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>
33 #include <linux/migrate.h>
36 #include <asm/pgalloc.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
45 #include "hugetlb_vmemmap.h"
47 int hugetlb_max_hstate __read_mostly
;
48 unsigned int default_hstate_idx
;
49 struct hstate hstates
[HUGE_MAX_HSTATE
];
52 static struct cma
*hugetlb_cma
[MAX_NUMNODES
];
54 static unsigned long hugetlb_cma_size __initdata
;
57 * Minimum page order among possible hugepage sizes, set to a proper value
60 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
62 __initdata
LIST_HEAD(huge_boot_pages
);
64 /* for command line parsing */
65 static struct hstate
* __initdata parsed_hstate
;
66 static unsigned long __initdata default_hstate_max_huge_pages
;
67 static bool __initdata parsed_valid_hugepagesz
= true;
68 static bool __initdata parsed_default_hugepagesz
;
71 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
74 DEFINE_SPINLOCK(hugetlb_lock
);
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
80 static int num_fault_mutexes
;
81 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
83 /* Forward declaration */
84 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
86 static inline bool subpool_is_free(struct hugepage_subpool
*spool
)
90 if (spool
->max_hpages
!= -1)
91 return spool
->used_hpages
== 0;
92 if (spool
->min_hpages
!= -1)
93 return spool
->rsv_hpages
== spool
->min_hpages
;
98 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
,
99 unsigned long irq_flags
)
101 spin_unlock_irqrestore(&spool
->lock
, irq_flags
);
103 /* If no pages are used, and no other handles to the subpool
104 * remain, give up any reservations based on minimum size and
105 * free the subpool */
106 if (subpool_is_free(spool
)) {
107 if (spool
->min_hpages
!= -1)
108 hugetlb_acct_memory(spool
->hstate
,
114 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
117 struct hugepage_subpool
*spool
;
119 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
123 spin_lock_init(&spool
->lock
);
125 spool
->max_hpages
= max_hpages
;
127 spool
->min_hpages
= min_hpages
;
129 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
133 spool
->rsv_hpages
= min_hpages
;
138 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
142 spin_lock_irqsave(&spool
->lock
, flags
);
143 BUG_ON(!spool
->count
);
145 unlock_or_release_subpool(spool
, flags
);
149 * Subpool accounting for allocating and reserving pages.
150 * Return -ENOMEM if there are not enough resources to satisfy the
151 * request. Otherwise, return the number of pages by which the
152 * global pools must be adjusted (upward). The returned value may
153 * only be different than the passed value (delta) in the case where
154 * a subpool minimum size must be maintained.
156 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
164 spin_lock_irq(&spool
->lock
);
166 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
167 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
168 spool
->used_hpages
+= delta
;
175 /* minimum size accounting */
176 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
177 if (delta
> spool
->rsv_hpages
) {
179 * Asking for more reserves than those already taken on
180 * behalf of subpool. Return difference.
182 ret
= delta
- spool
->rsv_hpages
;
183 spool
->rsv_hpages
= 0;
185 ret
= 0; /* reserves already accounted for */
186 spool
->rsv_hpages
-= delta
;
191 spin_unlock_irq(&spool
->lock
);
196 * Subpool accounting for freeing and unreserving pages.
197 * Return the number of global page reservations that must be dropped.
198 * The return value may only be different than the passed value (delta)
199 * in the case where a subpool minimum size must be maintained.
201 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
210 spin_lock_irqsave(&spool
->lock
, flags
);
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
, flags
);
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 hugetlb_resv_map_add(struct resv_map
*map
, struct file_region
*rg
, long from
,
362 long to
, struct hstate
*h
, struct hugetlb_cgroup
*cg
,
363 long *regions_needed
)
365 struct file_region
*nrg
;
367 if (!regions_needed
) {
368 nrg
= get_file_region_entry_from_cache(map
, from
, to
);
369 record_hugetlb_cgroup_uncharge_info(cg
, h
, map
, nrg
);
370 list_add(&nrg
->link
, rg
->link
.prev
);
371 coalesce_file_region(map
, nrg
);
373 *regions_needed
+= 1;
379 * Must be called with resv->lock held.
381 * Calling this with regions_needed != NULL will count the number of pages
382 * to be added but will not modify the linked list. And regions_needed will
383 * indicate the number of file_regions needed in the cache to carry out to add
384 * the regions for this range.
386 static long add_reservation_in_range(struct resv_map
*resv
, long f
, long t
,
387 struct hugetlb_cgroup
*h_cg
,
388 struct hstate
*h
, long *regions_needed
)
391 struct list_head
*head
= &resv
->regions
;
392 long last_accounted_offset
= f
;
393 struct file_region
*rg
= NULL
, *trg
= NULL
;
398 /* In this loop, we essentially handle an entry for the range
399 * [last_accounted_offset, rg->from), at every iteration, with some
402 list_for_each_entry_safe(rg
, trg
, head
, link
) {
403 /* Skip irrelevant regions that start before our range. */
405 /* If this region ends after the last accounted offset,
406 * then we need to update last_accounted_offset.
408 if (rg
->to
> last_accounted_offset
)
409 last_accounted_offset
= rg
->to
;
413 /* When we find a region that starts beyond our range, we've
419 /* Add an entry for last_accounted_offset -> rg->from, and
420 * update last_accounted_offset.
422 if (rg
->from
> last_accounted_offset
)
423 add
+= hugetlb_resv_map_add(resv
, rg
,
424 last_accounted_offset
,
428 last_accounted_offset
= rg
->to
;
431 /* Handle the case where our range extends beyond
432 * last_accounted_offset.
434 if (last_accounted_offset
< t
)
435 add
+= hugetlb_resv_map_add(resv
, rg
, last_accounted_offset
,
436 t
, h
, h_cg
, regions_needed
);
442 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
444 static int allocate_file_region_entries(struct resv_map
*resv
,
446 __must_hold(&resv
->lock
)
448 struct list_head allocated_regions
;
449 int to_allocate
= 0, i
= 0;
450 struct file_region
*trg
= NULL
, *rg
= NULL
;
452 VM_BUG_ON(regions_needed
< 0);
454 INIT_LIST_HEAD(&allocated_regions
);
457 * Check for sufficient descriptors in the cache to accommodate
458 * the number of in progress add operations plus regions_needed.
460 * This is a while loop because when we drop the lock, some other call
461 * to region_add or region_del may have consumed some region_entries,
462 * so we keep looping here until we finally have enough entries for
463 * (adds_in_progress + regions_needed).
465 while (resv
->region_cache_count
<
466 (resv
->adds_in_progress
+ regions_needed
)) {
467 to_allocate
= resv
->adds_in_progress
+ regions_needed
-
468 resv
->region_cache_count
;
470 /* At this point, we should have enough entries in the cache
471 * for all the existing adds_in_progress. We should only be
472 * needing to allocate for regions_needed.
474 VM_BUG_ON(resv
->region_cache_count
< resv
->adds_in_progress
);
476 spin_unlock(&resv
->lock
);
477 for (i
= 0; i
< to_allocate
; i
++) {
478 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
481 list_add(&trg
->link
, &allocated_regions
);
484 spin_lock(&resv
->lock
);
486 list_splice(&allocated_regions
, &resv
->region_cache
);
487 resv
->region_cache_count
+= to_allocate
;
493 list_for_each_entry_safe(rg
, trg
, &allocated_regions
, link
) {
501 * Add the huge page range represented by [f, t) to the reserve
502 * map. Regions will be taken from the cache to fill in this range.
503 * Sufficient regions should exist in the cache due to the previous
504 * call to region_chg with the same range, but in some cases the cache will not
505 * have sufficient entries due to races with other code doing region_add or
506 * region_del. The extra needed entries will be allocated.
508 * regions_needed is the out value provided by a previous call to region_chg.
510 * Return the number of new huge pages added to the map. This number is greater
511 * than or equal to zero. If file_region entries needed to be allocated for
512 * this operation and we were not able to allocate, it returns -ENOMEM.
513 * region_add of regions of length 1 never allocate file_regions and cannot
514 * fail; region_chg will always allocate at least 1 entry and a region_add for
515 * 1 page will only require at most 1 entry.
517 static long region_add(struct resv_map
*resv
, long f
, long t
,
518 long in_regions_needed
, struct hstate
*h
,
519 struct hugetlb_cgroup
*h_cg
)
521 long add
= 0, actual_regions_needed
= 0;
523 spin_lock(&resv
->lock
);
526 /* Count how many regions are actually needed to execute this add. */
527 add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
528 &actual_regions_needed
);
531 * Check for sufficient descriptors in the cache to accommodate
532 * this add operation. Note that actual_regions_needed may be greater
533 * than in_regions_needed, as the resv_map may have been modified since
534 * the region_chg call. In this case, we need to make sure that we
535 * allocate extra entries, such that we have enough for all the
536 * existing adds_in_progress, plus the excess needed for this
539 if (actual_regions_needed
> in_regions_needed
&&
540 resv
->region_cache_count
<
541 resv
->adds_in_progress
+
542 (actual_regions_needed
- in_regions_needed
)) {
543 /* region_add operation of range 1 should never need to
544 * allocate file_region entries.
546 VM_BUG_ON(t
- f
<= 1);
548 if (allocate_file_region_entries(
549 resv
, actual_regions_needed
- in_regions_needed
)) {
556 add
= add_reservation_in_range(resv
, f
, t
, h_cg
, h
, NULL
);
558 resv
->adds_in_progress
-= in_regions_needed
;
560 spin_unlock(&resv
->lock
);
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
584 static long region_chg(struct resv_map
*resv
, long f
, long t
,
585 long *out_regions_needed
)
589 spin_lock(&resv
->lock
);
591 /* Count how many hugepages in this range are NOT represented. */
592 chg
= add_reservation_in_range(resv
, f
, t
, NULL
, NULL
,
595 if (*out_regions_needed
== 0)
596 *out_regions_needed
= 1;
598 if (allocate_file_region_entries(resv
, *out_regions_needed
))
601 resv
->adds_in_progress
+= *out_regions_needed
;
603 spin_unlock(&resv
->lock
);
608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
620 static void region_abort(struct resv_map
*resv
, long f
, long t
,
623 spin_lock(&resv
->lock
);
624 VM_BUG_ON(!resv
->region_cache_count
);
625 resv
->adds_in_progress
-= regions_needed
;
626 spin_unlock(&resv
->lock
);
630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
643 static long region_del(struct resv_map
*resv
, long f
, long t
)
645 struct list_head
*head
= &resv
->regions
;
646 struct file_region
*rg
, *trg
;
647 struct file_region
*nrg
= NULL
;
651 spin_lock(&resv
->lock
);
652 list_for_each_entry_safe(rg
, trg
, head
, link
) {
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
660 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
666 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
672 resv
->region_cache_count
> resv
->adds_in_progress
) {
673 nrg
= list_first_entry(&resv
->region_cache
,
676 list_del(&nrg
->link
);
677 resv
->region_cache_count
--;
681 spin_unlock(&resv
->lock
);
682 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
689 hugetlb_cgroup_uncharge_file_region(
690 resv
, rg
, t
- f
, false);
692 /* New entry for end of split region */
696 copy_hugetlb_cgroup_uncharge_info(nrg
, rg
);
698 INIT_LIST_HEAD(&nrg
->link
);
700 /* Original entry is trimmed */
703 list_add(&nrg
->link
, &rg
->link
);
708 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
709 del
+= rg
->to
- rg
->from
;
710 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
711 rg
->to
- rg
->from
, true);
717 if (f
<= rg
->from
) { /* Trim beginning of region */
718 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
719 t
- rg
->from
, false);
723 } else { /* Trim end of region */
724 hugetlb_cgroup_uncharge_file_region(resv
, rg
,
732 spin_unlock(&resv
->lock
);
738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
746 void hugetlb_fix_reserve_counts(struct inode
*inode
)
748 struct hugepage_subpool
*spool
= subpool_inode(inode
);
750 bool reserved
= false;
752 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
753 if (rsv_adjust
> 0) {
754 struct hstate
*h
= hstate_inode(inode
);
756 if (!hugetlb_acct_memory(h
, 1))
758 } else if (!rsv_adjust
) {
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
770 static long region_count(struct resv_map
*resv
, long f
, long t
)
772 struct list_head
*head
= &resv
->regions
;
773 struct file_region
*rg
;
776 spin_lock(&resv
->lock
);
777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg
, head
, link
) {
787 seg_from
= max(rg
->from
, f
);
788 seg_to
= min(rg
->to
, t
);
790 chg
+= seg_to
- seg_from
;
792 spin_unlock(&resv
->lock
);
798 * Convert the address within this vma to the page offset within
799 * the mapping, in pagecache page units; huge pages here.
801 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
802 struct vm_area_struct
*vma
, unsigned long address
)
804 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
805 (vma
->vm_pgoff
>> huge_page_order(h
));
808 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
809 unsigned long address
)
811 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
813 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
819 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
821 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
822 return vma
->vm_ops
->pagesize(vma
);
825 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
833 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
835 return vma_kernel_pagesize(vma
);
839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
843 #define HPAGE_RESV_OWNER (1UL << 0)
844 #define HPAGE_RESV_UNMAPPED (1UL << 1)
845 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
866 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
868 return (unsigned long)vma
->vm_private_data
;
871 static void set_vma_private_data(struct vm_area_struct
*vma
,
874 vma
->vm_private_data
= (void *)value
;
878 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map
*resv_map
,
879 struct hugetlb_cgroup
*h_cg
,
882 #ifdef CONFIG_CGROUP_HUGETLB
884 resv_map
->reservation_counter
= NULL
;
885 resv_map
->pages_per_hpage
= 0;
886 resv_map
->css
= NULL
;
888 resv_map
->reservation_counter
=
889 &h_cg
->rsvd_hugepage
[hstate_index(h
)];
890 resv_map
->pages_per_hpage
= pages_per_huge_page(h
);
891 resv_map
->css
= &h_cg
->css
;
896 struct resv_map
*resv_map_alloc(void)
898 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
899 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
901 if (!resv_map
|| !rg
) {
907 kref_init(&resv_map
->refs
);
908 spin_lock_init(&resv_map
->lock
);
909 INIT_LIST_HEAD(&resv_map
->regions
);
911 resv_map
->adds_in_progress
= 0;
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, NULL
, NULL
);
920 INIT_LIST_HEAD(&resv_map
->region_cache
);
921 list_add(&rg
->link
, &resv_map
->region_cache
);
922 resv_map
->region_cache_count
= 1;
927 void resv_map_release(struct kref
*ref
)
929 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
930 struct list_head
*head
= &resv_map
->region_cache
;
931 struct file_region
*rg
, *trg
;
933 /* Clear out any active regions before we release the map. */
934 region_del(resv_map
, 0, LONG_MAX
);
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg
, trg
, head
, link
) {
942 VM_BUG_ON(resv_map
->adds_in_progress
);
947 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
957 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
960 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
963 if (vma
->vm_flags
& VM_MAYSHARE
) {
964 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
965 struct inode
*inode
= mapping
->host
;
967 return inode_resv_map(inode
);
970 return (struct resv_map
*)(get_vma_private_data(vma
) &
975 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
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
) &
981 HPAGE_RESV_MASK
) | (unsigned long)map
);
984 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
987 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
989 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
992 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
996 return (get_vma_private_data(vma
) & flag
) != 0;
999 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1000 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
1002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
1003 if (!(vma
->vm_flags
& VM_MAYSHARE
))
1004 vma
->vm_private_data
= (void *)0;
1007 /* Returns true if the VMA has associated reserve pages */
1008 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
1010 if (vma
->vm_flags
& VM_NORESERVE
) {
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1020 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
1026 /* Shared mappings always use reserves */
1027 if (vma
->vm_flags
& VM_MAYSHARE
) {
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
1032 * fallocate. In this case, there really are no reserves to
1033 * use. This situation is indicated if chg != 0.
1042 * Only the process that called mmap() has reserves for
1045 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1070 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
1072 int nid
= page_to_nid(page
);
1074 lockdep_assert_held(&hugetlb_lock
);
1075 VM_BUG_ON_PAGE(page_count(page
), page
);
1077 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
1078 h
->free_huge_pages
++;
1079 h
->free_huge_pages_node
[nid
]++;
1080 SetHPageFreed(page
);
1083 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
1086 bool pin
= !!(current
->flags
& PF_MEMALLOC_PIN
);
1088 lockdep_assert_held(&hugetlb_lock
);
1089 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
) {
1090 if (pin
&& !is_pinnable_page(page
))
1093 if (PageHWPoison(page
))
1096 list_move(&page
->lru
, &h
->hugepage_activelist
);
1097 set_page_refcounted(page
);
1098 ClearHPageFreed(page
);
1099 h
->free_huge_pages
--;
1100 h
->free_huge_pages_node
[nid
]--;
1107 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
1110 unsigned int cpuset_mems_cookie
;
1111 struct zonelist
*zonelist
;
1114 int node
= NUMA_NO_NODE
;
1116 zonelist
= node_zonelist(nid
, gfp_mask
);
1119 cpuset_mems_cookie
= read_mems_allowed_begin();
1120 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
1123 if (!cpuset_zone_allowed(zone
, gfp_mask
))
1126 * no need to ask again on the same node. Pool is node rather than
1129 if (zone_to_nid(zone
) == node
)
1131 node
= zone_to_nid(zone
);
1133 page
= dequeue_huge_page_node_exact(h
, node
);
1137 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
1143 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
1144 struct vm_area_struct
*vma
,
1145 unsigned long address
, int avoid_reserve
,
1148 struct page
*page
= NULL
;
1149 struct mempolicy
*mpol
;
1151 nodemask_t
*nodemask
;
1155 * A child process with MAP_PRIVATE mappings created by their parent
1156 * have no page reserves. This check ensures that reservations are
1157 * not "stolen". The child may still get SIGKILLed
1159 if (!vma_has_reserves(vma
, chg
) &&
1160 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1163 /* If reserves cannot be used, ensure enough pages are in the pool */
1164 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1167 gfp_mask
= htlb_alloc_mask(h
);
1168 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1170 if (mpol_is_preferred_many(mpol
)) {
1171 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
1173 /* Fallback to all nodes if page==NULL */
1178 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
1180 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
1181 SetHPageRestoreReserve(page
);
1182 h
->resv_huge_pages
--;
1185 mpol_cond_put(mpol
);
1193 * common helper functions for hstate_next_node_to_{alloc|free}.
1194 * We may have allocated or freed a huge page based on a different
1195 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1196 * be outside of *nodes_allowed. Ensure that we use an allowed
1197 * node for alloc or free.
1199 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1201 nid
= next_node_in(nid
, *nodes_allowed
);
1202 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1207 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1209 if (!node_isset(nid
, *nodes_allowed
))
1210 nid
= next_node_allowed(nid
, nodes_allowed
);
1215 * returns the previously saved node ["this node"] from which to
1216 * allocate a persistent huge page for the pool and advance the
1217 * next node from which to allocate, handling wrap at end of node
1220 static int hstate_next_node_to_alloc(struct hstate
*h
,
1221 nodemask_t
*nodes_allowed
)
1225 VM_BUG_ON(!nodes_allowed
);
1227 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1228 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1234 * helper for remove_pool_huge_page() - return the previously saved
1235 * node ["this node"] from which to free a huge page. Advance the
1236 * next node id whether or not we find a free huge page to free so
1237 * that the next attempt to free addresses the next node.
1239 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1243 VM_BUG_ON(!nodes_allowed
);
1245 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1246 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1251 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1252 for (nr_nodes = nodes_weight(*mask); \
1254 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1257 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1258 for (nr_nodes = nodes_weight(*mask); \
1260 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1263 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1264 static void destroy_compound_gigantic_page(struct page
*page
,
1268 int nr_pages
= 1 << order
;
1269 struct page
*p
= page
+ 1;
1271 atomic_set(compound_mapcount_ptr(page
), 0);
1272 atomic_set(compound_pincount_ptr(page
), 0);
1274 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1275 clear_compound_head(p
);
1276 set_page_refcounted(p
);
1279 set_compound_order(page
, 0);
1280 page
[1].compound_nr
= 0;
1281 __ClearPageHead(page
);
1284 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1287 * If the page isn't allocated using the cma allocator,
1288 * cma_release() returns false.
1291 if (cma_release(hugetlb_cma
[page_to_nid(page
)], page
, 1 << order
))
1295 free_contig_range(page_to_pfn(page
), 1 << order
);
1298 #ifdef CONFIG_CONTIG_ALLOC
1299 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1300 int nid
, nodemask_t
*nodemask
)
1302 unsigned long nr_pages
= pages_per_huge_page(h
);
1303 if (nid
== NUMA_NO_NODE
)
1304 nid
= numa_mem_id();
1311 if (hugetlb_cma
[nid
]) {
1312 page
= cma_alloc(hugetlb_cma
[nid
], nr_pages
,
1313 huge_page_order(h
), true);
1318 if (!(gfp_mask
& __GFP_THISNODE
)) {
1319 for_each_node_mask(node
, *nodemask
) {
1320 if (node
== nid
|| !hugetlb_cma
[node
])
1323 page
= cma_alloc(hugetlb_cma
[node
], nr_pages
,
1324 huge_page_order(h
), true);
1332 return alloc_contig_pages(nr_pages
, gfp_mask
, nid
, nodemask
);
1335 #else /* !CONFIG_CONTIG_ALLOC */
1336 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1337 int nid
, nodemask_t
*nodemask
)
1341 #endif /* CONFIG_CONTIG_ALLOC */
1343 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1344 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1345 int nid
, nodemask_t
*nodemask
)
1349 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1350 static inline void destroy_compound_gigantic_page(struct page
*page
,
1351 unsigned int order
) { }
1355 * Remove hugetlb page from lists, and update dtor so that page appears
1356 * as just a compound page. A reference is held on the page.
1358 * Must be called with hugetlb lock held.
1360 static void remove_hugetlb_page(struct hstate
*h
, struct page
*page
,
1361 bool adjust_surplus
)
1363 int nid
= page_to_nid(page
);
1365 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1366 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page
), page
);
1368 lockdep_assert_held(&hugetlb_lock
);
1369 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1372 list_del(&page
->lru
);
1374 if (HPageFreed(page
)) {
1375 h
->free_huge_pages
--;
1376 h
->free_huge_pages_node
[nid
]--;
1378 if (adjust_surplus
) {
1379 h
->surplus_huge_pages
--;
1380 h
->surplus_huge_pages_node
[nid
]--;
1386 * For non-gigantic pages set the destructor to the normal compound
1387 * page dtor. This is needed in case someone takes an additional
1388 * temporary ref to the page, and freeing is delayed until they drop
1391 * For gigantic pages set the destructor to the null dtor. This
1392 * destructor will never be called. Before freeing the gigantic
1393 * page destroy_compound_gigantic_page will turn the compound page
1394 * into a simple group of pages. After this the destructor does not
1397 * This handles the case where more than one ref is held when and
1398 * after update_and_free_page is called.
1400 set_page_refcounted(page
);
1401 if (hstate_is_gigantic(h
))
1402 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1404 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
1407 h
->nr_huge_pages_node
[nid
]--;
1410 static void add_hugetlb_page(struct hstate
*h
, struct page
*page
,
1411 bool adjust_surplus
)
1414 int nid
= page_to_nid(page
);
1416 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page
), page
);
1418 lockdep_assert_held(&hugetlb_lock
);
1420 INIT_LIST_HEAD(&page
->lru
);
1422 h
->nr_huge_pages_node
[nid
]++;
1424 if (adjust_surplus
) {
1425 h
->surplus_huge_pages
++;
1426 h
->surplus_huge_pages_node
[nid
]++;
1429 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1430 set_page_private(page
, 0);
1431 SetHPageVmemmapOptimized(page
);
1434 * This page is about to be managed by the hugetlb allocator and
1435 * should have no users. Drop our reference, and check for others
1438 zeroed
= put_page_testzero(page
);
1441 * It is VERY unlikely soneone else has taken a ref on
1442 * the page. In this case, we simply return as the
1443 * hugetlb destructor (free_huge_page) will be called
1444 * when this other ref is dropped.
1448 arch_clear_hugepage_flags(page
);
1449 enqueue_huge_page(h
, page
);
1452 static void __update_and_free_page(struct hstate
*h
, struct page
*page
)
1455 struct page
*subpage
= page
;
1457 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1460 if (alloc_huge_page_vmemmap(h
, page
)) {
1461 spin_lock_irq(&hugetlb_lock
);
1463 * If we cannot allocate vmemmap pages, just refuse to free the
1464 * page and put the page back on the hugetlb free list and treat
1465 * as a surplus page.
1467 add_hugetlb_page(h
, page
, true);
1468 spin_unlock_irq(&hugetlb_lock
);
1472 for (i
= 0; i
< pages_per_huge_page(h
);
1473 i
++, subpage
= mem_map_next(subpage
, page
, i
)) {
1474 subpage
->flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1475 1 << PG_referenced
| 1 << PG_dirty
|
1476 1 << PG_active
| 1 << PG_private
|
1479 if (hstate_is_gigantic(h
)) {
1480 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1481 free_gigantic_page(page
, huge_page_order(h
));
1483 __free_pages(page
, huge_page_order(h
));
1488 * As update_and_free_page() can be called under any context, so we cannot
1489 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1490 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1491 * the vmemmap pages.
1493 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1494 * freed and frees them one-by-one. As the page->mapping pointer is going
1495 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1496 * structure of a lockless linked list of huge pages to be freed.
1498 static LLIST_HEAD(hpage_freelist
);
1500 static void free_hpage_workfn(struct work_struct
*work
)
1502 struct llist_node
*node
;
1504 node
= llist_del_all(&hpage_freelist
);
1510 page
= container_of((struct address_space
**)node
,
1511 struct page
, mapping
);
1513 page
->mapping
= NULL
;
1515 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1516 * is going to trigger because a previous call to
1517 * remove_hugetlb_page() will set_compound_page_dtor(page,
1518 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1520 h
= size_to_hstate(page_size(page
));
1522 __update_and_free_page(h
, page
);
1527 static DECLARE_WORK(free_hpage_work
, free_hpage_workfn
);
1529 static inline void flush_free_hpage_work(struct hstate
*h
)
1531 if (free_vmemmap_pages_per_hpage(h
))
1532 flush_work(&free_hpage_work
);
1535 static void update_and_free_page(struct hstate
*h
, struct page
*page
,
1538 if (!HPageVmemmapOptimized(page
) || !atomic
) {
1539 __update_and_free_page(h
, page
);
1544 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1546 * Only call schedule_work() if hpage_freelist is previously
1547 * empty. Otherwise, schedule_work() had been called but the workfn
1548 * hasn't retrieved the list yet.
1550 if (llist_add((struct llist_node
*)&page
->mapping
, &hpage_freelist
))
1551 schedule_work(&free_hpage_work
);
1554 static void update_and_free_pages_bulk(struct hstate
*h
, struct list_head
*list
)
1556 struct page
*page
, *t_page
;
1558 list_for_each_entry_safe(page
, t_page
, list
, lru
) {
1559 update_and_free_page(h
, page
, false);
1564 struct hstate
*size_to_hstate(unsigned long size
)
1568 for_each_hstate(h
) {
1569 if (huge_page_size(h
) == size
)
1575 void free_huge_page(struct page
*page
)
1578 * Can't pass hstate in here because it is called from the
1579 * compound page destructor.
1581 struct hstate
*h
= page_hstate(page
);
1582 int nid
= page_to_nid(page
);
1583 struct hugepage_subpool
*spool
= hugetlb_page_subpool(page
);
1584 bool restore_reserve
;
1585 unsigned long flags
;
1587 VM_BUG_ON_PAGE(page_count(page
), page
);
1588 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1590 hugetlb_set_page_subpool(page
, NULL
);
1591 page
->mapping
= NULL
;
1592 restore_reserve
= HPageRestoreReserve(page
);
1593 ClearHPageRestoreReserve(page
);
1596 * If HPageRestoreReserve was set on page, page allocation consumed a
1597 * reservation. If the page was associated with a subpool, there
1598 * would have been a page reserved in the subpool before allocation
1599 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1600 * reservation, do not call hugepage_subpool_put_pages() as this will
1601 * remove the reserved page from the subpool.
1603 if (!restore_reserve
) {
1605 * A return code of zero implies that the subpool will be
1606 * under its minimum size if the reservation is not restored
1607 * after page is free. Therefore, force restore_reserve
1610 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1611 restore_reserve
= true;
1614 spin_lock_irqsave(&hugetlb_lock
, flags
);
1615 ClearHPageMigratable(page
);
1616 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1617 pages_per_huge_page(h
), page
);
1618 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
1619 pages_per_huge_page(h
), page
);
1620 if (restore_reserve
)
1621 h
->resv_huge_pages
++;
1623 if (HPageTemporary(page
)) {
1624 remove_hugetlb_page(h
, page
, false);
1625 spin_unlock_irqrestore(&hugetlb_lock
, flags
);
1626 update_and_free_page(h
, page
, true);
1627 } else if (h
->surplus_huge_pages_node
[nid
]) {
1628 /* remove the page from active list */
1629 remove_hugetlb_page(h
, page
, true);
1630 spin_unlock_irqrestore(&hugetlb_lock
, flags
);
1631 update_and_free_page(h
, page
, true);
1633 arch_clear_hugepage_flags(page
);
1634 enqueue_huge_page(h
, page
);
1635 spin_unlock_irqrestore(&hugetlb_lock
, flags
);
1640 * Must be called with the hugetlb lock held
1642 static void __prep_account_new_huge_page(struct hstate
*h
, int nid
)
1644 lockdep_assert_held(&hugetlb_lock
);
1646 h
->nr_huge_pages_node
[nid
]++;
1649 static void __prep_new_huge_page(struct hstate
*h
, struct page
*page
)
1651 free_huge_page_vmemmap(h
, page
);
1652 INIT_LIST_HEAD(&page
->lru
);
1653 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1654 hugetlb_set_page_subpool(page
, NULL
);
1655 set_hugetlb_cgroup(page
, NULL
);
1656 set_hugetlb_cgroup_rsvd(page
, NULL
);
1659 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1661 __prep_new_huge_page(h
, page
);
1662 spin_lock_irq(&hugetlb_lock
);
1663 __prep_account_new_huge_page(h
, nid
);
1664 spin_unlock_irq(&hugetlb_lock
);
1667 static bool prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1670 int nr_pages
= 1 << order
;
1671 struct page
*p
= page
+ 1;
1673 /* we rely on prep_new_huge_page to set the destructor */
1674 set_compound_order(page
, order
);
1675 __ClearPageReserved(page
);
1676 __SetPageHead(page
);
1677 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1679 * For gigantic hugepages allocated through bootmem at
1680 * boot, it's safer to be consistent with the not-gigantic
1681 * hugepages and clear the PG_reserved bit from all tail pages
1682 * too. Otherwise drivers using get_user_pages() to access tail
1683 * pages may get the reference counting wrong if they see
1684 * PG_reserved set on a tail page (despite the head page not
1685 * having PG_reserved set). Enforcing this consistency between
1686 * head and tail pages allows drivers to optimize away a check
1687 * on the head page when they need know if put_page() is needed
1688 * after get_user_pages().
1690 __ClearPageReserved(p
);
1692 * Subtle and very unlikely
1694 * Gigantic 'page allocators' such as memblock or cma will
1695 * return a set of pages with each page ref counted. We need
1696 * to turn this set of pages into a compound page with tail
1697 * page ref counts set to zero. Code such as speculative page
1698 * cache adding could take a ref on a 'to be' tail page.
1699 * We need to respect any increased ref count, and only set
1700 * the ref count to zero if count is currently 1. If count
1701 * is not 1, we return an error. An error return indicates
1702 * the set of pages can not be converted to a gigantic page.
1703 * The caller who allocated the pages should then discard the
1704 * pages using the appropriate free interface.
1706 if (!page_ref_freeze(p
, 1)) {
1707 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1710 set_page_count(p
, 0);
1711 set_compound_head(p
, page
);
1713 atomic_set(compound_mapcount_ptr(page
), -1);
1714 atomic_set(compound_pincount_ptr(page
), 0);
1718 /* undo tail page modifications made above */
1720 for (j
= 1; j
< i
; j
++, p
= mem_map_next(p
, page
, j
)) {
1721 clear_compound_head(p
);
1722 set_page_refcounted(p
);
1724 /* need to clear PG_reserved on remaining tail pages */
1725 for (; j
< nr_pages
; j
++, p
= mem_map_next(p
, page
, j
))
1726 __ClearPageReserved(p
);
1727 set_compound_order(page
, 0);
1728 page
[1].compound_nr
= 0;
1729 __ClearPageHead(page
);
1734 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1735 * transparent huge pages. See the PageTransHuge() documentation for more
1738 int PageHuge(struct page
*page
)
1740 if (!PageCompound(page
))
1743 page
= compound_head(page
);
1744 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1746 EXPORT_SYMBOL_GPL(PageHuge
);
1749 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1750 * normal or transparent huge pages.
1752 int PageHeadHuge(struct page
*page_head
)
1754 if (!PageHead(page_head
))
1757 return page_head
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1761 * Find and lock address space (mapping) in write mode.
1763 * Upon entry, the page is locked which means that page_mapping() is
1764 * stable. Due to locking order, we can only trylock_write. If we can
1765 * not get the lock, simply return NULL to caller.
1767 struct address_space
*hugetlb_page_mapping_lock_write(struct page
*hpage
)
1769 struct address_space
*mapping
= page_mapping(hpage
);
1774 if (i_mmap_trylock_write(mapping
))
1780 pgoff_t
hugetlb_basepage_index(struct page
*page
)
1782 struct page
*page_head
= compound_head(page
);
1783 pgoff_t index
= page_index(page_head
);
1784 unsigned long compound_idx
;
1786 if (compound_order(page_head
) >= MAX_ORDER
)
1787 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1789 compound_idx
= page
- page_head
;
1791 return (index
<< compound_order(page_head
)) + compound_idx
;
1794 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1795 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1796 nodemask_t
*node_alloc_noretry
)
1798 int order
= huge_page_order(h
);
1800 bool alloc_try_hard
= true;
1803 * By default we always try hard to allocate the page with
1804 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1805 * a loop (to adjust global huge page counts) and previous allocation
1806 * failed, do not continue to try hard on the same node. Use the
1807 * node_alloc_noretry bitmap to manage this state information.
1809 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1810 alloc_try_hard
= false;
1811 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1813 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1814 if (nid
== NUMA_NO_NODE
)
1815 nid
= numa_mem_id();
1816 page
= __alloc_pages(gfp_mask
, order
, nid
, nmask
);
1818 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1820 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1823 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1824 * indicates an overall state change. Clear bit so that we resume
1825 * normal 'try hard' allocations.
1827 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1828 node_clear(nid
, *node_alloc_noretry
);
1831 * If we tried hard to get a page but failed, set bit so that
1832 * subsequent attempts will not try as hard until there is an
1833 * overall state change.
1835 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1836 node_set(nid
, *node_alloc_noretry
);
1842 * Common helper to allocate a fresh hugetlb page. All specific allocators
1843 * should use this function to get new hugetlb pages
1845 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1846 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1847 nodemask_t
*node_alloc_noretry
)
1853 if (hstate_is_gigantic(h
))
1854 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1856 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1857 nid
, nmask
, node_alloc_noretry
);
1861 if (hstate_is_gigantic(h
)) {
1862 if (!prep_compound_gigantic_page(page
, huge_page_order(h
))) {
1864 * Rare failure to convert pages to compound page.
1865 * Free pages and try again - ONCE!
1867 free_gigantic_page(page
, huge_page_order(h
));
1875 prep_new_huge_page(h
, page
, page_to_nid(page
));
1881 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1884 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1885 nodemask_t
*node_alloc_noretry
)
1889 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1891 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1892 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1893 node_alloc_noretry
);
1901 put_page(page
); /* free it into the hugepage allocator */
1907 * Remove huge page from pool from next node to free. Attempt to keep
1908 * persistent huge pages more or less balanced over allowed nodes.
1909 * This routine only 'removes' the hugetlb page. The caller must make
1910 * an additional call to free the page to low level allocators.
1911 * Called with hugetlb_lock locked.
1913 static struct page
*remove_pool_huge_page(struct hstate
*h
,
1914 nodemask_t
*nodes_allowed
,
1918 struct page
*page
= NULL
;
1920 lockdep_assert_held(&hugetlb_lock
);
1921 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1923 * If we're returning unused surplus pages, only examine
1924 * nodes with surplus pages.
1926 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1927 !list_empty(&h
->hugepage_freelists
[node
])) {
1928 page
= list_entry(h
->hugepage_freelists
[node
].next
,
1930 remove_hugetlb_page(h
, page
, acct_surplus
);
1939 * Dissolve a given free hugepage into free buddy pages. This function does
1940 * nothing for in-use hugepages and non-hugepages.
1941 * This function returns values like below:
1943 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1944 * when the system is under memory pressure and the feature of
1945 * freeing unused vmemmap pages associated with each hugetlb page
1947 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1948 * (allocated or reserved.)
1949 * 0: successfully dissolved free hugepages or the page is not a
1950 * hugepage (considered as already dissolved)
1952 int dissolve_free_huge_page(struct page
*page
)
1957 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1958 if (!PageHuge(page
))
1961 spin_lock_irq(&hugetlb_lock
);
1962 if (!PageHuge(page
)) {
1967 if (!page_count(page
)) {
1968 struct page
*head
= compound_head(page
);
1969 struct hstate
*h
= page_hstate(head
);
1970 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1974 * We should make sure that the page is already on the free list
1975 * when it is dissolved.
1977 if (unlikely(!HPageFreed(head
))) {
1978 spin_unlock_irq(&hugetlb_lock
);
1982 * Theoretically, we should return -EBUSY when we
1983 * encounter this race. In fact, we have a chance
1984 * to successfully dissolve the page if we do a
1985 * retry. Because the race window is quite small.
1986 * If we seize this opportunity, it is an optimization
1987 * for increasing the success rate of dissolving page.
1992 remove_hugetlb_page(h
, head
, false);
1993 h
->max_huge_pages
--;
1994 spin_unlock_irq(&hugetlb_lock
);
1997 * Normally update_and_free_page will allocate required vmemmmap
1998 * before freeing the page. update_and_free_page will fail to
1999 * free the page if it can not allocate required vmemmap. We
2000 * need to adjust max_huge_pages if the page is not freed.
2001 * Attempt to allocate vmemmmap here so that we can take
2002 * appropriate action on failure.
2004 rc
= alloc_huge_page_vmemmap(h
, head
);
2007 * Move PageHWPoison flag from head page to the raw
2008 * error page, which makes any subpages rather than
2009 * the error page reusable.
2011 if (PageHWPoison(head
) && page
!= head
) {
2012 SetPageHWPoison(page
);
2013 ClearPageHWPoison(head
);
2015 update_and_free_page(h
, head
, false);
2017 spin_lock_irq(&hugetlb_lock
);
2018 add_hugetlb_page(h
, head
, false);
2019 h
->max_huge_pages
++;
2020 spin_unlock_irq(&hugetlb_lock
);
2026 spin_unlock_irq(&hugetlb_lock
);
2031 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2032 * make specified memory blocks removable from the system.
2033 * Note that this will dissolve a free gigantic hugepage completely, if any
2034 * part of it lies within the given range.
2035 * Also note that if dissolve_free_huge_page() returns with an error, all
2036 * free hugepages that were dissolved before that error are lost.
2038 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
2044 if (!hugepages_supported())
2047 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
2048 page
= pfn_to_page(pfn
);
2049 rc
= dissolve_free_huge_page(page
);
2058 * Allocates a fresh surplus page from the page allocator.
2060 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
2061 int nid
, nodemask_t
*nmask
, bool zero_ref
)
2063 struct page
*page
= NULL
;
2066 if (hstate_is_gigantic(h
))
2069 spin_lock_irq(&hugetlb_lock
);
2070 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
2072 spin_unlock_irq(&hugetlb_lock
);
2075 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
2079 spin_lock_irq(&hugetlb_lock
);
2081 * We could have raced with the pool size change.
2082 * Double check that and simply deallocate the new page
2083 * if we would end up overcommiting the surpluses. Abuse
2084 * temporary page to workaround the nasty free_huge_page
2087 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
2088 SetHPageTemporary(page
);
2089 spin_unlock_irq(&hugetlb_lock
);
2096 * Caller requires a page with zero ref count.
2097 * We will drop ref count here. If someone else is holding
2098 * a ref, the page will be freed when they drop it. Abuse
2099 * temporary page flag to accomplish this.
2101 SetHPageTemporary(page
);
2102 if (!put_page_testzero(page
)) {
2104 * Unexpected inflated ref count on freshly allocated
2107 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2108 spin_unlock_irq(&hugetlb_lock
);
2115 ClearHPageTemporary(page
);
2118 h
->surplus_huge_pages
++;
2119 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
2122 spin_unlock_irq(&hugetlb_lock
);
2127 static struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
2128 int nid
, nodemask_t
*nmask
)
2132 if (hstate_is_gigantic(h
))
2135 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
2140 * We do not account these pages as surplus because they are only
2141 * temporary and will be released properly on the last reference
2143 SetHPageTemporary(page
);
2149 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2152 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
2153 struct vm_area_struct
*vma
, unsigned long addr
)
2155 struct page
*page
= NULL
;
2156 struct mempolicy
*mpol
;
2157 gfp_t gfp_mask
= htlb_alloc_mask(h
);
2159 nodemask_t
*nodemask
;
2161 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
2162 if (mpol_is_preferred_many(mpol
)) {
2163 gfp_t gfp
= gfp_mask
| __GFP_NOWARN
;
2165 gfp
&= ~(__GFP_DIRECT_RECLAIM
| __GFP_NOFAIL
);
2166 page
= alloc_surplus_huge_page(h
, gfp
, nid
, nodemask
, false);
2168 /* Fallback to all nodes if page==NULL */
2173 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
, false);
2174 mpol_cond_put(mpol
);
2178 /* page migration callback function */
2179 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
2180 nodemask_t
*nmask
, gfp_t gfp_mask
)
2182 spin_lock_irq(&hugetlb_lock
);
2183 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
2186 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
2188 spin_unlock_irq(&hugetlb_lock
);
2192 spin_unlock_irq(&hugetlb_lock
);
2194 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
2197 /* mempolicy aware migration callback */
2198 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
2199 unsigned long address
)
2201 struct mempolicy
*mpol
;
2202 nodemask_t
*nodemask
;
2207 gfp_mask
= htlb_alloc_mask(h
);
2208 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
2209 page
= alloc_huge_page_nodemask(h
, node
, nodemask
, gfp_mask
);
2210 mpol_cond_put(mpol
);
2216 * Increase the hugetlb pool such that it can accommodate a reservation
2219 static int gather_surplus_pages(struct hstate
*h
, long delta
)
2220 __must_hold(&hugetlb_lock
)
2222 struct list_head surplus_list
;
2223 struct page
*page
, *tmp
;
2226 long needed
, allocated
;
2227 bool alloc_ok
= true;
2229 lockdep_assert_held(&hugetlb_lock
);
2230 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
2232 h
->resv_huge_pages
+= delta
;
2237 INIT_LIST_HEAD(&surplus_list
);
2241 spin_unlock_irq(&hugetlb_lock
);
2242 for (i
= 0; i
< needed
; i
++) {
2243 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
2244 NUMA_NO_NODE
, NULL
, true);
2249 list_add(&page
->lru
, &surplus_list
);
2255 * After retaking hugetlb_lock, we need to recalculate 'needed'
2256 * because either resv_huge_pages or free_huge_pages may have changed.
2258 spin_lock_irq(&hugetlb_lock
);
2259 needed
= (h
->resv_huge_pages
+ delta
) -
2260 (h
->free_huge_pages
+ allocated
);
2265 * We were not able to allocate enough pages to
2266 * satisfy the entire reservation so we free what
2267 * we've allocated so far.
2272 * The surplus_list now contains _at_least_ the number of extra pages
2273 * needed to accommodate the reservation. Add the appropriate number
2274 * of pages to the hugetlb pool and free the extras back to the buddy
2275 * allocator. Commit the entire reservation here to prevent another
2276 * process from stealing the pages as they are added to the pool but
2277 * before they are reserved.
2279 needed
+= allocated
;
2280 h
->resv_huge_pages
+= delta
;
2283 /* Free the needed pages to the hugetlb pool */
2284 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
2287 /* Add the page to the hugetlb allocator */
2288 enqueue_huge_page(h
, page
);
2291 spin_unlock_irq(&hugetlb_lock
);
2294 * Free unnecessary surplus pages to the buddy allocator.
2295 * Pages have no ref count, call free_huge_page directly.
2297 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
2298 free_huge_page(page
);
2299 spin_lock_irq(&hugetlb_lock
);
2305 * This routine has two main purposes:
2306 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2307 * in unused_resv_pages. This corresponds to the prior adjustments made
2308 * to the associated reservation map.
2309 * 2) Free any unused surplus pages that may have been allocated to satisfy
2310 * the reservation. As many as unused_resv_pages may be freed.
2312 static void return_unused_surplus_pages(struct hstate
*h
,
2313 unsigned long unused_resv_pages
)
2315 unsigned long nr_pages
;
2317 LIST_HEAD(page_list
);
2319 lockdep_assert_held(&hugetlb_lock
);
2320 /* Uncommit the reservation */
2321 h
->resv_huge_pages
-= unused_resv_pages
;
2323 /* Cannot return gigantic pages currently */
2324 if (hstate_is_gigantic(h
))
2328 * Part (or even all) of the reservation could have been backed
2329 * by pre-allocated pages. Only free surplus pages.
2331 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
2334 * We want to release as many surplus pages as possible, spread
2335 * evenly across all nodes with memory. Iterate across these nodes
2336 * until we can no longer free unreserved surplus pages. This occurs
2337 * when the nodes with surplus pages have no free pages.
2338 * remove_pool_huge_page() will balance the freed pages across the
2339 * on-line nodes with memory and will handle the hstate accounting.
2341 while (nr_pages
--) {
2342 page
= remove_pool_huge_page(h
, &node_states
[N_MEMORY
], 1);
2346 list_add(&page
->lru
, &page_list
);
2350 spin_unlock_irq(&hugetlb_lock
);
2351 update_and_free_pages_bulk(h
, &page_list
);
2352 spin_lock_irq(&hugetlb_lock
);
2357 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2358 * are used by the huge page allocation routines to manage reservations.
2360 * vma_needs_reservation is called to determine if the huge page at addr
2361 * within the vma has an associated reservation. If a reservation is
2362 * needed, the value 1 is returned. The caller is then responsible for
2363 * managing the global reservation and subpool usage counts. After
2364 * the huge page has been allocated, vma_commit_reservation is called
2365 * to add the page to the reservation map. If the page allocation fails,
2366 * the reservation must be ended instead of committed. vma_end_reservation
2367 * is called in such cases.
2369 * In the normal case, vma_commit_reservation returns the same value
2370 * as the preceding vma_needs_reservation call. The only time this
2371 * is not the case is if a reserve map was changed between calls. It
2372 * is the responsibility of the caller to notice the difference and
2373 * take appropriate action.
2375 * vma_add_reservation is used in error paths where a reservation must
2376 * be restored when a newly allocated huge page must be freed. It is
2377 * to be called after calling vma_needs_reservation to determine if a
2378 * reservation exists.
2380 * vma_del_reservation is used in error paths where an entry in the reserve
2381 * map was created during huge page allocation and must be removed. It is to
2382 * be called after calling vma_needs_reservation to determine if a reservation
2385 enum vma_resv_mode
{
2392 static long __vma_reservation_common(struct hstate
*h
,
2393 struct vm_area_struct
*vma
, unsigned long addr
,
2394 enum vma_resv_mode mode
)
2396 struct resv_map
*resv
;
2399 long dummy_out_regions_needed
;
2401 resv
= vma_resv_map(vma
);
2405 idx
= vma_hugecache_offset(h
, vma
, addr
);
2407 case VMA_NEEDS_RESV
:
2408 ret
= region_chg(resv
, idx
, idx
+ 1, &dummy_out_regions_needed
);
2409 /* We assume that vma_reservation_* routines always operate on
2410 * 1 page, and that adding to resv map a 1 page entry can only
2411 * ever require 1 region.
2413 VM_BUG_ON(dummy_out_regions_needed
!= 1);
2415 case VMA_COMMIT_RESV
:
2416 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2417 /* region_add calls of range 1 should never fail. */
2421 region_abort(resv
, idx
, idx
+ 1, 1);
2425 if (vma
->vm_flags
& VM_MAYSHARE
) {
2426 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2427 /* region_add calls of range 1 should never fail. */
2430 region_abort(resv
, idx
, idx
+ 1, 1);
2431 ret
= region_del(resv
, idx
, idx
+ 1);
2435 if (vma
->vm_flags
& VM_MAYSHARE
) {
2436 region_abort(resv
, idx
, idx
+ 1, 1);
2437 ret
= region_del(resv
, idx
, idx
+ 1);
2439 ret
= region_add(resv
, idx
, idx
+ 1, 1, NULL
, NULL
);
2440 /* region_add calls of range 1 should never fail. */
2448 if (vma
->vm_flags
& VM_MAYSHARE
|| mode
== VMA_DEL_RESV
)
2451 * We know private mapping must have HPAGE_RESV_OWNER set.
2453 * In most cases, reserves always exist for private mappings.
2454 * However, a file associated with mapping could have been
2455 * hole punched or truncated after reserves were consumed.
2456 * As subsequent fault on such a range will not use reserves.
2457 * Subtle - The reserve map for private mappings has the
2458 * opposite meaning than that of shared mappings. If NO
2459 * entry is in the reserve map, it means a reservation exists.
2460 * If an entry exists in the reserve map, it means the
2461 * reservation has already been consumed. As a result, the
2462 * return value of this routine is the opposite of the
2463 * value returned from reserve map manipulation routines above.
2472 static long vma_needs_reservation(struct hstate
*h
,
2473 struct vm_area_struct
*vma
, unsigned long addr
)
2475 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2478 static long vma_commit_reservation(struct hstate
*h
,
2479 struct vm_area_struct
*vma
, unsigned long addr
)
2481 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2484 static void vma_end_reservation(struct hstate
*h
,
2485 struct vm_area_struct
*vma
, unsigned long addr
)
2487 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2490 static long vma_add_reservation(struct hstate
*h
,
2491 struct vm_area_struct
*vma
, unsigned long addr
)
2493 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2496 static long vma_del_reservation(struct hstate
*h
,
2497 struct vm_area_struct
*vma
, unsigned long addr
)
2499 return __vma_reservation_common(h
, vma
, addr
, VMA_DEL_RESV
);
2503 * This routine is called to restore reservation information on error paths.
2504 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2505 * the hugetlb mutex should remain held when calling this routine.
2507 * It handles two specific cases:
2508 * 1) A reservation was in place and the page consumed the reservation.
2509 * HPageRestoreReserve is set in the page.
2510 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2511 * not set. However, alloc_huge_page always updates the reserve map.
2513 * In case 1, free_huge_page later in the error path will increment the
2514 * global reserve count. But, free_huge_page does not have enough context
2515 * to adjust the reservation map. This case deals primarily with private
2516 * mappings. Adjust the reserve map here to be consistent with global
2517 * reserve count adjustments to be made by free_huge_page. Make sure the
2518 * reserve map indicates there is a reservation present.
2520 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2522 void restore_reserve_on_error(struct hstate
*h
, struct vm_area_struct
*vma
,
2523 unsigned long address
, struct page
*page
)
2525 long rc
= vma_needs_reservation(h
, vma
, address
);
2527 if (HPageRestoreReserve(page
)) {
2528 if (unlikely(rc
< 0))
2530 * Rare out of memory condition in reserve map
2531 * manipulation. Clear HPageRestoreReserve so that
2532 * global reserve count will not be incremented
2533 * by free_huge_page. This will make it appear
2534 * as though the reservation for this page was
2535 * consumed. This may prevent the task from
2536 * faulting in the page at a later time. This
2537 * is better than inconsistent global huge page
2538 * accounting of reserve counts.
2540 ClearHPageRestoreReserve(page
);
2542 (void)vma_add_reservation(h
, vma
, address
);
2544 vma_end_reservation(h
, vma
, address
);
2548 * This indicates there is an entry in the reserve map
2549 * not added by alloc_huge_page. We know it was added
2550 * before the alloc_huge_page call, otherwise
2551 * HPageRestoreReserve would be set on the page.
2552 * Remove the entry so that a subsequent allocation
2553 * does not consume a reservation.
2555 rc
= vma_del_reservation(h
, vma
, address
);
2558 * VERY rare out of memory condition. Since
2559 * we can not delete the entry, set
2560 * HPageRestoreReserve so that the reserve
2561 * count will be incremented when the page
2562 * is freed. This reserve will be consumed
2563 * on a subsequent allocation.
2565 SetHPageRestoreReserve(page
);
2566 } else if (rc
< 0) {
2568 * Rare out of memory condition from
2569 * vma_needs_reservation call. Memory allocation is
2570 * only attempted if a new entry is needed. Therefore,
2571 * this implies there is not an entry in the
2574 * For shared mappings, no entry in the map indicates
2575 * no reservation. We are done.
2577 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2579 * For private mappings, no entry indicates
2580 * a reservation is present. Since we can
2581 * not add an entry, set SetHPageRestoreReserve
2582 * on the page so reserve count will be
2583 * incremented when freed. This reserve will
2584 * be consumed on a subsequent allocation.
2586 SetHPageRestoreReserve(page
);
2589 * No reservation present, do nothing
2591 vma_end_reservation(h
, vma
, address
);
2596 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2597 * @h: struct hstate old page belongs to
2598 * @old_page: Old page to dissolve
2599 * @list: List to isolate the page in case we need to
2600 * Returns 0 on success, otherwise negated error.
2602 static int alloc_and_dissolve_huge_page(struct hstate
*h
, struct page
*old_page
,
2603 struct list_head
*list
)
2605 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
2606 int nid
= page_to_nid(old_page
);
2607 bool alloc_retry
= false;
2608 struct page
*new_page
;
2612 * Before dissolving the page, we need to allocate a new one for the
2613 * pool to remain stable. Here, we allocate the page and 'prep' it
2614 * by doing everything but actually updating counters and adding to
2615 * the pool. This simplifies and let us do most of the processing
2619 new_page
= alloc_buddy_huge_page(h
, gfp_mask
, nid
, NULL
, NULL
);
2623 * If all goes well, this page will be directly added to the free
2624 * list in the pool. For this the ref count needs to be zero.
2625 * Attempt to drop now, and retry once if needed. It is VERY
2626 * unlikely there is another ref on the page.
2628 * If someone else has a reference to the page, it will be freed
2629 * when they drop their ref. Abuse temporary page flag to accomplish
2630 * this. Retry once if there is an inflated ref count.
2632 SetHPageTemporary(new_page
);
2633 if (!put_page_testzero(new_page
)) {
2640 ClearHPageTemporary(new_page
);
2642 __prep_new_huge_page(h
, new_page
);
2645 spin_lock_irq(&hugetlb_lock
);
2646 if (!PageHuge(old_page
)) {
2648 * Freed from under us. Drop new_page too.
2651 } else if (page_count(old_page
)) {
2653 * Someone has grabbed the page, try to isolate it here.
2654 * Fail with -EBUSY if not possible.
2656 spin_unlock_irq(&hugetlb_lock
);
2657 if (!isolate_huge_page(old_page
, list
))
2659 spin_lock_irq(&hugetlb_lock
);
2661 } else if (!HPageFreed(old_page
)) {
2663 * Page's refcount is 0 but it has not been enqueued in the
2664 * freelist yet. Race window is small, so we can succeed here if
2667 spin_unlock_irq(&hugetlb_lock
);
2672 * Ok, old_page is still a genuine free hugepage. Remove it from
2673 * the freelist and decrease the counters. These will be
2674 * incremented again when calling __prep_account_new_huge_page()
2675 * and enqueue_huge_page() for new_page. The counters will remain
2676 * stable since this happens under the lock.
2678 remove_hugetlb_page(h
, old_page
, false);
2681 * Ref count on new page is already zero as it was dropped
2682 * earlier. It can be directly added to the pool free list.
2684 __prep_account_new_huge_page(h
, nid
);
2685 enqueue_huge_page(h
, new_page
);
2688 * Pages have been replaced, we can safely free the old one.
2690 spin_unlock_irq(&hugetlb_lock
);
2691 update_and_free_page(h
, old_page
, false);
2697 spin_unlock_irq(&hugetlb_lock
);
2698 /* Page has a zero ref count, but needs a ref to be freed */
2699 set_page_refcounted(new_page
);
2700 update_and_free_page(h
, new_page
, false);
2705 int isolate_or_dissolve_huge_page(struct page
*page
, struct list_head
*list
)
2712 * The page might have been dissolved from under our feet, so make sure
2713 * to carefully check the state under the lock.
2714 * Return success when racing as if we dissolved the page ourselves.
2716 spin_lock_irq(&hugetlb_lock
);
2717 if (PageHuge(page
)) {
2718 head
= compound_head(page
);
2719 h
= page_hstate(head
);
2721 spin_unlock_irq(&hugetlb_lock
);
2724 spin_unlock_irq(&hugetlb_lock
);
2727 * Fence off gigantic pages as there is a cyclic dependency between
2728 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2729 * of bailing out right away without further retrying.
2731 if (hstate_is_gigantic(h
))
2734 if (page_count(head
) && isolate_huge_page(head
, list
))
2736 else if (!page_count(head
))
2737 ret
= alloc_and_dissolve_huge_page(h
, head
, list
);
2742 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2743 unsigned long addr
, int avoid_reserve
)
2745 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2746 struct hstate
*h
= hstate_vma(vma
);
2748 long map_chg
, map_commit
;
2751 struct hugetlb_cgroup
*h_cg
;
2752 bool deferred_reserve
;
2754 idx
= hstate_index(h
);
2756 * Examine the region/reserve map to determine if the process
2757 * has a reservation for the page to be allocated. A return
2758 * code of zero indicates a reservation exists (no change).
2760 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2762 return ERR_PTR(-ENOMEM
);
2765 * Processes that did not create the mapping will have no
2766 * reserves as indicated by the region/reserve map. Check
2767 * that the allocation will not exceed the subpool limit.
2768 * Allocations for MAP_NORESERVE mappings also need to be
2769 * checked against any subpool limit.
2771 if (map_chg
|| avoid_reserve
) {
2772 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2774 vma_end_reservation(h
, vma
, addr
);
2775 return ERR_PTR(-ENOSPC
);
2779 * Even though there was no reservation in the region/reserve
2780 * map, there could be reservations associated with the
2781 * subpool that can be used. This would be indicated if the
2782 * return value of hugepage_subpool_get_pages() is zero.
2783 * However, if avoid_reserve is specified we still avoid even
2784 * the subpool reservations.
2790 /* If this allocation is not consuming a reservation, charge it now.
2792 deferred_reserve
= map_chg
|| avoid_reserve
;
2793 if (deferred_reserve
) {
2794 ret
= hugetlb_cgroup_charge_cgroup_rsvd(
2795 idx
, pages_per_huge_page(h
), &h_cg
);
2797 goto out_subpool_put
;
2800 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2802 goto out_uncharge_cgroup_reservation
;
2804 spin_lock_irq(&hugetlb_lock
);
2806 * glb_chg is passed to indicate whether or not a page must be taken
2807 * from the global free pool (global change). gbl_chg == 0 indicates
2808 * a reservation exists for the allocation.
2810 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2812 spin_unlock_irq(&hugetlb_lock
);
2813 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2815 goto out_uncharge_cgroup
;
2816 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2817 SetHPageRestoreReserve(page
);
2818 h
->resv_huge_pages
--;
2820 spin_lock_irq(&hugetlb_lock
);
2821 list_add(&page
->lru
, &h
->hugepage_activelist
);
2824 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2825 /* If allocation is not consuming a reservation, also store the
2826 * hugetlb_cgroup pointer on the page.
2828 if (deferred_reserve
) {
2829 hugetlb_cgroup_commit_charge_rsvd(idx
, pages_per_huge_page(h
),
2833 spin_unlock_irq(&hugetlb_lock
);
2835 hugetlb_set_page_subpool(page
, spool
);
2837 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2838 if (unlikely(map_chg
> map_commit
)) {
2840 * The page was added to the reservation map between
2841 * vma_needs_reservation and vma_commit_reservation.
2842 * This indicates a race with hugetlb_reserve_pages.
2843 * Adjust for the subpool count incremented above AND
2844 * in hugetlb_reserve_pages for the same page. Also,
2845 * the reservation count added in hugetlb_reserve_pages
2846 * no longer applies.
2850 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2851 hugetlb_acct_memory(h
, -rsv_adjust
);
2852 if (deferred_reserve
)
2853 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h
),
2854 pages_per_huge_page(h
), page
);
2858 out_uncharge_cgroup
:
2859 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2860 out_uncharge_cgroup_reservation
:
2861 if (deferred_reserve
)
2862 hugetlb_cgroup_uncharge_cgroup_rsvd(idx
, pages_per_huge_page(h
),
2865 if (map_chg
|| avoid_reserve
)
2866 hugepage_subpool_put_pages(spool
, 1);
2867 vma_end_reservation(h
, vma
, addr
);
2868 return ERR_PTR(-ENOSPC
);
2871 int alloc_bootmem_huge_page(struct hstate
*h
)
2872 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2873 int __alloc_bootmem_huge_page(struct hstate
*h
)
2875 struct huge_bootmem_page
*m
;
2878 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2881 addr
= memblock_alloc_try_nid_raw(
2882 huge_page_size(h
), huge_page_size(h
),
2883 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2886 * Use the beginning of the huge page to store the
2887 * huge_bootmem_page struct (until gather_bootmem
2888 * puts them into the mem_map).
2897 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2898 /* Put them into a private list first because mem_map is not up yet */
2899 INIT_LIST_HEAD(&m
->list
);
2900 list_add(&m
->list
, &huge_boot_pages
);
2906 * Put bootmem huge pages into the standard lists after mem_map is up.
2907 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2909 static void __init
gather_bootmem_prealloc(void)
2911 struct huge_bootmem_page
*m
;
2913 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2914 struct page
*page
= virt_to_page(m
);
2915 struct hstate
*h
= m
->hstate
;
2917 VM_BUG_ON(!hstate_is_gigantic(h
));
2918 WARN_ON(page_count(page
) != 1);
2919 if (prep_compound_gigantic_page(page
, huge_page_order(h
))) {
2920 WARN_ON(PageReserved(page
));
2921 prep_new_huge_page(h
, page
, page_to_nid(page
));
2922 put_page(page
); /* add to the hugepage allocator */
2924 /* VERY unlikely inflated ref count on a tail page */
2925 free_gigantic_page(page
, huge_page_order(h
));
2929 * We need to restore the 'stolen' pages to totalram_pages
2930 * in order to fix confusing memory reports from free(1) and
2931 * other side-effects, like CommitLimit going negative.
2933 adjust_managed_page_count(page
, pages_per_huge_page(h
));
2938 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2941 nodemask_t
*node_alloc_noretry
;
2943 if (!hstate_is_gigantic(h
)) {
2945 * Bit mask controlling how hard we retry per-node allocations.
2946 * Ignore errors as lower level routines can deal with
2947 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2948 * time, we are likely in bigger trouble.
2950 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2953 /* allocations done at boot time */
2954 node_alloc_noretry
= NULL
;
2957 /* bit mask controlling how hard we retry per-node allocations */
2958 if (node_alloc_noretry
)
2959 nodes_clear(*node_alloc_noretry
);
2961 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2962 if (hstate_is_gigantic(h
)) {
2963 if (hugetlb_cma_size
) {
2964 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2967 if (!alloc_bootmem_huge_page(h
))
2969 } else if (!alloc_pool_huge_page(h
,
2970 &node_states
[N_MEMORY
],
2971 node_alloc_noretry
))
2975 if (i
< h
->max_huge_pages
) {
2978 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2979 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2980 h
->max_huge_pages
, buf
, i
);
2981 h
->max_huge_pages
= i
;
2984 kfree(node_alloc_noretry
);
2987 static void __init
hugetlb_init_hstates(void)
2991 for_each_hstate(h
) {
2992 if (minimum_order
> huge_page_order(h
))
2993 minimum_order
= huge_page_order(h
);
2995 /* oversize hugepages were init'ed in early boot */
2996 if (!hstate_is_gigantic(h
))
2997 hugetlb_hstate_alloc_pages(h
);
2999 VM_BUG_ON(minimum_order
== UINT_MAX
);
3002 static void __init
report_hugepages(void)
3006 for_each_hstate(h
) {
3009 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
3010 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3011 buf
, h
->free_huge_pages
);
3015 #ifdef CONFIG_HIGHMEM
3016 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
3017 nodemask_t
*nodes_allowed
)
3020 LIST_HEAD(page_list
);
3022 lockdep_assert_held(&hugetlb_lock
);
3023 if (hstate_is_gigantic(h
))
3027 * Collect pages to be freed on a list, and free after dropping lock
3029 for_each_node_mask(i
, *nodes_allowed
) {
3030 struct page
*page
, *next
;
3031 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
3032 list_for_each_entry_safe(page
, next
, freel
, lru
) {
3033 if (count
>= h
->nr_huge_pages
)
3035 if (PageHighMem(page
))
3037 remove_hugetlb_page(h
, page
, false);
3038 list_add(&page
->lru
, &page_list
);
3043 spin_unlock_irq(&hugetlb_lock
);
3044 update_and_free_pages_bulk(h
, &page_list
);
3045 spin_lock_irq(&hugetlb_lock
);
3048 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
3049 nodemask_t
*nodes_allowed
)
3055 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3056 * balanced by operating on them in a round-robin fashion.
3057 * Returns 1 if an adjustment was made.
3059 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
3064 lockdep_assert_held(&hugetlb_lock
);
3065 VM_BUG_ON(delta
!= -1 && delta
!= 1);
3068 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
3069 if (h
->surplus_huge_pages_node
[node
])
3073 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
3074 if (h
->surplus_huge_pages_node
[node
] <
3075 h
->nr_huge_pages_node
[node
])
3082 h
->surplus_huge_pages
+= delta
;
3083 h
->surplus_huge_pages_node
[node
] += delta
;
3087 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3088 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
3089 nodemask_t
*nodes_allowed
)
3091 unsigned long min_count
, ret
;
3093 LIST_HEAD(page_list
);
3094 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
3097 * Bit mask controlling how hard we retry per-node allocations.
3098 * If we can not allocate the bit mask, do not attempt to allocate
3099 * the requested huge pages.
3101 if (node_alloc_noretry
)
3102 nodes_clear(*node_alloc_noretry
);
3107 * resize_lock mutex prevents concurrent adjustments to number of
3108 * pages in hstate via the proc/sysfs interfaces.
3110 mutex_lock(&h
->resize_lock
);
3111 flush_free_hpage_work(h
);
3112 spin_lock_irq(&hugetlb_lock
);
3115 * Check for a node specific request.
3116 * Changing node specific huge page count may require a corresponding
3117 * change to the global count. In any case, the passed node mask
3118 * (nodes_allowed) will restrict alloc/free to the specified node.
3120 if (nid
!= NUMA_NO_NODE
) {
3121 unsigned long old_count
= count
;
3123 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
3125 * User may have specified a large count value which caused the
3126 * above calculation to overflow. In this case, they wanted
3127 * to allocate as many huge pages as possible. Set count to
3128 * largest possible value to align with their intention.
3130 if (count
< old_count
)
3135 * Gigantic pages runtime allocation depend on the capability for large
3136 * page range allocation.
3137 * If the system does not provide this feature, return an error when
3138 * the user tries to allocate gigantic pages but let the user free the
3139 * boottime allocated gigantic pages.
3141 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
3142 if (count
> persistent_huge_pages(h
)) {
3143 spin_unlock_irq(&hugetlb_lock
);
3144 mutex_unlock(&h
->resize_lock
);
3145 NODEMASK_FREE(node_alloc_noretry
);
3148 /* Fall through to decrease pool */
3152 * Increase the pool size
3153 * First take pages out of surplus state. Then make up the
3154 * remaining difference by allocating fresh huge pages.
3156 * We might race with alloc_surplus_huge_page() here and be unable
3157 * to convert a surplus huge page to a normal huge page. That is
3158 * not critical, though, it just means the overall size of the
3159 * pool might be one hugepage larger than it needs to be, but
3160 * within all the constraints specified by the sysctls.
3162 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
3163 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
3167 while (count
> persistent_huge_pages(h
)) {
3169 * If this allocation races such that we no longer need the
3170 * page, free_huge_page will handle it by freeing the page
3171 * and reducing the surplus.
3173 spin_unlock_irq(&hugetlb_lock
);
3175 /* yield cpu to avoid soft lockup */
3178 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
3179 node_alloc_noretry
);
3180 spin_lock_irq(&hugetlb_lock
);
3184 /* Bail for signals. Probably ctrl-c from user */
3185 if (signal_pending(current
))
3190 * Decrease the pool size
3191 * First return free pages to the buddy allocator (being careful
3192 * to keep enough around to satisfy reservations). Then place
3193 * pages into surplus state as needed so the pool will shrink
3194 * to the desired size as pages become free.
3196 * By placing pages into the surplus state independent of the
3197 * overcommit value, we are allowing the surplus pool size to
3198 * exceed overcommit. There are few sane options here. Since
3199 * alloc_surplus_huge_page() is checking the global counter,
3200 * though, we'll note that we're not allowed to exceed surplus
3201 * and won't grow the pool anywhere else. Not until one of the
3202 * sysctls are changed, or the surplus pages go out of use.
3204 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
3205 min_count
= max(count
, min_count
);
3206 try_to_free_low(h
, min_count
, nodes_allowed
);
3209 * Collect pages to be removed on list without dropping lock
3211 while (min_count
< persistent_huge_pages(h
)) {
3212 page
= remove_pool_huge_page(h
, nodes_allowed
, 0);
3216 list_add(&page
->lru
, &page_list
);
3218 /* free the pages after dropping lock */
3219 spin_unlock_irq(&hugetlb_lock
);
3220 update_and_free_pages_bulk(h
, &page_list
);
3221 flush_free_hpage_work(h
);
3222 spin_lock_irq(&hugetlb_lock
);
3224 while (count
< persistent_huge_pages(h
)) {
3225 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
3229 h
->max_huge_pages
= persistent_huge_pages(h
);
3230 spin_unlock_irq(&hugetlb_lock
);
3231 mutex_unlock(&h
->resize_lock
);
3233 NODEMASK_FREE(node_alloc_noretry
);
3238 #define HSTATE_ATTR_RO(_name) \
3239 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3241 #define HSTATE_ATTR(_name) \
3242 static struct kobj_attribute _name##_attr = \
3243 __ATTR(_name, 0644, _name##_show, _name##_store)
3245 static struct kobject
*hugepages_kobj
;
3246 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
3248 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
3250 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
3254 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
3255 if (hstate_kobjs
[i
] == kobj
) {
3257 *nidp
= NUMA_NO_NODE
;
3261 return kobj_to_node_hstate(kobj
, nidp
);
3264 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
3265 struct kobj_attribute
*attr
, char *buf
)
3268 unsigned long nr_huge_pages
;
3271 h
= kobj_to_hstate(kobj
, &nid
);
3272 if (nid
== NUMA_NO_NODE
)
3273 nr_huge_pages
= h
->nr_huge_pages
;
3275 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
3277 return sysfs_emit(buf
, "%lu\n", nr_huge_pages
);
3280 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
3281 struct hstate
*h
, int nid
,
3282 unsigned long count
, size_t len
)
3285 nodemask_t nodes_allowed
, *n_mask
;
3287 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
3290 if (nid
== NUMA_NO_NODE
) {
3292 * global hstate attribute
3294 if (!(obey_mempolicy
&&
3295 init_nodemask_of_mempolicy(&nodes_allowed
)))
3296 n_mask
= &node_states
[N_MEMORY
];
3298 n_mask
= &nodes_allowed
;
3301 * Node specific request. count adjustment happens in
3302 * set_max_huge_pages() after acquiring hugetlb_lock.
3304 init_nodemask_of_node(&nodes_allowed
, nid
);
3305 n_mask
= &nodes_allowed
;
3308 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
3310 return err
? err
: len
;
3313 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
3314 struct kobject
*kobj
, const char *buf
,
3318 unsigned long count
;
3322 err
= kstrtoul(buf
, 10, &count
);
3326 h
= kobj_to_hstate(kobj
, &nid
);
3327 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
3330 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
3331 struct kobj_attribute
*attr
, char *buf
)
3333 return nr_hugepages_show_common(kobj
, attr
, buf
);
3336 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
3337 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
3339 return nr_hugepages_store_common(false, kobj
, buf
, len
);
3341 HSTATE_ATTR(nr_hugepages
);
3346 * hstate attribute for optionally mempolicy-based constraint on persistent
3347 * huge page alloc/free.
3349 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
3350 struct kobj_attribute
*attr
,
3353 return nr_hugepages_show_common(kobj
, attr
, buf
);
3356 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
3357 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
3359 return nr_hugepages_store_common(true, kobj
, buf
, len
);
3361 HSTATE_ATTR(nr_hugepages_mempolicy
);
3365 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
3366 struct kobj_attribute
*attr
, char *buf
)
3368 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
3369 return sysfs_emit(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
3372 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
3373 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
3376 unsigned long input
;
3377 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
3379 if (hstate_is_gigantic(h
))
3382 err
= kstrtoul(buf
, 10, &input
);
3386 spin_lock_irq(&hugetlb_lock
);
3387 h
->nr_overcommit_huge_pages
= input
;
3388 spin_unlock_irq(&hugetlb_lock
);
3392 HSTATE_ATTR(nr_overcommit_hugepages
);
3394 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
3395 struct kobj_attribute
*attr
, char *buf
)
3398 unsigned long free_huge_pages
;
3401 h
= kobj_to_hstate(kobj
, &nid
);
3402 if (nid
== NUMA_NO_NODE
)
3403 free_huge_pages
= h
->free_huge_pages
;
3405 free_huge_pages
= h
->free_huge_pages_node
[nid
];
3407 return sysfs_emit(buf
, "%lu\n", free_huge_pages
);
3409 HSTATE_ATTR_RO(free_hugepages
);
3411 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
3412 struct kobj_attribute
*attr
, char *buf
)
3414 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
3415 return sysfs_emit(buf
, "%lu\n", h
->resv_huge_pages
);
3417 HSTATE_ATTR_RO(resv_hugepages
);
3419 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
3420 struct kobj_attribute
*attr
, char *buf
)
3423 unsigned long surplus_huge_pages
;
3426 h
= kobj_to_hstate(kobj
, &nid
);
3427 if (nid
== NUMA_NO_NODE
)
3428 surplus_huge_pages
= h
->surplus_huge_pages
;
3430 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
3432 return sysfs_emit(buf
, "%lu\n", surplus_huge_pages
);
3434 HSTATE_ATTR_RO(surplus_hugepages
);
3436 static struct attribute
*hstate_attrs
[] = {
3437 &nr_hugepages_attr
.attr
,
3438 &nr_overcommit_hugepages_attr
.attr
,
3439 &free_hugepages_attr
.attr
,
3440 &resv_hugepages_attr
.attr
,
3441 &surplus_hugepages_attr
.attr
,
3443 &nr_hugepages_mempolicy_attr
.attr
,
3448 static const struct attribute_group hstate_attr_group
= {
3449 .attrs
= hstate_attrs
,
3452 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
3453 struct kobject
**hstate_kobjs
,
3454 const struct attribute_group
*hstate_attr_group
)
3457 int hi
= hstate_index(h
);
3459 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
3460 if (!hstate_kobjs
[hi
])
3463 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
3465 kobject_put(hstate_kobjs
[hi
]);
3466 hstate_kobjs
[hi
] = NULL
;
3472 static void __init
hugetlb_sysfs_init(void)
3477 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
3478 if (!hugepages_kobj
)
3481 for_each_hstate(h
) {
3482 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
3483 hstate_kobjs
, &hstate_attr_group
);
3485 pr_err("HugeTLB: Unable to add hstate %s", h
->name
);
3492 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3493 * with node devices in node_devices[] using a parallel array. The array
3494 * index of a node device or _hstate == node id.
3495 * This is here to avoid any static dependency of the node device driver, in
3496 * the base kernel, on the hugetlb module.
3498 struct node_hstate
{
3499 struct kobject
*hugepages_kobj
;
3500 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
3502 static struct node_hstate node_hstates
[MAX_NUMNODES
];
3505 * A subset of global hstate attributes for node devices
3507 static struct attribute
*per_node_hstate_attrs
[] = {
3508 &nr_hugepages_attr
.attr
,
3509 &free_hugepages_attr
.attr
,
3510 &surplus_hugepages_attr
.attr
,
3514 static const struct attribute_group per_node_hstate_attr_group
= {
3515 .attrs
= per_node_hstate_attrs
,
3519 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3520 * Returns node id via non-NULL nidp.
3522 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3526 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
3527 struct node_hstate
*nhs
= &node_hstates
[nid
];
3529 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
3530 if (nhs
->hstate_kobjs
[i
] == kobj
) {
3542 * Unregister hstate attributes from a single node device.
3543 * No-op if no hstate attributes attached.
3545 static void hugetlb_unregister_node(struct node
*node
)
3548 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3550 if (!nhs
->hugepages_kobj
)
3551 return; /* no hstate attributes */
3553 for_each_hstate(h
) {
3554 int idx
= hstate_index(h
);
3555 if (nhs
->hstate_kobjs
[idx
]) {
3556 kobject_put(nhs
->hstate_kobjs
[idx
]);
3557 nhs
->hstate_kobjs
[idx
] = NULL
;
3561 kobject_put(nhs
->hugepages_kobj
);
3562 nhs
->hugepages_kobj
= NULL
;
3567 * Register hstate attributes for a single node device.
3568 * No-op if attributes already registered.
3570 static void hugetlb_register_node(struct node
*node
)
3573 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
3576 if (nhs
->hugepages_kobj
)
3577 return; /* already allocated */
3579 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
3581 if (!nhs
->hugepages_kobj
)
3584 for_each_hstate(h
) {
3585 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
3587 &per_node_hstate_attr_group
);
3589 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3590 h
->name
, node
->dev
.id
);
3591 hugetlb_unregister_node(node
);
3598 * hugetlb init time: register hstate attributes for all registered node
3599 * devices of nodes that have memory. All on-line nodes should have
3600 * registered their associated device by this time.
3602 static void __init
hugetlb_register_all_nodes(void)
3606 for_each_node_state(nid
, N_MEMORY
) {
3607 struct node
*node
= node_devices
[nid
];
3608 if (node
->dev
.id
== nid
)
3609 hugetlb_register_node(node
);
3613 * Let the node device driver know we're here so it can
3614 * [un]register hstate attributes on node hotplug.
3616 register_hugetlbfs_with_node(hugetlb_register_node
,
3617 hugetlb_unregister_node
);
3619 #else /* !CONFIG_NUMA */
3621 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
3629 static void hugetlb_register_all_nodes(void) { }
3633 static int __init
hugetlb_init(void)
3637 BUILD_BUG_ON(sizeof_field(struct page
, private) * BITS_PER_BYTE
<
3640 if (!hugepages_supported()) {
3641 if (hugetlb_max_hstate
|| default_hstate_max_huge_pages
)
3642 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3647 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3648 * architectures depend on setup being done here.
3650 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
3651 if (!parsed_default_hugepagesz
) {
3653 * If we did not parse a default huge page size, set
3654 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3655 * number of huge pages for this default size was implicitly
3656 * specified, set that here as well.
3657 * Note that the implicit setting will overwrite an explicit
3658 * setting. A warning will be printed in this case.
3660 default_hstate_idx
= hstate_index(size_to_hstate(HPAGE_SIZE
));
3661 if (default_hstate_max_huge_pages
) {
3662 if (default_hstate
.max_huge_pages
) {
3665 string_get_size(huge_page_size(&default_hstate
),
3666 1, STRING_UNITS_2
, buf
, 32);
3667 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3668 default_hstate
.max_huge_pages
, buf
);
3669 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3670 default_hstate_max_huge_pages
);
3672 default_hstate
.max_huge_pages
=
3673 default_hstate_max_huge_pages
;
3677 hugetlb_cma_check();
3678 hugetlb_init_hstates();
3679 gather_bootmem_prealloc();
3682 hugetlb_sysfs_init();
3683 hugetlb_register_all_nodes();
3684 hugetlb_cgroup_file_init();
3687 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
3689 num_fault_mutexes
= 1;
3691 hugetlb_fault_mutex_table
=
3692 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
3694 BUG_ON(!hugetlb_fault_mutex_table
);
3696 for (i
= 0; i
< num_fault_mutexes
; i
++)
3697 mutex_init(&hugetlb_fault_mutex_table
[i
]);
3700 subsys_initcall(hugetlb_init
);
3702 /* Overwritten by architectures with more huge page sizes */
3703 bool __init
__attribute((weak
)) arch_hugetlb_valid_size(unsigned long size
)
3705 return size
== HPAGE_SIZE
;
3708 void __init
hugetlb_add_hstate(unsigned int order
)
3713 if (size_to_hstate(PAGE_SIZE
<< order
)) {
3716 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
3718 h
= &hstates
[hugetlb_max_hstate
++];
3719 mutex_init(&h
->resize_lock
);
3721 h
->mask
= ~(huge_page_size(h
) - 1);
3722 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
3723 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
3724 INIT_LIST_HEAD(&h
->hugepage_activelist
);
3725 h
->next_nid_to_alloc
= first_memory_node
;
3726 h
->next_nid_to_free
= first_memory_node
;
3727 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
3728 huge_page_size(h
)/1024);
3729 hugetlb_vmemmap_init(h
);
3735 * hugepages command line processing
3736 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3737 * specification. If not, ignore the hugepages value. hugepages can also
3738 * be the first huge page command line option in which case it implicitly
3739 * specifies the number of huge pages for the default size.
3741 static int __init
hugepages_setup(char *s
)
3744 static unsigned long *last_mhp
;
3746 if (!parsed_valid_hugepagesz
) {
3747 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s
);
3748 parsed_valid_hugepagesz
= true;
3753 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3754 * yet, so this hugepages= parameter goes to the "default hstate".
3755 * Otherwise, it goes with the previously parsed hugepagesz or
3756 * default_hugepagesz.
3758 else if (!hugetlb_max_hstate
)
3759 mhp
= &default_hstate_max_huge_pages
;
3761 mhp
= &parsed_hstate
->max_huge_pages
;
3763 if (mhp
== last_mhp
) {
3764 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s
);
3768 if (sscanf(s
, "%lu", mhp
) <= 0)
3772 * Global state is always initialized later in hugetlb_init.
3773 * But we need to allocate gigantic hstates here early to still
3774 * use the bootmem allocator.
3776 if (hugetlb_max_hstate
&& hstate_is_gigantic(parsed_hstate
))
3777 hugetlb_hstate_alloc_pages(parsed_hstate
);
3783 __setup("hugepages=", hugepages_setup
);
3786 * hugepagesz command line processing
3787 * A specific huge page size can only be specified once with hugepagesz.
3788 * hugepagesz is followed by hugepages on the command line. The global
3789 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3790 * hugepagesz argument was valid.
3792 static int __init
hugepagesz_setup(char *s
)
3797 parsed_valid_hugepagesz
= false;
3798 size
= (unsigned long)memparse(s
, NULL
);
3800 if (!arch_hugetlb_valid_size(size
)) {
3801 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s
);
3805 h
= size_to_hstate(size
);
3808 * hstate for this size already exists. This is normally
3809 * an error, but is allowed if the existing hstate is the
3810 * default hstate. More specifically, it is only allowed if
3811 * the number of huge pages for the default hstate was not
3812 * previously specified.
3814 if (!parsed_default_hugepagesz
|| h
!= &default_hstate
||
3815 default_hstate
.max_huge_pages
) {
3816 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s
);
3821 * No need to call hugetlb_add_hstate() as hstate already
3822 * exists. But, do set parsed_hstate so that a following
3823 * hugepages= parameter will be applied to this hstate.
3826 parsed_valid_hugepagesz
= true;
3830 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3831 parsed_valid_hugepagesz
= true;
3834 __setup("hugepagesz=", hugepagesz_setup
);
3837 * default_hugepagesz command line input
3838 * Only one instance of default_hugepagesz allowed on command line.
3840 static int __init
default_hugepagesz_setup(char *s
)
3844 parsed_valid_hugepagesz
= false;
3845 if (parsed_default_hugepagesz
) {
3846 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s
);
3850 size
= (unsigned long)memparse(s
, NULL
);
3852 if (!arch_hugetlb_valid_size(size
)) {
3853 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s
);
3857 hugetlb_add_hstate(ilog2(size
) - PAGE_SHIFT
);
3858 parsed_valid_hugepagesz
= true;
3859 parsed_default_hugepagesz
= true;
3860 default_hstate_idx
= hstate_index(size_to_hstate(size
));
3863 * The number of default huge pages (for this size) could have been
3864 * specified as the first hugetlb parameter: hugepages=X. If so,
3865 * then default_hstate_max_huge_pages is set. If the default huge
3866 * page size is gigantic (>= MAX_ORDER), then the pages must be
3867 * allocated here from bootmem allocator.
3869 if (default_hstate_max_huge_pages
) {
3870 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
3871 if (hstate_is_gigantic(&default_hstate
))
3872 hugetlb_hstate_alloc_pages(&default_hstate
);
3873 default_hstate_max_huge_pages
= 0;
3878 __setup("default_hugepagesz=", default_hugepagesz_setup
);
3880 static unsigned int allowed_mems_nr(struct hstate
*h
)
3883 unsigned int nr
= 0;
3884 nodemask_t
*mpol_allowed
;
3885 unsigned int *array
= h
->free_huge_pages_node
;
3886 gfp_t gfp_mask
= htlb_alloc_mask(h
);
3888 mpol_allowed
= policy_nodemask_current(gfp_mask
);
3890 for_each_node_mask(node
, cpuset_current_mems_allowed
) {
3891 if (!mpol_allowed
|| node_isset(node
, *mpol_allowed
))
3898 #ifdef CONFIG_SYSCTL
3899 static int proc_hugetlb_doulongvec_minmax(struct ctl_table
*table
, int write
,
3900 void *buffer
, size_t *length
,
3901 loff_t
*ppos
, unsigned long *out
)
3903 struct ctl_table dup_table
;
3906 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3907 * can duplicate the @table and alter the duplicate of it.
3910 dup_table
.data
= out
;
3912 return proc_doulongvec_minmax(&dup_table
, write
, buffer
, length
, ppos
);
3915 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3916 struct ctl_table
*table
, int write
,
3917 void *buffer
, size_t *length
, loff_t
*ppos
)
3919 struct hstate
*h
= &default_hstate
;
3920 unsigned long tmp
= h
->max_huge_pages
;
3923 if (!hugepages_supported())
3926 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3932 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3933 NUMA_NO_NODE
, tmp
, *length
);
3938 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3939 void *buffer
, size_t *length
, loff_t
*ppos
)
3942 return hugetlb_sysctl_handler_common(false, table
, write
,
3943 buffer
, length
, ppos
);
3947 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3948 void *buffer
, size_t *length
, loff_t
*ppos
)
3950 return hugetlb_sysctl_handler_common(true, table
, write
,
3951 buffer
, length
, ppos
);
3953 #endif /* CONFIG_NUMA */
3955 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3956 void *buffer
, size_t *length
, loff_t
*ppos
)
3958 struct hstate
*h
= &default_hstate
;
3962 if (!hugepages_supported())
3965 tmp
= h
->nr_overcommit_huge_pages
;
3967 if (write
&& hstate_is_gigantic(h
))
3970 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3976 spin_lock_irq(&hugetlb_lock
);
3977 h
->nr_overcommit_huge_pages
= tmp
;
3978 spin_unlock_irq(&hugetlb_lock
);
3984 #endif /* CONFIG_SYSCTL */
3986 void hugetlb_report_meminfo(struct seq_file
*m
)
3989 unsigned long total
= 0;
3991 if (!hugepages_supported())
3994 for_each_hstate(h
) {
3995 unsigned long count
= h
->nr_huge_pages
;
3997 total
+= huge_page_size(h
) * count
;
3999 if (h
== &default_hstate
)
4001 "HugePages_Total: %5lu\n"
4002 "HugePages_Free: %5lu\n"
4003 "HugePages_Rsvd: %5lu\n"
4004 "HugePages_Surp: %5lu\n"
4005 "Hugepagesize: %8lu kB\n",
4009 h
->surplus_huge_pages
,
4010 huge_page_size(h
) / SZ_1K
);
4013 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ SZ_1K
);
4016 int hugetlb_report_node_meminfo(char *buf
, int len
, int nid
)
4018 struct hstate
*h
= &default_hstate
;
4020 if (!hugepages_supported())
4023 return sysfs_emit_at(buf
, len
,
4024 "Node %d HugePages_Total: %5u\n"
4025 "Node %d HugePages_Free: %5u\n"
4026 "Node %d HugePages_Surp: %5u\n",
4027 nid
, h
->nr_huge_pages_node
[nid
],
4028 nid
, h
->free_huge_pages_node
[nid
],
4029 nid
, h
->surplus_huge_pages_node
[nid
]);
4032 void hugetlb_show_meminfo(void)
4037 if (!hugepages_supported())
4040 for_each_node_state(nid
, N_MEMORY
)
4042 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4044 h
->nr_huge_pages_node
[nid
],
4045 h
->free_huge_pages_node
[nid
],
4046 h
->surplus_huge_pages_node
[nid
],
4047 huge_page_size(h
) / SZ_1K
);
4050 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
4052 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
4053 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
4056 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4057 unsigned long hugetlb_total_pages(void)
4060 unsigned long nr_total_pages
= 0;
4063 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
4064 return nr_total_pages
;
4067 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
4074 spin_lock_irq(&hugetlb_lock
);
4076 * When cpuset is configured, it breaks the strict hugetlb page
4077 * reservation as the accounting is done on a global variable. Such
4078 * reservation is completely rubbish in the presence of cpuset because
4079 * the reservation is not checked against page availability for the
4080 * current cpuset. Application can still potentially OOM'ed by kernel
4081 * with lack of free htlb page in cpuset that the task is in.
4082 * Attempt to enforce strict accounting with cpuset is almost
4083 * impossible (or too ugly) because cpuset is too fluid that
4084 * task or memory node can be dynamically moved between cpusets.
4086 * The change of semantics for shared hugetlb mapping with cpuset is
4087 * undesirable. However, in order to preserve some of the semantics,
4088 * we fall back to check against current free page availability as
4089 * a best attempt and hopefully to minimize the impact of changing
4090 * semantics that cpuset has.
4092 * Apart from cpuset, we also have memory policy mechanism that
4093 * also determines from which node the kernel will allocate memory
4094 * in a NUMA system. So similar to cpuset, we also should consider
4095 * the memory policy of the current task. Similar to the description
4099 if (gather_surplus_pages(h
, delta
) < 0)
4102 if (delta
> allowed_mems_nr(h
)) {
4103 return_unused_surplus_pages(h
, delta
);
4110 return_unused_surplus_pages(h
, (unsigned long) -delta
);
4113 spin_unlock_irq(&hugetlb_lock
);
4117 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
4119 struct resv_map
*resv
= vma_resv_map(vma
);
4122 * This new VMA should share its siblings reservation map if present.
4123 * The VMA will only ever have a valid reservation map pointer where
4124 * it is being copied for another still existing VMA. As that VMA
4125 * has a reference to the reservation map it cannot disappear until
4126 * after this open call completes. It is therefore safe to take a
4127 * new reference here without additional locking.
4129 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
4130 resv_map_dup_hugetlb_cgroup_uncharge_info(resv
);
4131 kref_get(&resv
->refs
);
4135 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
4137 struct hstate
*h
= hstate_vma(vma
);
4138 struct resv_map
*resv
= vma_resv_map(vma
);
4139 struct hugepage_subpool
*spool
= subpool_vma(vma
);
4140 unsigned long reserve
, start
, end
;
4143 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4146 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
4147 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
4149 reserve
= (end
- start
) - region_count(resv
, start
, end
);
4150 hugetlb_cgroup_uncharge_counter(resv
, start
, end
);
4153 * Decrement reserve counts. The global reserve count may be
4154 * adjusted if the subpool has a minimum size.
4156 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
4157 hugetlb_acct_memory(h
, -gbl_reserve
);
4160 kref_put(&resv
->refs
, resv_map_release
);
4163 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
4165 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
4170 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
4172 return huge_page_size(hstate_vma(vma
));
4176 * We cannot handle pagefaults against hugetlb pages at all. They cause
4177 * handle_mm_fault() to try to instantiate regular-sized pages in the
4178 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4181 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
4188 * When a new function is introduced to vm_operations_struct and added
4189 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4190 * This is because under System V memory model, mappings created via
4191 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4192 * their original vm_ops are overwritten with shm_vm_ops.
4194 const struct vm_operations_struct hugetlb_vm_ops
= {
4195 .fault
= hugetlb_vm_op_fault
,
4196 .open
= hugetlb_vm_op_open
,
4197 .close
= hugetlb_vm_op_close
,
4198 .may_split
= hugetlb_vm_op_split
,
4199 .pagesize
= hugetlb_vm_op_pagesize
,
4202 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
4206 unsigned int shift
= huge_page_shift(hstate_vma(vma
));
4209 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
4210 vma
->vm_page_prot
)));
4212 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
4213 vma
->vm_page_prot
));
4215 entry
= pte_mkyoung(entry
);
4216 entry
= pte_mkhuge(entry
);
4217 entry
= arch_make_huge_pte(entry
, shift
, vma
->vm_flags
);
4222 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
4223 unsigned long address
, pte_t
*ptep
)
4227 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
4228 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
4229 update_mmu_cache(vma
, address
, ptep
);
4232 bool is_hugetlb_entry_migration(pte_t pte
)
4236 if (huge_pte_none(pte
) || pte_present(pte
))
4238 swp
= pte_to_swp_entry(pte
);
4239 if (is_migration_entry(swp
))
4245 static bool is_hugetlb_entry_hwpoisoned(pte_t pte
)
4249 if (huge_pte_none(pte
) || pte_present(pte
))
4251 swp
= pte_to_swp_entry(pte
);
4252 if (is_hwpoison_entry(swp
))
4259 hugetlb_install_page(struct vm_area_struct
*vma
, pte_t
*ptep
, unsigned long addr
,
4260 struct page
*new_page
)
4262 __SetPageUptodate(new_page
);
4263 set_huge_pte_at(vma
->vm_mm
, addr
, ptep
, make_huge_pte(vma
, new_page
, 1));
4264 hugepage_add_new_anon_rmap(new_page
, vma
, addr
);
4265 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma
)), vma
->vm_mm
);
4266 ClearHPageRestoreReserve(new_page
);
4267 SetHPageMigratable(new_page
);
4270 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
4271 struct vm_area_struct
*vma
)
4273 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
4274 struct page
*ptepage
;
4276 bool cow
= is_cow_mapping(vma
->vm_flags
);
4277 struct hstate
*h
= hstate_vma(vma
);
4278 unsigned long sz
= huge_page_size(h
);
4279 unsigned long npages
= pages_per_huge_page(h
);
4280 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4281 struct mmu_notifier_range range
;
4285 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
4288 mmu_notifier_invalidate_range_start(&range
);
4291 * For shared mappings i_mmap_rwsem must be held to call
4292 * huge_pte_alloc, otherwise the returned ptep could go
4293 * away if part of a shared pmd and another thread calls
4296 i_mmap_lock_read(mapping
);
4299 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
4300 spinlock_t
*src_ptl
, *dst_ptl
;
4301 src_pte
= huge_pte_offset(src
, addr
, sz
);
4304 dst_pte
= huge_pte_alloc(dst
, vma
, addr
, sz
);
4311 * If the pagetables are shared don't copy or take references.
4312 * dst_pte == src_pte is the common case of src/dest sharing.
4314 * However, src could have 'unshared' and dst shares with
4315 * another vma. If dst_pte !none, this implies sharing.
4316 * Check here before taking page table lock, and once again
4317 * after taking the lock below.
4319 dst_entry
= huge_ptep_get(dst_pte
);
4320 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
4323 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
4324 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
4325 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
4326 entry
= huge_ptep_get(src_pte
);
4327 dst_entry
= huge_ptep_get(dst_pte
);
4329 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
4331 * Skip if src entry none. Also, skip in the
4332 * unlikely case dst entry !none as this implies
4333 * sharing with another vma.
4336 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
4337 is_hugetlb_entry_hwpoisoned(entry
))) {
4338 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
4340 if (is_writable_migration_entry(swp_entry
) && cow
) {
4342 * COW mappings require pages in both
4343 * parent and child to be set to read.
4345 swp_entry
= make_readable_migration_entry(
4346 swp_offset(swp_entry
));
4347 entry
= swp_entry_to_pte(swp_entry
);
4348 set_huge_swap_pte_at(src
, addr
, src_pte
,
4351 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
4353 entry
= huge_ptep_get(src_pte
);
4354 ptepage
= pte_page(entry
);
4358 * This is a rare case where we see pinned hugetlb
4359 * pages while they're prone to COW. We need to do the
4360 * COW earlier during fork.
4362 * When pre-allocating the page or copying data, we
4363 * need to be without the pgtable locks since we could
4364 * sleep during the process.
4366 if (unlikely(page_needs_cow_for_dma(vma
, ptepage
))) {
4367 pte_t src_pte_old
= entry
;
4370 spin_unlock(src_ptl
);
4371 spin_unlock(dst_ptl
);
4372 /* Do not use reserve as it's private owned */
4373 new = alloc_huge_page(vma
, addr
, 1);
4379 copy_user_huge_page(new, ptepage
, addr
, vma
,
4383 /* Install the new huge page if src pte stable */
4384 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
4385 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
4386 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
4387 entry
= huge_ptep_get(src_pte
);
4388 if (!pte_same(src_pte_old
, entry
)) {
4389 restore_reserve_on_error(h
, vma
, addr
,
4392 /* dst_entry won't change as in child */
4395 hugetlb_install_page(vma
, dst_pte
, addr
, new);
4396 spin_unlock(src_ptl
);
4397 spin_unlock(dst_ptl
);
4403 * No need to notify as we are downgrading page
4404 * table protection not changing it to point
4407 * See Documentation/vm/mmu_notifier.rst
4409 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
4410 entry
= huge_pte_wrprotect(entry
);
4413 page_dup_rmap(ptepage
, true);
4414 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
4415 hugetlb_count_add(npages
, dst
);
4417 spin_unlock(src_ptl
);
4418 spin_unlock(dst_ptl
);
4422 mmu_notifier_invalidate_range_end(&range
);
4424 i_mmap_unlock_read(mapping
);
4429 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
4430 unsigned long start
, unsigned long end
,
4431 struct page
*ref_page
)
4433 struct mm_struct
*mm
= vma
->vm_mm
;
4434 unsigned long address
;
4439 struct hstate
*h
= hstate_vma(vma
);
4440 unsigned long sz
= huge_page_size(h
);
4441 struct mmu_notifier_range range
;
4443 WARN_ON(!is_vm_hugetlb_page(vma
));
4444 BUG_ON(start
& ~huge_page_mask(h
));
4445 BUG_ON(end
& ~huge_page_mask(h
));
4448 * This is a hugetlb vma, all the pte entries should point
4451 tlb_change_page_size(tlb
, sz
);
4452 tlb_start_vma(tlb
, vma
);
4455 * If sharing possible, alert mmu notifiers of worst case.
4457 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
4459 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
4460 mmu_notifier_invalidate_range_start(&range
);
4462 for (; address
< end
; address
+= sz
) {
4463 ptep
= huge_pte_offset(mm
, address
, sz
);
4467 ptl
= huge_pte_lock(h
, mm
, ptep
);
4468 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
4471 * We just unmapped a page of PMDs by clearing a PUD.
4472 * The caller's TLB flush range should cover this area.
4477 pte
= huge_ptep_get(ptep
);
4478 if (huge_pte_none(pte
)) {
4484 * Migrating hugepage or HWPoisoned hugepage is already
4485 * unmapped and its refcount is dropped, so just clear pte here.
4487 if (unlikely(!pte_present(pte
))) {
4488 huge_pte_clear(mm
, address
, ptep
, sz
);
4493 page
= pte_page(pte
);
4495 * If a reference page is supplied, it is because a specific
4496 * page is being unmapped, not a range. Ensure the page we
4497 * are about to unmap is the actual page of interest.
4500 if (page
!= ref_page
) {
4505 * Mark the VMA as having unmapped its page so that
4506 * future faults in this VMA will fail rather than
4507 * looking like data was lost
4509 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
4512 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4513 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
4514 if (huge_pte_dirty(pte
))
4515 set_page_dirty(page
);
4517 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
4518 page_remove_rmap(page
, true);
4521 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
4523 * Bail out after unmapping reference page if supplied
4528 mmu_notifier_invalidate_range_end(&range
);
4529 tlb_end_vma(tlb
, vma
);
4532 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
4533 struct vm_area_struct
*vma
, unsigned long start
,
4534 unsigned long end
, struct page
*ref_page
)
4536 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
4539 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4540 * test will fail on a vma being torn down, and not grab a page table
4541 * on its way out. We're lucky that the flag has such an appropriate
4542 * name, and can in fact be safely cleared here. We could clear it
4543 * before the __unmap_hugepage_range above, but all that's necessary
4544 * is to clear it before releasing the i_mmap_rwsem. This works
4545 * because in the context this is called, the VMA is about to be
4546 * destroyed and the i_mmap_rwsem is held.
4548 vma
->vm_flags
&= ~VM_MAYSHARE
;
4551 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
4552 unsigned long end
, struct page
*ref_page
)
4554 struct mmu_gather tlb
;
4556 tlb_gather_mmu(&tlb
, vma
->vm_mm
);
4557 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
4558 tlb_finish_mmu(&tlb
);
4562 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4563 * mapping it owns the reserve page for. The intention is to unmap the page
4564 * from other VMAs and let the children be SIGKILLed if they are faulting the
4567 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4568 struct page
*page
, unsigned long address
)
4570 struct hstate
*h
= hstate_vma(vma
);
4571 struct vm_area_struct
*iter_vma
;
4572 struct address_space
*mapping
;
4576 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4577 * from page cache lookup which is in HPAGE_SIZE units.
4579 address
= address
& huge_page_mask(h
);
4580 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
4582 mapping
= vma
->vm_file
->f_mapping
;
4585 * Take the mapping lock for the duration of the table walk. As
4586 * this mapping should be shared between all the VMAs,
4587 * __unmap_hugepage_range() is called as the lock is already held
4589 i_mmap_lock_write(mapping
);
4590 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
4591 /* Do not unmap the current VMA */
4592 if (iter_vma
== vma
)
4596 * Shared VMAs have their own reserves and do not affect
4597 * MAP_PRIVATE accounting but it is possible that a shared
4598 * VMA is using the same page so check and skip such VMAs.
4600 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
4604 * Unmap the page from other VMAs without their own reserves.
4605 * They get marked to be SIGKILLed if they fault in these
4606 * areas. This is because a future no-page fault on this VMA
4607 * could insert a zeroed page instead of the data existing
4608 * from the time of fork. This would look like data corruption
4610 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
4611 unmap_hugepage_range(iter_vma
, address
,
4612 address
+ huge_page_size(h
), page
);
4614 i_mmap_unlock_write(mapping
);
4618 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4619 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4620 * cannot race with other handlers or page migration.
4621 * Keep the pte_same checks anyway to make transition from the mutex easier.
4623 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4624 unsigned long address
, pte_t
*ptep
,
4625 struct page
*pagecache_page
, spinlock_t
*ptl
)
4628 struct hstate
*h
= hstate_vma(vma
);
4629 struct page
*old_page
, *new_page
;
4630 int outside_reserve
= 0;
4632 unsigned long haddr
= address
& huge_page_mask(h
);
4633 struct mmu_notifier_range range
;
4635 pte
= huge_ptep_get(ptep
);
4636 old_page
= pte_page(pte
);
4639 /* If no-one else is actually using this page, avoid the copy
4640 * and just make the page writable */
4641 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
4642 page_move_anon_rmap(old_page
, vma
);
4643 set_huge_ptep_writable(vma
, haddr
, ptep
);
4648 * If the process that created a MAP_PRIVATE mapping is about to
4649 * perform a COW due to a shared page count, attempt to satisfy
4650 * the allocation without using the existing reserves. The pagecache
4651 * page is used to determine if the reserve at this address was
4652 * consumed or not. If reserves were used, a partial faulted mapping
4653 * at the time of fork() could consume its reserves on COW instead
4654 * of the full address range.
4656 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
4657 old_page
!= pagecache_page
)
4658 outside_reserve
= 1;
4663 * Drop page table lock as buddy allocator may be called. It will
4664 * be acquired again before returning to the caller, as expected.
4667 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
4669 if (IS_ERR(new_page
)) {
4671 * If a process owning a MAP_PRIVATE mapping fails to COW,
4672 * it is due to references held by a child and an insufficient
4673 * huge page pool. To guarantee the original mappers
4674 * reliability, unmap the page from child processes. The child
4675 * may get SIGKILLed if it later faults.
4677 if (outside_reserve
) {
4678 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4683 BUG_ON(huge_pte_none(pte
));
4685 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4686 * unmapping. unmapping needs to hold i_mmap_rwsem
4687 * in write mode. Dropping i_mmap_rwsem in read mode
4688 * here is OK as COW mappings do not interact with
4691 * Reacquire both after unmap operation.
4693 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4694 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4695 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4696 i_mmap_unlock_read(mapping
);
4698 unmap_ref_private(mm
, vma
, old_page
, haddr
);
4700 i_mmap_lock_read(mapping
);
4701 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4703 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4705 pte_same(huge_ptep_get(ptep
), pte
)))
4706 goto retry_avoidcopy
;
4708 * race occurs while re-acquiring page table
4709 * lock, and our job is done.
4714 ret
= vmf_error(PTR_ERR(new_page
));
4715 goto out_release_old
;
4719 * When the original hugepage is shared one, it does not have
4720 * anon_vma prepared.
4722 if (unlikely(anon_vma_prepare(vma
))) {
4724 goto out_release_all
;
4727 copy_user_huge_page(new_page
, old_page
, address
, vma
,
4728 pages_per_huge_page(h
));
4729 __SetPageUptodate(new_page
);
4731 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
4732 haddr
+ huge_page_size(h
));
4733 mmu_notifier_invalidate_range_start(&range
);
4736 * Retake the page table lock to check for racing updates
4737 * before the page tables are altered
4740 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4741 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
4742 ClearHPageRestoreReserve(new_page
);
4745 huge_ptep_clear_flush(vma
, haddr
, ptep
);
4746 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
4747 set_huge_pte_at(mm
, haddr
, ptep
,
4748 make_huge_pte(vma
, new_page
, 1));
4749 page_remove_rmap(old_page
, true);
4750 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
4751 SetHPageMigratable(new_page
);
4752 /* Make the old page be freed below */
4753 new_page
= old_page
;
4756 mmu_notifier_invalidate_range_end(&range
);
4758 /* No restore in case of successful pagetable update (Break COW) */
4759 if (new_page
!= old_page
)
4760 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
4765 spin_lock(ptl
); /* Caller expects lock to be held */
4769 /* Return the pagecache page at a given address within a VMA */
4770 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
4771 struct vm_area_struct
*vma
, unsigned long address
)
4773 struct address_space
*mapping
;
4776 mapping
= vma
->vm_file
->f_mapping
;
4777 idx
= vma_hugecache_offset(h
, vma
, address
);
4779 return find_lock_page(mapping
, idx
);
4783 * Return whether there is a pagecache page to back given address within VMA.
4784 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4786 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
4787 struct vm_area_struct
*vma
, unsigned long address
)
4789 struct address_space
*mapping
;
4793 mapping
= vma
->vm_file
->f_mapping
;
4794 idx
= vma_hugecache_offset(h
, vma
, address
);
4796 page
= find_get_page(mapping
, idx
);
4799 return page
!= NULL
;
4802 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
4805 struct inode
*inode
= mapping
->host
;
4806 struct hstate
*h
= hstate_inode(inode
);
4807 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
4811 ClearHPageRestoreReserve(page
);
4814 * set page dirty so that it will not be removed from cache/file
4815 * by non-hugetlbfs specific code paths.
4817 set_page_dirty(page
);
4819 spin_lock(&inode
->i_lock
);
4820 inode
->i_blocks
+= blocks_per_huge_page(h
);
4821 spin_unlock(&inode
->i_lock
);
4825 static inline vm_fault_t
hugetlb_handle_userfault(struct vm_area_struct
*vma
,
4826 struct address_space
*mapping
,
4829 unsigned long haddr
,
4830 unsigned long reason
)
4834 struct vm_fault vmf
= {
4840 * Hard to debug if it ends up being
4841 * used by a callee that assumes
4842 * something about the other
4843 * uninitialized fields... same as in
4849 * hugetlb_fault_mutex and i_mmap_rwsem must be
4850 * dropped before handling userfault. Reacquire
4851 * after handling fault to make calling code simpler.
4853 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
4854 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4855 i_mmap_unlock_read(mapping
);
4856 ret
= handle_userfault(&vmf
, reason
);
4857 i_mmap_lock_read(mapping
);
4858 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4863 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
4864 struct vm_area_struct
*vma
,
4865 struct address_space
*mapping
, pgoff_t idx
,
4866 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
4868 struct hstate
*h
= hstate_vma(vma
);
4869 vm_fault_t ret
= VM_FAULT_SIGBUS
;
4875 unsigned long haddr
= address
& huge_page_mask(h
);
4876 bool new_page
, new_pagecache_page
= false;
4879 * Currently, we are forced to kill the process in the event the
4880 * original mapper has unmapped pages from the child due to a failed
4881 * COW. Warn that such a situation has occurred as it may not be obvious
4883 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
4884 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4890 * We can not race with truncation due to holding i_mmap_rwsem.
4891 * i_size is modified when holding i_mmap_rwsem, so check here
4892 * once for faults beyond end of file.
4894 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4900 page
= find_lock_page(mapping
, idx
);
4902 /* Check for page in userfault range */
4903 if (userfaultfd_missing(vma
)) {
4904 ret
= hugetlb_handle_userfault(vma
, mapping
, idx
,
4910 page
= alloc_huge_page(vma
, haddr
, 0);
4913 * Returning error will result in faulting task being
4914 * sent SIGBUS. The hugetlb fault mutex prevents two
4915 * tasks from racing to fault in the same page which
4916 * could result in false unable to allocate errors.
4917 * Page migration does not take the fault mutex, but
4918 * does a clear then write of pte's under page table
4919 * lock. Page fault code could race with migration,
4920 * notice the clear pte and try to allocate a page
4921 * here. Before returning error, get ptl and make
4922 * sure there really is no pte entry.
4924 ptl
= huge_pte_lock(h
, mm
, ptep
);
4926 if (huge_pte_none(huge_ptep_get(ptep
)))
4927 ret
= vmf_error(PTR_ERR(page
));
4931 clear_huge_page(page
, address
, pages_per_huge_page(h
));
4932 __SetPageUptodate(page
);
4935 if (vma
->vm_flags
& VM_MAYSHARE
) {
4936 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
4943 new_pagecache_page
= true;
4946 if (unlikely(anon_vma_prepare(vma
))) {
4948 goto backout_unlocked
;
4954 * If memory error occurs between mmap() and fault, some process
4955 * don't have hwpoisoned swap entry for errored virtual address.
4956 * So we need to block hugepage fault by PG_hwpoison bit check.
4958 if (unlikely(PageHWPoison(page
))) {
4959 ret
= VM_FAULT_HWPOISON_LARGE
|
4960 VM_FAULT_SET_HINDEX(hstate_index(h
));
4961 goto backout_unlocked
;
4964 /* Check for page in userfault range. */
4965 if (userfaultfd_minor(vma
)) {
4968 ret
= hugetlb_handle_userfault(vma
, mapping
, idx
,
4976 * If we are going to COW a private mapping later, we examine the
4977 * pending reservations for this page now. This will ensure that
4978 * any allocations necessary to record that reservation occur outside
4981 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4982 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4984 goto backout_unlocked
;
4986 /* Just decrements count, does not deallocate */
4987 vma_end_reservation(h
, vma
, haddr
);
4990 ptl
= huge_pte_lock(h
, mm
, ptep
);
4992 if (!huge_pte_none(huge_ptep_get(ptep
)))
4996 ClearHPageRestoreReserve(page
);
4997 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4999 page_dup_rmap(page
, true);
5000 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
5001 && (vma
->vm_flags
& VM_SHARED
)));
5002 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
5004 hugetlb_count_add(pages_per_huge_page(h
), mm
);
5005 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
5006 /* Optimization, do the COW without a second fault */
5007 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
5013 * Only set HPageMigratable in newly allocated pages. Existing pages
5014 * found in the pagecache may not have HPageMigratableset if they have
5015 * been isolated for migration.
5018 SetHPageMigratable(page
);
5028 /* restore reserve for newly allocated pages not in page cache */
5029 if (new_page
&& !new_pagecache_page
)
5030 restore_reserve_on_error(h
, vma
, haddr
, page
);
5036 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
5038 unsigned long key
[2];
5041 key
[0] = (unsigned long) mapping
;
5044 hash
= jhash2((u32
*)&key
, sizeof(key
)/(sizeof(u32
)), 0);
5046 return hash
& (num_fault_mutexes
- 1);
5050 * For uniprocessor systems we always use a single mutex, so just
5051 * return 0 and avoid the hashing overhead.
5053 u32
hugetlb_fault_mutex_hash(struct address_space
*mapping
, pgoff_t idx
)
5059 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5060 unsigned long address
, unsigned int flags
)
5067 struct page
*page
= NULL
;
5068 struct page
*pagecache_page
= NULL
;
5069 struct hstate
*h
= hstate_vma(vma
);
5070 struct address_space
*mapping
;
5071 int need_wait_lock
= 0;
5072 unsigned long haddr
= address
& huge_page_mask(h
);
5074 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
5077 * Since we hold no locks, ptep could be stale. That is
5078 * OK as we are only making decisions based on content and
5079 * not actually modifying content here.
5081 entry
= huge_ptep_get(ptep
);
5082 if (unlikely(is_hugetlb_entry_migration(entry
))) {
5083 migration_entry_wait_huge(vma
, mm
, ptep
);
5085 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
5086 return VM_FAULT_HWPOISON_LARGE
|
5087 VM_FAULT_SET_HINDEX(hstate_index(h
));
5091 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5092 * until finished with ptep. This serves two purposes:
5093 * 1) It prevents huge_pmd_unshare from being called elsewhere
5094 * and making the ptep no longer valid.
5095 * 2) It synchronizes us with i_size modifications during truncation.
5097 * ptep could have already be assigned via huge_pte_offset. That
5098 * is OK, as huge_pte_alloc will return the same value unless
5099 * something has changed.
5101 mapping
= vma
->vm_file
->f_mapping
;
5102 i_mmap_lock_read(mapping
);
5103 ptep
= huge_pte_alloc(mm
, vma
, haddr
, huge_page_size(h
));
5105 i_mmap_unlock_read(mapping
);
5106 return VM_FAULT_OOM
;
5110 * Serialize hugepage allocation and instantiation, so that we don't
5111 * get spurious allocation failures if two CPUs race to instantiate
5112 * the same page in the page cache.
5114 idx
= vma_hugecache_offset(h
, vma
, haddr
);
5115 hash
= hugetlb_fault_mutex_hash(mapping
, idx
);
5116 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
5118 entry
= huge_ptep_get(ptep
);
5119 if (huge_pte_none(entry
)) {
5120 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
5127 * entry could be a migration/hwpoison entry at this point, so this
5128 * check prevents the kernel from going below assuming that we have
5129 * an active hugepage in pagecache. This goto expects the 2nd page
5130 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5131 * properly handle it.
5133 if (!pte_present(entry
))
5137 * If we are going to COW the mapping later, we examine the pending
5138 * reservations for this page now. This will ensure that any
5139 * allocations necessary to record that reservation occur outside the
5140 * spinlock. For private mappings, we also lookup the pagecache
5141 * page now as it is used to determine if a reservation has been
5144 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
5145 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
5149 /* Just decrements count, does not deallocate */
5150 vma_end_reservation(h
, vma
, haddr
);
5152 if (!(vma
->vm_flags
& VM_MAYSHARE
))
5153 pagecache_page
= hugetlbfs_pagecache_page(h
,
5157 ptl
= huge_pte_lock(h
, mm
, ptep
);
5159 /* Check for a racing update before calling hugetlb_cow */
5160 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
5164 * hugetlb_cow() requires page locks of pte_page(entry) and
5165 * pagecache_page, so here we need take the former one
5166 * when page != pagecache_page or !pagecache_page.
5168 page
= pte_page(entry
);
5169 if (page
!= pagecache_page
)
5170 if (!trylock_page(page
)) {
5177 if (flags
& FAULT_FLAG_WRITE
) {
5178 if (!huge_pte_write(entry
)) {
5179 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
5180 pagecache_page
, ptl
);
5183 entry
= huge_pte_mkdirty(entry
);
5185 entry
= pte_mkyoung(entry
);
5186 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
5187 flags
& FAULT_FLAG_WRITE
))
5188 update_mmu_cache(vma
, haddr
, ptep
);
5190 if (page
!= pagecache_page
)
5196 if (pagecache_page
) {
5197 unlock_page(pagecache_page
);
5198 put_page(pagecache_page
);
5201 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
5202 i_mmap_unlock_read(mapping
);
5204 * Generally it's safe to hold refcount during waiting page lock. But
5205 * here we just wait to defer the next page fault to avoid busy loop and
5206 * the page is not used after unlocked before returning from the current
5207 * page fault. So we are safe from accessing freed page, even if we wait
5208 * here without taking refcount.
5211 wait_on_page_locked(page
);
5215 #ifdef CONFIG_USERFAULTFD
5217 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5218 * modifications for huge pages.
5220 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
5222 struct vm_area_struct
*dst_vma
,
5223 unsigned long dst_addr
,
5224 unsigned long src_addr
,
5225 enum mcopy_atomic_mode mode
,
5226 struct page
**pagep
)
5228 bool is_continue
= (mode
== MCOPY_ATOMIC_CONTINUE
);
5229 struct hstate
*h
= hstate_vma(dst_vma
);
5230 struct address_space
*mapping
= dst_vma
->vm_file
->f_mapping
;
5231 pgoff_t idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
5233 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
5239 bool new_pagecache_page
= false;
5243 page
= find_lock_page(mapping
, idx
);
5246 } else if (!*pagep
) {
5247 /* If a page already exists, then it's UFFDIO_COPY for
5248 * a non-missing case. Return -EEXIST.
5251 hugetlbfs_pagecache_present(h
, dst_vma
, dst_addr
)) {
5256 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
5262 ret
= copy_huge_page_from_user(page
,
5263 (const void __user
*) src_addr
,
5264 pages_per_huge_page(h
), false);
5266 /* fallback to copy_from_user outside mmap_lock */
5267 if (unlikely(ret
)) {
5269 /* Free the allocated page which may have
5270 * consumed a reservation.
5272 restore_reserve_on_error(h
, dst_vma
, dst_addr
, page
);
5275 /* Allocate a temporary page to hold the copied
5278 page
= alloc_huge_page_vma(h
, dst_vma
, dst_addr
);
5284 /* Set the outparam pagep and return to the caller to
5285 * copy the contents outside the lock. Don't free the
5292 hugetlbfs_pagecache_present(h
, dst_vma
, dst_addr
)) {
5299 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
5305 copy_huge_page(page
, *pagep
);
5311 * The memory barrier inside __SetPageUptodate makes sure that
5312 * preceding stores to the page contents become visible before
5313 * the set_pte_at() write.
5315 __SetPageUptodate(page
);
5317 /* Add shared, newly allocated pages to the page cache. */
5318 if (vm_shared
&& !is_continue
) {
5319 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
5322 goto out_release_nounlock
;
5325 * Serialization between remove_inode_hugepages() and
5326 * huge_add_to_page_cache() below happens through the
5327 * hugetlb_fault_mutex_table that here must be hold by
5330 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
5332 goto out_release_nounlock
;
5333 new_pagecache_page
= true;
5336 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
5340 * Recheck the i_size after holding PT lock to make sure not
5341 * to leave any page mapped (as page_mapped()) beyond the end
5342 * of the i_size (remove_inode_hugepages() is strict about
5343 * enforcing that). If we bail out here, we'll also leave a
5344 * page in the radix tree in the vm_shared case beyond the end
5345 * of the i_size, but remove_inode_hugepages() will take care
5346 * of it as soon as we drop the hugetlb_fault_mutex_table.
5348 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
5351 goto out_release_unlock
;
5354 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
5355 goto out_release_unlock
;
5358 page_dup_rmap(page
, true);
5360 ClearHPageRestoreReserve(page
);
5361 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
5364 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5365 if (is_continue
&& !vm_shared
)
5368 writable
= dst_vma
->vm_flags
& VM_WRITE
;
5370 _dst_pte
= make_huge_pte(dst_vma
, page
, writable
);
5372 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
5373 _dst_pte
= pte_mkyoung(_dst_pte
);
5375 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
5377 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
5378 dst_vma
->vm_flags
& VM_WRITE
);
5379 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
5381 /* No need to invalidate - it was non-present before */
5382 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
5386 SetHPageMigratable(page
);
5387 if (vm_shared
|| is_continue
)
5394 if (vm_shared
|| is_continue
)
5396 out_release_nounlock
:
5397 if (!new_pagecache_page
)
5398 restore_reserve_on_error(h
, dst_vma
, dst_addr
, page
);
5402 #endif /* CONFIG_USERFAULTFD */
5404 static void record_subpages_vmas(struct page
*page
, struct vm_area_struct
*vma
,
5405 int refs
, struct page
**pages
,
5406 struct vm_area_struct
**vmas
)
5410 for (nr
= 0; nr
< refs
; nr
++) {
5412 pages
[nr
] = mem_map_offset(page
, nr
);
5418 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5419 struct page
**pages
, struct vm_area_struct
**vmas
,
5420 unsigned long *position
, unsigned long *nr_pages
,
5421 long i
, unsigned int flags
, int *locked
)
5423 unsigned long pfn_offset
;
5424 unsigned long vaddr
= *position
;
5425 unsigned long remainder
= *nr_pages
;
5426 struct hstate
*h
= hstate_vma(vma
);
5427 int err
= -EFAULT
, refs
;
5429 while (vaddr
< vma
->vm_end
&& remainder
) {
5431 spinlock_t
*ptl
= NULL
;
5436 * If we have a pending SIGKILL, don't keep faulting pages and
5437 * potentially allocating memory.
5439 if (fatal_signal_pending(current
)) {
5445 * Some archs (sparc64, sh*) have multiple pte_ts to
5446 * each hugepage. We have to make sure we get the
5447 * first, for the page indexing below to work.
5449 * Note that page table lock is not held when pte is null.
5451 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
5454 ptl
= huge_pte_lock(h
, mm
, pte
);
5455 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
5458 * When coredumping, it suits get_dump_page if we just return
5459 * an error where there's an empty slot with no huge pagecache
5460 * to back it. This way, we avoid allocating a hugepage, and
5461 * the sparse dumpfile avoids allocating disk blocks, but its
5462 * huge holes still show up with zeroes where they need to be.
5464 if (absent
&& (flags
& FOLL_DUMP
) &&
5465 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
5473 * We need call hugetlb_fault for both hugepages under migration
5474 * (in which case hugetlb_fault waits for the migration,) and
5475 * hwpoisoned hugepages (in which case we need to prevent the
5476 * caller from accessing to them.) In order to do this, we use
5477 * here is_swap_pte instead of is_hugetlb_entry_migration and
5478 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5479 * both cases, and because we can't follow correct pages
5480 * directly from any kind of swap entries.
5482 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
5483 ((flags
& FOLL_WRITE
) &&
5484 !huge_pte_write(huge_ptep_get(pte
)))) {
5486 unsigned int fault_flags
= 0;
5490 if (flags
& FOLL_WRITE
)
5491 fault_flags
|= FAULT_FLAG_WRITE
;
5493 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
5494 FAULT_FLAG_KILLABLE
;
5495 if (flags
& FOLL_NOWAIT
)
5496 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
5497 FAULT_FLAG_RETRY_NOWAIT
;
5498 if (flags
& FOLL_TRIED
) {
5500 * Note: FAULT_FLAG_ALLOW_RETRY and
5501 * FAULT_FLAG_TRIED can co-exist
5503 fault_flags
|= FAULT_FLAG_TRIED
;
5505 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
5506 if (ret
& VM_FAULT_ERROR
) {
5507 err
= vm_fault_to_errno(ret
, flags
);
5511 if (ret
& VM_FAULT_RETRY
) {
5513 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
5517 * VM_FAULT_RETRY must not return an
5518 * error, it will return zero
5521 * No need to update "position" as the
5522 * caller will not check it after
5523 * *nr_pages is set to 0.
5530 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
5531 page
= pte_page(huge_ptep_get(pte
));
5534 * If subpage information not requested, update counters
5535 * and skip the same_page loop below.
5537 if (!pages
&& !vmas
&& !pfn_offset
&&
5538 (vaddr
+ huge_page_size(h
) < vma
->vm_end
) &&
5539 (remainder
>= pages_per_huge_page(h
))) {
5540 vaddr
+= huge_page_size(h
);
5541 remainder
-= pages_per_huge_page(h
);
5542 i
+= pages_per_huge_page(h
);
5547 /* vaddr may not be aligned to PAGE_SIZE */
5548 refs
= min3(pages_per_huge_page(h
) - pfn_offset
, remainder
,
5549 (vma
->vm_end
- ALIGN_DOWN(vaddr
, PAGE_SIZE
)) >> PAGE_SHIFT
);
5552 record_subpages_vmas(mem_map_offset(page
, pfn_offset
),
5554 likely(pages
) ? pages
+ i
: NULL
,
5555 vmas
? vmas
+ i
: NULL
);
5559 * try_grab_compound_head() should always succeed here,
5560 * because: a) we hold the ptl lock, and b) we've just
5561 * checked that the huge page is present in the page
5562 * tables. If the huge page is present, then the tail
5563 * pages must also be present. The ptl prevents the
5564 * head page and tail pages from being rearranged in
5565 * any way. So this page must be available at this
5566 * point, unless the page refcount overflowed:
5568 if (WARN_ON_ONCE(!try_grab_compound_head(pages
[i
],
5578 vaddr
+= (refs
<< PAGE_SHIFT
);
5584 *nr_pages
= remainder
;
5586 * setting position is actually required only if remainder is
5587 * not zero but it's faster not to add a "if (remainder)"
5595 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
5596 unsigned long address
, unsigned long end
, pgprot_t newprot
)
5598 struct mm_struct
*mm
= vma
->vm_mm
;
5599 unsigned long start
= address
;
5602 struct hstate
*h
= hstate_vma(vma
);
5603 unsigned long pages
= 0;
5604 bool shared_pmd
= false;
5605 struct mmu_notifier_range range
;
5608 * In the case of shared PMDs, the area to flush could be beyond
5609 * start/end. Set range.start/range.end to cover the maximum possible
5610 * range if PMD sharing is possible.
5612 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
5613 0, vma
, mm
, start
, end
);
5614 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
5616 BUG_ON(address
>= end
);
5617 flush_cache_range(vma
, range
.start
, range
.end
);
5619 mmu_notifier_invalidate_range_start(&range
);
5620 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
5621 for (; address
< end
; address
+= huge_page_size(h
)) {
5623 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
5626 ptl
= huge_pte_lock(h
, mm
, ptep
);
5627 if (huge_pmd_unshare(mm
, vma
, &address
, ptep
)) {
5633 pte
= huge_ptep_get(ptep
);
5634 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
5638 if (unlikely(is_hugetlb_entry_migration(pte
))) {
5639 swp_entry_t entry
= pte_to_swp_entry(pte
);
5641 if (is_writable_migration_entry(entry
)) {
5644 entry
= make_readable_migration_entry(
5646 newpte
= swp_entry_to_pte(entry
);
5647 set_huge_swap_pte_at(mm
, address
, ptep
,
5648 newpte
, huge_page_size(h
));
5654 if (!huge_pte_none(pte
)) {
5656 unsigned int shift
= huge_page_shift(hstate_vma(vma
));
5658 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
5659 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
5660 pte
= arch_make_huge_pte(pte
, shift
, vma
->vm_flags
);
5661 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
5667 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5668 * may have cleared our pud entry and done put_page on the page table:
5669 * once we release i_mmap_rwsem, another task can do the final put_page
5670 * and that page table be reused and filled with junk. If we actually
5671 * did unshare a page of pmds, flush the range corresponding to the pud.
5674 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
5676 flush_hugetlb_tlb_range(vma
, start
, end
);
5678 * No need to call mmu_notifier_invalidate_range() we are downgrading
5679 * page table protection not changing it to point to a new page.
5681 * See Documentation/vm/mmu_notifier.rst
5683 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
5684 mmu_notifier_invalidate_range_end(&range
);
5686 return pages
<< h
->order
;
5689 /* Return true if reservation was successful, false otherwise. */
5690 bool hugetlb_reserve_pages(struct inode
*inode
,
5692 struct vm_area_struct
*vma
,
5693 vm_flags_t vm_flags
)
5696 struct hstate
*h
= hstate_inode(inode
);
5697 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5698 struct resv_map
*resv_map
;
5699 struct hugetlb_cgroup
*h_cg
= NULL
;
5700 long gbl_reserve
, regions_needed
= 0;
5702 /* This should never happen */
5704 VM_WARN(1, "%s called with a negative range\n", __func__
);
5709 * Only apply hugepage reservation if asked. At fault time, an
5710 * attempt will be made for VM_NORESERVE to allocate a page
5711 * without using reserves
5713 if (vm_flags
& VM_NORESERVE
)
5717 * Shared mappings base their reservation on the number of pages that
5718 * are already allocated on behalf of the file. Private mappings need
5719 * to reserve the full area even if read-only as mprotect() may be
5720 * called to make the mapping read-write. Assume !vma is a shm mapping
5722 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5724 * resv_map can not be NULL as hugetlb_reserve_pages is only
5725 * called for inodes for which resv_maps were created (see
5726 * hugetlbfs_get_inode).
5728 resv_map
= inode_resv_map(inode
);
5730 chg
= region_chg(resv_map
, from
, to
, ®ions_needed
);
5733 /* Private mapping. */
5734 resv_map
= resv_map_alloc();
5740 set_vma_resv_map(vma
, resv_map
);
5741 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
5747 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h
),
5748 chg
* pages_per_huge_page(h
), &h_cg
) < 0)
5751 if (vma
&& !(vma
->vm_flags
& VM_MAYSHARE
) && h_cg
) {
5752 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5755 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map
, h_cg
, h
);
5759 * There must be enough pages in the subpool for the mapping. If
5760 * the subpool has a minimum size, there may be some global
5761 * reservations already in place (gbl_reserve).
5763 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
5764 if (gbl_reserve
< 0)
5765 goto out_uncharge_cgroup
;
5768 * Check enough hugepages are available for the reservation.
5769 * Hand the pages back to the subpool if there are not
5771 if (hugetlb_acct_memory(h
, gbl_reserve
) < 0)
5775 * Account for the reservations made. Shared mappings record regions
5776 * that have reservations as they are shared by multiple VMAs.
5777 * When the last VMA disappears, the region map says how much
5778 * the reservation was and the page cache tells how much of
5779 * the reservation was consumed. Private mappings are per-VMA and
5780 * only the consumed reservations are tracked. When the VMA
5781 * disappears, the original reservation is the VMA size and the
5782 * consumed reservations are stored in the map. Hence, nothing
5783 * else has to be done for private mappings here
5785 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
5786 add
= region_add(resv_map
, from
, to
, regions_needed
, h
, h_cg
);
5788 if (unlikely(add
< 0)) {
5789 hugetlb_acct_memory(h
, -gbl_reserve
);
5791 } else if (unlikely(chg
> add
)) {
5793 * pages in this range were added to the reserve
5794 * map between region_chg and region_add. This
5795 * indicates a race with alloc_huge_page. Adjust
5796 * the subpool and reserve counts modified above
5797 * based on the difference.
5802 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5803 * reference to h_cg->css. See comment below for detail.
5805 hugetlb_cgroup_uncharge_cgroup_rsvd(
5807 (chg
- add
) * pages_per_huge_page(h
), h_cg
);
5809 rsv_adjust
= hugepage_subpool_put_pages(spool
,
5811 hugetlb_acct_memory(h
, -rsv_adjust
);
5814 * The file_regions will hold their own reference to
5815 * h_cg->css. So we should release the reference held
5816 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5819 hugetlb_cgroup_put_rsvd_cgroup(h_cg
);
5825 /* put back original number of pages, chg */
5826 (void)hugepage_subpool_put_pages(spool
, chg
);
5827 out_uncharge_cgroup
:
5828 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h
),
5829 chg
* pages_per_huge_page(h
), h_cg
);
5831 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
5832 /* Only call region_abort if the region_chg succeeded but the
5833 * region_add failed or didn't run.
5835 if (chg
>= 0 && add
< 0)
5836 region_abort(resv_map
, from
, to
, regions_needed
);
5837 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
5838 kref_put(&resv_map
->refs
, resv_map_release
);
5842 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
5845 struct hstate
*h
= hstate_inode(inode
);
5846 struct resv_map
*resv_map
= inode_resv_map(inode
);
5848 struct hugepage_subpool
*spool
= subpool_inode(inode
);
5852 * Since this routine can be called in the evict inode path for all
5853 * hugetlbfs inodes, resv_map could be NULL.
5856 chg
= region_del(resv_map
, start
, end
);
5858 * region_del() can fail in the rare case where a region
5859 * must be split and another region descriptor can not be
5860 * allocated. If end == LONG_MAX, it will not fail.
5866 spin_lock(&inode
->i_lock
);
5867 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
5868 spin_unlock(&inode
->i_lock
);
5871 * If the subpool has a minimum size, the number of global
5872 * reservations to be released may be adjusted.
5874 * Note that !resv_map implies freed == 0. So (chg - freed)
5875 * won't go negative.
5877 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
5878 hugetlb_acct_memory(h
, -gbl_reserve
);
5883 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5884 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
5885 struct vm_area_struct
*vma
,
5886 unsigned long addr
, pgoff_t idx
)
5888 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
5890 unsigned long sbase
= saddr
& PUD_MASK
;
5891 unsigned long s_end
= sbase
+ PUD_SIZE
;
5893 /* Allow segments to share if only one is marked locked */
5894 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5895 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
5898 * match the virtual addresses, permission and the alignment of the
5901 if (pmd_index(addr
) != pmd_index(saddr
) ||
5902 vm_flags
!= svm_flags
||
5903 !range_in_vma(svma
, sbase
, s_end
))
5909 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
5911 unsigned long base
= addr
& PUD_MASK
;
5912 unsigned long end
= base
+ PUD_SIZE
;
5915 * check on proper vm_flags and page table alignment
5917 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
5922 bool want_pmd_share(struct vm_area_struct
*vma
, unsigned long addr
)
5924 #ifdef CONFIG_USERFAULTFD
5925 if (uffd_disable_huge_pmd_share(vma
))
5928 return vma_shareable(vma
, addr
);
5932 * Determine if start,end range within vma could be mapped by shared pmd.
5933 * If yes, adjust start and end to cover range associated with possible
5934 * shared pmd mappings.
5936 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5937 unsigned long *start
, unsigned long *end
)
5939 unsigned long v_start
= ALIGN(vma
->vm_start
, PUD_SIZE
),
5940 v_end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
5943 * vma needs to span at least one aligned PUD size, and the range
5944 * must be at least partially within in.
5946 if (!(vma
->vm_flags
& VM_MAYSHARE
) || !(v_end
> v_start
) ||
5947 (*end
<= v_start
) || (*start
>= v_end
))
5950 /* Extend the range to be PUD aligned for a worst case scenario */
5951 if (*start
> v_start
)
5952 *start
= ALIGN_DOWN(*start
, PUD_SIZE
);
5955 *end
= ALIGN(*end
, PUD_SIZE
);
5959 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5960 * and returns the corresponding pte. While this is not necessary for the
5961 * !shared pmd case because we can allocate the pmd later as well, it makes the
5962 * code much cleaner.
5964 * This routine must be called with i_mmap_rwsem held in at least read mode if
5965 * sharing is possible. For hugetlbfs, this prevents removal of any page
5966 * table entries associated with the address space. This is important as we
5967 * are setting up sharing based on existing page table entries (mappings).
5969 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5970 * huge_pte_alloc know that sharing is not possible and do not take
5971 * i_mmap_rwsem as a performance optimization. This is handled by the
5972 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5973 * only required for subsequent processing.
5975 pte_t
*huge_pmd_share(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
5976 unsigned long addr
, pud_t
*pud
)
5978 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
5979 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
5981 struct vm_area_struct
*svma
;
5982 unsigned long saddr
;
5987 i_mmap_assert_locked(mapping
);
5988 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
5992 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
5994 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
5995 vma_mmu_pagesize(svma
));
5997 get_page(virt_to_page(spte
));
6006 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
6007 if (pud_none(*pud
)) {
6008 pud_populate(mm
, pud
,
6009 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
6012 put_page(virt_to_page(spte
));
6016 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
6021 * unmap huge page backed by shared pte.
6023 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6024 * indicated by page_count > 1, unmap is achieved by clearing pud and
6025 * decrementing the ref count. If count == 1, the pte page is not shared.
6027 * Called with page table lock held and i_mmap_rwsem held in write mode.
6029 * returns: 1 successfully unmapped a shared pte page
6030 * 0 the underlying pte page is not shared, or it is the last user
6032 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
6033 unsigned long *addr
, pte_t
*ptep
)
6035 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
6036 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
6037 pud_t
*pud
= pud_offset(p4d
, *addr
);
6039 i_mmap_assert_write_locked(vma
->vm_file
->f_mapping
);
6040 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
6041 if (page_count(virt_to_page(ptep
)) == 1)
6045 put_page(virt_to_page(ptep
));
6047 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
6051 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6052 pte_t
*huge_pmd_share(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
6053 unsigned long addr
, pud_t
*pud
)
6058 int huge_pmd_unshare(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
6059 unsigned long *addr
, pte_t
*ptep
)
6064 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
6065 unsigned long *start
, unsigned long *end
)
6069 bool want_pmd_share(struct vm_area_struct
*vma
, unsigned long addr
)
6073 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6075 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6076 pte_t
*huge_pte_alloc(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
6077 unsigned long addr
, unsigned long sz
)
6084 pgd
= pgd_offset(mm
, addr
);
6085 p4d
= p4d_alloc(mm
, pgd
, addr
);
6088 pud
= pud_alloc(mm
, p4d
, addr
);
6090 if (sz
== PUD_SIZE
) {
6093 BUG_ON(sz
!= PMD_SIZE
);
6094 if (want_pmd_share(vma
, addr
) && pud_none(*pud
))
6095 pte
= huge_pmd_share(mm
, vma
, addr
, pud
);
6097 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
6100 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
6106 * huge_pte_offset() - Walk the page table to resolve the hugepage
6107 * entry at address @addr
6109 * Return: Pointer to page table entry (PUD or PMD) for
6110 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6111 * size @sz doesn't match the hugepage size at this level of the page
6114 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
6115 unsigned long addr
, unsigned long sz
)
6122 pgd
= pgd_offset(mm
, addr
);
6123 if (!pgd_present(*pgd
))
6125 p4d
= p4d_offset(pgd
, addr
);
6126 if (!p4d_present(*p4d
))
6129 pud
= pud_offset(p4d
, addr
);
6131 /* must be pud huge, non-present or none */
6132 return (pte_t
*)pud
;
6133 if (!pud_present(*pud
))
6135 /* must have a valid entry and size to go further */
6137 pmd
= pmd_offset(pud
, addr
);
6138 /* must be pmd huge, non-present or none */
6139 return (pte_t
*)pmd
;
6142 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6145 * These functions are overwritable if your architecture needs its own
6148 struct page
* __weak
6149 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
6152 return ERR_PTR(-EINVAL
);
6155 struct page
* __weak
6156 follow_huge_pd(struct vm_area_struct
*vma
,
6157 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
6159 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6163 struct page
* __weak
6164 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
6165 pmd_t
*pmd
, int flags
)
6167 struct page
*page
= NULL
;
6171 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6172 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
6173 (FOLL_PIN
| FOLL_GET
)))
6177 ptl
= pmd_lockptr(mm
, pmd
);
6180 * make sure that the address range covered by this pmd is not
6181 * unmapped from other threads.
6183 if (!pmd_huge(*pmd
))
6185 pte
= huge_ptep_get((pte_t
*)pmd
);
6186 if (pte_present(pte
)) {
6187 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
6189 * try_grab_page() should always succeed here, because: a) we
6190 * hold the pmd (ptl) lock, and b) we've just checked that the
6191 * huge pmd (head) page is present in the page tables. The ptl
6192 * prevents the head page and tail pages from being rearranged
6193 * in any way. So this page must be available at this point,
6194 * unless the page refcount overflowed:
6196 if (WARN_ON_ONCE(!try_grab_page(page
, flags
))) {
6201 if (is_hugetlb_entry_migration(pte
)) {
6203 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
6207 * hwpoisoned entry is treated as no_page_table in
6208 * follow_page_mask().
6216 struct page
* __weak
6217 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
6218 pud_t
*pud
, int flags
)
6220 if (flags
& (FOLL_GET
| FOLL_PIN
))
6223 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
6226 struct page
* __weak
6227 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
6229 if (flags
& (FOLL_GET
| FOLL_PIN
))
6232 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
6235 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
6239 spin_lock_irq(&hugetlb_lock
);
6240 if (!PageHeadHuge(page
) ||
6241 !HPageMigratable(page
) ||
6242 !get_page_unless_zero(page
)) {
6246 ClearHPageMigratable(page
);
6247 list_move_tail(&page
->lru
, list
);
6249 spin_unlock_irq(&hugetlb_lock
);
6253 int get_hwpoison_huge_page(struct page
*page
, bool *hugetlb
)
6258 spin_lock_irq(&hugetlb_lock
);
6259 if (PageHeadHuge(page
)) {
6261 if (HPageFreed(page
) || HPageMigratable(page
))
6262 ret
= get_page_unless_zero(page
);
6266 spin_unlock_irq(&hugetlb_lock
);
6270 void putback_active_hugepage(struct page
*page
)
6272 spin_lock_irq(&hugetlb_lock
);
6273 SetHPageMigratable(page
);
6274 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
6275 spin_unlock_irq(&hugetlb_lock
);
6279 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
6281 struct hstate
*h
= page_hstate(oldpage
);
6283 hugetlb_cgroup_migrate(oldpage
, newpage
);
6284 set_page_owner_migrate_reason(newpage
, reason
);
6287 * transfer temporary state of the new huge page. This is
6288 * reverse to other transitions because the newpage is going to
6289 * be final while the old one will be freed so it takes over
6290 * the temporary status.
6292 * Also note that we have to transfer the per-node surplus state
6293 * here as well otherwise the global surplus count will not match
6296 if (HPageTemporary(newpage
)) {
6297 int old_nid
= page_to_nid(oldpage
);
6298 int new_nid
= page_to_nid(newpage
);
6300 SetHPageTemporary(oldpage
);
6301 ClearHPageTemporary(newpage
);
6304 * There is no need to transfer the per-node surplus state
6305 * when we do not cross the node.
6307 if (new_nid
== old_nid
)
6309 spin_lock_irq(&hugetlb_lock
);
6310 if (h
->surplus_huge_pages_node
[old_nid
]) {
6311 h
->surplus_huge_pages_node
[old_nid
]--;
6312 h
->surplus_huge_pages_node
[new_nid
]++;
6314 spin_unlock_irq(&hugetlb_lock
);
6319 * This function will unconditionally remove all the shared pmd pgtable entries
6320 * within the specific vma for a hugetlbfs memory range.
6322 void hugetlb_unshare_all_pmds(struct vm_area_struct
*vma
)
6324 struct hstate
*h
= hstate_vma(vma
);
6325 unsigned long sz
= huge_page_size(h
);
6326 struct mm_struct
*mm
= vma
->vm_mm
;
6327 struct mmu_notifier_range range
;
6328 unsigned long address
, start
, end
;
6332 if (!(vma
->vm_flags
& VM_MAYSHARE
))
6335 start
= ALIGN(vma
->vm_start
, PUD_SIZE
);
6336 end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
6342 * No need to call adjust_range_if_pmd_sharing_possible(), because
6343 * we have already done the PUD_SIZE alignment.
6345 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
6347 mmu_notifier_invalidate_range_start(&range
);
6348 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
6349 for (address
= start
; address
< end
; address
+= PUD_SIZE
) {
6350 unsigned long tmp
= address
;
6352 ptep
= huge_pte_offset(mm
, address
, sz
);
6355 ptl
= huge_pte_lock(h
, mm
, ptep
);
6356 /* We don't want 'address' to be changed */
6357 huge_pmd_unshare(mm
, vma
, &tmp
, ptep
);
6360 flush_hugetlb_tlb_range(vma
, start
, end
);
6361 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
6363 * No need to call mmu_notifier_invalidate_range(), see
6364 * Documentation/vm/mmu_notifier.rst.
6366 mmu_notifier_invalidate_range_end(&range
);
6370 static bool cma_reserve_called __initdata
;
6372 static int __init
cmdline_parse_hugetlb_cma(char *p
)
6374 hugetlb_cma_size
= memparse(p
, &p
);
6378 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma
);
6380 void __init
hugetlb_cma_reserve(int order
)
6382 unsigned long size
, reserved
, per_node
;
6385 cma_reserve_called
= true;
6387 if (!hugetlb_cma_size
)
6390 if (hugetlb_cma_size
< (PAGE_SIZE
<< order
)) {
6391 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6392 (PAGE_SIZE
<< order
) / SZ_1M
);
6397 * If 3 GB area is requested on a machine with 4 numa nodes,
6398 * let's allocate 1 GB on first three nodes and ignore the last one.
6400 per_node
= DIV_ROUND_UP(hugetlb_cma_size
, nr_online_nodes
);
6401 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6402 hugetlb_cma_size
/ SZ_1M
, per_node
/ SZ_1M
);
6405 for_each_node_state(nid
, N_ONLINE
) {
6407 char name
[CMA_MAX_NAME
];
6409 size
= min(per_node
, hugetlb_cma_size
- reserved
);
6410 size
= round_up(size
, PAGE_SIZE
<< order
);
6412 snprintf(name
, sizeof(name
), "hugetlb%d", nid
);
6413 res
= cma_declare_contiguous_nid(0, size
, 0, PAGE_SIZE
<< order
,
6415 &hugetlb_cma
[nid
], nid
);
6417 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6423 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6426 if (reserved
>= hugetlb_cma_size
)
6431 void __init
hugetlb_cma_check(void)
6433 if (!hugetlb_cma_size
|| cma_reserve_called
)
6436 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6439 #endif /* CONFIG_CMA */