in the radix tree in the vm_shared case beyond the end
4783 * of the i_size, but remove_inode_hugepages() will take care
4784 * of it as soon as we drop the hugetlb_fault_mutex_table.
4786 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4789 goto out_release_unlock
;
4792 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4793 goto out_release_unlock
;
4796 page_dup_rmap(page
, true);
4798 ClearPagePrivate(page
);
4799 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4802 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4803 if (dst_vma
->vm_flags
& VM_WRITE
)
4804 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4805 _dst_pte
= pte_mkyoung(_dst_pte
);
4807 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4809 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4810 dst_vma
->vm_flags
& VM_WRITE
);
4811 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4813 /* No need to invalidate - it was non-present before */
4814 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4817 set_page_huge_active(page
);
4827 out_release_nounlock
:
4832 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4833 struct page
**pages
, struct vm_area_struct
**vmas
,
4834 unsigned long *position
, unsigned long *nr_pages
,
4835 long i
, unsigned int flags
, int *locked
)
4837 unsigned long pfn_offset
;
4838 unsigned long vaddr
= *position
;
4839 unsigned long remainder
= *nr_pages
;
4840 struct hstate
*h
= hstate_vma(vma
);
4843 while (vaddr
< vma
->vm_end
&& remainder
) {
4845 spinlock_t
*ptl
= NULL
;
4850 * If we have a pending SIGKILL, don't keep faulting pages and
4851 * potentially allocating memory.
4853 if (fatal_signal_pending(current
)) {
4859 * Some archs (sparc64, sh*) have multiple pte_ts to
4860 * each hugepage. We have to make sure we get the
4861 * first, for the page indexing below to work.
4863 * Note that page table lock is not held when pte is null.
4865 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4868 ptl
= huge_pte_lock(h
, mm
, pte
);
4869 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4872 * When coredumping, it suits get_dump_page if we just return
4873 * an error where there's an empty slot with no huge pagecache
4874 * to back it. This way, we avoid allocating a hugepage, and
4875 * the sparse dumpfile avoids allocating disk blocks, but its
4876 * huge holes still show up with zeroes where they need to be.
4878 if (absent
&& (flags
& FOLL_DUMP
) &&
4879 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4887 * We need call hugetlb_fault for both hugepages under migration
4888 * (in which case hugetlb_fault waits for the migration,) and
4889 * hwpoisoned hugepages (in which case we need to prevent the
4890 * caller from accessing to them.) In order to do this, we use
4891 * here is_swap_pte instead of is_hugetlb_entry_migration and
4892 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4893 * both cases, and because we can't follow correct pages
4894 * directly from any kind of swap entries.
4896 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4897 ((flags
& FOLL_WRITE
) &&
4898 !huge_pte_write(huge_ptep_get(pte
)))) {
4900 unsigned int fault_flags
= 0;
4904 if (flags
& FOLL_WRITE
)
4905 fault_flags
|= FAULT_FLAG_WRITE
;
4907 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4908 FAULT_FLAG_KILLABLE
;
4909 if (flags
& FOLL_NOWAIT
)
4910 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4911 FAULT_FLAG_RETRY_NOWAIT
;
4912 if (flags
& FOLL_TRIED
) {
4914 * Note: FAULT_FLAG_ALLOW_RETRY and
4915 * FAULT_FLAG_TRIED can co-exist
4917 fault_flags
|= FAULT_FLAG_TRIED
;
4919 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4920 if (ret
& VM_FAULT_ERROR
) {
4921 err
= vm_fault_to_errno(ret
, flags
);
4925 if (ret
& VM_FAULT_RETRY
) {
4927 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
4931 * VM_FAULT_RETRY must not return an
4932 * error, it will return zero
4935 * No need to update "position" as the
4936 * caller will not check it after
4937 * *nr_pages is set to 0.
4944 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4945 page
= pte_page(huge_ptep_get(pte
));
4948 * If subpage information not requested, update counters
4949 * and skip the same_page loop below.
4951 if (!pages
&& !vmas
&& !pfn_offset
&&
4952 (vaddr
+ huge_page_size(h
) < vma
->vm_end
) &&
4953 (remainder
>= pages_per_huge_page(h
))) {
4954 vaddr
+= huge_page_size(h
);
4955 remainder
-= pages_per_huge_page(h
);
4956 i
+= pages_per_huge_page(h
);
4963 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4965 * try_grab_page() should always succeed here, because:
4966 * a) we hold the ptl lock, and b) we've just checked
4967 * that the huge page is present in the page tables. If
4968 * the huge page is present, then the tail pages must
4969 * also be present. The ptl prevents the head page and
4970 * tail pages from being rearranged in any way. So this
4971 * page must be available at this point, unless the page
4972 * refcount overflowed:
4974 if (WARN_ON_ONCE(!try_grab_page(pages
[i
], flags
))) {
4989 if (vaddr
< vma
->vm_end
&& remainder
&&
4990 pfn_offset
< pages_per_huge_page(h
)) {
4992 * We use pfn_offset to avoid touching the pageframes
4993 * of this compound page.
4999 *nr_pages
= remainder
;
5001 * setting position is actually required only if remainder is
5002 * not zero but it's faster not to add a "if (remainder)"
5010 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
5012 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
5015 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5018 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
5019 unsigned long address
, unsigned long end
, pgprot_t newprot
)
5021 struct mm_struct
*mm
= vma
->vm_mm
;
5022 unsigned long start
= address
;
5025 struct hstate
*h
= hstate_vma(vma
);
5026 unsigned long pages
= 0;
5027 bool shared_pmd
= false;
5028 struct mmu_notifier_range range
;
5031 * In the case of shared PMDs, the area to flush could be beyond
5032 * start/end. Set range.start/range.end to cover the maximum possible
5033 * range if PMD sharing is possible.
5035 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
5036 0, vma
, mm
, start
, end
);
5037 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
5039 BUG_ON(address
>= end
);
5040 flush_cache_range(vma
, range
.start
, range
.end
);
5042 mmu_notifier_invalidate_range_start(&range
);
5043 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
5044 for (; address
< end
; address
+= huge_page_size(h
)) {
5046 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
5049 ptl
= huge_pte_lock(h
, mm
, ptep
);
5050 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
5056 pte
= huge_ptep_get(ptep
);
5057 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
5061 if (unlikely(is_hugetlb_entry_migration(pte
))) {
5062 swp_entry_t entry
= pte_to_swp_entry(pte
);
5064 if (is_write_migration_entry(entry
)) {
5067 make_migration_entry_read(&entry
);
5068 newpte
= swp_entry_to_pte(entry
);
5069 set_huge_swap_pte_at(mm
, address
, ptep
,
5070 newpte
, huge_page_size(h
));
5076 if (!huge_pte_none(pte
)) {
5079 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
5080 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
5081 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
5082 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
5088 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5089 * may have cleared our pud entry and done put_page on the page table:
5090 * once we release i_mmap_rwsem, another task can do the final put_page
5091 * and that page table be reused and filled with junk. If we actually
5092 * did unshare a page of pmds, flush the range corresponding to the pud.
5095 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
5097 flush_hugetlb_tlb_range(vma
, start
, end
);
5099 * No need to call mmu_notifier_invalidate_range() we are downgrading
5100 * page table protection not changing it to point to a new page.
5102 * See Documentation/vm/mmu_notifier.rst
5104 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
5105 mmu_notifier_invalidate_range_end(&range
);
5107 return pages
<< h
->order
;
5110 int hugetlb_reserve_pages(struct inode
*inode
,
5112 struct vm_area_struct
*vma
,
5113 vm_flags_t vm_flags
)
5115 long ret
, chg
, add
= -1;
5116 struct hstate
*h
= hstate_inode(inode
);
5117 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5118 struct resv_map
*resv_map
;
5119 struct hugetlb_cgroup
*h_cg
= NULL
;
5120 long gbl_reserve
, regions_needed
= 0;
5122 /* This should never happen */
5124 VM_WARN(1, "%s called with a negative range\n", __func__
);
5129 * Only apply hugepage reservation if asked. At fault time, an
5130 * attempt will be made for VM_NORESERVE to allocate a page
5131 * without using reserves
5133 if (vm_flags
& VM_NORESERVE
)
5137 * Shared mappings base their reservation on the number of pages that
5138 * are already allocated on behalf of the file. Private mappings need
5139 * to reserve the full area even if read-only as mprotect() may be
5140 * called to make the mapping read-write. Assume !vma is a shm mapping
5142 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5144 * resv_map can not be NULL as hugetlb_reserve_pages is only
5145 * called for inodes for which resv_maps were created (see
5146 * hugetlbfs_get_inode).
5148 resv_map
= inode_resv_map(inode
);
5150 chg
= region_chg(resv_map
, from
, to
, ®ions_needed
);
5153 /* Private mapping. */
5154 resv_map
= resv_map_alloc();
5160 set_vma_resv_map(vma
, resv_map
);
5161 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
5169 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
5170 hstate_index(h
), chg
* pages_per_huge_page(h
), &h_cg
);
5177 if (vma
&& !(vma
->vm_flags
& VM_MAYSHARE
) && h_cg
) {
5178 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5181 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, h_cg
, h
);
5185 * There must be enough pages in the subpool for the mapping. If
5186 * the subpool has a minimum size, there may be some global
5187 * reservations already in place (gbl_reserve).
5189 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
5190 if (gbl_reserve
< 0) {
5192 goto out_uncharge_cgroup
;
5196 * Check enough hugepages are available for the reservation.
5197 * Hand the pages back to the subpool if there are not
5199 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
5205 * Account for the reservations made. Shared mappings record regions
5206 * that have reservations as they are shared by multiple VMAs.
5207 * When the last VMA disappears, the region map says how much
5208 * the reservation was and the page cache tells how much of
5209 * the reservation was consumed. Private mappings are per-VMA and
5210 * only the consumed reservations are tracked. When the VMA
5211 * disappears, the original reservation is the VMA size and the
5212 * consumed reservations are stored in the map. Hence, nothing
5213 * else has to be done for private mappings here
5215 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5216 add
= region_add(resv_map
, from
, to
, regions_needed
, h
, h_cg
);
5218 if (unlikely(add
< 0)) {
5219 hugetlb_acct_memory(h
, -gbl_reserve
);
5222 } else if (unlikely(chg
> add
)) {
5224 * pages in this range were added to the reserve
5225 * map between region_chg and region_add. This
5226 * indicates a race with alloc_huge_page. Adjust
5227 * the subpool and reserve counts modified above
5228 * based on the difference.
5233 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5234 * reference to h_cg->css. See comment below for detail.
5236 hugetlb_cgroup_uncharge_cgroup_rsvd(
5238 (chg
- add
) * pages_per_huge_page(h
), h_cg
);
5240 rsv_adjust
= hugepage_subpool_put_pages(spool
,
5242 hugetlb_acct_memory(h
, -rsv_adjust
);
5245 * The file_regions will hold their own reference to
5246 * h_cg->css. So we should release the reference held
5247 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5250 hugetlb_cgroup_put_rsvd_cgroup(h_cg
);
5255 /* put back original number of pages, chg */
5256 (void)hugepage_subpool_put_pages(spool
, chg
);
5257 out_uncharge_cgroup
:
5258 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h
),
5259 chg
* pages_per_huge_page(h
), h_cg
);
5261 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
5262 /* Only call region_abort if the region_chg succeeded but the
5263 * region_add failed or didn't run.
5265 if (chg
>= 0 && add
< 0)
5266 region_abort(resv_map
, from
, to
, regions_needed
);
5267 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
5268 kref_put(&resv_map
->refs
, resv_map_release
);
5272 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
5275 struct hstate
*h
= hstate_inode(inode
);
5276 struct resv_map
*resv_map
= inode_resv_map(inode
);
5278 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5282 * Since this routine can be called in the evict inode path for all
5283 * hugetlbfs inodes, resv_map could be NULL.
5286 chg
= region_del(resv_map
, start
, end
);
5288 * region_del() can fail in the rare case where a region
5289 * must be split and another region descriptor can not be
5290 * allocated. If end == LONG_MAX, it will not fail.
5296 spin_lock(&inode
->i_lock
);
5297 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
5298 spin_unlock(&inode
->i_lock
);
5301 * If the subpool has a minimum size, the number of global
5302 * reservations to be released may be adjusted.
5304 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
5305 hugetlb_acct_memory(h
, -gbl_reserve
);
5310 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5311 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
5312 struct vm_area_struct
*vma
,
5313 unsigned long addr
, pgoff_t idx
)
5315 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
5317 unsigned long sbase
= saddr
& PUD_MASK
;
5318 unsigned long s_end
= sbase
+ PUD_SIZE
;
5320 /* Allow segments to share if only one is marked locked */
5321 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5322 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5325 * match the virtual addresses, permission and the alignment of the
5328 if (pmd_index(addr
) != pmd_index(saddr
) ||
5329 vm_flags
!= svm_flags
||
5330 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
5336 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
5338 unsigned long base
= addr
& PUD_MASK
;
5339 unsigned long end
= base
+ PUD_SIZE
;
5342 * check on proper vm_flags and page table alignment
5344 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
5350 * Determine if start,end range within vma could be mapped by shared pmd.
5351 * If yes, adjust start and end to cover range associated with possible
5352 * shared pmd mappings.
5354 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5355 unsigned long *start
, unsigned long *end
)
5357 unsigned long v_start
= ALIGN(vma
->vm_start
, PUD_SIZE
),
5358 v_end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
5361 * vma need span at least one aligned PUD size and the start,end range
5362 * must at least partialy within it.
5364 if (!(vma
->vm_flags
& VM_MAYSHARE
) || !(v_end
> v_start
) ||
5365 (*end
<= v_start
) || (*start
>= v_end
))
5368 /* Extend the range to be PUD aligned for a worst case scenario */
5369 if (*start
> v_start
)
5370 *start
= ALIGN_DOWN(*start
, PUD_SIZE
);
5373 *end
= ALIGN(*end
, PUD_SIZE
);
5377 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5378 * and returns the corresponding pte. While this is not necessary for the
5379 * !shared pmd case because we can allocate the pmd later as well, it makes the
5380 * code much cleaner.
5382 * This routine must be called with i_mmap_rwsem held in at least read mode if
5383 * sharing is possible. For hugetlbfs, this prevents removal of any page
5384 * table entries associated with the address space. This is important as we
5385 * are setting up sharing based on existing page table entries (mappings).
5387 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5388 * huge_pte_alloc know that sharing is not possible and do not take
5389 * i_mmap_rwsem as a performance optimization. This is handled by the
5390 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5391 * only required for subsequent processing.
5393 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5395 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
5396 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
5397 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
5399 struct vm_area_struct
*svma
;
5400 unsigned long saddr
;
5405 if (!vma_shareable(vma
, addr
))
5406 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5408 i_mmap_assert_locked(mapping
);
5409 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
5413 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
5415 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
5416 vma_mmu_pagesize(svma
));
5418 get_page(virt_to_page(spte
));
5427 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
5428 if (pud_none(*pud
)) {
5429 pud_populate(mm
, pud
,
5430 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
5433 put_page(virt_to_page(spte
));
5437 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5442 * unmap huge page backed by shared pte.
5444 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5445 * indicated by page_count > 1, unmap is achieved by clearing pud and
5446 * decrementing the ref count. If count == 1, the pte page is not shared.
5448 * Called with page table lock held and i_mmap_rwsem held in write mode.
5450 * returns: 1 successfully unmapped a shared pte page
5451 * 0 the underlying pte page is not shared, or it is the last user
5453 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5454 unsigned long *addr
, pte_t
*ptep
)
5456 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
5457 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
5458 pud_t
*pud
= pud_offset(p4d
, *addr
);
5460 i_mmap_assert_write_locked(vma
->vm_file
->f_mapping
);
5461 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
5462 if (page_count(virt_to_page(ptep
)) == 1)
5466 put_page(virt_to_page(ptep
));
5468 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
5471 #define want_pmd_share() (1)
5472 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5473 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5478 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5479 unsigned long *addr
, pte_t
*ptep
)
5484 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5485 unsigned long *start
, unsigned long *end
)
5488 #define want_pmd_share() (0)
5489 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5491 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5492 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
5493 unsigned long addr
, unsigned long sz
)
5500 pgd
= pgd_offset(mm
, addr
);
5501 p4d
= p4d_alloc(mm
, pgd
, addr
);
5504 pud
= pud_alloc(mm
, p4d
, addr
);
5506 if (sz
== PUD_SIZE
) {
5509 BUG_ON(sz
!= PMD_SIZE
);
5510 if (want_pmd_share() && pud_none(*pud
))
5511 pte
= huge_pmd_share(mm
, addr
, pud
);
5513 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5516 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
5522 * huge_pte_offset() - Walk the page table to resolve the hugepage
5523 * entry at address @addr
5525 * Return: Pointer to page table entry (PUD or PMD) for
5526 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5527 * size @sz doesn't match the hugepage size at this level of the page
5530 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
5531 unsigned long addr
, unsigned long sz
)
5538 pgd
= pgd_offset(mm
, addr
);
5539 if (!pgd_present(*pgd
))
5541 p4d
= p4d_offset(pgd
, addr
);
5542 if (!p4d_present(*p4d
))
5545 pud
= pud_offset(p4d
, addr
);
5547 /* must be pud huge, non-present or none */
5548 return (pte_t
*)pud
;
5549 if (!pud_present(*pud
))
5551 /* must have a valid entry and size to go further */
5553 pmd
= pmd_offset(pud
, addr
);
5554 /* must be pmd huge, non-present or none */
5555 return (pte_t
*)pmd
;
5558 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5561 * These functions are overwritable if your architecture needs its own
5564 struct page
* __weak
5565 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
5568 return ERR_PTR(-EINVAL
);
5571 struct page
* __weak
5572 follow_huge_pd(struct vm_area_struct
*vma
,
5573 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
5575 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5579 struct page
* __weak
5580 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
5581 pmd_t
*pmd
, int flags
)
5583 struct page
*page
= NULL
;
5587 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5588 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
5589 (FOLL_PIN
| FOLL_GET
)))
5593 ptl
= pmd_lockptr(mm
, pmd
);
5596 * make sure that the address range covered by this pmd is not
5597 * unmapped from other threads.
5599 if (!pmd_huge(*pmd
))
5601 pte
= huge_ptep_get((pte_t
*)pmd
);
5602 if (pte_present(pte
)) {
5603 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
5605 * try_grab_page() should always succeed here, because: a) we
5606 * hold the pmd (ptl) lock, and b) we've just checked that the
5607 * huge pmd (head) page is present in the page tables. The ptl
5608 * prevents the head page and tail pages from being rearranged
5609 * in any way. So this page must be available at this point,
5610 * unless the page refcount overflowed:
5612 if (WARN_ON_ONCE(!try_grab_page(page
, flags
))) {
5617 if (is_hugetlb_entry_migration(pte
)) {
5619 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
5623 * hwpoisoned entry is treated as no_page_table in
5624 * follow_page_mask().
5632 struct page
* __weak
5633 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
5634 pud_t
*pud
, int flags
)
5636 if (flags
& (FOLL_GET
| FOLL_PIN
))
5639 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
5642 struct page
* __weak
5643 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
5645 if (flags
& (FOLL_GET
| FOLL_PIN
))
5648 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
5651 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
5655 spin_lock(&hugetlb_lock
);
5656 if (!PageHeadHuge(page
) || !page_huge_active(page
) ||
5657 !get_page_unless_zero(page
)) {
5661 clear_page_huge_active(page
);
5662 list_move_tail(&page
->lru
, list
);
5664 spin_unlock(&hugetlb_lock
);
5668 void putback_active_hugepage(struct page
*page
)
5670 VM_BUG_ON_PAGE(!PageHead(page
), page
);
5671 spin_lock(&hugetlb_lock
);
5672 set_page_huge_active(page
);
5673 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
5674 spin_unlock(&hugetlb_lock
);
5678 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
5680 struct hstate
*h
= page_hstate(oldpage
);
5682 hugetlb_cgroup_migrate(oldpage
, newpage
);
5683 set_page_owner_migrate_reason(newpage
, reason
);
5686 * transfer temporary state of the new huge page. This is
5687 * reverse to other transitions because the newpage is going to
5688 * be final while the old one will be freed so it takes over
5689 * the temporary status.
5691 * Also note that we have to transfer the per-node surplus state
5692 * here as well otherwise the global surplus count will not match
5695 if (PageHugeTemporary(newpage
)) {
5696 int old_nid
= page_to_nid(oldpage
);
5697 int new_nid
= page_to_nid(newpage
);
5699 SetPageHugeTemporary(oldpage
);
5700 ClearPageHugeTemporary(newpage
);
5702 spin_lock(&hugetlb_lock
);
5703 if (h
->surplus_huge_pages_node
[old_nid
]) {
5704 h
->surplus_huge_pages_node
[old_nid
]--;
5705 h
->surplus_huge_pages_node
[new_nid
]++;
5707 spin_unlock(&hugetlb_lock
);
5712 static bool cma_reserve_called __initdata
;
5714 static int __init
cmdline_parse_hugetlb_cma(char *p
)
5716 hugetlb_cma_size
= memparse(p
, &p
);
5720 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma
);
5722 void __init
hugetlb_cma_reserve(int order
)
5724 unsigned long size
, reserved
, per_node
;
5727 cma_reserve_called
= true;
5729 if (!hugetlb_cma_size
)
5732 if (hugetlb_cma_size
< (PAGE_SIZE
<< order
)) {
5733 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5734 (PAGE_SIZE
<< order
) / SZ_1M
);
5739 * If 3 GB area is requested on a machine with 4 numa nodes,
5740 * let's allocate 1 GB on first three nodes and ignore the last one.
5742 per_node
= DIV_ROUND_UP(hugetlb_cma_size
, nr_online_nodes
);
5743 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5744 hugetlb_cma_size
/ SZ_1M
, per_node
/ SZ_1M
);
5747 for_each_node_state(nid
, N_ONLINE
) {
5749 char name
[CMA_MAX_NAME
];
5751 size
= min(per_node
, hugetlb_cma_size
- reserved
);
5752 size
= round_up(size
, PAGE_SIZE
<< order
);
5754 snprintf(name
, sizeof(name
), "hugetlb%d", nid
);
5755 res
= cma_declare_contiguous_nid(0, size
, 0, PAGE_SIZE
<< order
,
5757 &hugetlb_cma
[nid
], nid
);
5759 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5765 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5768 if (reserved
>= hugetlb_cma_size
)
5773 void __init
hugetlb_cma_check(void)
5775 if (!hugetlb_cma_size
|| cma_reserve_called
)
5778 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5781 #endif /* CONFIG_CMA */