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
);
749 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
751 struct hstate
*h
= hstate_inode(inode
);
753 hugetlb_acct_memory(h
, 1);
758 * Count and return the number of huge pages in the reserve map
759 * that intersect with the range [f, t).
761 static long region_count(struct resv_map
*resv
, long f
, long t
)
763 struct list_head
*head
= &resv
->regions
;
764 struct file_region
*rg
;
767 spin_lock(&resv
->lock
);
768 /* Locate each segment we overlap with, and count that overlap. */
769 list_for_each_entry(rg
, head
, link
) {
778 seg_from
= max(rg
->from
, f
);
779 seg_to
= min(rg
->to
, t
);
781 chg
+= seg_to
- seg_from
;
783 spin_unlock(&resv
->lock
);
789 * Convert the address within this vma to the page offset within
790 * the mapping, in pagecache page units; huge pages here.
792 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
793 struct vm_area_struct
*vma
, unsigned long address
)
795 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
796 (vma
->vm_pgoff
>> huge_page_order(h
));
799 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
800 unsigned long address
)
802 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
804 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
807 * Return the size of the pages allocated when backing a VMA. In the majority
808 * cases this will be same size as used by the page table entries.
810 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
812 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
813 return vma
->vm_ops
->pagesize(vma
);
816 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
819 * Return the page size being used by the MMU to back a VMA. In the majority
820 * of cases, the page size used by the kernel matches the MMU size. On
821 * architectures where it differs, an architecture-specific 'strong'
822 * version of this symbol is required.
824 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
826 return vma_kernel_pagesize(vma
);
830 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
831 * bits of the reservation map pointer, which are always clear due to
834 #define HPAGE_RESV_OWNER (1UL << 0)
835 #define HPAGE_RESV_UNMAPPED (1UL << 1)
836 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
839 * These helpers are used to track how many pages are reserved for
840 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
841 * is guaranteed to have their future faults succeed.
843 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
844 * the reserve counters are updated with the hugetlb_lock held. It is safe
845 * to reset the VMA at fork() time as it is not in use yet and there is no
846 * chance of the global counters getting corrupted as a result of the values.
848 * The private mapping reservation is represented in a subtly different
849 * manner to a shared mapping. A shared mapping has a region map associated
850 * with the underlying file, this region map represents the backing file
851 * pages which have ever had a reservation assigned which this persists even
852 * after the page is instantiated. A private mapping has a region map
853 * associated with the original mmap which is attached to all VMAs which
854 * reference it, this region map represents those offsets which have consumed
855 * reservation ie. where pages have been instantiated.
857 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
859 return (unsigned long)vma
->vm_private_data
;
862 static void set_vma_private_data(struct vm_area_struct
*vma
,
865 vma
->vm_private_data
= (void *)value
;
869 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map
*resv_map
,
870 struct hugetlb_cgroup
*h_cg
,
873 #ifdef CONFIG_CGROUP_HUGETLB
875 resv_map
->reservation_counter
= NULL
;
876 resv_map
->pages_per_hpage
= 0;
877 resv_map
->css
= NULL
;
879 resv_map
->reservation_counter
=
880 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
881 resv_map
->pages_per_hpage
= pages_per_huge_page(h
);
882 resv_map
->css
= &h_cg
->css
;
887 struct resv_map
*resv_map_alloc(void)
889 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
890 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
892 if (!resv_map
|| !rg
) {
898 kref_init(&resv_map
->refs
);
899 spin_lock_init(&resv_map
->lock
);
900 INIT_LIST_HEAD(&resv_map
->regions
);
902 resv_map
->adds_in_progress
= 0;
904 * Initialize these to 0. On shared mappings, 0's here indicate these
905 * fields don't do cgroup accounting. On private mappings, these will be
906 * re-initialized to the proper values, to indicate that hugetlb cgroup
907 * reservations are to be un-charged from here.
909 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, NULL
, NULL
);
911 INIT_LIST_HEAD(&resv_map
->region_cache
);
912 list_add(&rg
->link
, &resv_map
->region_cache
);
913 resv_map
->region_cache_count
= 1;
918 void resv_map_release(struct kref
*ref
)
920 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
921 struct list_head
*head
= &resv_map
->region_cache
;
922 struct file_region
*rg
, *trg
;
924 /* Clear out any active regions before we release the map. */
925 region_del(resv_map
, 0, LONG_MAX
);
927 /* ... and any entries left in the cache */
928 list_for_each_entry_safe(rg
, trg
, head
, link
) {
933 VM_BUG_ON(resv_map
->adds_in_progress
);
938 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
941 * At inode evict time, i_mapping may not point to the original
942 * address space within the inode. This original address space
943 * contains the pointer to the resv_map. So, always use the
944 * address space embedded within the inode.
945 * The VERY common case is inode->mapping == &inode->i_data but,
946 * this may not be true for device special inodes.
948 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
951 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
953 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
954 if (vma
->vm_flags
& VM_MAYSHARE
) {
955 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
956 struct inode
*inode
= mapping
->host
;
958 return inode_resv_map(inode
);
961 return (struct resv_map
*)(get_vma_private_data(vma
) &
966 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
968 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
969 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
971 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
972 HPAGE_RESV_MASK
) | (unsigned long)map
);
975 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
978 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
980 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
983 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
985 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
987 return (get_vma_private_data(vma
) & flag
) != 0;
990 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
991 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
993 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
994 if (!(vma
->vm_flags
& VM_MAYSHARE
))
995 vma
->vm_private_data
= (void *)0;
998 /* Returns true if the VMA has associated reserve pages */
999 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
1001 if (vma
->vm_flags
& VM_NORESERVE
) {
1003 * This address is already reserved by other process(chg == 0),
1004 * so, we should decrement reserved count. Without decrementing,
1005 * reserve count remains after releasing inode, because this
1006 * allocated page will go into page cache and is regarded as
1007 * coming from reserved pool in releasing step. Currently, we
1008 * don't have any other solution to deal with this situation
1009 * properly, so add work-around here.
1011 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
1017 /* Shared mappings always use reserves */
1018 if (vma
->vm_flags
& VM_MAYSHARE
) {
1020 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1021 * be a region map for all pages. The only situation where
1022 * there is no region map is if a hole was punched via
1023 * fallocate. In this case, there really are no reserves to
1024 * use. This situation is indicated if chg != 0.
1033 * Only the process that called mmap() has reserves for
1036 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1038 * Like the shared case above, a hole punch or truncate
1039 * could have been performed on the private mapping.
1040 * Examine the value of chg to determine if reserves
1041 * actually exist or were previously consumed.
1042 * Very Subtle - The value of chg comes from a previous
1043 * call to vma_needs_reserves(). The reserve map for
1044 * private mappings has different (opposite) semantics
1045 * than that of shared mappings. vma_needs_reserves()
1046 * has already taken this difference in semantics into
1047 * account. Therefore, the meaning of chg is the same
1048 * as in the shared case above. Code could easily be
1049 * combined, but keeping it separate draws attention to
1050 * subtle differences.
1061 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
1063 int nid
= page_to_nid(page
);
1064 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
1065 h
->free_huge_pages
++;
1066 h
->free_huge_pages_node
[nid
]++;
1067 SetPageHugeFreed(page
);
1070 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
1073 bool nocma
= !!(current
->flags
& PF_MEMALLOC_NOCMA
);
1075 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
) {
1076 if (nocma
&& is_migrate_cma_page(page
))
1079 if (PageHWPoison(page
))
1082 list_move(&page
->lru
, &h
->hugepage_activelist
);
1083 set_page_refcounted(page
);
1084 ClearPageHugeFreed(page
);
1085 h
->free_huge_pages
--;
1086 h
->free_huge_pages_node
[nid
]--;
1093 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
1096 unsigned int cpuset_mems_cookie
;
1097 struct zonelist
*zonelist
;
1100 int node
= NUMA_NO_NODE
;
1102 zonelist
= node_zonelist(nid
, gfp_mask
);
1105 cpuset_mems_cookie
= read_mems_allowed_begin();
1106 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
1109 if (!cpuset_zone_allowed(zone
, gfp_mask
))
1112 * no need to ask again on the same node. Pool is node rather than
1115 if (zone_to_nid(zone
) == node
)
1117 node
= zone_to_nid(zone
);
1119 page
= dequeue_huge_page_node_exact(h
, node
);
1123 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
1129 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
1130 struct vm_area_struct
*vma
,
1131 unsigned long address
, int avoid_reserve
,
1135 struct mempolicy
*mpol
;
1137 nodemask_t
*nodemask
;
1141 * A child process with MAP_PRIVATE mappings created by their parent
1142 * have no page reserves. This check ensures that reservations are
1143 * not "stolen". The child may still get SIGKILLed
1145 if (!vma_has_reserves(vma
, chg
) &&
1146 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1149 /* If reserves cannot be used, ensure enough pages are in the pool */
1150 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1153 gfp_mask
= htlb_alloc_mask(h
);
1154 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1155 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
1156 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
1157 SetPagePrivate(page
);
1158 h
->resv_huge_pages
--;
1161 mpol_cond_put(mpol
);
1169 * common helper functions for hstate_next_node_to_{alloc|free}.
1170 * We may have allocated or freed a huge page based on a different
1171 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1172 * be outside of *nodes_allowed. Ensure that we use an allowed
1173 * node for alloc or free.
1175 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1177 nid
= next_node_in(nid
, *nodes_allowed
);
1178 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1183 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1185 if (!node_isset(nid
, *nodes_allowed
))
1186 nid
= next_node_allowed(nid
, nodes_allowed
);
1191 * returns the previously saved node ["this node"] from which to
1192 * allocate a persistent huge page for the pool and advance the
1193 * next node from which to allocate, handling wrap at end of node
1196 static int hstate_next_node_to_alloc(struct hstate
*h
,
1197 nodemask_t
*nodes_allowed
)
1201 VM_BUG_ON(!nodes_allowed
);
1203 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1204 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1210 * helper for free_pool_huge_page() - return the previously saved
1211 * node ["this node"] from which to free a huge page. Advance the
1212 * next node id whether or not we find a free huge page to free so
1213 * that the next attempt to free addresses the next node.
1215 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1219 VM_BUG_ON(!nodes_allowed
);
1221 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1222 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1227 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1228 for (nr_nodes = nodes_weight(*mask); \
1230 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1233 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1234 for (nr_nodes = nodes_weight(*mask); \
1236 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1239 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1240 static void destroy_compound_gigantic_page(struct page
*page
,
1244 int nr_pages
= 1 << order
;
1245 struct page
*p
= page
+ 1;
1247 atomic_set(compound_mapcount_ptr(page
), 0);
1248 if (hpage_pincount_available(page
))
1249 atomic_set(compound_pincount_ptr(page
), 0);
1251 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1252 clear_compound_head(p
);
1253 set_page_refcounted(p
);
1256 set_compound_order(page
, 0);
1257 page
[1].compound_nr
= 0;
1258 __ClearPageHead(page
);
1261 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1264 * If the page isn't allocated using the cma allocator,
1265 * cma_release() returns false.
1268 if (cma_release(hugetlb_cma
[page_to_nid(page
)], page
, 1 << order
))
1272 free_contig_range(page_to_pfn(page
), 1 << order
);
1275 #ifdef CONFIG_CONTIG_ALLOC
1276 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1277 int nid
, nodemask_t
*nodemask
)
1279 unsigned long nr_pages
= 1UL << huge_page_order(h
);
1280 if (nid
== NUMA_NO_NODE
)
1281 nid
= numa_mem_id();
1288 if (hugetlb_cma
[nid
]) {
1289 page
= cma_alloc(hugetlb_cma
[nid
], nr_pages
,
1290 huge_page_order(h
), true);
1295 if (!(gfp_mask
& __GFP_THISNODE
)) {
1296 for_each_node_mask(node
, *nodemask
) {
1297 if (node
== nid
|| !hugetlb_cma
[node
])
1300 page
= cma_alloc(hugetlb_cma
[node
], nr_pages
,
1301 huge_page_order(h
), true);
1309 return alloc_contig_pages(nr_pages
, gfp_mask
, nid
, nodemask
);
1312 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1313 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1314 #else /* !CONFIG_CONTIG_ALLOC */
1315 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1316 int nid
, nodemask_t
*nodemask
)
1320 #endif /* CONFIG_CONTIG_ALLOC */
1322 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1323 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1324 int nid
, nodemask_t
*nodemask
)
1328 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1329 static inline void destroy_compound_gigantic_page(struct page
*page
,
1330 unsigned int order
) { }
1333 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1336 struct page
*subpage
= page
;
1338 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1342 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1343 for (i
= 0; i
< pages_per_huge_page(h
);
1344 i
++, subpage
= mem_map_next(subpage
, page
, i
)) {
1345 subpage
->flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1346 1 << PG_referenced
| 1 << PG_dirty
|
1347 1 << PG_active
| 1 << PG_private
|
1350 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1351 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page
), page
);
1352 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1353 set_page_refcounted(page
);
1354 if (hstate_is_gigantic(h
)) {
1356 * Temporarily drop the hugetlb_lock, because
1357 * we might block in free_gigantic_page().
1359 spin_unlock(&hugetlb_lock
);
1360 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1361 free_gigantic_page(page
, huge_page_order(h
));
1362 spin_lock(&hugetlb_lock
);
1364 __free_pages(page
, huge_page_order(h
));
1368 struct hstate
*size_to_hstate(unsigned long size
)
1372 for_each_hstate(h
) {
1373 if (huge_page_size(h
) == size
)
1380 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1381 * to hstate->hugepage_activelist.)
1383 * This function can be called for tail pages, but never returns true for them.
1385 bool page_huge_active(struct page
*page
)
1387 return PageHeadHuge(page
) && PagePrivate(&page
[1]);
1390 /* never called for tail page */
1391 void set_page_huge_active(struct page
*page
)
1393 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1394 SetPagePrivate(&page
[1]);
1397 static void clear_page_huge_active(struct page
*page
)
1399 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1400 ClearPagePrivate(&page
[1]);
1404 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1407 static inline bool PageHugeTemporary(struct page
*page
)
1409 if (!PageHuge(page
))
1412 return (unsigned long)page
[2].mapping
== -1U;
1415 static inline void SetPageHugeTemporary(struct page
*page
)
1417 page
[2].mapping
= (void *)-1U;
1420 static inline void ClearPageHugeTemporary(struct page
*page
)
1422 page
[2].mapping
= NULL
;
1425 static void __free_huge_page(struct page
*page
)
1428 * Can't pass hstate in here because it is called from the
1429 * compound page destructor.
1431 struct hstate
*h
= page_hstate(page
);
1432 int nid
= page_to_nid(page
);
1433 struct hugepage_subpool
*spool
=
1434 (struct hugepage_subpool
*)page_private(page
);
1435 bool restore_reserve
;
1437 VM_BUG_ON_PAGE(page_count(page
), page
);
1438 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1440 set_page_private(page
, 0);
1441 page
->mapping
= NULL
;
1442 restore_reserve
= PagePrivate(page
);
1443 ClearPagePrivate(page
);
1446 * If PagePrivate() was set on page, page allocation consumed a
1447 * reservation. If the page was associated with a subpool, there
1448 * would have been a page reserved in the subpool before allocation
1449 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1450 * reservtion, do not call hugepage_subpool_put_pages() as this will
1451 * remove the reserved page from the subpool.
1453 if (!restore_reserve
) {
1455 * A return code of zero implies that the subpool will be
1456 * under its minimum size if the reservation is not restored
1457 * after page is free. Therefore, force restore_reserve
1460 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1461 restore_reserve
= true;
1464 spin_lock(&hugetlb_lock
);
1465 clear_page_huge_active(page
);
1466 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1467 pages_per_huge_page(h
), page
);
1468 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
1469 pages_per_huge_page(h
), page
);
1470 if (restore_reserve
)
1471 h
->resv_huge_pages
++;
1473 if (PageHugeTemporary(page
)) {
1474 list_del(&page
->lru
);
1475 ClearPageHugeTemporary(page
);
1476 update_and_free_page(h
, page
);
1477 } else if (h
->surplus_huge_pages_node
[nid
]) {
1478 /* remove the page from active list */
1479 list_del(&page
->lru
);
1480 update_and_free_page(h
, page
);
1481 h
->surplus_huge_pages
--;
1482 h
->surplus_huge_pages_node
[nid
]--;
1484 arch_clear_hugepage_flags(page
);
1485 enqueue_huge_page(h
, page
);
1487 spin_unlock(&hugetlb_lock
);
1491 * As free_huge_page() can be called from a non-task context, we have
1492 * to defer the actual freeing in a workqueue to prevent potential
1493 * hugetlb_lock deadlock.
1495 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1496 * be freed and frees them one-by-one. As the page->mapping pointer is
1497 * going to be cleared in __free_huge_page() anyway, it is reused as the
1498 * llist_node structure of a lockless linked list of huge pages to be freed.
1500 static LLIST_HEAD(hpage_freelist
);
1502 static void free_hpage_workfn(struct work_struct
*work
)
1504 struct llist_node
*node
;
1507 node
= llist_del_all(&hpage_freelist
);
1510 page
= container_of((struct address_space
**)node
,
1511 struct page
, mapping
);
1513 __free_huge_page(page
);
1516 static DECLARE_WORK(free_hpage_work
, free_hpage_workfn
);
1518 void free_huge_page(struct page
*page
)
1521 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1525 * Only call schedule_work() if hpage_freelist is previously
1526 * empty. Otherwise, schedule_work() had been called but the
1527 * workfn hasn't retrieved the list yet.
1529 if (llist_add((struct llist_node
*)&page
->mapping
,
1531 schedule_work(&free_hpage_work
);
1535 __free_huge_page(page
);
1538 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1540 INIT_LIST_HEAD(&page
->lru
);
1541 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1542 set_hugetlb_cgroup(page
, NULL
);
1543 set_hugetlb_cgroup_rsvd(page
, NULL
);
1544 spin_lock(&hugetlb_lock
);
1546 h
->nr_huge_pages_node
[nid
]++;
1547 ClearPageHugeFreed(page
);
1548 spin_unlock(&hugetlb_lock
);
1551 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1554 int nr_pages
= 1 << order
;
1555 struct page
*p
= page
+ 1;
1557 /* we rely on prep_new_huge_page to set the destructor */
1558 set_compound_order(page
, order
);
1559 __ClearPageReserved(page
);
1560 __SetPageHead(page
);
1561 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1563 * For gigantic hugepages allocated through bootmem at
1564 * boot, it's safer to be consistent with the not-gigantic
1565 * hugepages and clear the PG_reserved bit from all tail pages
1566 * too. Otherwise drivers using get_user_pages() to access tail
1567 * pages may get the reference counting wrong if they see
1568 * PG_reserved set on a tail page (despite the head page not
1569 * having PG_reserved set). Enforcing this consistency between
1570 * head and tail pages allows drivers to optimize away a check
1571 * on the head page when they need know if put_page() is needed
1572 * after get_user_pages().
1574 __ClearPageReserved(p
);
1575 set_page_count(p
, 0);
1576 set_compound_head(p
, page
);
1578 atomic_set(compound_mapcount_ptr(page
), -1);
1580 if (hpage_pincount_available(page
))
1581 atomic_set(compound_pincount_ptr(page
), 0);
1585 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1586 * transparent huge pages. See the PageTransHuge() documentation for more
1589 int PageHuge(struct page
*page
)
1591 if (!PageCompound(page
))
1594 page
= compound_head(page
);
1595 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1597 EXPORT_SYMBOL_GPL(PageHuge
);
1600 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1601 * normal or transparent huge pages.
1603 int PageHeadHuge(struct page
*page_head
)
1605 if (!PageHead(page_head
))
1608 return page_head
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1612 * Find and lock address space (mapping) in write mode.
1614 * Upon entry, the page is locked which means that page_mapping() is
1615 * stable. Due to locking order, we can only trylock_write. If we can
1616 * not get the lock, simply return NULL to caller.
1618 struct address_space
*hugetlb_page_mapping_lock_write(struct page
*hpage
)
1620 struct address_space
*mapping
= page_mapping(hpage
);
1625 if (i_mmap_trylock_write(mapping
))
1631 pgoff_t
__basepage_index(struct page
*page
)
1633 struct page
*page_head
= compound_head(page
);
1634 pgoff_t index
= page_index(page_head
);
1635 unsigned long compound_idx
;
1637 if (!PageHuge(page_head
))
1638 return page_index(page
);
1640 if (compound_order(page_head
) >= MAX_ORDER
)
1641 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1643 compound_idx
= page
- page_head
;
1645 return (index
<< compound_order(page_head
)) + compound_idx
;
1648 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1649 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1650 nodemask_t
*node_alloc_noretry
)
1652 int order
= huge_page_order(h
);
1654 bool alloc_try_hard
= true;
1657 * By default we always try hard to allocate the page with
1658 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1659 * a loop (to adjust global huge page counts) and previous allocation
1660 * failed, do not continue to try hard on the same node. Use the
1661 * node_alloc_noretry bitmap to manage this state information.
1663 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1664 alloc_try_hard
= false;
1665 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1667 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1668 if (nid
== NUMA_NO_NODE
)
1669 nid
= numa_mem_id();
1670 page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, nmask
);
1672 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1674 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1677 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1678 * indicates an overall state change. Clear bit so that we resume
1679 * normal 'try hard' allocations.
1681 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1682 node_clear(nid
, *node_alloc_noretry
);
1685 * If we tried hard to get a page but failed, set bit so that
1686 * subsequent attempts will not try as hard until there is an
1687 * overall state change.
1689 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1690 node_set(nid
, *node_alloc_noretry
);
1696 * Common helper to allocate a fresh hugetlb page. All specific allocators
1697 * should use this function to get new hugetlb pages
1699 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1700 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1701 nodemask_t
*node_alloc_noretry
)
1705 if (hstate_is_gigantic(h
))
1706 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1708 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1709 nid
, nmask
, node_alloc_noretry
);
1713 if (hstate_is_gigantic(h
))
1714 prep_compound_gigantic_page(page
, huge_page_order(h
));
1715 prep_new_huge_page(h
, page
, page_to_nid(page
));
1721 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1724 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1725 nodemask_t
*node_alloc_noretry
)
1729 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1731 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1732 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1733 node_alloc_noretry
);
1741 put_page(page
); /* free it into the hugepage allocator */
1747 * Free huge page from pool from next node to free.
1748 * Attempt to keep persistent huge pages more or less
1749 * balanced over allowed nodes.
1750 * Called with hugetlb_lock locked.
1752 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1758 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1760 * If we're returning unused surplus pages, only examine
1761 * nodes with surplus pages.
1763 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1764 !list_empty(&h
->hugepage_freelists
[node
])) {
1766 list_entry(h
->hugepage_freelists
[node
].next
,
1768 list_del(&page
->lru
);
1769 h
->free_huge_pages
--;
1770 h
->free_huge_pages_node
[node
]--;
1772 h
->surplus_huge_pages
--;
1773 h
->surplus_huge_pages_node
[node
]--;
1775 update_and_free_page(h
, page
);
1785 * Dissolve a given free hugepage into free buddy pages. This function does
1786 * nothing for in-use hugepages and non-hugepages.
1787 * This function returns values like below:
1789 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1790 * (allocated or reserved.)
1791 * 0: successfully dissolved free hugepages or the page is not a
1792 * hugepage (considered as already dissolved)
1794 int dissolve_free_huge_page(struct page
*page
)
1799 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1800 if (!PageHuge(page
))
1803 spin_lock(&hugetlb_lock
);
1804 if (!PageHuge(page
)) {
1809 if (!page_count(page
)) {
1810 struct page
*head
= compound_head(page
);
1811 struct hstate
*h
= page_hstate(head
);
1812 int nid
= page_to_nid(head
);
1813 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1817 * We should make sure that the page is already on the free list
1818 * when it is dissolved.
1820 if (unlikely(!PageHugeFreed(head
))) {
1821 spin_unlock(&hugetlb_lock
);
1825 * Theoretically, we should return -EBUSY when we
1826 * encounter this race. In fact, we have a chance
1827 * to successfully dissolve the page if we do a
1828 * retry. Because the race window is quite small.
1829 * If we seize this opportunity, it is an optimization
1830 * for increasing the success rate of dissolving page.
1836 * Move PageHWPoison flag from head page to the raw error page,
1837 * which makes any subpages rather than the error page reusable.
1839 if (PageHWPoison(head
) && page
!= head
) {
1840 SetPageHWPoison(page
);
1841 ClearPageHWPoison(head
);
1843 list_del(&head
->lru
);
1844 h
->free_huge_pages
--;
1845 h
->free_huge_pages_node
[nid
]--;
1846 h
->max_huge_pages
--;
1847 update_and_free_page(h
, head
);
1851 spin_unlock(&hugetlb_lock
);
1856 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1857 * make specified memory blocks removable from the system.
1858 * Note that this will dissolve a free gigantic hugepage completely, if any
1859 * part of it lies within the given range.
1860 * Also note that if dissolve_free_huge_page() returns with an error, all
1861 * free hugepages that were dissolved before that error are lost.
1863 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1869 if (!hugepages_supported())
1872 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1873 page
= pfn_to_page(pfn
);
1874 rc
= dissolve_free_huge_page(page
);
1883 * Allocates a fresh surplus page from the page allocator.
1885 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1886 int nid
, nodemask_t
*nmask
)
1888 struct page
*page
= NULL
;
1890 if (hstate_is_gigantic(h
))
1893 spin_lock(&hugetlb_lock
);
1894 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
1896 spin_unlock(&hugetlb_lock
);
1898 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1902 spin_lock(&hugetlb_lock
);
1904 * We could have raced with the pool size change.
1905 * Double check that and simply deallocate the new page
1906 * if we would end up overcommiting the surpluses. Abuse
1907 * temporary page to workaround the nasty free_huge_page
1910 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1911 SetPageHugeTemporary(page
);
1912 spin_unlock(&hugetlb_lock
);
1916 h
->surplus_huge_pages
++;
1917 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
1921 spin_unlock(&hugetlb_lock
);
1926 static struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1927 int nid
, nodemask_t
*nmask
)
1931 if (hstate_is_gigantic(h
))
1934 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1939 * We do not account these pages as surplus because they are only
1940 * temporary and will be released properly on the last reference
1942 SetPageHugeTemporary(page
);
1948 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1951 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1952 struct vm_area_struct
*vma
, unsigned long addr
)
1955 struct mempolicy
*mpol
;
1956 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1958 nodemask_t
*nodemask
;
1960 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
1961 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
);
1962 mpol_cond_put(mpol
);
1967 /* page migration callback function */
1968 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
1969 nodemask_t
*nmask
, gfp_t gfp_mask
)
1971 spin_lock(&hugetlb_lock
);
1972 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1975 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
1977 spin_unlock(&hugetlb_lock
);
1981 spin_unlock(&hugetlb_lock
);
1983 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
1986 /* mempolicy aware migration callback */
1987 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
1988 unsigned long address
)
1990 struct mempolicy
*mpol
;
1991 nodemask_t
*nodemask
;
1996 gfp_mask
= htlb_alloc_mask(h
);
1997 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1998 page
= alloc_huge_page_nodemask(h
, node
, nodemask
, gfp_mask
);
1999 mpol_cond_put(mpol
);
2005 * Increase the hugetlb pool such that it can accommodate a reservation
2008 static int gather_surplus_pages(struct hstate
*h
, long delta
)
2009 __must_hold(&hugetlb_lock
)
2011 struct list_head surplus_list
;
2012 struct page
*page
, *tmp
;
2015 long needed
, allocated
;
2016 bool alloc_ok
= true;
2018 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
2020 h
->resv_huge_pages
+= delta
;
2025 INIT_LIST_HEAD(&surplus_list
);
2029 spin_unlock(&hugetlb_lock
);
2030 for (i
= 0; i
< needed
; i
++) {
2031 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
2032 NUMA_NO_NODE
, NULL
);
2037 list_add(&page
->lru
, &surplus_list
);
2043 * After retaking hugetlb_lock, we need to recalculate 'needed'
2044 * because either resv_huge_pages or free_huge_pages may have changed.
2046 spin_lock(&hugetlb_lock
);
2047 needed
= (h
->resv_huge_pages
+ delta
) -
2048 (h
->free_huge_pages
+ allocated
);
2053 * We were not able to allocate enough pages to
2054 * satisfy the entire reservation so we free what
2055 * we've allocated so far.
2060 * The surplus_list now contains _at_least_ the number of extra pages
2061 * needed to accommodate the reservation. Add the appropriate number
2062 * of pages to the hugetlb pool and free the extras back to the buddy
2063 * allocator. Commit the entire reservation here to prevent another
2064 * process from stealing the pages as they are added to the pool but
2065 * before they are reserved.
2067 needed
+= allocated
;
2068 h
->resv_huge_pages
+= delta
;
2071 /* Free the needed pages to the hugetlb pool */
2072 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
2078 * This page is now managed by the hugetlb allocator and has
2079 * no users -- drop the buddy allocator's reference.
2081 zeroed
= put_page_testzero(page
);
2082 VM_BUG_ON_PAGE(!zeroed
, page
);
2083 enqueue_huge_page(h
, page
);
2086 spin_unlock(&hugetlb_lock
);
2088 /* Free unnecessary surplus pages to the buddy allocator */
2089 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
2091 spin_lock(&hugetlb_lock
);
2097 * This routine has two main purposes:
2098 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2099 * in unused_resv_pages. This corresponds to the prior adjustments made
2100 * to the associated reservation map.
2101 * 2) Free any unused surplus pages that may have been allocated to satisfy
2102 * the reservation. As many as unused_resv_pages may be freed.
2104 * Called with hugetlb_lock held. However, the lock could be dropped (and
2105 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2106 * we must make sure nobody else can claim pages we are in the process of
2107 * freeing. Do this by ensuring resv_huge_page always is greater than the
2108 * number of huge pages we plan to free when dropping the lock.
2110 static void return_unused_surplus_pages(struct hstate
*h
,
2111 unsigned long unused_resv_pages
)
2113 unsigned long nr_pages
;
2115 /* Cannot return gigantic pages currently */
2116 if (hstate_is_gigantic(h
))
2120 * Part (or even all) of the reservation could have been backed
2121 * by pre-allocated pages. Only free surplus pages.
2123 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
2126 * We want to release as many surplus pages as possible, spread
2127 * evenly across all nodes with memory. Iterate across these nodes
2128 * until we can no longer free unreserved surplus pages. This occurs
2129 * when the nodes with surplus pages have no free pages.
2130 * free_pool_huge_page() will balance the freed pages across the
2131 * on-line nodes with memory and will handle the hstate accounting.
2133 * Note that we decrement resv_huge_pages as we free the pages. If
2134 * we drop the lock, resv_huge_pages will still be sufficiently large
2135 * to cover subsequent pages we may free.
2137 while (nr_pages
--) {
2138 h
->resv_huge_pages
--;
2139 unused_resv_pages
--;
2140 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
2142 cond_resched_lock(&hugetlb_lock
);
2146 /* Fully uncommit the reservation */
2147 h
->resv_huge_pages
-= unused_resv_pages
;
2152 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2153 * are used by the huge page allocation routines to manage reservations.
2155 * vma_needs_reservation is called to determine if the huge page at addr
2156 * within the vma has an associated reservation. If a reservation is
2157 * needed, the value 1 is returned. The caller is then responsible for
2158 * managing the global reservation and subpool usage counts. After
2159 * the huge page has been allocated, vma_commit_reservation is called
2160 * to add the page to the reservation map. If the page allocation fails,
2161 * the reservation must be ended instead of committed. vma_end_reservation
2162 * is called in such cases.
2164 * In the normal case, vma_commit_reservation returns the same value
2165 * as the preceding vma_needs_reservation call. The only time this
2166 * is not the case is if a reserve map was changed between calls. It
2167 * is the responsibility of the caller to notice the difference and
2168 * take appropriate action.
2170 * vma_add_reservation is used in error paths where a reservation must
2171 * be restored when a newly allocated huge page must be freed. It is
2172 * to be called after calling vma_needs_reservation to determine if a
2173 * reservation exists.
2175 enum vma_resv_mode
{
2181 static long __vma_reservation_common(struct hstate
*h
,
2182 struct vm_area_struct
*vma
, unsigned long addr
,
2183 enum vma_resv_mode mode
)
2185 struct resv_map
*resv
;
2188 long dummy_out_regions_needed
;
2190 resv
= vma_resv_map(vma
);
2194 idx
= vma_hugecache_offset(h
, vma
, addr
);
2196 case VMA_NEEDS_RESV
:
2197 ret
= region_chg(resv
, idx
, idx
+ 1, &dummy_out_regions_needed
);
2198 /* We assume that vma_reservation_* routines always operate on
2199 * 1 page, and that adding to resv map a 1 page entry can only
2200 * ever require 1 region.
2202 VM_BUG_ON(dummy_out_regions_needed
!= 1);
2204 case VMA_COMMIT_RESV
:
2205 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2206 /* region_add calls of range 1 should never fail. */
2210 region_abort(resv
, idx
, idx
+ 1, 1);
2214 if (vma
->vm_flags
& VM_MAYSHARE
) {
2215 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2216 /* region_add calls of range 1 should never fail. */
2219 region_abort(resv
, idx
, idx
+ 1, 1);
2220 ret
= region_del(resv
, idx
, idx
+ 1);
2227 if (vma
->vm_flags
& VM_MAYSHARE
)
2229 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
2231 * In most cases, reserves always exist for private mappings.
2232 * However, a file associated with mapping could have been
2233 * hole punched or truncated after reserves were consumed.
2234 * As subsequent fault on such a range will not use reserves.
2235 * Subtle - The reserve map for private mappings has the
2236 * opposite meaning than that of shared mappings. If NO
2237 * entry is in the reserve map, it means a reservation exists.
2238 * If an entry exists in the reserve map, it means the
2239 * reservation has already been consumed. As a result, the
2240 * return value of this routine is the opposite of the
2241 * value returned from reserve map manipulation routines above.
2249 return ret
< 0 ? ret
: 0;
2252 static long vma_needs_reservation(struct hstate
*h
,
2253 struct vm_area_struct
*vma
, unsigned long addr
)
2255 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2258 static long vma_commit_reservation(struct hstate
*h
,
2259 struct vm_area_struct
*vma
, unsigned long addr
)
2261 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2264 static void vma_end_reservation(struct hstate
*h
,
2265 struct vm_area_struct
*vma
, unsigned long addr
)
2267 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2270 static long vma_add_reservation(struct hstate
*h
,
2271 struct vm_area_struct
*vma
, unsigned long addr
)
2273 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2277 * This routine is called to restore a reservation on error paths. In the
2278 * specific error paths, a huge page was allocated (via alloc_huge_page)
2279 * and is about to be freed. If a reservation for the page existed,
2280 * alloc_huge_page would have consumed the reservation and set PagePrivate
2281 * in the newly allocated page. When the page is freed via free_huge_page,
2282 * the global reservation count will be incremented if PagePrivate is set.
2283 * However, free_huge_page can not adjust the reserve map. Adjust the
2284 * reserve map here to be consistent with global reserve count adjustments
2285 * to be made by free_huge_page.
2287 static void restore_reserve_on_error(struct hstate
*h
,
2288 struct vm_area_struct
*vma
, unsigned long address
,
2291 if (unlikely(PagePrivate(page
))) {
2292 long rc
= vma_needs_reservation(h
, vma
, address
);
2294 if (unlikely(rc
< 0)) {
2296 * Rare out of memory condition in reserve map
2297 * manipulation. Clear PagePrivate so that
2298 * global reserve count will not be incremented
2299 * by free_huge_page. This will make it appear
2300 * as though the reservation for this page was
2301 * consumed. This may prevent the task from
2302 * faulting in the page at a later time. This
2303 * is better than inconsistent global huge page
2304 * accounting of reserve counts.
2306 ClearPagePrivate(page
);
2308 rc
= vma_add_reservation(h
, vma
, address
);
2309 if (unlikely(rc
< 0))
2311 * See above comment about rare out of
2314 ClearPagePrivate(page
);
2316 vma_end_reservation(h
, vma
, address
);
2320 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2321 unsigned long addr
, int avoid_reserve
)
2323 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2324 struct hstate
*h
= hstate_vma(vma
);
2326 long map_chg
, map_commit
;
2329 struct hugetlb_cgroup
*h_cg
;
2330 bool deferred_reserve
;
2332 idx
= hstate_index(h
);
2334 * Examine the region/reserve map to determine if the process
2335 * has a reservation for the page to be allocated. A return
2336 * code of zero indicates a reservation exists (no change).
2338 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2340 return ERR_PTR(-ENOMEM
);
2343 * Processes that did not create the mapping will have no
2344 * reserves as indicated by the region/reserve map. Check
2345 * that the allocation will not exceed the subpool limit.
2346 * Allocations for MAP_NORESERVE mappings also need to be
2347 * checked against any subpool limit.
2349 if (map_chg
|| avoid_reserve
) {
2350 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2352 vma_end_reservation(h
, vma
, addr
);
2353 return ERR_PTR(-ENOSPC
);
2357 * Even though there was no reservation in the region/reserve
2358 * map, there could be reservations associated with the
2359 * subpool that can be used. This would be indicated if the
2360 * return value of hugepage_subpool_get_pages() is zero.
2361 * However, if avoid_reserve is specified we still avoid even
2362 * the subpool reservations.
2368 /* If this allocation is not consuming a reservation, charge it now.
2370 deferred_reserve
= map_chg
|| avoid_reserve
|| !vma_resv_map(vma
);
2371 if (deferred_reserve
) {
2372 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
2373 idx
, pages_per_huge_page(h
), &h_cg
);
2375 goto out_subpool_put
;
2378 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2380 goto out_uncharge_cgroup_reservation
;
2382 spin_lock(&hugetlb_lock
);
2384 * glb_chg is passed to indicate whether or not a page must be taken
2385 * from the global free pool (global change). gbl_chg == 0 indicates
2386 * a reservation exists for the allocation.
2388 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2390 spin_unlock(&hugetlb_lock
);
2391 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2393 goto out_uncharge_cgroup
;
2394 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2395 SetPagePrivate(page
);
2396 h
->resv_huge_pages
--;
2398 spin_lock(&hugetlb_lock
);
2399 list_add(&page
->lru
, &h
->hugepage_activelist
);
2402 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2403 /* If allocation is not consuming a reservation, also store the
2404 * hugetlb_cgroup pointer on the page.
2406 if (deferred_reserve
) {
2407 hugetlb_cgroup_commit_charge_rsvd(idx
, pages_per_huge_page(h
),
2411 spin_unlock(&hugetlb_lock
);
2413 set_page_private(page
, (unsigned long)spool
);
2415 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2416 if (unlikely(map_chg
> map_commit
)) {
2418 * The page was added to the reservation map between
2419 * vma_needs_reservation and vma_commit_reservation.
2420 * This indicates a race with hugetlb_reserve_pages.
2421 * Adjust for the subpool count incremented above AND
2422 * in hugetlb_reserve_pages for the same page. Also,
2423 * the reservation count added in hugetlb_reserve_pages
2424 * no longer applies.
2428 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2429 hugetlb_acct_memory(h
, -rsv_adjust
);
2430 if (deferred_reserve
)
2431 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
2432 pages_per_huge_page(h
), page
);
2436 out_uncharge_cgroup
:
2437 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2438 out_uncharge_cgroup_reservation
:
2439 if (deferred_reserve
)
2440 hugetlb_cgroup_uncharge_cgroup_rsvd(idx
, pages_per_huge_page(h
),
2443 if (map_chg
|| avoid_reserve
)
2444 hugepage_subpool_put_pages(spool
, 1);
2445 vma_end_reservation(h
, vma
, addr
);
2446 return ERR_PTR(-ENOSPC
);
2449 int alloc_bootmem_huge_page(struct hstate
*h
)
2450 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2451 int __alloc_bootmem_huge_page(struct hstate
*h
)
2453 struct huge_bootmem_page
*m
;
2456 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2459 addr
= memblock_alloc_try_nid_raw(
2460 huge_page_size(h
), huge_page_size(h
),
2461 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2464 * Use the beginning of the huge page to store the
2465 * huge_bootmem_page struct (until gather_bootmem
2466 * puts them into the mem_map).
2475 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2476 /* Put them into a private list first because mem_map is not up yet */
2477 INIT_LIST_HEAD(&m
->list
);
2478 list_add(&m
->list
, &huge_boot_pages
);
2483 static void __init
prep_compound_huge_page(struct page
*page
,
2486 if (unlikely(order
> (MAX_ORDER
- 1)))
2487 prep_compound_gigantic_page(page
, order
);
2489 prep_compound_page(page
, order
);
2492 /* Put bootmem huge pages into the standard lists after mem_map is up */
2493 static void __init
gather_bootmem_prealloc(void)
2495 struct huge_bootmem_page
*m
;
2497 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2498 struct page
*page
= virt_to_page(m
);
2499 struct hstate
*h
= m
->hstate
;
2501 WARN_ON(page_count(page
) != 1);
2502 prep_compound_huge_page(page
, h
->order
);
2503 WARN_ON(PageReserved(page
));
2504 prep_new_huge_page(h
, page
, page_to_nid(page
));
2505 put_page(page
); /* free it into the hugepage allocator */
2508 * If we had gigantic hugepages allocated at boot time, we need
2509 * to restore the 'stolen' pages to totalram_pages in order to
2510 * fix confusing memory reports from free(1) and another
2511 * side-effects, like CommitLimit going negative.
2513 if (hstate_is_gigantic(h
))
2514 adjust_managed_page_count(page
, 1 << h
->order
);
2519 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2522 nodemask_t
*node_alloc_noretry
;
2524 if (!hstate_is_gigantic(h
)) {
2526 * Bit mask controlling how hard we retry per-node allocations.
2527 * Ignore errors as lower level routines can deal with
2528 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2529 * time, we are likely in bigger trouble.
2531 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2534 /* allocations done at boot time */
2535 node_alloc_noretry
= NULL
;
2538 /* bit mask controlling how hard we retry per-node allocations */
2539 if (node_alloc_noretry
)
2540 nodes_clear(*node_alloc_noretry
);
2542 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2543 if (hstate_is_gigantic(h
)) {
2544 if (hugetlb_cma_size
) {
2545 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2548 if (!alloc_bootmem_huge_page(h
))
2550 } else if (!alloc_pool_huge_page(h
,
2551 &node_states
[N_MEMORY
],
2552 node_alloc_noretry
))
2556 if (i
< h
->max_huge_pages
) {
2559 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2560 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2561 h
->max_huge_pages
, buf
, i
);
2562 h
->max_huge_pages
= i
;
2565 kfree(node_alloc_noretry
);
2568 static void __init
hugetlb_init_hstates(void)
2572 for_each_hstate(h
) {
2573 if (minimum_order
> huge_page_order(h
))
2574 minimum_order
= huge_page_order(h
);
2576 /* oversize hugepages were init'ed in early boot */
2577 if (!hstate_is_gigantic(h
))
2578 hugetlb_hstate_alloc_pages(h
);
2580 VM_BUG_ON(minimum_order
== UINT_MAX
);
2583 static void __init
report_hugepages(void)
2587 for_each_hstate(h
) {
2590 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2591 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2592 buf
, h
->free_huge_pages
);
2596 #ifdef CONFIG_HIGHMEM
2597 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2598 nodemask_t
*nodes_allowed
)
2602 if (hstate_is_gigantic(h
))
2605 for_each_node_mask(i
, *nodes_allowed
) {
2606 struct page
*page
, *next
;
2607 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2608 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2609 if (count
>= h
->nr_huge_pages
)
2611 if (PageHighMem(page
))
2613 list_del(&page
->lru
);
2614 update_and_free_page(h
, page
);
2615 h
->free_huge_pages
--;
2616 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2621 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2622 nodemask_t
*nodes_allowed
)
2628 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2629 * balanced by operating on them in a round-robin fashion.
2630 * Returns 1 if an adjustment was made.
2632 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2637 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2640 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2641 if (h
->surplus_huge_pages_node
[node
])
2645 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2646 if (h
->surplus_huge_pages_node
[node
] <
2647 h
->nr_huge_pages_node
[node
])
2654 h
->surplus_huge_pages
+= delta
;
2655 h
->surplus_huge_pages_node
[node
] += delta
;
2659 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2660 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
2661 nodemask_t
*nodes_allowed
)
2663 unsigned long min_count
, ret
;
2664 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
2667 * Bit mask controlling how hard we retry per-node allocations.
2668 * If we can not allocate the bit mask, do not attempt to allocate
2669 * the requested huge pages.
2671 if (node_alloc_noretry
)
2672 nodes_clear(*node_alloc_noretry
);
2676 spin_lock(&hugetlb_lock
);
2679 * Check for a node specific request.
2680 * Changing node specific huge page count may require a corresponding
2681 * change to the global count. In any case, the passed node mask
2682 * (nodes_allowed) will restrict alloc/free to the specified node.
2684 if (nid
!= NUMA_NO_NODE
) {
2685 unsigned long old_count
= count
;
2687 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2689 * User may have specified a large count value which caused the
2690 * above calculation to overflow. In this case, they wanted
2691 * to allocate as many huge pages as possible. Set count to
2692 * largest possible value to align with their intention.
2694 if (count
< old_count
)
2699 * Gigantic pages runtime allocation depend on the capability for large
2700 * page range allocation.
2701 * If the system does not provide this feature, return an error when
2702 * the user tries to allocate gigantic pages but let the user free the
2703 * boottime allocated gigantic pages.
2705 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
2706 if (count
> persistent_huge_pages(h
)) {
2707 spin_unlock(&hugetlb_lock
);
2708 NODEMASK_FREE(node_alloc_noretry
);
2711 /* Fall through to decrease pool */
2715 * Increase the pool size
2716 * First take pages out of surplus state. Then make up the
2717 * remaining difference by allocating fresh huge pages.
2719 * We might race with alloc_surplus_huge_page() here and be unable
2720 * to convert a surplus huge page to a normal huge page. That is
2721 * not critical, though, it just means the overall size of the
2722 * pool might be one hugepage larger than it needs to be, but
2723 * within all the constraints specified by the sysctls.
2725 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2726 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2730 while (count
> persistent_huge_pages(h
)) {
2732 * If this allocation races such that we no longer need the
2733 * page, free_huge_page will handle it by freeing the page
2734 * and reducing the surplus.
2736 spin_unlock(&hugetlb_lock
);
2738 /* yield cpu to avoid soft lockup */
2741 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
2742 node_alloc_noretry
);
2743 spin_lock(&hugetlb_lock
);
2747 /* Bail for signals. Probably ctrl-c from user */
2748 if (signal_pending(current
))
2753 * Decrease the pool size
2754 * First return free pages to the buddy allocator (being careful
2755 * to keep enough around to satisfy reservations). Then place
2756 * pages into surplus state as needed so the pool will shrink
2757 * to the desired size as pages become free.
2759 * By placing pages into the surplus state independent of the
2760 * overcommit value, we are allowing the surplus pool size to
2761 * exceed overcommit. There are few sane options here. Since
2762 * alloc_surplus_huge_page() is checking the global counter,
2763 * though, we'll note that we're not allowed to exceed surplus
2764 * and won't grow the pool anywhere else. Not until one of the
2765 * sysctls are changed, or the surplus pages go out of use.
2767 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2768 min_count
= max(count
, min_count
);
2769 try_to_free_low(h
, min_count
, nodes_allowed
);
2770 while (min_count
< persistent_huge_pages(h
)) {
2771 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2773 cond_resched_lock(&hugetlb_lock
);
2775 while (count
< persistent_huge_pages(h
)) {
2776 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2780 h
->max_huge_pages
= persistent_huge_pages(h
);
2781 spin_unlock(&hugetlb_lock
);
2783 NODEMASK_FREE(node_alloc_noretry
);
2788 #define HSTATE_ATTR_RO(_name) \
2789 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2791 #define HSTATE_ATTR(_name) \
2792 static struct kobj_attribute _name##_attr = \
2793 __ATTR(_name, 0644, _name##_show, _name##_store)
2795 static struct kobject
*hugepages_kobj
;
2796 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2798 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2800 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2804 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2805 if (hstate_kobjs
[i
] == kobj
) {
2807 *nidp
= NUMA_NO_NODE
;
2811 return kobj_to_node_hstate(kobj
, nidp
);
2814 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2815 struct kobj_attribute
*attr
, char *buf
)
2818 unsigned long nr_huge_pages
;
2821 h
= kobj_to_hstate(kobj
, &nid
);
2822 if (nid
== NUMA_NO_NODE
)
2823 nr_huge_pages
= h
->nr_huge_pages
;
2825 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2827 return sysfs_emit(buf
, "%lu\n", nr_huge_pages
);
2830 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2831 struct hstate
*h
, int nid
,
2832 unsigned long count
, size_t len
)
2835 nodemask_t nodes_allowed
, *n_mask
;
2837 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
2840 if (nid
== NUMA_NO_NODE
) {
2842 * global hstate attribute
2844 if (!(obey_mempolicy
&&
2845 init_nodemask_of_mempolicy(&nodes_allowed
)))
2846 n_mask
= &node_states
[N_MEMORY
];
2848 n_mask
= &nodes_allowed
;
2851 * Node specific request. count adjustment happens in
2852 * set_max_huge_pages() after acquiring hugetlb_lock.
2854 init_nodemask_of_node(&nodes_allowed
, nid
);
2855 n_mask
= &nodes_allowed
;
2858 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
2860 return err
? err
: len
;
2863 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2864 struct kobject
*kobj
, const char *buf
,
2868 unsigned long count
;
2872 err
= kstrtoul(buf
, 10, &count
);
2876 h
= kobj_to_hstate(kobj
, &nid
);
2877 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2880 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2881 struct kobj_attribute
*attr
, char *buf
)
2883 return nr_hugepages_show_common(kobj
, attr
, buf
);
2886 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2887 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2889 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2891 HSTATE_ATTR(nr_hugepages
);
2896 * hstate attribute for optionally mempolicy-based constraint on persistent
2897 * huge page alloc/free.
2899 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2900 struct kobj_attribute
*attr
,
2903 return nr_hugepages_show_common(kobj
, attr
, buf
);
2906 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2907 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2909 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2911 HSTATE_ATTR(nr_hugepages_mempolicy
);
2915 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2916 struct kobj_attribute
*attr
, char *buf
)
2918 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2919 return sysfs_emit(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2922 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2923 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2926 unsigned long input
;
2927 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2929 if (hstate_is_gigantic(h
))
2932 err
= kstrtoul(buf
, 10, &input
);
2936 spin_lock(&hugetlb_lock
);
2937 h
->nr_overcommit_huge_pages
= input
;
2938 spin_unlock(&hugetlb_lock
);
2942 HSTATE_ATTR(nr_overcommit_hugepages
);
2944 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2945 struct kobj_attribute
*attr
, char *buf
)
2948 unsigned long free_huge_pages
;
2951 h
= kobj_to_hstate(kobj
, &nid
);
2952 if (nid
== NUMA_NO_NODE
)
2953 free_huge_pages
= h
->free_huge_pages
;
2955 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2957 return sysfs_emit(buf
, "%lu\n", free_huge_pages
);
2959 HSTATE_ATTR_RO(free_hugepages
);
2961 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2962 struct kobj_attribute
*attr
, char *buf
)
2964 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2965 return sysfs_emit(buf
, "%lu\n", h
->resv_huge_pages
);
2967 HSTATE_ATTR_RO(resv_hugepages
);
2969 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2970 struct kobj_attribute
*attr
, char *buf
)
2973 unsigned long surplus_huge_pages
;
2976 h
= kobj_to_hstate(kobj
, &nid
);
2977 if (nid
== NUMA_NO_NODE
)
2978 surplus_huge_pages
= h
->surplus_huge_pages
;
2980 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2982 return sysfs_emit(buf
, "%lu\n", surplus_huge_pages
);
2984 HSTATE_ATTR_RO(surplus_hugepages
);
2986 static struct attribute
*hstate_attrs
[] = {
2987 &nr_hugepages_attr
.attr
,
2988 &nr_overcommit_hugepages_attr
.attr
,
2989 &free_hugepages_attr
.attr
,
2990 &resv_hugepages_attr
.attr
,
2991 &surplus_hugepages_attr
.attr
,
2993 &nr_hugepages_mempolicy_attr
.attr
,
2998 static const struct attribute_group hstate_attr_group
= {
2999 .attrs
= hstate_attrs
,
3002 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
3003 struct kobject
**hstate_kobjs
,
3004 const struct attribute_group
*hstate_attr_group
)
3007 int hi
= hstate_index(h
);
3009 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
3010 if (!hstate_kobjs
[hi
])
3013 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
3015 kobject_put(hstate_kobjs
[hi
]);
3016 hstate_kobjs
[hi
] = NULL
;
3022 static void __init
hugetlb_sysfs_init(void)
3027 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
3028 if (!hugepages_kobj
)
3031 for_each_hstate(h
) {
3032 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
3033 hstate_kobjs
, &hstate_attr_group
);
3035 pr_err("HugeTLB: Unable to add hstate %s", h
->name
);
3042 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3043 * with node devices in node_devices[] using a parallel array. The array
3044 * index of a node device or _hstate == node id.
3045 * This is here to avoid any static dependency of the node device driver, in
3046 * the base kernel, on the hugetlb module.
3048 struct node_hstate
{
3049 struct kobject
*hugepages_kobj
;
3050 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
3052 static struct node_hstate node_hstates
[MAX_NUMNODES
];
3055 * A subset of global hstate attributes for node devices
3057 static struct attribute
*per_node_hstate_attrs
[] = {
3058 &nr_hugepages_attr
.attr
,
3059 &free_hugepages_attr
.attr
,
3060 &surplus_hugepages_attr
.attr
,
3064 static const struct attribute_group per_node_hstate_attr_group
= {
3065 .attrs
= per_node_hstate_attrs
,
3069 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3070 * Returns node id via non-NULL nidp.
3072 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3076 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
3077 struct node_hstate
*nhs
= &node_hstates
[nid
];
3079 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
3080 if (nhs
->hstate_kobjs
[i
] == kobj
) {
3092 * Unregister hstate attributes from a single node device.
3093 * No-op if no hstate attributes attached.
3095 static void hugetlb_unregister_node(struct node
*node
)
3098 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3100 if (!nhs
->hugepages_kobj
)
3101 return; /* no hstate attributes */
3103 for_each_hstate(h
) {
3104 int idx
= hstate_index(h
);
3105 if (nhs
->hstate_kobjs
[idx
]) {
3106 kobject_put(nhs
->hstate_kobjs
[idx
]);
3107 nhs
->hstate_kobjs
[idx
] = NULL
;
3111 kobject_put(nhs
->hugepages_kobj
);
3112 nhs
->hugepages_kobj
= NULL
;
3117 * Register hstate attributes for a single node device.
3118 * No-op if attributes already registered.
3120 static void hugetlb_register_node(struct node
*node
)
3123 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3126 if (nhs
->hugepages_kobj
)
3127 return; /* already allocated */
3129 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
3131 if (!nhs
->hugepages_kobj
)
3134 for_each_hstate(h
) {
3135 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
3137 &per_node_hstate_attr_group
);
3139 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3140 h
->name
, node
->dev
.id
);
3141 hugetlb_unregister_node(node
);
3148 * hugetlb init time: register hstate attributes for all registered node
3149 * devices of nodes that have memory. All on-line nodes should have
3150 * registered their associated device by this time.
3152 static void __init
hugetlb_register_all_nodes(void)
3156 for_each_node_state(nid
, N_MEMORY
) {
3157 struct node
*node
= node_devices
[nid
];
3158 if (node
->dev
.id
== nid
)
3159 hugetlb_register_node(node
);
3163 * Let the node device driver know we're here so it can
3164 * [un]register hstate attributes on node hotplug.
3166 register_hugetlbfs_with_node(hugetlb_register_node
,
3167 hugetlb_unregister_node
);
3169 #else /* !CONFIG_NUMA */
3171 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3179 static void hugetlb_register_all_nodes(void) { }
3183 static int __init
hugetlb_init(void)
3187 if (!hugepages_supported()) {
3188 if (hugetlb_max_hstate
|| default_hstate_max_huge_pages
)
3189 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3194 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3195 * architectures depend on setup being done here.
3197 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
3198 if (!parsed_default_hugepagesz
) {
3200 * If we did not parse a default huge page size, set
3201 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3202 * number of huge pages for this default size was implicitly
3203 * specified, set that here as well.
3204 * Note that the implicit setting will overwrite an explicit
3205 * setting. A warning will be printed in this case.
3207 default_hstate_idx
= hstate_index(size_to_hstate(HPAGE_SIZE
));
3208 if (default_hstate_max_huge_pages
) {
3209 if (default_hstate
.max_huge_pages
) {
3212 string_get_size(huge_page_size(&default_hstate
),
3213 1, STRING_UNITS_2
, buf
, 32);
3214 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3215 default_hstate
.max_huge_pages
, buf
);
3216 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3217 default_hstate_max_huge_pages
);
3219 default_hstate
.max_huge_pages
=
3220 default_hstate_max_huge_pages
;
3224 hugetlb_cma_check();
3225 hugetlb_init_hstates();
3226 gather_bootmem_prealloc();
3229 hugetlb_sysfs_init();
3230 hugetlb_register_all_nodes();
3231 hugetlb_cgroup_file_init();
3234 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
3236 num_fault_mutexes
= 1;
3238 hugetlb_fault_mutex_table
=
3239 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
3241 BUG_ON(!hugetlb_fault_mutex_table
);
3243 for (i
= 0; i
< num_fault_mutexes
; i
++)
3244 mutex_init(&hugetlb_fault_mutex_table
[i
]);
3247 subsys_initcall(hugetlb_init
);
3249 /* Overwritten by architectures with more huge page sizes */
3250 bool __init
__attribute((weak
)) arch_hugetlb_valid_size(unsigned long size
)
3252 return size
== HPAGE_SIZE
;
3255 void __init
hugetlb_add_hstate(unsigned int order
)
3260 if (size_to_hstate(PAGE_SIZE
<< order
)) {
3263 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
3265 h
= &hstates
[hugetlb_max_hstate
++];
3267 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
3268 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
3269 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
3270 INIT_LIST_HEAD(&h
->hugepage_activelist
);
3271 h
->next_nid_to_alloc
= first_memory_node
;
3272 h
->next_nid_to_free
= first_memory_node
;
3273 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
3274 huge_page_size(h
)/1024);
3280 * hugepages command line processing
3281 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3282 * specification. If not, ignore the hugepages value. hugepages can also
3283 * be the first huge page command line option in which case it implicitly
3284 * specifies the number of huge pages for the default size.
3286 static int __init
hugepages_setup(char *s
)
3289 static unsigned long *last_mhp
;
3291 if (!parsed_valid_hugepagesz
) {
3292 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s
);
3293 parsed_valid_hugepagesz
= true;
3298 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3299 * yet, so this hugepages= parameter goes to the "default hstate".
3300 * Otherwise, it goes with the previously parsed hugepagesz or
3301 * default_hugepagesz.
3303 else if (!hugetlb_max_hstate
)
3304 mhp
= &default_hstate_max_huge_pages
;
3306 mhp
= &parsed_hstate
->max_huge_pages
;
3308 if (mhp
== last_mhp
) {
3309 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s
);
3313 if (sscanf(s
, "%lu", mhp
) <= 0)
3317 * Global state is always initialized later in hugetlb_init.
3318 * But we need to allocate >= MAX_ORDER hstates here early to still
3319 * use the bootmem allocator.
3321 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
3322 hugetlb_hstate_alloc_pages(parsed_hstate
);
3328 __setup("hugepages=", hugepages_setup
);
3331 * hugepagesz command line processing
3332 * A specific huge page size can only be specified once with hugepagesz.
3333 * hugepagesz is followed by hugepages on the command line. The global
3334 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3335 * hugepagesz argument was valid.
3337 static int __init
hugepagesz_setup(char *s
)
3342 parsed_valid_hugepagesz
= false;
3343 size
= (unsigned long)memparse(s
, NULL
);
3345 if (!arch_hugetlb_valid_size(size
)) {
3346 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s
);
3350 h
= size_to_hstate(size
);
3353 * hstate for this size already exists. This is normally
3354 * an error, but is allowed if the existing hstate is the
3355 * default hstate. More specifically, it is only allowed if
3356 * the number of huge pages for the default hstate was not
3357 * previously specified.
3359 if (!parsed_default_hugepagesz
|| h
!= &default_hstate
||
3360 default_hstate
.max_huge_pages
) {
3361 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s
);
3366 * No need to call hugetlb_add_hstate() as hstate already
3367 * exists. But, do set parsed_hstate so that a following
3368 * hugepages= parameter will be applied to this hstate.
3371 parsed_valid_hugepagesz
= true;
3375 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3376 parsed_valid_hugepagesz
= true;
3379 __setup("hugepagesz=", hugepagesz_setup
);
3382 * default_hugepagesz command line input
3383 * Only one instance of default_hugepagesz allowed on command line.
3385 static int __init
default_hugepagesz_setup(char *s
)
3389 parsed_valid_hugepagesz
= false;
3390 if (parsed_default_hugepagesz
) {
3391 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s
);
3395 size
= (unsigned long)memparse(s
, NULL
);
3397 if (!arch_hugetlb_valid_size(size
)) {
3398 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s
);
3402 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3403 parsed_valid_hugepagesz
= true;
3404 parsed_default_hugepagesz
= true;
3405 default_hstate_idx
= hstate_index(size_to_hstate(size
));
3408 * The number of default huge pages (for this size) could have been
3409 * specified as the first hugetlb parameter: hugepages=X. If so,
3410 * then default_hstate_max_huge_pages is set. If the default huge
3411 * page size is gigantic (>= MAX_ORDER), then the pages must be
3412 * allocated here from bootmem allocator.
3414 if (default_hstate_max_huge_pages
) {
3415 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
3416 if (hstate_is_gigantic(&default_hstate
))
3417 hugetlb_hstate_alloc_pages(&default_hstate
);
3418 default_hstate_max_huge_pages
= 0;
3423 __setup("default_hugepagesz=", default_hugepagesz_setup
);
3425 static unsigned int allowed_mems_nr(struct hstate
*h
)
3428 unsigned int nr
= 0;
3429 nodemask_t
*mpol_allowed
;
3430 unsigned int *array
= h
->free_huge_pages_node
;
3431 gfp_t gfp_mask
= htlb_alloc_mask(h
);
3433 mpol_allowed
= policy_nodemask_current(gfp_mask
);
3435 for_each_node_mask(node
, cpuset_current_mems_allowed
) {
3436 if (!mpol_allowed
||
3437 (mpol_allowed
&& node_isset(node
, *mpol_allowed
)))
3444 #ifdef CONFIG_SYSCTL
3445 static int proc_hugetlb_doulongvec_minmax(struct ctl_table
*table
, int write
,
3446 void *buffer
, size_t *length
,
3447 loff_t
*ppos
, unsigned long *out
)
3449 struct ctl_table dup_table
;
3452 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3453 * can duplicate the @table and alter the duplicate of it.
3456 dup_table
.data
= out
;
3458 return proc_doulongvec_minmax(&dup_table
, write
, buffer
, length
, ppos
);
3461 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3462 struct ctl_table
*table
, int write
,
3463 void *buffer
, size_t *length
, loff_t
*ppos
)
3465 struct hstate
*h
= &default_hstate
;
3466 unsigned long tmp
= h
->max_huge_pages
;
3469 if (!hugepages_supported())
3472 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3478 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3479 NUMA_NO_NODE
, tmp
, *length
);
3484 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3485 void *buffer
, size_t *length
, loff_t
*ppos
)
3488 return hugetlb_sysctl_handler_common(false, table
, write
,
3489 buffer
, length
, ppos
);
3493 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3494 void *buffer
, size_t *length
, loff_t
*ppos
)
3496 return hugetlb_sysctl_handler_common(true, table
, write
,
3497 buffer
, length
, ppos
);
3499 #endif /* CONFIG_NUMA */
3501 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3502 void *buffer
, size_t *length
, loff_t
*ppos
)
3504 struct hstate
*h
= &default_hstate
;
3508 if (!hugepages_supported())
3511 tmp
= h
->nr_overcommit_huge_pages
;
3513 if (write
&& hstate_is_gigantic(h
))
3516 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3522 spin_lock(&hugetlb_lock
);
3523 h
->nr_overcommit_huge_pages
= tmp
;
3524 spin_unlock(&hugetlb_lock
);
3530 #endif /* CONFIG_SYSCTL */
3532 void hugetlb_report_meminfo(struct seq_file
*m
)
3535 unsigned long total
= 0;
3537 if (!hugepages_supported())
3540 for_each_hstate(h
) {
3541 unsigned long count
= h
->nr_huge_pages
;
3543 total
+= (PAGE_SIZE
<< huge_page_order(h
)) * count
;
3545 if (h
== &default_hstate
)
3547 "HugePages_Total: %5lu\n"
3548 "HugePages_Free: %5lu\n"
3549 "HugePages_Rsvd: %5lu\n"
3550 "HugePages_Surp: %5lu\n"
3551 "Hugepagesize: %8lu kB\n",
3555 h
->surplus_huge_pages
,
3556 (PAGE_SIZE
<< huge_page_order(h
)) / 1024);
3559 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ 1024);
3562 int hugetlb_report_node_meminfo(char *buf
, int len
, int nid
)
3564 struct hstate
*h
= &default_hstate
;
3566 if (!hugepages_supported())
3569 return sysfs_emit_at(buf
, len
,
3570 "Node %d HugePages_Total: %5u\n"
3571 "Node %d HugePages_Free: %5u\n"
3572 "Node %d HugePages_Surp: %5u\n",
3573 nid
, h
->nr_huge_pages_node
[nid
],
3574 nid
, h
->free_huge_pages_node
[nid
],
3575 nid
, h
->surplus_huge_pages_node
[nid
]);
3578 void hugetlb_show_meminfo(void)
3583 if (!hugepages_supported())
3586 for_each_node_state(nid
, N_MEMORY
)
3588 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3590 h
->nr_huge_pages_node
[nid
],
3591 h
->free_huge_pages_node
[nid
],
3592 h
->surplus_huge_pages_node
[nid
],
3593 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3596 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3598 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3599 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3602 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3603 unsigned long hugetlb_total_pages(void)
3606 unsigned long nr_total_pages
= 0;
3609 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3610 return nr_total_pages
;
3613 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3617 spin_lock(&hugetlb_lock
);
3619 * When cpuset is configured, it breaks the strict hugetlb page
3620 * reservation as the accounting is done on a global variable. Such
3621 * reservation is completely rubbish in the presence of cpuset because
3622 * the reservation is not checked against page availability for the
3623 * current cpuset. Application can still potentially OOM'ed by kernel
3624 * with lack of free htlb page in cpuset that the task is in.
3625 * Attempt to enforce strict accounting with cpuset is almost
3626 * impossible (or too ugly) because cpuset is too fluid that
3627 * task or memory node can be dynamically moved between cpusets.
3629 * The change of semantics for shared hugetlb mapping with cpuset is
3630 * undesirable. However, in order to preserve some of the semantics,
3631 * we fall back to check against current free page availability as
3632 * a best attempt and hopefully to minimize the impact of changing
3633 * semantics that cpuset has.
3635 * Apart from cpuset, we also have memory policy mechanism that
3636 * also determines from which node the kernel will allocate memory
3637 * in a NUMA system. So similar to cpuset, we also should consider
3638 * the memory policy of the current task. Similar to the description
3642 if (gather_surplus_pages(h
, delta
) < 0)
3645 if (delta
> allowed_mems_nr(h
)) {
3646 return_unused_surplus_pages(h
, delta
);
3653 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3656 spin_unlock(&hugetlb_lock
);
3660 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3662 struct resv_map
*resv
= vma_resv_map(vma
);
3665 * This new VMA should share its siblings reservation map if present.
3666 * The VMA will only ever have a valid reservation map pointer where
3667 * it is being copied for another still existing VMA. As that VMA
3668 * has a reference to the reservation map it cannot disappear until
3669 * after this open call completes. It is therefore safe to take a
3670 * new reference here without additional locking.
3672 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3673 kref_get(&resv
->refs
);
3676 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3678 struct hstate
*h
= hstate_vma(vma
);
3679 struct resv_map
*resv
= vma_resv_map(vma
);
3680 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3681 unsigned long reserve
, start
, end
;
3684 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3687 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3688 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3690 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3691 hugetlb_cgroup_uncharge_counter(resv
, start
, end
);
3694 * Decrement reserve counts. The global reserve count may be
3695 * adjusted if the subpool has a minimum size.
3697 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3698 hugetlb_acct_memory(h
, -gbl_reserve
);
3701 kref_put(&resv
->refs
, resv_map_release
);
3704 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
3706 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
3711 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
3713 struct hstate
*hstate
= hstate_vma(vma
);
3715 return 1UL << huge_page_shift(hstate
);
3719 * We cannot handle pagefaults against hugetlb pages at all. They cause
3720 * handle_mm_fault() to try to instantiate regular-sized pages in the
3721 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3724 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3731 * When a new function is introduced to vm_operations_struct and added
3732 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3733 * This is because under System V memory model, mappings created via
3734 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3735 * their original vm_ops are overwritten with shm_vm_ops.
3737 const struct vm_operations_struct hugetlb_vm_ops
= {
3738 .fault
= hugetlb_vm_op_fault
,
3739 .open
= hugetlb_vm_op_open
,
3740 .close
= hugetlb_vm_op_close
,
3741 .may_split
= hugetlb_vm_op_split
,
3742 .pagesize
= hugetlb_vm_op_pagesize
,
3745 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3751 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3752 vma
->vm_page_prot
)));
3754 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3755 vma
->vm_page_prot
));
3757 entry
= pte_mkyoung(entry
);
3758 entry
= pte_mkhuge(entry
);
3759 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3764 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3765 unsigned long address
, pte_t
*ptep
)
3769 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3770 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3771 update_mmu_cache(vma
, address
, ptep
);
3774 bool is_hugetlb_entry_migration(pte_t pte
)
3778 if (huge_pte_none(pte
) || pte_present(pte
))
3780 swp
= pte_to_swp_entry(pte
);
3781 if (is_migration_entry(swp
))
3787 static bool is_hugetlb_entry_hwpoisoned(pte_t pte
)
3791 if (huge_pte_none(pte
) || pte_present(pte
))
3793 swp
= pte_to_swp_entry(pte
);
3794 if (is_hwpoison_entry(swp
))
3800 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3801 struct vm_area_struct
*vma
)
3803 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
3804 struct page
*ptepage
;
3807 struct hstate
*h
= hstate_vma(vma
);
3808 unsigned long sz
= huge_page_size(h
);
3809 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3810 struct mmu_notifier_range range
;
3813 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3816 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
3819 mmu_notifier_invalidate_range_start(&range
);
3822 * For shared mappings i_mmap_rwsem must be held to call
3823 * huge_pte_alloc, otherwise the returned ptep could go
3824 * away if part of a shared pmd and another thread calls
3827 i_mmap_lock_read(mapping
);
3830 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3831 spinlock_t
*src_ptl
, *dst_ptl
;
3832 src_pte
= huge_pte_offset(src
, addr
, sz
);
3835 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3842 * If the pagetables are shared don't copy or take references.
3843 * dst_pte == src_pte is the common case of src/dest sharing.
3845 * However, src could have 'unshared' and dst shares with
3846 * another vma. If dst_pte !none, this implies sharing.
3847 * Check here before taking page table lock, and once again
3848 * after taking the lock below.
3850 dst_entry
= huge_ptep_get(dst_pte
);
3851 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
3854 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3855 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3856 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3857 entry
= huge_ptep_get(src_pte
);
3858 dst_entry
= huge_ptep_get(dst_pte
);
3859 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
3861 * Skip if src entry none. Also, skip in the
3862 * unlikely case dst entry !none as this implies
3863 * sharing with another vma.
3866 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3867 is_hugetlb_entry_hwpoisoned(entry
))) {
3868 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3870 if (is_write_migration_entry(swp_entry
) && cow
) {
3872 * COW mappings require pages in both
3873 * parent and child to be set to read.
3875 make_migration_entry_read(&swp_entry
);
3876 entry
= swp_entry_to_pte(swp_entry
);
3877 set_huge_swap_pte_at(src
, addr
, src_pte
,
3880 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3884 * No need to notify as we are downgrading page
3885 * table protection not changing it to point
3888 * See Documentation/vm/mmu_notifier.rst
3890 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3892 entry
= huge_ptep_get(src_pte
);
3893 ptepage
= pte_page(entry
);
3895 page_dup_rmap(ptepage
, true);
3896 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3897 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3899 spin_unlock(src_ptl
);
3900 spin_unlock(dst_ptl
);
3904 mmu_notifier_invalidate_range_end(&range
);
3906 i_mmap_unlock_read(mapping
);
3911 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3912 unsigned long start
, unsigned long end
,
3913 struct page
*ref_page
)
3915 struct mm_struct
*mm
= vma
->vm_mm
;
3916 unsigned long address
;
3921 struct hstate
*h
= hstate_vma(vma
);
3922 unsigned long sz
= huge_page_size(h
);
3923 struct mmu_notifier_range range
;
3925 WARN_ON(!is_vm_hugetlb_page(vma
));
3926 BUG_ON(start
& ~huge_page_mask(h
));
3927 BUG_ON(end
& ~huge_page_mask(h
));
3930 * This is a hugetlb vma, all the pte entries should point
3933 tlb_change_page_size(tlb
, sz
);
3934 tlb_start_vma(tlb
, vma
);
3937 * If sharing possible, alert mmu notifiers of worst case.
3939 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
3941 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
3942 mmu_notifier_invalidate_range_start(&range
);
3944 for (; address
< end
; address
+= sz
) {
3945 ptep
= huge_pte_offset(mm
, address
, sz
);
3949 ptl
= huge_pte_lock(h
, mm
, ptep
);
3950 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
3953 * We just unmapped a page of PMDs by clearing a PUD.
3954 * The caller's TLB flush range should cover this area.
3959 pte
= huge_ptep_get(ptep
);
3960 if (huge_pte_none(pte
)) {
3966 * Migrating hugepage or HWPoisoned hugepage is already
3967 * unmapped and its refcount is dropped, so just clear pte here.
3969 if (unlikely(!pte_present(pte
))) {
3970 huge_pte_clear(mm
, address
, ptep
, sz
);
3975 page
= pte_page(pte
);
3977 * If a reference page is supplied, it is because a specific
3978 * page is being unmapped, not a range. Ensure the page we
3979 * are about to unmap is the actual page of interest.
3982 if (page
!= ref_page
) {
3987 * Mark the VMA as having unmapped its page so that
3988 * future faults in this VMA will fail rather than
3989 * looking like data was lost
3991 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3994 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3995 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3996 if (huge_pte_dirty(pte
))
3997 set_page_dirty(page
);
3999 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
4000 page_remove_rmap(page
, true);
4003 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
4005 * Bail out after unmapping reference page if supplied
4010 mmu_notifier_invalidate_range_end(&range
);
4011 tlb_end_vma(tlb
, vma
);
4014 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
4015 struct vm_area_struct
*vma
, unsigned long start
,
4016 unsigned long end
, struct page
*ref_page
)
4018 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
4021 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4022 * test will fail on a vma being torn down, and not grab a page table
4023 * on its way out. We're lucky that the flag has such an appropriate
4024 * name, and can in fact be safely cleared here. We could clear it
4025 * before the __unmap_hugepage_range above, but all that's necessary
4026 * is to clear it before releasing the i_mmap_rwsem. This works
4027 * because in the context this is called, the VMA is about to be
4028 * destroyed and the i_mmap_rwsem is held.
4030 vma
->vm_flags
&= ~VM_MAYSHARE
;
4033 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
4034 unsigned long end
, struct page
*ref_page
)
4036 struct mm_struct
*mm
;
4037 struct mmu_gather tlb
;
4038 unsigned long tlb_start
= start
;
4039 unsigned long tlb_end
= end
;
4042 * If shared PMDs were possibly used within this vma range, adjust
4043 * start/end for worst case tlb flushing.
4044 * Note that we can not be sure if PMDs are shared until we try to
4045 * unmap pages. However, we want to make sure TLB flushing covers
4046 * the largest possible range.
4048 adjust_range_if_pmd_sharing_possible(vma
, &tlb_start
, &tlb_end
);
4052 tlb_gather_mmu(&tlb
, mm
, tlb_start
, tlb_end
);
4053 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
4054 tlb_finish_mmu(&tlb
, tlb_start
, tlb_end
);
4058 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4059 * mappping it owns the reserve page for. The intention is to unmap the page
4060 * from other VMAs and let the children be SIGKILLed if they are faulting the
4063 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4064 struct page
*page
, unsigned long address
)
4066 struct hstate
*h
= hstate_vma(vma
);
4067 struct vm_area_struct
*iter_vma
;
4068 struct address_space
*mapping
;
4072 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4073 * from page cache lookup which is in HPAGE_SIZE units.
4075 address
= address
& huge_page_mask(h
);
4076 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
4078 mapping
= vma
->vm_file
->f_mapping
;
4081 * Take the mapping lock for the duration of the table walk. As
4082 * this mapping should be shared between all the VMAs,
4083 * __unmap_hugepage_range() is called as the lock is already held
4085 i_mmap_lock_write(mapping
);
4086 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
4087 /* Do not unmap the current VMA */
4088 if (iter_vma
== vma
)
4092 * Shared VMAs have their own reserves and do not affect
4093 * MAP_PRIVATE accounting but it is possible that a shared
4094 * VMA is using the same page so check and skip such VMAs.
4096 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
4100 * Unmap the page from other VMAs without their own reserves.
4101 * They get marked to be SIGKILLed if they fault in these
4102 * areas. This is because a future no-page fault on this VMA
4103 * could insert a zeroed page instead of the data existing
4104 * from the time of fork. This would look like data corruption
4106 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
4107 unmap_hugepage_range(iter_vma
, address
,
4108 address
+ huge_page_size(h
), page
);
4110 i_mmap_unlock_write(mapping
);
4114 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4115 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4116 * cannot race with other handlers or page migration.
4117 * Keep the pte_same checks anyway to make transition from the mutex easier.
4119 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4120 unsigned long address
, pte_t
*ptep
,
4121 struct page
*pagecache_page
, spinlock_t
*ptl
)
4124 struct hstate
*h
= hstate_vma(vma
);
4125 struct page
*old_page
, *new_page
;
4126 int outside_reserve
= 0;
4128 unsigned long haddr
= address
& huge_page_mask(h
);
4129 struct mmu_notifier_range range
;
4131 pte
= huge_ptep_get(ptep
);
4132 old_page
= pte_page(pte
);
4135 /* If no-one else is actually using this page, avoid the copy
4136 * and just make the page writable */
4137 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
4138 page_move_anon_rmap(old_page
, vma
);
4139 set_huge_ptep_writable(vma
, haddr
, ptep
);
4144 * If the process that created a MAP_PRIVATE mapping is about to
4145 * perform a COW due to a shared page count, attempt to satisfy
4146 * the allocation without using the existing reserves. The pagecache
4147 * page is used to determine if the reserve at this address was
4148 * consumed or not. If reserves were used, a partial faulted mapping
4149 * at the time of fork() could consume its reserves on COW instead
4150 * of the full address range.
4152 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
4153 old_page
!= pagecache_page
)
4154 outside_reserve
= 1;
4159 * Drop page table lock as buddy allocator may be called. It will
4160 * be acquired again before returning to the caller, as expected.
4163 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
4165 if (IS_ERR(new_page
)) {
4167 * If a process owning a MAP_PRIVATE mapping fails to COW,
4168 * it is due to references held by a child and an insufficient
4169 * huge page pool. To guarantee the original mappers
4170 * reliability, unmap the page from child processes. The child
4171 * may get SIGKILLed if it later faults.
4173 if (outside_reserve
) {
4174 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4179 BUG_ON(huge_pte_none(pte
));
4181 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4182 * unmapping. unmapping needs to hold i_mmap_rwsem
4183 * in write mode. Dropping i_mmap_rwsem in read mode
4184 * here is OK as COW mappings do not interact with
4187 * Reacquire both after unmap operation.
4189 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4190 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4191 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4192 i_mmap_unlock_read(mapping
);
4194 unmap_ref_private(mm
, vma
, old_page
, haddr
);
4196 i_mmap_lock_read(mapping
);
4197 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4199 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4201 pte_same(huge_ptep_get(ptep
), pte
)))
4202 goto retry_avoidcopy
;
4204 * race occurs while re-acquiring page table
4205 * lock, and our job is done.
4210 ret
= vmf_error(PTR_ERR(new_page
));
4211 goto out_release_old
;
4215 * When the original hugepage is shared one, it does not have
4216 * anon_vma prepared.
4218 if (unlikely(anon_vma_prepare(vma
))) {
4220 goto out_release_all
;
4223 copy_user_huge_page(new_page
, old_page
, address
, vma
,
4224 pages_per_huge_page(h
));
4225 __SetPageUptodate(new_page
);
4227 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
4228 haddr
+ huge_page_size(h
));
4229 mmu_notifier_invalidate_range_start(&range
);
4232 * Retake the page table lock to check for racing updates
4233 * before the page tables are altered
4236 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4237 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
4238 ClearPagePrivate(new_page
);
4241 huge_ptep_clear_flush(vma
, haddr
, ptep
);
4242 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
4243 set_huge_pte_at(mm
, haddr
, ptep
,
4244 make_huge_pte(vma
, new_page
, 1));
4245 page_remove_rmap(old_page
, true);
4246 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
4247 set_page_huge_active(new_page
);
4248 /* Make the old page be freed below */
4249 new_page
= old_page
;
4252 mmu_notifier_invalidate_range_end(&range
);
4254 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
4259 spin_lock(ptl
); /* Caller expects lock to be held */
4263 /* Return the pagecache page at a given address within a VMA */
4264 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
4265 struct vm_area_struct
*vma
, unsigned long address
)
4267 struct address_space
*mapping
;
4270 mapping
= vma
->vm_file
->f_mapping
;
4271 idx
= vma_hugecache_offset(h
, vma
, address
);
4273 return find_lock_page(mapping
, idx
);
4277 * Return whether there is a pagecache page to back given address within VMA.
4278 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4280 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
4281 struct vm_area_struct
*vma
, unsigned long address
)
4283 struct address_space
*mapping
;
4287 mapping
= vma
->vm_file
->f_mapping
;
4288 idx
= vma_hugecache_offset(h
, vma
, address
);
4290 page
= find_get_page(mapping
, idx
);
4293 return page
!= NULL
;
4296 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
4299 struct inode
*inode
= mapping
->host
;
4300 struct hstate
*h
= hstate_inode(inode
);
4301 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
4305 ClearPagePrivate(page
);
4308 * set page dirty so that it will not be removed from cache/file
4309 * by non-hugetlbfs specific code paths.
4311 set_page_dirty(page
);
4313 spin_lock(&inode
->i_lock
);
4314 inode
->i_blocks
+= blocks_per_huge_page(h
);
4315 spin_unlock(&inode
->i_lock
);
4319 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
4320 struct vm_area_struct
*vma
,
4321 struct address_space
*mapping
, pgoff_t idx
,
4322 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
4324 struct hstate
*h
= hstate_vma(vma
);
4325 vm_fault_t ret
= VM_FAULT_SIGBUS
;
4331 unsigned long haddr
= address
& huge_page_mask(h
);
4332 bool new_page
= false;
4335 * Currently, we are forced to kill the process in the event the
4336 * original mapper has unmapped pages from the child due to a failed
4337 * COW. Warn that such a situation has occurred as it may not be obvious
4339 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
4340 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4346 * We can not race with truncation due to holding i_mmap_rwsem.
4347 * i_size is modified when holding i_mmap_rwsem, so check here
4348 * once for faults beyond end of file.
4350 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4355 page
= find_lock_page(mapping
, idx
);
4358 * Check for page in userfault range
4360 if (userfaultfd_missing(vma
)) {
4362 struct vm_fault vmf
= {
4367 * Hard to debug if it ends up being
4368 * used by a callee that assumes
4369 * something about the other
4370 * uninitialized fields... same as in
4376 * hugetlb_fault_mutex and i_mmap_rwsem must be
4377 * dropped before handling userfault. Reacquire
4378 * after handling fault to make calling code simpler.
4380 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4381 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4382 i_mmap_unlock_read(mapping
);
4383 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
4384 i_mmap_lock_read(mapping
);
4385 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4389 page
= alloc_huge_page(vma
, haddr
, 0);
4392 * Returning error will result in faulting task being
4393 * sent SIGBUS. The hugetlb fault mutex prevents two
4394 * tasks from racing to fault in the same page which
4395 * could result in false unable to allocate errors.
4396 * Page migration does not take the fault mutex, but
4397 * does a clear then write of pte's under page table
4398 * lock. Page fault code could race with migration,
4399 * notice the clear pte and try to allocate a page
4400 * here. Before returning error, get ptl and make
4401 * sure there really is no pte entry.
4403 ptl
= huge_pte_lock(h
, mm
, ptep
);
4404 if (!huge_pte_none(huge_ptep_get(ptep
))) {
4410 ret
= vmf_error(PTR_ERR(page
));
4413 clear_huge_page(page
, address
, pages_per_huge_page(h
));
4414 __SetPageUptodate(page
);
4417 if (vma
->vm_flags
& VM_MAYSHARE
) {
4418 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
4427 if (unlikely(anon_vma_prepare(vma
))) {
4429 goto backout_unlocked
;
4435 * If memory error occurs between mmap() and fault, some process
4436 * don't have hwpoisoned swap entry for errored virtual address.
4437 * So we need to block hugepage fault by PG_hwpoison bit check.
4439 if (unlikely(PageHWPoison(page
))) {
4440 ret
= VM_FAULT_HWPOISON_LARGE
|
4441 VM_FAULT_SET_HINDEX(hstate_index(h
));
4442 goto backout_unlocked
;
4447 * If we are going to COW a private mapping later, we examine the
4448 * pending reservations for this page now. This will ensure that
4449 * any allocations necessary to record that reservation occur outside
4452 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4453 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4455 goto backout_unlocked
;
4457 /* Just decrements count, does not deallocate */
4458 vma_end_reservation(h
, vma
, haddr
);
4461 ptl
= huge_pte_lock(h
, mm
, ptep
);
4463 if (!huge_pte_none(huge_ptep_get(ptep
)))
4467 ClearPagePrivate(page
);
4468 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4470 page_dup_rmap(page
, true);
4471 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
4472 && (vma
->vm_flags
& VM_SHARED
)));
4473 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
4475 hugetlb_count_add(pages_per_huge_page(h
), mm
);
4476 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4477 /* Optimization, do the COW without a second fault */
4478 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
4484 * Only make newly allocated pages active. Existing pages found
4485 * in the pagecache could be !page_huge_active() if they have been
4486 * isolated for migration.
4489 set_page_huge_active(page
);
4499 restore_reserve_on_error(h
, vma
, haddr
, page
);
4505 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4507 unsigned long key
[2];
4510 key
[0] = (unsigned long) mapping
;
4513 hash
= jhash2((u32
*)&key
, sizeof(key
)/(sizeof(u32
)), 0);
4515 return hash
& (num_fault_mutexes
- 1);
4519 * For uniprocesor systems we always use a single mutex, so just
4520 * return 0 and avoid the hashing overhead.
4522 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
4528 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4529 unsigned long address
, unsigned int flags
)
4536 struct page
*page
= NULL
;
4537 struct page
*pagecache_page
= NULL
;
4538 struct hstate
*h
= hstate_vma(vma
);
4539 struct address_space
*mapping
;
4540 int need_wait_lock
= 0;
4541 unsigned long haddr
= address
& huge_page_mask(h
);
4543 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4546 * Since we hold no locks, ptep could be stale. That is
4547 * OK as we are only making decisions based on content and
4548 * not actually modifying content here.
4550 entry
= huge_ptep_get(ptep
);
4551 if (unlikely(is_hugetlb_entry_migration(entry
))) {
4552 migration_entry_wait_huge(vma
, mm
, ptep
);
4554 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
4555 return VM_FAULT_HWPOISON_LARGE
|
4556 VM_FAULT_SET_HINDEX(hstate_index(h
));
4560 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4561 * until finished with ptep. This serves two purposes:
4562 * 1) It prevents huge_pmd_unshare from being called elsewhere
4563 * and making the ptep no longer valid.
4564 * 2) It synchronizes us with i_size modifications during truncation.
4566 * ptep could have already be assigned via huge_pte_offset. That
4567 * is OK, as huge_pte_alloc will return the same value unless
4568 * something has changed.
4570 mapping
= vma
->vm_file
->f_mapping
;
4571 i_mmap_lock_read(mapping
);
4572 ptep
= huge_pte_alloc(mm
, haddr
, huge_page_size(h
));
4574 i_mmap_unlock_read(mapping
);
4575 return VM_FAULT_OOM
;
4579 * Serialize hugepage allocation and instantiation, so that we don't
4580 * get spurious allocation failures if two CPUs race to instantiate
4581 * the same page in the page cache.
4583 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4584 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4585 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4587 entry
= huge_ptep_get(ptep
);
4588 if (huge_pte_none(entry
)) {
4589 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
4596 * entry could be a migration/hwpoison entry at this point, so this
4597 * check prevents the kernel from going below assuming that we have
4598 * an active hugepage in pagecache. This goto expects the 2nd page
4599 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4600 * properly handle it.
4602 if (!pte_present(entry
))
4606 * If we are going to COW the mapping later, we examine the pending
4607 * reservations for this page now. This will ensure that any
4608 * allocations necessary to record that reservation occur outside the
4609 * spinlock. For private mappings, we also lookup the pagecache
4610 * page now as it is used to determine if a reservation has been
4613 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
4614 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4618 /* Just decrements count, does not deallocate */
4619 vma_end_reservation(h
, vma
, haddr
);
4621 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4622 pagecache_page
= hugetlbfs_pagecache_page(h
,
4626 ptl
= huge_pte_lock(h
, mm
, ptep
);
4628 /* Check for a racing update before calling hugetlb_cow */
4629 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4633 * hugetlb_cow() requires page locks of pte_page(entry) and
4634 * pagecache_page, so here we need take the former one
4635 * when page != pagecache_page or !pagecache_page.
4637 page
= pte_page(entry
);
4638 if (page
!= pagecache_page
)
4639 if (!trylock_page(page
)) {
4646 if (flags
& FAULT_FLAG_WRITE
) {
4647 if (!huge_pte_write(entry
)) {
4648 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4649 pagecache_page
, ptl
);
4652 entry
= huge_pte_mkdirty(entry
);
4654 entry
= pte_mkyoung(entry
);
4655 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
4656 flags
& FAULT_FLAG_WRITE
))
4657 update_mmu_cache(vma
, haddr
, ptep
);
4659 if (page
!= pagecache_page
)
4665 if (pagecache_page
) {
4666 unlock_page(pagecache_page
);
4667 put_page(pagecache_page
);
4670 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4671 i_mmap_unlock_read(mapping
);
4673 * Generally it's safe to hold refcount during waiting page lock. But
4674 * here we just wait to defer the next page fault to avoid busy loop and
4675 * the page is not used after unlocked before returning from the current
4676 * page fault. So we are safe from accessing freed page, even if we wait
4677 * here without taking refcount.
4680 wait_on_page_locked(page
);
4685 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4686 * modifications for huge pages.
4688 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4690 struct vm_area_struct
*dst_vma
,
4691 unsigned long dst_addr
,
4692 unsigned long src_addr
,
4693 struct page
**pagep
)
4695 struct address_space
*mapping
;
4698 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4699 struct hstate
*h
= hstate_vma(dst_vma
);
4707 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4711 ret
= copy_huge_page_from_user(page
,
4712 (const void __user
*) src_addr
,
4713 pages_per_huge_page(h
), false);
4715 /* fallback to copy_from_user outside mmap_lock */
4716 if (unlikely(ret
)) {
4719 /* don't free the page */
4728 * The memory barrier inside __SetPageUptodate makes sure that
4729 * preceding stores to the page contents become visible before
4730 * the set_pte_at() write.
4732 __SetPageUptodate(page
);
4734 mapping
= dst_vma
->vm_file
->f_mapping
;
4735 idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4738 * If shared, add to page cache
4741 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4744 goto out_release_nounlock
;
4747 * Serialization between remove_inode_hugepages() and
4748 * huge_add_to_page_cache() below happens through the
4749 * hugetlb_fault_mutex_table that here must be hold by
4752 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4754 goto out_release_nounlock
;
4757 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4761 * Recheck the i_size after holding PT lock to make sure not
4762 * to leave any page mapped (as page_mapped()) beyond the end
4763 * of the i_size (remove_inode_hugepages() is strict about
4764 * enforcing that). If we bail out here, we'll also leave a
4765 * page in the radix tree in the vm_shared case beyond the end
4766 * of the i_size, but remove_inode_hugepages() will take care
4767 * of it as soon as we drop the hugetlb_fault_mutex_table.
4769 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4772 goto out_release_unlock
;
4775 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4776 goto out_release_unlock
;
4779 page_dup_rmap(page
, true);
4781 ClearPagePrivate(page
);
4782 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4785 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4786 if (dst_vma
->vm_flags
& VM_WRITE
)
4787 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4788 _dst_pte
= pte_mkyoung(_dst_pte
);
4790 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4792 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4793 dst_vma
->vm_flags
& VM_WRITE
);
4794 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4796 /* No need to invalidate - it was non-present before */
4797 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4800 set_page_huge_active(page
);
4810 out_release_nounlock
:
4815 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4816 struct page
**pages
, struct vm_area_struct
**vmas
,
4817 unsigned long *position
, unsigned long *nr_pages
,
4818 long i
, unsigned int flags
, int *locked
)
4820 unsigned long pfn_offset
;
4821 unsigned long vaddr
= *position
;
4822 unsigned long remainder
= *nr_pages
;
4823 struct hstate
*h
= hstate_vma(vma
);
4826 while (vaddr
< vma
->vm_end
&& remainder
) {
4828 spinlock_t
*ptl
= NULL
;
4833 * If we have a pending SIGKILL, don't keep faulting pages and
4834 * potentially allocating memory.
4836 if (fatal_signal_pending(current
)) {
4842 * Some archs (sparc64, sh*) have multiple pte_ts to
4843 * each hugepage. We have to make sure we get the
4844 * first, for the page indexing below to work.
4846 * Note that page table lock is not held when pte is null.
4848 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4851 ptl
= huge_pte_lock(h
, mm
, pte
);
4852 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4855 * When coredumping, it suits get_dump_page if we just return
4856 * an error where there's an empty slot with no huge pagecache
4857 * to back it. This way, we avoid allocating a hugepage, and
4858 * the sparse dumpfile avoids allocating disk blocks, but its
4859 * huge holes still show up with zeroes where they need to be.
4861 if (absent
&& (flags
& FOLL_DUMP
) &&
4862 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4870 * We need call hugetlb_fault for both hugepages under migration
4871 * (in which case hugetlb_fault waits for the migration,) and
4872 * hwpoisoned hugepages (in which case we need to prevent the
4873 * caller from accessing to them.) In order to do this, we use
4874 * here is_swap_pte instead of is_hugetlb_entry_migration and
4875 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4876 * both cases, and because we can't follow correct pages
4877 * directly from any kind of swap entries.
4879 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4880 ((flags
& FOLL_WRITE
) &&
4881 !huge_pte_write(huge_ptep_get(pte
)))) {
4883 unsigned int fault_flags
= 0;
4887 if (flags
& FOLL_WRITE
)
4888 fault_flags
|= FAULT_FLAG_WRITE
;
4890 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4891 FAULT_FLAG_KILLABLE
;
4892 if (flags
& FOLL_NOWAIT
)
4893 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4894 FAULT_FLAG_RETRY_NOWAIT
;
4895 if (flags
& FOLL_TRIED
) {
4897 * Note: FAULT_FLAG_ALLOW_RETRY and
4898 * FAULT_FLAG_TRIED can co-exist
4900 fault_flags
|= FAULT_FLAG_TRIED
;
4902 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4903 if (ret
& VM_FAULT_ERROR
) {
4904 err
= vm_fault_to_errno(ret
, flags
);
4908 if (ret
& VM_FAULT_RETRY
) {
4910 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
4914 * VM_FAULT_RETRY must not return an
4915 * error, it will return zero
4918 * No need to update "position" as the
4919 * caller will not check it after
4920 * *nr_pages is set to 0.
4927 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4928 page
= pte_page(huge_ptep_get(pte
));
4931 * If subpage information not requested, update counters
4932 * and skip the same_page loop below.
4934 if (!pages
&& !vmas
&& !pfn_offset
&&
4935 (vaddr
+ huge_page_size(h
) < vma
->vm_end
) &&
4936 (remainder
>= pages_per_huge_page(h
))) {
4937 vaddr
+= huge_page_size(h
);
4938 remainder
-= pages_per_huge_page(h
);
4939 i
+= pages_per_huge_page(h
);
4946 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4948 * try_grab_page() should always succeed here, because:
4949 * a) we hold the ptl lock, and b) we've just checked
4950 * that the huge page is present in the page tables. If
4951 * the huge page is present, then the tail pages must
4952 * also be present. The ptl prevents the head page and
4953 * tail pages from being rearranged in any way. So this
4954 * page must be available at this point, unless the page
4955 * refcount overflowed:
4957 if (WARN_ON_ONCE(!try_grab_page(pages
[i
], flags
))) {
4972 if (vaddr
< vma
->vm_end
&& remainder
&&
4973 pfn_offset
< pages_per_huge_page(h
)) {
4975 * We use pfn_offset to avoid touching the pageframes
4976 * of this compound page.
4982 *nr_pages
= remainder
;
4984 * setting position is actually required only if remainder is
4985 * not zero but it's faster not to add a "if (remainder)"
4993 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4995 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4998 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5001 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
5002 unsigned long address
, unsigned long end
, pgprot_t newprot
)
5004 struct mm_struct
*mm
= vma
->vm_mm
;
5005 unsigned long start
= address
;
5008 struct hstate
*h
= hstate_vma(vma
);
5009 unsigned long pages
= 0;
5010 bool shared_pmd
= false;
5011 struct mmu_notifier_range range
;
5014 * In the case of shared PMDs, the area to flush could be beyond
5015 * start/end. Set range.start/range.end to cover the maximum possible
5016 * range if PMD sharing is possible.
5018 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
5019 0, vma
, mm
, start
, end
);
5020 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
5022 BUG_ON(address
>= end
);
5023 flush_cache_range(vma
, range
.start
, range
.end
);
5025 mmu_notifier_invalidate_range_start(&range
);
5026 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
5027 for (; address
< end
; address
+= huge_page_size(h
)) {
5029 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
5032 ptl
= huge_pte_lock(h
, mm
, ptep
);
5033 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
5039 pte
= huge_ptep_get(ptep
);
5040 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
5044 if (unlikely(is_hugetlb_entry_migration(pte
))) {
5045 swp_entry_t entry
= pte_to_swp_entry(pte
);
5047 if (is_write_migration_entry(entry
)) {
5050 make_migration_entry_read(&entry
);
5051 newpte
= swp_entry_to_pte(entry
);
5052 set_huge_swap_pte_at(mm
, address
, ptep
,
5053 newpte
, huge_page_size(h
));
5059 if (!huge_pte_none(pte
)) {
5062 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
5063 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
5064 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
5065 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
5071 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5072 * may have cleared our pud entry and done put_page on the page table:
5073 * once we release i_mmap_rwsem, another task can do the final put_page
5074 * and that page table be reused and filled with junk. If we actually
5075 * did unshare a page of pmds, flush the range corresponding to the pud.
5078 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
5080 flush_hugetlb_tlb_range(vma
, start
, end
);
5082 * No need to call mmu_notifier_invalidate_range() we are downgrading
5083 * page table protection not changing it to point to a new page.
5085 * See Documentation/vm/mmu_notifier.rst
5087 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
5088 mmu_notifier_invalidate_range_end(&range
);
5090 return pages
<< h
->order
;
5093 int hugetlb_reserve_pages(struct inode
*inode
,
5095 struct vm_area_struct
*vma
,
5096 vm_flags_t vm_flags
)
5098 long ret
, chg
, add
= -1;
5099 struct hstate
*h
= hstate_inode(inode
);
5100 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5101 struct resv_map
*resv_map
;
5102 struct hugetlb_cgroup
*h_cg
= NULL
;
5103 long gbl_reserve
, regions_needed
= 0;
5105 /* This should never happen */
5107 VM_WARN(1, "%s called with a negative range\n", __func__
);
5112 * Only apply hugepage reservation if asked. At fault time, an
5113 * attempt will be made for VM_NORESERVE to allocate a page
5114 * without using reserves
5116 if (vm_flags
& VM_NORESERVE
)
5120 * Shared mappings base their reservation on the number of pages that
5121 * are already allocated on behalf of the file. Private mappings need
5122 * to reserve the full area even if read-only as mprotect() may be
5123 * called to make the mapping read-write. Assume !vma is a shm mapping
5125 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5127 * resv_map can not be NULL as hugetlb_reserve_pages is only
5128 * called for inodes for which resv_maps were created (see
5129 * hugetlbfs_get_inode).
5131 resv_map
= inode_resv_map(inode
);
5133 chg
= region_chg(resv_map
, from
, to
, ®ions_needed
);
5136 /* Private mapping. */
5137 resv_map
= resv_map_alloc();
5143 set_vma_resv_map(vma
, resv_map
);
5144 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
5152 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
5153 hstate_index(h
), chg
* pages_per_huge_page(h
), &h_cg
);
5160 if (vma
&& !(vma
->vm_flags
& VM_MAYSHARE
) && h_cg
) {
5161 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5164 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, h_cg
, h
);
5168 * There must be enough pages in the subpool for the mapping. If
5169 * the subpool has a minimum size, there may be some global
5170 * reservations already in place (gbl_reserve).
5172 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
5173 if (gbl_reserve
< 0) {
5175 goto out_uncharge_cgroup
;
5179 * Check enough hugepages are available for the reservation.
5180 * Hand the pages back to the subpool if there are not
5182 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
5188 * Account for the reservations made. Shared mappings record regions
5189 * that have reservations as they are shared by multiple VMAs.
5190 * When the last VMA disappears, the region map says how much
5191 * the reservation was and the page cache tells how much of
5192 * the reservation was consumed. Private mappings are per-VMA and
5193 * only the consumed reservations are tracked. When the VMA
5194 * disappears, the original reservation is the VMA size and the
5195 * consumed reservations are stored in the map. Hence, nothing
5196 * else has to be done for private mappings here
5198 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5199 add
= region_add(resv_map
, from
, to
, regions_needed
, h
, h_cg
);
5201 if (unlikely(add
< 0)) {
5202 hugetlb_acct_memory(h
, -gbl_reserve
);
5205 } else if (unlikely(chg
> add
)) {
5207 * pages in this range were added to the reserve
5208 * map between region_chg and region_add. This
5209 * indicates a race with alloc_huge_page. Adjust
5210 * the subpool and reserve counts modified above
5211 * based on the difference.
5216 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5217 * reference to h_cg->css. See comment below for detail.
5219 hugetlb_cgroup_uncharge_cgroup_rsvd(
5221 (chg
- add
) * pages_per_huge_page(h
), h_cg
);
5223 rsv_adjust
= hugepage_subpool_put_pages(spool
,
5225 hugetlb_acct_memory(h
, -rsv_adjust
);
5228 * The file_regions will hold their own reference to
5229 * h_cg->css. So we should release the reference held
5230 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5233 hugetlb_cgroup_put_rsvd_cgroup(h_cg
);
5238 /* put back original number of pages, chg */
5239 (void)hugepage_subpool_put_pages(spool
, chg
);
5240 out_uncharge_cgroup
:
5241 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h
),
5242 chg
* pages_per_huge_page(h
), h_cg
);
5244 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
5245 /* Only call region_abort if the region_chg succeeded but the
5246 * region_add failed or didn't run.
5248 if (chg
>= 0 && add
< 0)
5249 region_abort(resv_map
, from
, to
, regions_needed
);
5250 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
5251 kref_put(&resv_map
->refs
, resv_map_release
);
5255 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
5258 struct hstate
*h
= hstate_inode(inode
);
5259 struct resv_map
*resv_map
= inode_resv_map(inode
);
5261 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5265 * Since this routine can be called in the evict inode path for all
5266 * hugetlbfs inodes, resv_map could be NULL.
5269 chg
= region_del(resv_map
, start
, end
);
5271 * region_del() can fail in the rare case where a region
5272 * must be split and another region descriptor can not be
5273 * allocated. If end == LONG_MAX, it will not fail.
5279 spin_lock(&inode
->i_lock
);
5280 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
5281 spin_unlock(&inode
->i_lock
);
5284 * If the subpool has a minimum size, the number of global
5285 * reservations to be released may be adjusted.
5287 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
5288 hugetlb_acct_memory(h
, -gbl_reserve
);
5293 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5294 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
5295 struct vm_area_struct
*vma
,
5296 unsigned long addr
, pgoff_t idx
)
5298 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
5300 unsigned long sbase
= saddr
& PUD_MASK
;
5301 unsigned long s_end
= sbase
+ PUD_SIZE
;
5303 /* Allow segments to share if only one is marked locked */
5304 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5305 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5308 * match the virtual addresses, permission and the alignment of the
5311 if (pmd_index(addr
) != pmd_index(saddr
) ||
5312 vm_flags
!= svm_flags
||
5313 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
5319 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
5321 unsigned long base
= addr
& PUD_MASK
;
5322 unsigned long end
= base
+ PUD_SIZE
;
5325 * check on proper vm_flags and page table alignment
5327 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
5333 * Determine if start,end range within vma could be mapped by shared pmd.
5334 * If yes, adjust start and end to cover range associated with possible
5335 * shared pmd mappings.
5337 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5338 unsigned long *start
, unsigned long *end
)
5340 unsigned long v_start
= ALIGN(vma
->vm_start
, PUD_SIZE
),
5341 v_end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
5344 * vma need span at least one aligned PUD size and the start,end range
5345 * must at least partialy within it.
5347 if (!(vma
->vm_flags
& VM_MAYSHARE
) || !(v_end
> v_start
) ||
5348 (*end
<= v_start
) || (*start
>= v_end
))
5351 /* Extend the range to be PUD aligned for a worst case scenario */
5352 if (*start
> v_start
)
5353 *start
= ALIGN_DOWN(*start
, PUD_SIZE
);
5356 *end
= ALIGN(*end
, PUD_SIZE
);
5360 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5361 * and returns the corresponding pte. While this is not necessary for the
5362 * !shared pmd case because we can allocate the pmd later as well, it makes the
5363 * code much cleaner.
5365 * This routine must be called with i_mmap_rwsem held in at least read mode if
5366 * sharing is possible. For hugetlbfs, this prevents removal of any page
5367 * table entries associated with the address space. This is important as we
5368 * are setting up sharing based on existing page table entries (mappings).
5370 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5371 * huge_pte_alloc know that sharing is not possible and do not take
5372 * i_mmap_rwsem as a performance optimization. This is handled by the
5373 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5374 * only required for subsequent processing.
5376 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5378 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
5379 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
5380 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
5382 struct vm_area_struct
*svma
;
5383 unsigned long saddr
;
5388 if (!vma_shareable(vma
, addr
))
5389 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5391 i_mmap_assert_locked(mapping
);
5392 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
5396 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
5398 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
5399 vma_mmu_pagesize(svma
));
5401 get_page(virt_to_page(spte
));
5410 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
5411 if (pud_none(*pud
)) {
5412 pud_populate(mm
, pud
,
5413 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
5416 put_page(virt_to_page(spte
));
5420 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5425 * unmap huge page backed by shared pte.
5427 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5428 * indicated by page_count > 1, unmap is achieved by clearing pud and
5429 * decrementing the ref count. If count == 1, the pte page is not shared.
5431 * Called with page table lock held and i_mmap_rwsem held in write mode.
5433 * returns: 1 successfully unmapped a shared pte page
5434 * 0 the underlying pte page is not shared, or it is the last user
5436 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5437 unsigned long *addr
, pte_t
*ptep
)
5439 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
5440 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
5441 pud_t
*pud
= pud_offset(p4d
, *addr
);
5443 i_mmap_assert_write_locked(vma
->vm_file
->f_mapping
);
5444 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
5445 if (page_count(virt_to_page(ptep
)) == 1)
5449 put_page(virt_to_page(ptep
));
5451 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
5454 #define want_pmd_share() (1)
5455 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5456 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5461 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5462 unsigned long *addr
, pte_t
*ptep
)
5467 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5468 unsigned long *start
, unsigned long *end
)
5471 #define want_pmd_share() (0)
5472 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5474 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5475 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
5476 unsigned long addr
, unsigned long sz
)
5483 pgd
= pgd_offset(mm
, addr
);
5484 p4d
= p4d_alloc(mm
, pgd
, addr
);
5487 pud
= pud_alloc(mm
, p4d
, addr
);
5489 if (sz
== PUD_SIZE
) {
5492 BUG_ON(sz
!= PMD_SIZE
);
5493 if (want_pmd_share() && pud_none(*pud
))
5494 pte
= huge_pmd_share(mm
, addr
, pud
);
5496 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5499 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
5505 * huge_pte_offset() - Walk the page table to resolve the hugepage
5506 * entry at address @addr
5508 * Return: Pointer to page table entry (PUD or PMD) for
5509 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5510 * size @sz doesn't match the hugepage size at this level of the page
5513 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
5514 unsigned long addr
, unsigned long sz
)
5521 pgd
= pgd_offset(mm
, addr
);
5522 if (!pgd_present(*pgd
))
5524 p4d
= p4d_offset(pgd
, addr
);
5525 if (!p4d_present(*p4d
))
5528 pud
= pud_offset(p4d
, addr
);
5530 /* must be pud huge, non-present or none */
5531 return (pte_t
*)pud
;
5532 if (!pud_present(*pud
))
5534 /* must have a valid entry and size to go further */
5536 pmd
= pmd_offset(pud
, addr
);
5537 /* must be pmd huge, non-present or none */
5538 return (pte_t
*)pmd
;
5541 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5544 * These functions are overwritable if your architecture needs its own
5547 struct page
* __weak
5548 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
5551 return ERR_PTR(-EINVAL
);
5554 struct page
* __weak
5555 follow_huge_pd(struct vm_area_struct
*vma
,
5556 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
5558 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5562 struct page
* __weak
5563 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
5564 pmd_t
*pmd
, int flags
)
5566 struct page
*page
= NULL
;
5570 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5571 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
5572 (FOLL_PIN
| FOLL_GET
)))
5576 ptl
= pmd_lockptr(mm
, pmd
);
5579 * make sure that the address range covered by this pmd is not
5580 * unmapped from other threads.
5582 if (!pmd_huge(*pmd
))
5584 pte
= huge_ptep_get((pte_t
*)pmd
);
5585 if (pte_present(pte
)) {
5586 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
5588 * try_grab_page() should always succeed here, because: a) we
5589 * hold the pmd (ptl) lock, and b) we've just checked that the
5590 * huge pmd (head) page is present in the page tables. The ptl
5591 * prevents the head page and tail pages from being rearranged
5592 * in any way. So this page must be available at this point,
5593 * unless the page refcount overflowed:
5595 if (WARN_ON_ONCE(!try_grab_page(page
, flags
))) {
5600 if (is_hugetlb_entry_migration(pte
)) {
5602 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
5606 * hwpoisoned entry is treated as no_page_table in
5607 * follow_page_mask().
5615 struct page
* __weak
5616 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
5617 pud_t
*pud
, int flags
)
5619 if (flags
& (FOLL_GET
| FOLL_PIN
))
5622 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
5625 struct page
* __weak
5626 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
5628 if (flags
& (FOLL_GET
| FOLL_PIN
))
5631 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
5634 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
5638 spin_lock(&hugetlb_lock
);
5639 if (!PageHeadHuge(page
) || !page_huge_active(page
) ||
5640 !get_page_unless_zero(page
)) {
5644 clear_page_huge_active(page
);
5645 list_move_tail(&page
->lru
, list
);
5647 spin_unlock(&hugetlb_lock
);
5651 void putback_active_hugepage(struct page
*page
)
5653 VM_BUG_ON_PAGE(!PageHead(page
), page
);
5654 spin_lock(&hugetlb_lock
);
5655 set_page_huge_active(page
);
5656 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
5657 spin_unlock(&hugetlb_lock
);
5661 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
5663 struct hstate
*h
= page_hstate(oldpage
);
5665 hugetlb_cgroup_migrate(oldpage
, newpage
);
5666 set_page_owner_migrate_reason(newpage
, reason
);
5669 * transfer temporary state of the new huge page. This is
5670 * reverse to other transitions because the newpage is going to
5671 * be final while the old one will be freed so it takes over
5672 * the temporary status.
5674 * Also note that we have to transfer the per-node surplus state
5675 * here as well otherwise the global surplus count will not match
5678 if (PageHugeTemporary(newpage
)) {
5679 int old_nid
= page_to_nid(oldpage
);
5680 int new_nid
= page_to_nid(newpage
);
5682 SetPageHugeTemporary(oldpage
);
5683 ClearPageHugeTemporary(newpage
);
5685 spin_lock(&hugetlb_lock
);
5686 if (h
->surplus_huge_pages_node
[old_nid
]) {
5687 h
->surplus_huge_pages_node
[old_nid
]--;
5688 h
->surplus_huge_pages_node
[new_nid
]++;
5690 spin_unlock(&hugetlb_lock
);
5695 static bool cma_reserve_called __initdata
;
5697 static int __init
cmdline_parse_hugetlb_cma(char *p
)
5699 hugetlb_cma_size
= memparse(p
, &p
);
5703 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma
);
5705 void __init
hugetlb_cma_reserve(int order
)
5707 unsigned long size
, reserved
, per_node
;
5710 cma_reserve_called
= true;
5712 if (!hugetlb_cma_size
)
5715 if (hugetlb_cma_size
< (PAGE_SIZE
<< order
)) {
5716 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5717 (PAGE_SIZE
<< order
) / SZ_1M
);
5722 * If 3 GB area is requested on a machine with 4 numa nodes,
5723 * let's allocate 1 GB on first three nodes and ignore the last one.
5725 per_node
= DIV_ROUND_UP(hugetlb_cma_size
, nr_online_nodes
);
5726 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5727 hugetlb_cma_size
/ SZ_1M
, per_node
/ SZ_1M
);
5730 for_each_node_state(nid
, N_ONLINE
) {
5732 char name
[CMA_MAX_NAME
];
5734 size
= min(per_node
, hugetlb_cma_size
- reserved
);
5735 size
= round_up(size
, PAGE_SIZE
<< order
);
5737 snprintf(name
, sizeof(name
), "hugetlb%d", nid
);
5738 res
= cma_declare_contiguous_nid(0, size
, 0, PAGE_SIZE
<< order
,
5740 &hugetlb_cma
[nid
], nid
);
5742 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5748 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5751 if (reserved
>= hugetlb_cma_size
)
5756 void __init
hugetlb_cma_check(void)
5758 if (!hugetlb_cma_size
|| cma_reserve_called
)
5761 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5764 #endif /* CONFIG_CMA */