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>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
38 #include <linux/padata.h>
41 #include <asm/pgalloc.h>
45 #include <linux/hugetlb.h>
46 #include <linux/hugetlb_cgroup.h>
47 #include <linux/node.h>
48 #include <linux/page_owner.h>
50 #include "hugetlb_vmemmap.h"
52 int hugetlb_max_hstate __read_mostly;
53 unsigned int default_hstate_idx;
54 struct hstate hstates[HUGE_MAX_HSTATE];
57 static struct cma *hugetlb_cma[MAX_NUMNODES];
58 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
59 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
61 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
65 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
70 static unsigned long hugetlb_cma_size __initdata;
72 __initdata struct list_head huge_boot_pages[MAX_NUMNODES];
74 /* for command line parsing */
75 static struct hstate * __initdata parsed_hstate;
76 static unsigned long __initdata default_hstate_max_huge_pages;
77 static bool __initdata parsed_valid_hugepagesz = true;
78 static bool __initdata parsed_default_hugepagesz;
79 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
82 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
83 * free_huge_pages, and surplus_huge_pages.
85 DEFINE_SPINLOCK(hugetlb_lock);
88 * Serializes faults on the same logical page. This is used to
89 * prevent spurious OOMs when the hugepage pool is fully utilized.
91 static int num_fault_mutexes;
92 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
94 /* Forward declaration */
95 static int hugetlb_acct_memory(struct hstate *h, long delta);
96 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
97 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
98 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
99 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
100 unsigned long start, unsigned long end);
101 static struct resv_map *vma_resv_map(struct vm_area_struct *vma);
103 static inline bool subpool_is_free(struct hugepage_subpool *spool)
107 if (spool->max_hpages != -1)
108 return spool->used_hpages == 0;
109 if (spool->min_hpages != -1)
110 return spool->rsv_hpages == spool->min_hpages;
115 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
116 unsigned long irq_flags)
118 spin_unlock_irqrestore(&spool->lock, irq_flags);
120 /* If no pages are used, and no other handles to the subpool
121 * remain, give up any reservations based on minimum size and
122 * free the subpool */
123 if (subpool_is_free(spool)) {
124 if (spool->min_hpages != -1)
125 hugetlb_acct_memory(spool->hstate,
131 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
134 struct hugepage_subpool *spool;
136 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
140 spin_lock_init(&spool->lock);
142 spool->max_hpages = max_hpages;
144 spool->min_hpages = min_hpages;
146 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
150 spool->rsv_hpages = min_hpages;
155 void hugepage_put_subpool(struct hugepage_subpool *spool)
159 spin_lock_irqsave(&spool->lock, flags);
160 BUG_ON(!spool->count);
162 unlock_or_release_subpool(spool, flags);
166 * Subpool accounting for allocating and reserving pages.
167 * Return -ENOMEM if there are not enough resources to satisfy the
168 * request. Otherwise, return the number of pages by which the
169 * global pools must be adjusted (upward). The returned value may
170 * only be different than the passed value (delta) in the case where
171 * a subpool minimum size must be maintained.
173 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
181 spin_lock_irq(&spool->lock);
183 if (spool->max_hpages != -1) { /* maximum size accounting */
184 if ((spool->used_hpages + delta) <= spool->max_hpages)
185 spool->used_hpages += delta;
192 /* minimum size accounting */
193 if (spool->min_hpages != -1 && spool->rsv_hpages) {
194 if (delta > spool->rsv_hpages) {
196 * Asking for more reserves than those already taken on
197 * behalf of subpool. Return difference.
199 ret = delta - spool->rsv_hpages;
200 spool->rsv_hpages = 0;
202 ret = 0; /* reserves already accounted for */
203 spool->rsv_hpages -= delta;
208 spin_unlock_irq(&spool->lock);
213 * Subpool accounting for freeing and unreserving pages.
214 * Return the number of global page reservations that must be dropped.
215 * The return value may only be different than the passed value (delta)
216 * in the case where a subpool minimum size must be maintained.
218 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
227 spin_lock_irqsave(&spool->lock, flags);
229 if (spool->max_hpages != -1) /* maximum size accounting */
230 spool->used_hpages -= delta;
232 /* minimum size accounting */
233 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
234 if (spool->rsv_hpages + delta <= spool->min_hpages)
237 ret = spool->rsv_hpages + delta - spool->min_hpages;
239 spool->rsv_hpages += delta;
240 if (spool->rsv_hpages > spool->min_hpages)
241 spool->rsv_hpages = spool->min_hpages;
245 * If hugetlbfs_put_super couldn't free spool due to an outstanding
246 * quota reference, free it now.
248 unlock_or_release_subpool(spool, flags);
253 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
255 return HUGETLBFS_SB(inode->i_sb)->spool;
258 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
260 return subpool_inode(file_inode(vma->vm_file));
264 * hugetlb vma_lock helper routines
266 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
268 if (__vma_shareable_lock(vma)) {
269 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
271 down_read(&vma_lock->rw_sema);
272 } else if (__vma_private_lock(vma)) {
273 struct resv_map *resv_map = vma_resv_map(vma);
275 down_read(&resv_map->rw_sema);
279 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
281 if (__vma_shareable_lock(vma)) {
282 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
284 up_read(&vma_lock->rw_sema);
285 } else if (__vma_private_lock(vma)) {
286 struct resv_map *resv_map = vma_resv_map(vma);
288 up_read(&resv_map->rw_sema);
292 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
294 if (__vma_shareable_lock(vma)) {
295 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
297 down_write(&vma_lock->rw_sema);
298 } else if (__vma_private_lock(vma)) {
299 struct resv_map *resv_map = vma_resv_map(vma);
301 down_write(&resv_map->rw_sema);
305 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
307 if (__vma_shareable_lock(vma)) {
308 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
310 up_write(&vma_lock->rw_sema);
311 } else if (__vma_private_lock(vma)) {
312 struct resv_map *resv_map = vma_resv_map(vma);
314 up_write(&resv_map->rw_sema);
318 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
321 if (__vma_shareable_lock(vma)) {
322 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
324 return down_write_trylock(&vma_lock->rw_sema);
325 } else if (__vma_private_lock(vma)) {
326 struct resv_map *resv_map = vma_resv_map(vma);
328 return down_write_trylock(&resv_map->rw_sema);
334 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
336 if (__vma_shareable_lock(vma)) {
337 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
339 lockdep_assert_held(&vma_lock->rw_sema);
340 } else if (__vma_private_lock(vma)) {
341 struct resv_map *resv_map = vma_resv_map(vma);
343 lockdep_assert_held(&resv_map->rw_sema);
347 void hugetlb_vma_lock_release(struct kref *kref)
349 struct hugetlb_vma_lock *vma_lock = container_of(kref,
350 struct hugetlb_vma_lock, refs);
355 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
357 struct vm_area_struct *vma = vma_lock->vma;
360 * vma_lock structure may or not be released as a result of put,
361 * it certainly will no longer be attached to vma so clear pointer.
362 * Semaphore synchronizes access to vma_lock->vma field.
364 vma_lock->vma = NULL;
365 vma->vm_private_data = NULL;
366 up_write(&vma_lock->rw_sema);
367 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
370 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
372 if (__vma_shareable_lock(vma)) {
373 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
375 __hugetlb_vma_unlock_write_put(vma_lock);
376 } else if (__vma_private_lock(vma)) {
377 struct resv_map *resv_map = vma_resv_map(vma);
379 /* no free for anon vmas, but still need to unlock */
380 up_write(&resv_map->rw_sema);
384 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
387 * Only present in sharable vmas.
389 if (!vma || !__vma_shareable_lock(vma))
392 if (vma->vm_private_data) {
393 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
395 down_write(&vma_lock->rw_sema);
396 __hugetlb_vma_unlock_write_put(vma_lock);
400 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
402 struct hugetlb_vma_lock *vma_lock;
404 /* Only establish in (flags) sharable vmas */
405 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
408 /* Should never get here with non-NULL vm_private_data */
409 if (vma->vm_private_data)
412 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
415 * If we can not allocate structure, then vma can not
416 * participate in pmd sharing. This is only a possible
417 * performance enhancement and memory saving issue.
418 * However, the lock is also used to synchronize page
419 * faults with truncation. If the lock is not present,
420 * unlikely races could leave pages in a file past i_size
421 * until the file is removed. Warn in the unlikely case of
422 * allocation failure.
424 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
428 kref_init(&vma_lock->refs);
429 init_rwsem(&vma_lock->rw_sema);
431 vma->vm_private_data = vma_lock;
434 /* Helper that removes a struct file_region from the resv_map cache and returns
437 static struct file_region *
438 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
440 struct file_region *nrg;
442 VM_BUG_ON(resv->region_cache_count <= 0);
444 resv->region_cache_count--;
445 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
446 list_del(&nrg->link);
454 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
455 struct file_region *rg)
457 #ifdef CONFIG_CGROUP_HUGETLB
458 nrg->reservation_counter = rg->reservation_counter;
465 /* Helper that records hugetlb_cgroup uncharge info. */
466 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
468 struct resv_map *resv,
469 struct file_region *nrg)
471 #ifdef CONFIG_CGROUP_HUGETLB
473 nrg->reservation_counter =
474 &h_cg->rsvd_hugepage[hstate_index(h)];
475 nrg->css = &h_cg->css;
477 * The caller will hold exactly one h_cg->css reference for the
478 * whole contiguous reservation region. But this area might be
479 * scattered when there are already some file_regions reside in
480 * it. As a result, many file_regions may share only one css
481 * reference. In order to ensure that one file_region must hold
482 * exactly one h_cg->css reference, we should do css_get for
483 * each file_region and leave the reference held by caller
487 if (!resv->pages_per_hpage)
488 resv->pages_per_hpage = pages_per_huge_page(h);
489 /* pages_per_hpage should be the same for all entries in
492 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
494 nrg->reservation_counter = NULL;
500 static void put_uncharge_info(struct file_region *rg)
502 #ifdef CONFIG_CGROUP_HUGETLB
508 static bool has_same_uncharge_info(struct file_region *rg,
509 struct file_region *org)
511 #ifdef CONFIG_CGROUP_HUGETLB
512 return rg->reservation_counter == org->reservation_counter &&
520 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
522 struct file_region *nrg, *prg;
524 prg = list_prev_entry(rg, link);
525 if (&prg->link != &resv->regions && prg->to == rg->from &&
526 has_same_uncharge_info(prg, rg)) {
530 put_uncharge_info(rg);
536 nrg = list_next_entry(rg, link);
537 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
538 has_same_uncharge_info(nrg, rg)) {
539 nrg->from = rg->from;
542 put_uncharge_info(rg);
548 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
549 long to, struct hstate *h, struct hugetlb_cgroup *cg,
550 long *regions_needed)
552 struct file_region *nrg;
554 if (!regions_needed) {
555 nrg = get_file_region_entry_from_cache(map, from, to);
556 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
557 list_add(&nrg->link, rg);
558 coalesce_file_region(map, nrg);
560 *regions_needed += 1;
566 * Must be called with resv->lock held.
568 * Calling this with regions_needed != NULL will count the number of pages
569 * to be added but will not modify the linked list. And regions_needed will
570 * indicate the number of file_regions needed in the cache to carry out to add
571 * the regions for this range.
573 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
574 struct hugetlb_cgroup *h_cg,
575 struct hstate *h, long *regions_needed)
578 struct list_head *head = &resv->regions;
579 long last_accounted_offset = f;
580 struct file_region *iter, *trg = NULL;
581 struct list_head *rg = NULL;
586 /* In this loop, we essentially handle an entry for the range
587 * [last_accounted_offset, iter->from), at every iteration, with some
590 list_for_each_entry_safe(iter, trg, head, link) {
591 /* Skip irrelevant regions that start before our range. */
592 if (iter->from < f) {
593 /* If this region ends after the last accounted offset,
594 * then we need to update last_accounted_offset.
596 if (iter->to > last_accounted_offset)
597 last_accounted_offset = iter->to;
601 /* When we find a region that starts beyond our range, we've
604 if (iter->from >= t) {
605 rg = iter->link.prev;
609 /* Add an entry for last_accounted_offset -> iter->from, and
610 * update last_accounted_offset.
612 if (iter->from > last_accounted_offset)
613 add += hugetlb_resv_map_add(resv, iter->link.prev,
614 last_accounted_offset,
618 last_accounted_offset = iter->to;
621 /* Handle the case where our range extends beyond
622 * last_accounted_offset.
626 if (last_accounted_offset < t)
627 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
628 t, h, h_cg, regions_needed);
633 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
635 static int allocate_file_region_entries(struct resv_map *resv,
637 __must_hold(&resv->lock)
639 LIST_HEAD(allocated_regions);
640 int to_allocate = 0, i = 0;
641 struct file_region *trg = NULL, *rg = NULL;
643 VM_BUG_ON(regions_needed < 0);
646 * Check for sufficient descriptors in the cache to accommodate
647 * the number of in progress add operations plus regions_needed.
649 * This is a while loop because when we drop the lock, some other call
650 * to region_add or region_del may have consumed some region_entries,
651 * so we keep looping here until we finally have enough entries for
652 * (adds_in_progress + regions_needed).
654 while (resv->region_cache_count <
655 (resv->adds_in_progress + regions_needed)) {
656 to_allocate = resv->adds_in_progress + regions_needed -
657 resv->region_cache_count;
659 /* At this point, we should have enough entries in the cache
660 * for all the existing adds_in_progress. We should only be
661 * needing to allocate for regions_needed.
663 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
665 spin_unlock(&resv->lock);
666 for (i = 0; i < to_allocate; i++) {
667 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
670 list_add(&trg->link, &allocated_regions);
673 spin_lock(&resv->lock);
675 list_splice(&allocated_regions, &resv->region_cache);
676 resv->region_cache_count += to_allocate;
682 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
690 * Add the huge page range represented by [f, t) to the reserve
691 * map. Regions will be taken from the cache to fill in this range.
692 * Sufficient regions should exist in the cache due to the previous
693 * call to region_chg with the same range, but in some cases the cache will not
694 * have sufficient entries due to races with other code doing region_add or
695 * region_del. The extra needed entries will be allocated.
697 * regions_needed is the out value provided by a previous call to region_chg.
699 * Return the number of new huge pages added to the map. This number is greater
700 * than or equal to zero. If file_region entries needed to be allocated for
701 * this operation and we were not able to allocate, it returns -ENOMEM.
702 * region_add of regions of length 1 never allocate file_regions and cannot
703 * fail; region_chg will always allocate at least 1 entry and a region_add for
704 * 1 page will only require at most 1 entry.
706 static long region_add(struct resv_map *resv, long f, long t,
707 long in_regions_needed, struct hstate *h,
708 struct hugetlb_cgroup *h_cg)
710 long add = 0, actual_regions_needed = 0;
712 spin_lock(&resv->lock);
715 /* Count how many regions are actually needed to execute this add. */
716 add_reservation_in_range(resv, f, t, NULL, NULL,
717 &actual_regions_needed);
720 * Check for sufficient descriptors in the cache to accommodate
721 * this add operation. Note that actual_regions_needed may be greater
722 * than in_regions_needed, as the resv_map may have been modified since
723 * the region_chg call. In this case, we need to make sure that we
724 * allocate extra entries, such that we have enough for all the
725 * existing adds_in_progress, plus the excess needed for this
728 if (actual_regions_needed > in_regions_needed &&
729 resv->region_cache_count <
730 resv->adds_in_progress +
731 (actual_regions_needed - in_regions_needed)) {
732 /* region_add operation of range 1 should never need to
733 * allocate file_region entries.
735 VM_BUG_ON(t - f <= 1);
737 if (allocate_file_region_entries(
738 resv, actual_regions_needed - in_regions_needed)) {
745 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
747 resv->adds_in_progress -= in_regions_needed;
749 spin_unlock(&resv->lock);
754 * Examine the existing reserve map and determine how many
755 * huge pages in the specified range [f, t) are NOT currently
756 * represented. This routine is called before a subsequent
757 * call to region_add that will actually modify the reserve
758 * map to add the specified range [f, t). region_chg does
759 * not change the number of huge pages represented by the
760 * map. A number of new file_region structures is added to the cache as a
761 * placeholder, for the subsequent region_add call to use. At least 1
762 * file_region structure is added.
764 * out_regions_needed is the number of regions added to the
765 * resv->adds_in_progress. This value needs to be provided to a follow up call
766 * to region_add or region_abort for proper accounting.
768 * Returns the number of huge pages that need to be added to the existing
769 * reservation map for the range [f, t). This number is greater or equal to
770 * zero. -ENOMEM is returned if a new file_region structure or cache entry
771 * is needed and can not be allocated.
773 static long region_chg(struct resv_map *resv, long f, long t,
774 long *out_regions_needed)
778 spin_lock(&resv->lock);
780 /* Count how many hugepages in this range are NOT represented. */
781 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
784 if (*out_regions_needed == 0)
785 *out_regions_needed = 1;
787 if (allocate_file_region_entries(resv, *out_regions_needed))
790 resv->adds_in_progress += *out_regions_needed;
792 spin_unlock(&resv->lock);
797 * Abort the in progress add operation. The adds_in_progress field
798 * of the resv_map keeps track of the operations in progress between
799 * calls to region_chg and region_add. Operations are sometimes
800 * aborted after the call to region_chg. In such cases, region_abort
801 * is called to decrement the adds_in_progress counter. regions_needed
802 * is the value returned by the region_chg call, it is used to decrement
803 * the adds_in_progress counter.
805 * NOTE: The range arguments [f, t) are not needed or used in this
806 * routine. They are kept to make reading the calling code easier as
807 * arguments will match the associated region_chg call.
809 static void region_abort(struct resv_map *resv, long f, long t,
812 spin_lock(&resv->lock);
813 VM_BUG_ON(!resv->region_cache_count);
814 resv->adds_in_progress -= regions_needed;
815 spin_unlock(&resv->lock);
819 * Delete the specified range [f, t) from the reserve map. If the
820 * t parameter is LONG_MAX, this indicates that ALL regions after f
821 * should be deleted. Locate the regions which intersect [f, t)
822 * and either trim, delete or split the existing regions.
824 * Returns the number of huge pages deleted from the reserve map.
825 * In the normal case, the return value is zero or more. In the
826 * case where a region must be split, a new region descriptor must
827 * be allocated. If the allocation fails, -ENOMEM will be returned.
828 * NOTE: If the parameter t == LONG_MAX, then we will never split
829 * a region and possibly return -ENOMEM. Callers specifying
830 * t == LONG_MAX do not need to check for -ENOMEM error.
832 static long region_del(struct resv_map *resv, long f, long t)
834 struct list_head *head = &resv->regions;
835 struct file_region *rg, *trg;
836 struct file_region *nrg = NULL;
840 spin_lock(&resv->lock);
841 list_for_each_entry_safe(rg, trg, head, link) {
843 * Skip regions before the range to be deleted. file_region
844 * ranges are normally of the form [from, to). However, there
845 * may be a "placeholder" entry in the map which is of the form
846 * (from, to) with from == to. Check for placeholder entries
847 * at the beginning of the range to be deleted.
849 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
855 if (f > rg->from && t < rg->to) { /* Must split region */
857 * Check for an entry in the cache before dropping
858 * lock and attempting allocation.
861 resv->region_cache_count > resv->adds_in_progress) {
862 nrg = list_first_entry(&resv->region_cache,
865 list_del(&nrg->link);
866 resv->region_cache_count--;
870 spin_unlock(&resv->lock);
871 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
878 hugetlb_cgroup_uncharge_file_region(
879 resv, rg, t - f, false);
881 /* New entry for end of split region */
885 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
887 INIT_LIST_HEAD(&nrg->link);
889 /* Original entry is trimmed */
892 list_add(&nrg->link, &rg->link);
897 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
898 del += rg->to - rg->from;
899 hugetlb_cgroup_uncharge_file_region(resv, rg,
900 rg->to - rg->from, true);
906 if (f <= rg->from) { /* Trim beginning of region */
907 hugetlb_cgroup_uncharge_file_region(resv, rg,
908 t - rg->from, false);
912 } else { /* Trim end of region */
913 hugetlb_cgroup_uncharge_file_region(resv, rg,
921 spin_unlock(&resv->lock);
927 * A rare out of memory error was encountered which prevented removal of
928 * the reserve map region for a page. The huge page itself was free'ed
929 * and removed from the page cache. This routine will adjust the subpool
930 * usage count, and the global reserve count if needed. By incrementing
931 * these counts, the reserve map entry which could not be deleted will
932 * appear as a "reserved" entry instead of simply dangling with incorrect
935 void hugetlb_fix_reserve_counts(struct inode *inode)
937 struct hugepage_subpool *spool = subpool_inode(inode);
939 bool reserved = false;
941 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
942 if (rsv_adjust > 0) {
943 struct hstate *h = hstate_inode(inode);
945 if (!hugetlb_acct_memory(h, 1))
947 } else if (!rsv_adjust) {
952 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
956 * Count and return the number of huge pages in the reserve map
957 * that intersect with the range [f, t).
959 static long region_count(struct resv_map *resv, long f, long t)
961 struct list_head *head = &resv->regions;
962 struct file_region *rg;
965 spin_lock(&resv->lock);
966 /* Locate each segment we overlap with, and count that overlap. */
967 list_for_each_entry(rg, head, link) {
976 seg_from = max(rg->from, f);
977 seg_to = min(rg->to, t);
979 chg += seg_to - seg_from;
981 spin_unlock(&resv->lock);
987 * Convert the address within this vma to the page offset within
988 * the mapping, huge page units here.
990 static pgoff_t vma_hugecache_offset(struct hstate *h,
991 struct vm_area_struct *vma, unsigned long address)
993 return ((address - vma->vm_start) >> huge_page_shift(h)) +
994 (vma->vm_pgoff >> huge_page_order(h));
998 * vma_kernel_pagesize - Page size granularity for this VMA.
999 * @vma: The user mapping.
1001 * Folios in this VMA will be aligned to, and at least the size of the
1002 * number of bytes returned by this function.
1004 * Return: The default size of the folios allocated when backing a VMA.
1006 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
1008 if (vma->vm_ops && vma->vm_ops->pagesize)
1009 return vma->vm_ops->pagesize(vma);
1012 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
1015 * Return the page size being used by the MMU to back a VMA. In the majority
1016 * of cases, the page size used by the kernel matches the MMU size. On
1017 * architectures where it differs, an architecture-specific 'strong'
1018 * version of this symbol is required.
1020 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
1022 return vma_kernel_pagesize(vma);
1026 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1027 * bits of the reservation map pointer, which are always clear due to
1030 #define HPAGE_RESV_OWNER (1UL << 0)
1031 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1032 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1035 * These helpers are used to track how many pages are reserved for
1036 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1037 * is guaranteed to have their future faults succeed.
1039 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1040 * the reserve counters are updated with the hugetlb_lock held. It is safe
1041 * to reset the VMA at fork() time as it is not in use yet and there is no
1042 * chance of the global counters getting corrupted as a result of the values.
1044 * The private mapping reservation is represented in a subtly different
1045 * manner to a shared mapping. A shared mapping has a region map associated
1046 * with the underlying file, this region map represents the backing file
1047 * pages which have ever had a reservation assigned which this persists even
1048 * after the page is instantiated. A private mapping has a region map
1049 * associated with the original mmap which is attached to all VMAs which
1050 * reference it, this region map represents those offsets which have consumed
1051 * reservation ie. where pages have been instantiated.
1053 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1055 return (unsigned long)vma->vm_private_data;
1058 static void set_vma_private_data(struct vm_area_struct *vma,
1059 unsigned long value)
1061 vma->vm_private_data = (void *)value;
1065 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1066 struct hugetlb_cgroup *h_cg,
1069 #ifdef CONFIG_CGROUP_HUGETLB
1071 resv_map->reservation_counter = NULL;
1072 resv_map->pages_per_hpage = 0;
1073 resv_map->css = NULL;
1075 resv_map->reservation_counter =
1076 &h_cg->rsvd_hugepage[hstate_index(h)];
1077 resv_map->pages_per_hpage = pages_per_huge_page(h);
1078 resv_map->css = &h_cg->css;
1083 struct resv_map *resv_map_alloc(void)
1085 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1086 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1088 if (!resv_map || !rg) {
1094 kref_init(&resv_map->refs);
1095 spin_lock_init(&resv_map->lock);
1096 INIT_LIST_HEAD(&resv_map->regions);
1097 init_rwsem(&resv_map->rw_sema);
1099 resv_map->adds_in_progress = 0;
1101 * Initialize these to 0. On shared mappings, 0's here indicate these
1102 * fields don't do cgroup accounting. On private mappings, these will be
1103 * re-initialized to the proper values, to indicate that hugetlb cgroup
1104 * reservations are to be un-charged from here.
1106 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1108 INIT_LIST_HEAD(&resv_map->region_cache);
1109 list_add(&rg->link, &resv_map->region_cache);
1110 resv_map->region_cache_count = 1;
1115 void resv_map_release(struct kref *ref)
1117 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1118 struct list_head *head = &resv_map->region_cache;
1119 struct file_region *rg, *trg;
1121 /* Clear out any active regions before we release the map. */
1122 region_del(resv_map, 0, LONG_MAX);
1124 /* ... and any entries left in the cache */
1125 list_for_each_entry_safe(rg, trg, head, link) {
1126 list_del(&rg->link);
1130 VM_BUG_ON(resv_map->adds_in_progress);
1135 static inline struct resv_map *inode_resv_map(struct inode *inode)
1138 * At inode evict time, i_mapping may not point to the original
1139 * address space within the inode. This original address space
1140 * contains the pointer to the resv_map. So, always use the
1141 * address space embedded within the inode.
1142 * The VERY common case is inode->mapping == &inode->i_data but,
1143 * this may not be true for device special inodes.
1145 return (struct resv_map *)(&inode->i_data)->i_private_data;
1148 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1150 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1151 if (vma->vm_flags & VM_MAYSHARE) {
1152 struct address_space *mapping = vma->vm_file->f_mapping;
1153 struct inode *inode = mapping->host;
1155 return inode_resv_map(inode);
1158 return (struct resv_map *)(get_vma_private_data(vma) &
1163 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1165 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1166 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1168 set_vma_private_data(vma, (unsigned long)map);
1171 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1173 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1174 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1176 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1179 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1181 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1183 return (get_vma_private_data(vma) & flag) != 0;
1186 bool __vma_private_lock(struct vm_area_struct *vma)
1188 return !(vma->vm_flags & VM_MAYSHARE) &&
1189 get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&
1190 is_vma_resv_set(vma, HPAGE_RESV_OWNER);
1193 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1195 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1197 * Clear vm_private_data
1198 * - For shared mappings this is a per-vma semaphore that may be
1199 * allocated in a subsequent call to hugetlb_vm_op_open.
1200 * Before clearing, make sure pointer is not associated with vma
1201 * as this will leak the structure. This is the case when called
1202 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1203 * been called to allocate a new structure.
1204 * - For MAP_PRIVATE mappings, this is the reserve map which does
1205 * not apply to children. Faults generated by the children are
1206 * not guaranteed to succeed, even if read-only.
1208 if (vma->vm_flags & VM_MAYSHARE) {
1209 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1211 if (vma_lock && vma_lock->vma != vma)
1212 vma->vm_private_data = NULL;
1214 vma->vm_private_data = NULL;
1218 * Reset and decrement one ref on hugepage private reservation.
1219 * Called with mm->mmap_lock writer semaphore held.
1220 * This function should be only used by move_vma() and operate on
1221 * same sized vma. It should never come here with last ref on the
1224 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1227 * Clear the old hugetlb private page reservation.
1228 * It has already been transferred to new_vma.
1230 * During a mremap() operation of a hugetlb vma we call move_vma()
1231 * which copies vma into new_vma and unmaps vma. After the copy
1232 * operation both new_vma and vma share a reference to the resv_map
1233 * struct, and at that point vma is about to be unmapped. We don't
1234 * want to return the reservation to the pool at unmap of vma because
1235 * the reservation still lives on in new_vma, so simply decrement the
1236 * ref here and remove the resv_map reference from this vma.
1238 struct resv_map *reservations = vma_resv_map(vma);
1240 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1241 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1242 kref_put(&reservations->refs, resv_map_release);
1245 hugetlb_dup_vma_private(vma);
1248 /* Returns true if the VMA has associated reserve pages */
1249 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1251 if (vma->vm_flags & VM_NORESERVE) {
1253 * This address is already reserved by other process(chg == 0),
1254 * so, we should decrement reserved count. Without decrementing,
1255 * reserve count remains after releasing inode, because this
1256 * allocated page will go into page cache and is regarded as
1257 * coming from reserved pool in releasing step. Currently, we
1258 * don't have any other solution to deal with this situation
1259 * properly, so add work-around here.
1261 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1267 /* Shared mappings always use reserves */
1268 if (vma->vm_flags & VM_MAYSHARE) {
1270 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1271 * be a region map for all pages. The only situation where
1272 * there is no region map is if a hole was punched via
1273 * fallocate. In this case, there really are no reserves to
1274 * use. This situation is indicated if chg != 0.
1283 * Only the process that called mmap() has reserves for
1286 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1288 * Like the shared case above, a hole punch or truncate
1289 * could have been performed on the private mapping.
1290 * Examine the value of chg to determine if reserves
1291 * actually exist or were previously consumed.
1292 * Very Subtle - The value of chg comes from a previous
1293 * call to vma_needs_reserves(). The reserve map for
1294 * private mappings has different (opposite) semantics
1295 * than that of shared mappings. vma_needs_reserves()
1296 * has already taken this difference in semantics into
1297 * account. Therefore, the meaning of chg is the same
1298 * as in the shared case above. Code could easily be
1299 * combined, but keeping it separate draws attention to
1300 * subtle differences.
1311 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1313 int nid = folio_nid(folio);
1315 lockdep_assert_held(&hugetlb_lock);
1316 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1318 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1319 h->free_huge_pages++;
1320 h->free_huge_pages_node[nid]++;
1321 folio_set_hugetlb_freed(folio);
1324 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1327 struct folio *folio;
1328 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1330 lockdep_assert_held(&hugetlb_lock);
1331 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1332 if (pin && !folio_is_longterm_pinnable(folio))
1335 if (folio_test_hwpoison(folio))
1338 list_move(&folio->lru, &h->hugepage_activelist);
1339 folio_ref_unfreeze(folio, 1);
1340 folio_clear_hugetlb_freed(folio);
1341 h->free_huge_pages--;
1342 h->free_huge_pages_node[nid]--;
1349 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1350 int nid, nodemask_t *nmask)
1352 unsigned int cpuset_mems_cookie;
1353 struct zonelist *zonelist;
1356 int node = NUMA_NO_NODE;
1358 zonelist = node_zonelist(nid, gfp_mask);
1361 cpuset_mems_cookie = read_mems_allowed_begin();
1362 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1363 struct folio *folio;
1365 if (!cpuset_zone_allowed(zone, gfp_mask))
1368 * no need to ask again on the same node. Pool is node rather than
1371 if (zone_to_nid(zone) == node)
1373 node = zone_to_nid(zone);
1375 folio = dequeue_hugetlb_folio_node_exact(h, node);
1379 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1385 static unsigned long available_huge_pages(struct hstate *h)
1387 return h->free_huge_pages - h->resv_huge_pages;
1390 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1391 struct vm_area_struct *vma,
1392 unsigned long address, int avoid_reserve,
1395 struct folio *folio = NULL;
1396 struct mempolicy *mpol;
1398 nodemask_t *nodemask;
1402 * A child process with MAP_PRIVATE mappings created by their parent
1403 * have no page reserves. This check ensures that reservations are
1404 * not "stolen". The child may still get SIGKILLed
1406 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1409 /* If reserves cannot be used, ensure enough pages are in the pool */
1410 if (avoid_reserve && !available_huge_pages(h))
1413 gfp_mask = htlb_alloc_mask(h);
1414 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1416 if (mpol_is_preferred_many(mpol)) {
1417 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1420 /* Fallback to all nodes if page==NULL */
1425 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1428 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1429 folio_set_hugetlb_restore_reserve(folio);
1430 h->resv_huge_pages--;
1433 mpol_cond_put(mpol);
1441 * common helper functions for hstate_next_node_to_{alloc|free}.
1442 * We may have allocated or freed a huge page based on a different
1443 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1444 * be outside of *nodes_allowed. Ensure that we use an allowed
1445 * node for alloc or free.
1447 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1449 nid = next_node_in(nid, *nodes_allowed);
1450 VM_BUG_ON(nid >= MAX_NUMNODES);
1455 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1457 if (!node_isset(nid, *nodes_allowed))
1458 nid = next_node_allowed(nid, nodes_allowed);
1463 * returns the previously saved node ["this node"] from which to
1464 * allocate a persistent huge page for the pool and advance the
1465 * next node from which to allocate, handling wrap at end of node
1468 static int hstate_next_node_to_alloc(int *next_node,
1469 nodemask_t *nodes_allowed)
1473 VM_BUG_ON(!nodes_allowed);
1475 nid = get_valid_node_allowed(*next_node, nodes_allowed);
1476 *next_node = next_node_allowed(nid, nodes_allowed);
1482 * helper for remove_pool_hugetlb_folio() - return the previously saved
1483 * node ["this node"] from which to free a huge page. Advance the
1484 * next node id whether or not we find a free huge page to free so
1485 * that the next attempt to free addresses the next node.
1487 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1491 VM_BUG_ON(!nodes_allowed);
1493 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1494 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1499 #define for_each_node_mask_to_alloc(next_node, nr_nodes, node, mask) \
1500 for (nr_nodes = nodes_weight(*mask); \
1502 ((node = hstate_next_node_to_alloc(next_node, mask)) || 1); \
1505 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1506 for (nr_nodes = nodes_weight(*mask); \
1508 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1511 /* used to demote non-gigantic_huge pages as well */
1512 static void __destroy_compound_gigantic_folio(struct folio *folio,
1513 unsigned int order, bool demote)
1516 int nr_pages = 1 << order;
1519 atomic_set(&folio->_entire_mapcount, 0);
1520 atomic_set(&folio->_nr_pages_mapped, 0);
1521 atomic_set(&folio->_pincount, 0);
1523 for (i = 1; i < nr_pages; i++) {
1524 p = folio_page(folio, i);
1525 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1527 clear_compound_head(p);
1529 set_page_refcounted(p);
1532 __folio_clear_head(folio);
1535 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1538 __destroy_compound_gigantic_folio(folio, order, true);
1541 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1542 static void destroy_compound_gigantic_folio(struct folio *folio,
1545 __destroy_compound_gigantic_folio(folio, order, false);
1548 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1551 * If the page isn't allocated using the cma allocator,
1552 * cma_release() returns false.
1555 int nid = folio_nid(folio);
1557 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1561 free_contig_range(folio_pfn(folio), 1 << order);
1564 #ifdef CONFIG_CONTIG_ALLOC
1565 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1566 int nid, nodemask_t *nodemask)
1569 unsigned long nr_pages = pages_per_huge_page(h);
1570 if (nid == NUMA_NO_NODE)
1571 nid = numa_mem_id();
1577 if (hugetlb_cma[nid]) {
1578 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1579 huge_page_order(h), true);
1581 return page_folio(page);
1584 if (!(gfp_mask & __GFP_THISNODE)) {
1585 for_each_node_mask(node, *nodemask) {
1586 if (node == nid || !hugetlb_cma[node])
1589 page = cma_alloc(hugetlb_cma[node], nr_pages,
1590 huge_page_order(h), true);
1592 return page_folio(page);
1598 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1599 return page ? page_folio(page) : NULL;
1602 #else /* !CONFIG_CONTIG_ALLOC */
1603 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1604 int nid, nodemask_t *nodemask)
1608 #endif /* CONFIG_CONTIG_ALLOC */
1610 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1611 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1612 int nid, nodemask_t *nodemask)
1616 static inline void free_gigantic_folio(struct folio *folio,
1617 unsigned int order) { }
1618 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1619 unsigned int order) { }
1622 static inline void __clear_hugetlb_destructor(struct hstate *h,
1623 struct folio *folio)
1625 lockdep_assert_held(&hugetlb_lock);
1627 __folio_clear_hugetlb(folio);
1631 * Remove hugetlb folio from lists.
1632 * If vmemmap exists for the folio, update dtor so that the folio appears
1633 * as just a compound page. Otherwise, wait until after allocating vmemmap
1636 * A reference is held on the folio, except in the case of demote.
1638 * Must be called with hugetlb lock held.
1640 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1641 bool adjust_surplus,
1644 int nid = folio_nid(folio);
1646 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1647 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1649 lockdep_assert_held(&hugetlb_lock);
1650 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1653 list_del(&folio->lru);
1655 if (folio_test_hugetlb_freed(folio)) {
1656 h->free_huge_pages--;
1657 h->free_huge_pages_node[nid]--;
1659 if (adjust_surplus) {
1660 h->surplus_huge_pages--;
1661 h->surplus_huge_pages_node[nid]--;
1665 * We can only clear the hugetlb destructor after allocating vmemmap
1666 * pages. Otherwise, someone (memory error handling) may try to write
1667 * to tail struct pages.
1669 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1670 __clear_hugetlb_destructor(h, folio);
1673 * In the case of demote we do not ref count the page as it will soon
1674 * be turned into a page of smaller size.
1677 folio_ref_unfreeze(folio, 1);
1680 h->nr_huge_pages_node[nid]--;
1683 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1684 bool adjust_surplus)
1686 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1689 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1690 bool adjust_surplus)
1692 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1695 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1696 bool adjust_surplus)
1699 int nid = folio_nid(folio);
1701 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1703 lockdep_assert_held(&hugetlb_lock);
1705 INIT_LIST_HEAD(&folio->lru);
1707 h->nr_huge_pages_node[nid]++;
1709 if (adjust_surplus) {
1710 h->surplus_huge_pages++;
1711 h->surplus_huge_pages_node[nid]++;
1714 __folio_set_hugetlb(folio);
1715 folio_change_private(folio, NULL);
1717 * We have to set hugetlb_vmemmap_optimized again as above
1718 * folio_change_private(folio, NULL) cleared it.
1720 folio_set_hugetlb_vmemmap_optimized(folio);
1723 * This folio is about to be managed by the hugetlb allocator and
1724 * should have no users. Drop our reference, and check for others
1727 zeroed = folio_put_testzero(folio);
1728 if (unlikely(!zeroed))
1730 * It is VERY unlikely soneone else has taken a ref
1731 * on the folio. In this case, we simply return as
1732 * free_huge_folio() will be called when this other ref
1737 arch_clear_hugepage_flags(&folio->page);
1738 enqueue_hugetlb_folio(h, folio);
1741 static void __update_and_free_hugetlb_folio(struct hstate *h,
1742 struct folio *folio)
1744 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1746 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1750 * If we don't know which subpages are hwpoisoned, we can't free
1751 * the hugepage, so it's leaked intentionally.
1753 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1757 * If folio is not vmemmap optimized (!clear_dtor), then the folio
1758 * is no longer identified as a hugetlb page. hugetlb_vmemmap_restore_folio
1759 * can only be passed hugetlb pages and will BUG otherwise.
1761 if (clear_dtor && hugetlb_vmemmap_restore_folio(h, folio)) {
1762 spin_lock_irq(&hugetlb_lock);
1764 * If we cannot allocate vmemmap pages, just refuse to free the
1765 * page and put the page back on the hugetlb free list and treat
1766 * as a surplus page.
1768 add_hugetlb_folio(h, folio, true);
1769 spin_unlock_irq(&hugetlb_lock);
1774 * Move PageHWPoison flag from head page to the raw error pages,
1775 * which makes any healthy subpages reusable.
1777 if (unlikely(folio_test_hwpoison(folio)))
1778 folio_clear_hugetlb_hwpoison(folio);
1781 * If vmemmap pages were allocated above, then we need to clear the
1782 * hugetlb destructor under the hugetlb lock.
1785 spin_lock_irq(&hugetlb_lock);
1786 __clear_hugetlb_destructor(h, folio);
1787 spin_unlock_irq(&hugetlb_lock);
1791 * Non-gigantic pages demoted from CMA allocated gigantic pages
1792 * need to be given back to CMA in free_gigantic_folio.
1794 if (hstate_is_gigantic(h) ||
1795 hugetlb_cma_folio(folio, huge_page_order(h))) {
1796 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1797 free_gigantic_folio(folio, huge_page_order(h));
1799 __free_pages(&folio->page, huge_page_order(h));
1804 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1805 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1806 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1807 * the vmemmap pages.
1809 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1810 * freed and frees them one-by-one. As the page->mapping pointer is going
1811 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1812 * structure of a lockless linked list of huge pages to be freed.
1814 static LLIST_HEAD(hpage_freelist);
1816 static void free_hpage_workfn(struct work_struct *work)
1818 struct llist_node *node;
1820 node = llist_del_all(&hpage_freelist);
1823 struct folio *folio;
1826 folio = container_of((struct address_space **)node,
1827 struct folio, mapping);
1829 folio->mapping = NULL;
1831 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1832 * folio_hstate() is going to trigger because a previous call to
1833 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1834 * not use folio_hstate() directly.
1836 h = size_to_hstate(folio_size(folio));
1838 __update_and_free_hugetlb_folio(h, folio);
1843 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1845 static inline void flush_free_hpage_work(struct hstate *h)
1847 if (hugetlb_vmemmap_optimizable(h))
1848 flush_work(&free_hpage_work);
1851 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1854 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1855 __update_and_free_hugetlb_folio(h, folio);
1860 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1862 * Only call schedule_work() if hpage_freelist is previously
1863 * empty. Otherwise, schedule_work() had been called but the workfn
1864 * hasn't retrieved the list yet.
1866 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1867 schedule_work(&free_hpage_work);
1870 static void bulk_vmemmap_restore_error(struct hstate *h,
1871 struct list_head *folio_list,
1872 struct list_head *non_hvo_folios)
1874 struct folio *folio, *t_folio;
1876 if (!list_empty(non_hvo_folios)) {
1878 * Free any restored hugetlb pages so that restore of the
1879 * entire list can be retried.
1880 * The idea is that in the common case of ENOMEM errors freeing
1881 * hugetlb pages with vmemmap we will free up memory so that we
1882 * can allocate vmemmap for more hugetlb pages.
1884 list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) {
1885 list_del(&folio->lru);
1886 spin_lock_irq(&hugetlb_lock);
1887 __clear_hugetlb_destructor(h, folio);
1888 spin_unlock_irq(&hugetlb_lock);
1889 update_and_free_hugetlb_folio(h, folio, false);
1894 * In the case where there are no folios which can be
1895 * immediately freed, we loop through the list trying to restore
1896 * vmemmap individually in the hope that someone elsewhere may
1897 * have done something to cause success (such as freeing some
1898 * memory). If unable to restore a hugetlb page, the hugetlb
1899 * page is made a surplus page and removed from the list.
1900 * If are able to restore vmemmap and free one hugetlb page, we
1901 * quit processing the list to retry the bulk operation.
1903 list_for_each_entry_safe(folio, t_folio, folio_list, lru)
1904 if (hugetlb_vmemmap_restore_folio(h, folio)) {
1905 list_del(&folio->lru);
1906 spin_lock_irq(&hugetlb_lock);
1907 add_hugetlb_folio(h, folio, true);
1908 spin_unlock_irq(&hugetlb_lock);
1910 list_del(&folio->lru);
1911 spin_lock_irq(&hugetlb_lock);
1912 __clear_hugetlb_destructor(h, folio);
1913 spin_unlock_irq(&hugetlb_lock);
1914 update_and_free_hugetlb_folio(h, folio, false);
1921 static void update_and_free_pages_bulk(struct hstate *h,
1922 struct list_head *folio_list)
1925 struct folio *folio, *t_folio;
1926 LIST_HEAD(non_hvo_folios);
1929 * First allocate required vmemmmap (if necessary) for all folios.
1930 * Carefully handle errors and free up any available hugetlb pages
1931 * in an effort to make forward progress.
1934 ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios);
1936 bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios);
1941 * At this point, list should be empty, ret should be >= 0 and there
1942 * should only be pages on the non_hvo_folios list.
1943 * Do note that the non_hvo_folios list could be empty.
1944 * Without HVO enabled, ret will be 0 and there is no need to call
1945 * __clear_hugetlb_destructor as this was done previously.
1947 VM_WARN_ON(!list_empty(folio_list));
1948 VM_WARN_ON(ret < 0);
1949 if (!list_empty(&non_hvo_folios) && ret) {
1950 spin_lock_irq(&hugetlb_lock);
1951 list_for_each_entry(folio, &non_hvo_folios, lru)
1952 __clear_hugetlb_destructor(h, folio);
1953 spin_unlock_irq(&hugetlb_lock);
1956 list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) {
1957 update_and_free_hugetlb_folio(h, folio, false);
1962 struct hstate *size_to_hstate(unsigned long size)
1966 for_each_hstate(h) {
1967 if (huge_page_size(h) == size)
1973 void free_huge_folio(struct folio *folio)
1976 * Can't pass hstate in here because it is called from the
1977 * compound page destructor.
1979 struct hstate *h = folio_hstate(folio);
1980 int nid = folio_nid(folio);
1981 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1982 bool restore_reserve;
1983 unsigned long flags;
1985 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1986 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1988 hugetlb_set_folio_subpool(folio, NULL);
1989 if (folio_test_anon(folio))
1990 __ClearPageAnonExclusive(&folio->page);
1991 folio->mapping = NULL;
1992 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1993 folio_clear_hugetlb_restore_reserve(folio);
1996 * If HPageRestoreReserve was set on page, page allocation consumed a
1997 * reservation. If the page was associated with a subpool, there
1998 * would have been a page reserved in the subpool before allocation
1999 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
2000 * reservation, do not call hugepage_subpool_put_pages() as this will
2001 * remove the reserved page from the subpool.
2003 if (!restore_reserve) {
2005 * A return code of zero implies that the subpool will be
2006 * under its minimum size if the reservation is not restored
2007 * after page is free. Therefore, force restore_reserve
2010 if (hugepage_subpool_put_pages(spool, 1) == 0)
2011 restore_reserve = true;
2014 spin_lock_irqsave(&hugetlb_lock, flags);
2015 folio_clear_hugetlb_migratable(folio);
2016 hugetlb_cgroup_uncharge_folio(hstate_index(h),
2017 pages_per_huge_page(h), folio);
2018 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
2019 pages_per_huge_page(h), folio);
2020 mem_cgroup_uncharge(folio);
2021 if (restore_reserve)
2022 h->resv_huge_pages++;
2024 if (folio_test_hugetlb_temporary(folio)) {
2025 remove_hugetlb_folio(h, folio, false);
2026 spin_unlock_irqrestore(&hugetlb_lock, flags);
2027 update_and_free_hugetlb_folio(h, folio, true);
2028 } else if (h->surplus_huge_pages_node[nid]) {
2029 /* remove the page from active list */
2030 remove_hugetlb_folio(h, folio, true);
2031 spin_unlock_irqrestore(&hugetlb_lock, flags);
2032 update_and_free_hugetlb_folio(h, folio, true);
2034 arch_clear_hugepage_flags(&folio->page);
2035 enqueue_hugetlb_folio(h, folio);
2036 spin_unlock_irqrestore(&hugetlb_lock, flags);
2041 * Must be called with the hugetlb lock held
2043 static void __prep_account_new_huge_page(struct hstate *h, int nid)
2045 lockdep_assert_held(&hugetlb_lock);
2047 h->nr_huge_pages_node[nid]++;
2050 static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2052 __folio_set_hugetlb(folio);
2053 INIT_LIST_HEAD(&folio->lru);
2054 hugetlb_set_folio_subpool(folio, NULL);
2055 set_hugetlb_cgroup(folio, NULL);
2056 set_hugetlb_cgroup_rsvd(folio, NULL);
2059 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2061 init_new_hugetlb_folio(h, folio);
2062 hugetlb_vmemmap_optimize_folio(h, folio);
2065 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
2067 __prep_new_hugetlb_folio(h, folio);
2068 spin_lock_irq(&hugetlb_lock);
2069 __prep_account_new_huge_page(h, nid);
2070 spin_unlock_irq(&hugetlb_lock);
2073 static bool __prep_compound_gigantic_folio(struct folio *folio,
2074 unsigned int order, bool demote)
2077 int nr_pages = 1 << order;
2080 __folio_clear_reserved(folio);
2081 for (i = 0; i < nr_pages; i++) {
2082 p = folio_page(folio, i);
2085 * For gigantic hugepages allocated through bootmem at
2086 * boot, it's safer to be consistent with the not-gigantic
2087 * hugepages and clear the PG_reserved bit from all tail pages
2088 * too. Otherwise drivers using get_user_pages() to access tail
2089 * pages may get the reference counting wrong if they see
2090 * PG_reserved set on a tail page (despite the head page not
2091 * having PG_reserved set). Enforcing this consistency between
2092 * head and tail pages allows drivers to optimize away a check
2093 * on the head page when they need know if put_page() is needed
2094 * after get_user_pages().
2096 if (i != 0) /* head page cleared above */
2097 __ClearPageReserved(p);
2099 * Subtle and very unlikely
2101 * Gigantic 'page allocators' such as memblock or cma will
2102 * return a set of pages with each page ref counted. We need
2103 * to turn this set of pages into a compound page with tail
2104 * page ref counts set to zero. Code such as speculative page
2105 * cache adding could take a ref on a 'to be' tail page.
2106 * We need to respect any increased ref count, and only set
2107 * the ref count to zero if count is currently 1. If count
2108 * is not 1, we return an error. An error return indicates
2109 * the set of pages can not be converted to a gigantic page.
2110 * The caller who allocated the pages should then discard the
2111 * pages using the appropriate free interface.
2113 * In the case of demote, the ref count will be zero.
2116 if (!page_ref_freeze(p, 1)) {
2117 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2121 VM_BUG_ON_PAGE(page_count(p), p);
2124 set_compound_head(p, &folio->page);
2126 __folio_set_head(folio);
2127 /* we rely on prep_new_hugetlb_folio to set the destructor */
2128 folio_set_order(folio, order);
2129 atomic_set(&folio->_entire_mapcount, -1);
2130 atomic_set(&folio->_nr_pages_mapped, 0);
2131 atomic_set(&folio->_pincount, 0);
2135 /* undo page modifications made above */
2136 for (j = 0; j < i; j++) {
2137 p = folio_page(folio, j);
2139 clear_compound_head(p);
2140 set_page_refcounted(p);
2142 /* need to clear PG_reserved on remaining tail pages */
2143 for (; j < nr_pages; j++) {
2144 p = folio_page(folio, j);
2145 __ClearPageReserved(p);
2150 static bool prep_compound_gigantic_folio(struct folio *folio,
2153 return __prep_compound_gigantic_folio(folio, order, false);
2156 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2159 return __prep_compound_gigantic_folio(folio, order, true);
2163 * Find and lock address space (mapping) in write mode.
2165 * Upon entry, the page is locked which means that page_mapping() is
2166 * stable. Due to locking order, we can only trylock_write. If we can
2167 * not get the lock, simply return NULL to caller.
2169 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2171 struct address_space *mapping = page_mapping(hpage);
2176 if (i_mmap_trylock_write(mapping))
2182 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2183 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2184 nodemask_t *node_alloc_noretry)
2186 int order = huge_page_order(h);
2188 bool alloc_try_hard = true;
2192 * By default we always try hard to allocate the page with
2193 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2194 * a loop (to adjust global huge page counts) and previous allocation
2195 * failed, do not continue to try hard on the same node. Use the
2196 * node_alloc_noretry bitmap to manage this state information.
2198 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2199 alloc_try_hard = false;
2200 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2202 gfp_mask |= __GFP_RETRY_MAYFAIL;
2203 if (nid == NUMA_NO_NODE)
2204 nid = numa_mem_id();
2206 page = __alloc_pages(gfp_mask, order, nid, nmask);
2208 /* Freeze head page */
2209 if (page && !page_ref_freeze(page, 1)) {
2210 __free_pages(page, order);
2211 if (retry) { /* retry once */
2215 /* WOW! twice in a row. */
2216 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2221 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2222 * indicates an overall state change. Clear bit so that we resume
2223 * normal 'try hard' allocations.
2225 if (node_alloc_noretry && page && !alloc_try_hard)
2226 node_clear(nid, *node_alloc_noretry);
2229 * If we tried hard to get a page but failed, set bit so that
2230 * subsequent attempts will not try as hard until there is an
2231 * overall state change.
2233 if (node_alloc_noretry && !page && alloc_try_hard)
2234 node_set(nid, *node_alloc_noretry);
2237 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2241 __count_vm_event(HTLB_BUDDY_PGALLOC);
2242 return page_folio(page);
2245 static struct folio *__alloc_fresh_hugetlb_folio(struct hstate *h,
2246 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2247 nodemask_t *node_alloc_noretry)
2249 struct folio *folio;
2253 if (hstate_is_gigantic(h))
2254 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2256 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2257 nid, nmask, node_alloc_noretry);
2261 if (hstate_is_gigantic(h)) {
2262 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2264 * Rare failure to convert pages to compound page.
2265 * Free pages and try again - ONCE!
2267 free_gigantic_folio(folio, huge_page_order(h));
2279 static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h,
2280 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2281 nodemask_t *node_alloc_noretry)
2283 struct folio *folio;
2285 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2286 node_alloc_noretry);
2288 init_new_hugetlb_folio(h, folio);
2293 * Common helper to allocate a fresh hugetlb page. All specific allocators
2294 * should use this function to get new hugetlb pages
2296 * Note that returned page is 'frozen': ref count of head page and all tail
2299 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2300 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2301 nodemask_t *node_alloc_noretry)
2303 struct folio *folio;
2305 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2306 node_alloc_noretry);
2310 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2314 static void prep_and_add_allocated_folios(struct hstate *h,
2315 struct list_head *folio_list)
2317 unsigned long flags;
2318 struct folio *folio, *tmp_f;
2320 /* Send list for bulk vmemmap optimization processing */
2321 hugetlb_vmemmap_optimize_folios(h, folio_list);
2323 /* Add all new pool pages to free lists in one lock cycle */
2324 spin_lock_irqsave(&hugetlb_lock, flags);
2325 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
2326 __prep_account_new_huge_page(h, folio_nid(folio));
2327 enqueue_hugetlb_folio(h, folio);
2329 spin_unlock_irqrestore(&hugetlb_lock, flags);
2333 * Allocates a fresh hugetlb page in a node interleaved manner. The page
2334 * will later be added to the appropriate hugetlb pool.
2336 static struct folio *alloc_pool_huge_folio(struct hstate *h,
2337 nodemask_t *nodes_allowed,
2338 nodemask_t *node_alloc_noretry,
2341 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2344 for_each_node_mask_to_alloc(next_node, nr_nodes, node, nodes_allowed) {
2345 struct folio *folio;
2347 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2348 nodes_allowed, node_alloc_noretry);
2357 * Remove huge page from pool from next node to free. Attempt to keep
2358 * persistent huge pages more or less balanced over allowed nodes.
2359 * This routine only 'removes' the hugetlb page. The caller must make
2360 * an additional call to free the page to low level allocators.
2361 * Called with hugetlb_lock locked.
2363 static struct folio *remove_pool_hugetlb_folio(struct hstate *h,
2364 nodemask_t *nodes_allowed, bool acct_surplus)
2367 struct folio *folio = NULL;
2369 lockdep_assert_held(&hugetlb_lock);
2370 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2372 * If we're returning unused surplus pages, only examine
2373 * nodes with surplus pages.
2375 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2376 !list_empty(&h->hugepage_freelists[node])) {
2377 folio = list_entry(h->hugepage_freelists[node].next,
2379 remove_hugetlb_folio(h, folio, acct_surplus);
2388 * Dissolve a given free hugepage into free buddy pages. This function does
2389 * nothing for in-use hugepages and non-hugepages.
2390 * This function returns values like below:
2392 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2393 * when the system is under memory pressure and the feature of
2394 * freeing unused vmemmap pages associated with each hugetlb page
2396 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2397 * (allocated or reserved.)
2398 * 0: successfully dissolved free hugepages or the page is not a
2399 * hugepage (considered as already dissolved)
2401 int dissolve_free_huge_page(struct page *page)
2404 struct folio *folio = page_folio(page);
2407 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2408 if (!folio_test_hugetlb(folio))
2411 spin_lock_irq(&hugetlb_lock);
2412 if (!folio_test_hugetlb(folio)) {
2417 if (!folio_ref_count(folio)) {
2418 struct hstate *h = folio_hstate(folio);
2419 if (!available_huge_pages(h))
2423 * We should make sure that the page is already on the free list
2424 * when it is dissolved.
2426 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2427 spin_unlock_irq(&hugetlb_lock);
2431 * Theoretically, we should return -EBUSY when we
2432 * encounter this race. In fact, we have a chance
2433 * to successfully dissolve the page if we do a
2434 * retry. Because the race window is quite small.
2435 * If we seize this opportunity, it is an optimization
2436 * for increasing the success rate of dissolving page.
2441 remove_hugetlb_folio(h, folio, false);
2442 h->max_huge_pages--;
2443 spin_unlock_irq(&hugetlb_lock);
2446 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2447 * before freeing the page. update_and_free_hugtlb_folio will fail to
2448 * free the page if it can not allocate required vmemmap. We
2449 * need to adjust max_huge_pages if the page is not freed.
2450 * Attempt to allocate vmemmmap here so that we can take
2451 * appropriate action on failure.
2453 * The folio_test_hugetlb check here is because
2454 * remove_hugetlb_folio will clear hugetlb folio flag for
2455 * non-vmemmap optimized hugetlb folios.
2457 if (folio_test_hugetlb(folio)) {
2458 rc = hugetlb_vmemmap_restore_folio(h, folio);
2460 spin_lock_irq(&hugetlb_lock);
2461 add_hugetlb_folio(h, folio, false);
2462 h->max_huge_pages++;
2468 update_and_free_hugetlb_folio(h, folio, false);
2472 spin_unlock_irq(&hugetlb_lock);
2477 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2478 * make specified memory blocks removable from the system.
2479 * Note that this will dissolve a free gigantic hugepage completely, if any
2480 * part of it lies within the given range.
2481 * Also note that if dissolve_free_huge_page() returns with an error, all
2482 * free hugepages that were dissolved before that error are lost.
2484 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2492 if (!hugepages_supported())
2495 order = huge_page_order(&default_hstate);
2497 order = min(order, huge_page_order(h));
2499 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2500 page = pfn_to_page(pfn);
2501 rc = dissolve_free_huge_page(page);
2510 * Allocates a fresh surplus page from the page allocator.
2512 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2513 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2515 struct folio *folio = NULL;
2517 if (hstate_is_gigantic(h))
2520 spin_lock_irq(&hugetlb_lock);
2521 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2523 spin_unlock_irq(&hugetlb_lock);
2525 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2529 spin_lock_irq(&hugetlb_lock);
2531 * We could have raced with the pool size change.
2532 * Double check that and simply deallocate the new page
2533 * if we would end up overcommiting the surpluses. Abuse
2534 * temporary page to workaround the nasty free_huge_folio
2537 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2538 folio_set_hugetlb_temporary(folio);
2539 spin_unlock_irq(&hugetlb_lock);
2540 free_huge_folio(folio);
2544 h->surplus_huge_pages++;
2545 h->surplus_huge_pages_node[folio_nid(folio)]++;
2548 spin_unlock_irq(&hugetlb_lock);
2553 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2554 int nid, nodemask_t *nmask)
2556 struct folio *folio;
2558 if (hstate_is_gigantic(h))
2561 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2565 /* fresh huge pages are frozen */
2566 folio_ref_unfreeze(folio, 1);
2568 * We do not account these pages as surplus because they are only
2569 * temporary and will be released properly on the last reference
2571 folio_set_hugetlb_temporary(folio);
2577 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2580 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2581 struct vm_area_struct *vma, unsigned long addr)
2583 struct folio *folio = NULL;
2584 struct mempolicy *mpol;
2585 gfp_t gfp_mask = htlb_alloc_mask(h);
2587 nodemask_t *nodemask;
2589 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2590 if (mpol_is_preferred_many(mpol)) {
2591 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2593 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2594 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2596 /* Fallback to all nodes if page==NULL */
2601 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2602 mpol_cond_put(mpol);
2606 /* folio migration callback function */
2607 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2608 nodemask_t *nmask, gfp_t gfp_mask)
2610 spin_lock_irq(&hugetlb_lock);
2611 if (available_huge_pages(h)) {
2612 struct folio *folio;
2614 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2615 preferred_nid, nmask);
2617 spin_unlock_irq(&hugetlb_lock);
2621 spin_unlock_irq(&hugetlb_lock);
2623 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2627 * Increase the hugetlb pool such that it can accommodate a reservation
2630 static int gather_surplus_pages(struct hstate *h, long delta)
2631 __must_hold(&hugetlb_lock)
2633 LIST_HEAD(surplus_list);
2634 struct folio *folio, *tmp;
2637 long needed, allocated;
2638 bool alloc_ok = true;
2640 lockdep_assert_held(&hugetlb_lock);
2641 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2643 h->resv_huge_pages += delta;
2651 spin_unlock_irq(&hugetlb_lock);
2652 for (i = 0; i < needed; i++) {
2653 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2654 NUMA_NO_NODE, NULL);
2659 list_add(&folio->lru, &surplus_list);
2665 * After retaking hugetlb_lock, we need to recalculate 'needed'
2666 * because either resv_huge_pages or free_huge_pages may have changed.
2668 spin_lock_irq(&hugetlb_lock);
2669 needed = (h->resv_huge_pages + delta) -
2670 (h->free_huge_pages + allocated);
2675 * We were not able to allocate enough pages to
2676 * satisfy the entire reservation so we free what
2677 * we've allocated so far.
2682 * The surplus_list now contains _at_least_ the number of extra pages
2683 * needed to accommodate the reservation. Add the appropriate number
2684 * of pages to the hugetlb pool and free the extras back to the buddy
2685 * allocator. Commit the entire reservation here to prevent another
2686 * process from stealing the pages as they are added to the pool but
2687 * before they are reserved.
2689 needed += allocated;
2690 h->resv_huge_pages += delta;
2693 /* Free the needed pages to the hugetlb pool */
2694 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2697 /* Add the page to the hugetlb allocator */
2698 enqueue_hugetlb_folio(h, folio);
2701 spin_unlock_irq(&hugetlb_lock);
2704 * Free unnecessary surplus pages to the buddy allocator.
2705 * Pages have no ref count, call free_huge_folio directly.
2707 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2708 free_huge_folio(folio);
2709 spin_lock_irq(&hugetlb_lock);
2715 * This routine has two main purposes:
2716 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2717 * in unused_resv_pages. This corresponds to the prior adjustments made
2718 * to the associated reservation map.
2719 * 2) Free any unused surplus pages that may have been allocated to satisfy
2720 * the reservation. As many as unused_resv_pages may be freed.
2722 static void return_unused_surplus_pages(struct hstate *h,
2723 unsigned long unused_resv_pages)
2725 unsigned long nr_pages;
2726 LIST_HEAD(page_list);
2728 lockdep_assert_held(&hugetlb_lock);
2729 /* Uncommit the reservation */
2730 h->resv_huge_pages -= unused_resv_pages;
2732 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2736 * Part (or even all) of the reservation could have been backed
2737 * by pre-allocated pages. Only free surplus pages.
2739 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2742 * We want to release as many surplus pages as possible, spread
2743 * evenly across all nodes with memory. Iterate across these nodes
2744 * until we can no longer free unreserved surplus pages. This occurs
2745 * when the nodes with surplus pages have no free pages.
2746 * remove_pool_hugetlb_folio() will balance the freed pages across the
2747 * on-line nodes with memory and will handle the hstate accounting.
2749 while (nr_pages--) {
2750 struct folio *folio;
2752 folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1);
2756 list_add(&folio->lru, &page_list);
2760 spin_unlock_irq(&hugetlb_lock);
2761 update_and_free_pages_bulk(h, &page_list);
2762 spin_lock_irq(&hugetlb_lock);
2767 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2768 * are used by the huge page allocation routines to manage reservations.
2770 * vma_needs_reservation is called to determine if the huge page at addr
2771 * within the vma has an associated reservation. If a reservation is
2772 * needed, the value 1 is returned. The caller is then responsible for
2773 * managing the global reservation and subpool usage counts. After
2774 * the huge page has been allocated, vma_commit_reservation is called
2775 * to add the page to the reservation map. If the page allocation fails,
2776 * the reservation must be ended instead of committed. vma_end_reservation
2777 * is called in such cases.
2779 * In the normal case, vma_commit_reservation returns the same value
2780 * as the preceding vma_needs_reservation call. The only time this
2781 * is not the case is if a reserve map was changed between calls. It
2782 * is the responsibility of the caller to notice the difference and
2783 * take appropriate action.
2785 * vma_add_reservation is used in error paths where a reservation must
2786 * be restored when a newly allocated huge page must be freed. It is
2787 * to be called after calling vma_needs_reservation to determine if a
2788 * reservation exists.
2790 * vma_del_reservation is used in error paths where an entry in the reserve
2791 * map was created during huge page allocation and must be removed. It is to
2792 * be called after calling vma_needs_reservation to determine if a reservation
2795 enum vma_resv_mode {
2802 static long __vma_reservation_common(struct hstate *h,
2803 struct vm_area_struct *vma, unsigned long addr,
2804 enum vma_resv_mode mode)
2806 struct resv_map *resv;
2809 long dummy_out_regions_needed;
2811 resv = vma_resv_map(vma);
2815 idx = vma_hugecache_offset(h, vma, addr);
2817 case VMA_NEEDS_RESV:
2818 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2819 /* We assume that vma_reservation_* routines always operate on
2820 * 1 page, and that adding to resv map a 1 page entry can only
2821 * ever require 1 region.
2823 VM_BUG_ON(dummy_out_regions_needed != 1);
2825 case VMA_COMMIT_RESV:
2826 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2827 /* region_add calls of range 1 should never fail. */
2831 region_abort(resv, idx, idx + 1, 1);
2835 if (vma->vm_flags & VM_MAYSHARE) {
2836 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2837 /* region_add calls of range 1 should never fail. */
2840 region_abort(resv, idx, idx + 1, 1);
2841 ret = region_del(resv, idx, idx + 1);
2845 if (vma->vm_flags & VM_MAYSHARE) {
2846 region_abort(resv, idx, idx + 1, 1);
2847 ret = region_del(resv, idx, idx + 1);
2849 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2850 /* region_add calls of range 1 should never fail. */
2858 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2861 * We know private mapping must have HPAGE_RESV_OWNER set.
2863 * In most cases, reserves always exist for private mappings.
2864 * However, a file associated with mapping could have been
2865 * hole punched or truncated after reserves were consumed.
2866 * As subsequent fault on such a range will not use reserves.
2867 * Subtle - The reserve map for private mappings has the
2868 * opposite meaning than that of shared mappings. If NO
2869 * entry is in the reserve map, it means a reservation exists.
2870 * If an entry exists in the reserve map, it means the
2871 * reservation has already been consumed. As a result, the
2872 * return value of this routine is the opposite of the
2873 * value returned from reserve map manipulation routines above.
2882 static long vma_needs_reservation(struct hstate *h,
2883 struct vm_area_struct *vma, unsigned long addr)
2885 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2888 static long vma_commit_reservation(struct hstate *h,
2889 struct vm_area_struct *vma, unsigned long addr)
2891 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2894 static void vma_end_reservation(struct hstate *h,
2895 struct vm_area_struct *vma, unsigned long addr)
2897 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2900 static long vma_add_reservation(struct hstate *h,
2901 struct vm_area_struct *vma, unsigned long addr)
2903 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2906 static long vma_del_reservation(struct hstate *h,
2907 struct vm_area_struct *vma, unsigned long addr)
2909 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2913 * This routine is called to restore reservation information on error paths.
2914 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2915 * and the hugetlb mutex should remain held when calling this routine.
2917 * It handles two specific cases:
2918 * 1) A reservation was in place and the folio consumed the reservation.
2919 * hugetlb_restore_reserve is set in the folio.
2920 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2921 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2923 * In case 1, free_huge_folio later in the error path will increment the
2924 * global reserve count. But, free_huge_folio does not have enough context
2925 * to adjust the reservation map. This case deals primarily with private
2926 * mappings. Adjust the reserve map here to be consistent with global
2927 * reserve count adjustments to be made by free_huge_folio. Make sure the
2928 * reserve map indicates there is a reservation present.
2930 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2932 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2933 unsigned long address, struct folio *folio)
2935 long rc = vma_needs_reservation(h, vma, address);
2937 if (folio_test_hugetlb_restore_reserve(folio)) {
2938 if (unlikely(rc < 0))
2940 * Rare out of memory condition in reserve map
2941 * manipulation. Clear hugetlb_restore_reserve so
2942 * that global reserve count will not be incremented
2943 * by free_huge_folio. This will make it appear
2944 * as though the reservation for this folio was
2945 * consumed. This may prevent the task from
2946 * faulting in the folio at a later time. This
2947 * is better than inconsistent global huge page
2948 * accounting of reserve counts.
2950 folio_clear_hugetlb_restore_reserve(folio);
2952 (void)vma_add_reservation(h, vma, address);
2954 vma_end_reservation(h, vma, address);
2958 * This indicates there is an entry in the reserve map
2959 * not added by alloc_hugetlb_folio. We know it was added
2960 * before the alloc_hugetlb_folio call, otherwise
2961 * hugetlb_restore_reserve would be set on the folio.
2962 * Remove the entry so that a subsequent allocation
2963 * does not consume a reservation.
2965 rc = vma_del_reservation(h, vma, address);
2968 * VERY rare out of memory condition. Since
2969 * we can not delete the entry, set
2970 * hugetlb_restore_reserve so that the reserve
2971 * count will be incremented when the folio
2972 * is freed. This reserve will be consumed
2973 * on a subsequent allocation.
2975 folio_set_hugetlb_restore_reserve(folio);
2976 } else if (rc < 0) {
2978 * Rare out of memory condition from
2979 * vma_needs_reservation call. Memory allocation is
2980 * only attempted if a new entry is needed. Therefore,
2981 * this implies there is not an entry in the
2984 * For shared mappings, no entry in the map indicates
2985 * no reservation. We are done.
2987 if (!(vma->vm_flags & VM_MAYSHARE))
2989 * For private mappings, no entry indicates
2990 * a reservation is present. Since we can
2991 * not add an entry, set hugetlb_restore_reserve
2992 * on the folio so reserve count will be
2993 * incremented when freed. This reserve will
2994 * be consumed on a subsequent allocation.
2996 folio_set_hugetlb_restore_reserve(folio);
2999 * No reservation present, do nothing
3001 vma_end_reservation(h, vma, address);
3006 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
3008 * @h: struct hstate old page belongs to
3009 * @old_folio: Old folio to dissolve
3010 * @list: List to isolate the page in case we need to
3011 * Returns 0 on success, otherwise negated error.
3013 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
3014 struct folio *old_folio, struct list_head *list)
3016 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3017 int nid = folio_nid(old_folio);
3018 struct folio *new_folio = NULL;
3022 spin_lock_irq(&hugetlb_lock);
3023 if (!folio_test_hugetlb(old_folio)) {
3025 * Freed from under us. Drop new_folio too.
3028 } else if (folio_ref_count(old_folio)) {
3032 * Someone has grabbed the folio, try to isolate it here.
3033 * Fail with -EBUSY if not possible.
3035 spin_unlock_irq(&hugetlb_lock);
3036 isolated = isolate_hugetlb(old_folio, list);
3037 ret = isolated ? 0 : -EBUSY;
3038 spin_lock_irq(&hugetlb_lock);
3040 } else if (!folio_test_hugetlb_freed(old_folio)) {
3042 * Folio's refcount is 0 but it has not been enqueued in the
3043 * freelist yet. Race window is small, so we can succeed here if
3046 spin_unlock_irq(&hugetlb_lock);
3051 spin_unlock_irq(&hugetlb_lock);
3052 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid,
3056 __prep_new_hugetlb_folio(h, new_folio);
3061 * Ok, old_folio is still a genuine free hugepage. Remove it from
3062 * the freelist and decrease the counters. These will be
3063 * incremented again when calling __prep_account_new_huge_page()
3064 * and enqueue_hugetlb_folio() for new_folio. The counters will
3065 * remain stable since this happens under the lock.
3067 remove_hugetlb_folio(h, old_folio, false);
3070 * Ref count on new_folio is already zero as it was dropped
3071 * earlier. It can be directly added to the pool free list.
3073 __prep_account_new_huge_page(h, nid);
3074 enqueue_hugetlb_folio(h, new_folio);
3077 * Folio has been replaced, we can safely free the old one.
3079 spin_unlock_irq(&hugetlb_lock);
3080 update_and_free_hugetlb_folio(h, old_folio, false);
3086 spin_unlock_irq(&hugetlb_lock);
3088 /* Folio has a zero ref count, but needs a ref to be freed */
3089 folio_ref_unfreeze(new_folio, 1);
3090 update_and_free_hugetlb_folio(h, new_folio, false);
3096 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3099 struct folio *folio = page_folio(page);
3103 * The page might have been dissolved from under our feet, so make sure
3104 * to carefully check the state under the lock.
3105 * Return success when racing as if we dissolved the page ourselves.
3107 spin_lock_irq(&hugetlb_lock);
3108 if (folio_test_hugetlb(folio)) {
3109 h = folio_hstate(folio);
3111 spin_unlock_irq(&hugetlb_lock);
3114 spin_unlock_irq(&hugetlb_lock);
3117 * Fence off gigantic pages as there is a cyclic dependency between
3118 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3119 * of bailing out right away without further retrying.
3121 if (hstate_is_gigantic(h))
3124 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3126 else if (!folio_ref_count(folio))
3127 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3132 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3133 unsigned long addr, int avoid_reserve)
3135 struct hugepage_subpool *spool = subpool_vma(vma);
3136 struct hstate *h = hstate_vma(vma);
3137 struct folio *folio;
3138 long map_chg, map_commit, nr_pages = pages_per_huge_page(h);
3140 int memcg_charge_ret, ret, idx;
3141 struct hugetlb_cgroup *h_cg = NULL;
3142 struct mem_cgroup *memcg;
3143 bool deferred_reserve;
3144 gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL;
3146 memcg = get_mem_cgroup_from_current();
3147 memcg_charge_ret = mem_cgroup_hugetlb_try_charge(memcg, gfp, nr_pages);
3148 if (memcg_charge_ret == -ENOMEM) {
3149 mem_cgroup_put(memcg);
3150 return ERR_PTR(-ENOMEM);
3153 idx = hstate_index(h);
3155 * Examine the region/reserve map to determine if the process
3156 * has a reservation for the page to be allocated. A return
3157 * code of zero indicates a reservation exists (no change).
3159 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3161 if (!memcg_charge_ret)
3162 mem_cgroup_cancel_charge(memcg, nr_pages);
3163 mem_cgroup_put(memcg);
3164 return ERR_PTR(-ENOMEM);
3168 * Processes that did not create the mapping will have no
3169 * reserves as indicated by the region/reserve map. Check
3170 * that the allocation will not exceed the subpool limit.
3171 * Allocations for MAP_NORESERVE mappings also need to be
3172 * checked against any subpool limit.
3174 if (map_chg || avoid_reserve) {
3175 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3177 goto out_end_reservation;
3180 * Even though there was no reservation in the region/reserve
3181 * map, there could be reservations associated with the
3182 * subpool that can be used. This would be indicated if the
3183 * return value of hugepage_subpool_get_pages() is zero.
3184 * However, if avoid_reserve is specified we still avoid even
3185 * the subpool reservations.
3191 /* If this allocation is not consuming a reservation, charge it now.
3193 deferred_reserve = map_chg || avoid_reserve;
3194 if (deferred_reserve) {
3195 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3196 idx, pages_per_huge_page(h), &h_cg);
3198 goto out_subpool_put;
3201 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3203 goto out_uncharge_cgroup_reservation;
3205 spin_lock_irq(&hugetlb_lock);
3207 * glb_chg is passed to indicate whether or not a page must be taken
3208 * from the global free pool (global change). gbl_chg == 0 indicates
3209 * a reservation exists for the allocation.
3211 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3213 spin_unlock_irq(&hugetlb_lock);
3214 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3216 goto out_uncharge_cgroup;
3217 spin_lock_irq(&hugetlb_lock);
3218 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3219 folio_set_hugetlb_restore_reserve(folio);
3220 h->resv_huge_pages--;
3222 list_add(&folio->lru, &h->hugepage_activelist);
3223 folio_ref_unfreeze(folio, 1);
3227 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3228 /* If allocation is not consuming a reservation, also store the
3229 * hugetlb_cgroup pointer on the page.
3231 if (deferred_reserve) {
3232 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3236 spin_unlock_irq(&hugetlb_lock);
3238 hugetlb_set_folio_subpool(folio, spool);
3240 map_commit = vma_commit_reservation(h, vma, addr);
3241 if (unlikely(map_chg > map_commit)) {
3243 * The page was added to the reservation map between
3244 * vma_needs_reservation and vma_commit_reservation.
3245 * This indicates a race with hugetlb_reserve_pages.
3246 * Adjust for the subpool count incremented above AND
3247 * in hugetlb_reserve_pages for the same page. Also,
3248 * the reservation count added in hugetlb_reserve_pages
3249 * no longer applies.
3253 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3254 hugetlb_acct_memory(h, -rsv_adjust);
3255 if (deferred_reserve) {
3256 spin_lock_irq(&hugetlb_lock);
3257 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3258 pages_per_huge_page(h), folio);
3259 spin_unlock_irq(&hugetlb_lock);
3263 if (!memcg_charge_ret)
3264 mem_cgroup_commit_charge(folio, memcg);
3265 mem_cgroup_put(memcg);
3269 out_uncharge_cgroup:
3270 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3271 out_uncharge_cgroup_reservation:
3272 if (deferred_reserve)
3273 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3276 if (map_chg || avoid_reserve)
3277 hugepage_subpool_put_pages(spool, 1);
3278 out_end_reservation:
3279 vma_end_reservation(h, vma, addr);
3280 if (!memcg_charge_ret)
3281 mem_cgroup_cancel_charge(memcg, nr_pages);
3282 mem_cgroup_put(memcg);
3283 return ERR_PTR(-ENOSPC);
3286 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3287 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3288 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3290 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3291 int nr_nodes, node = nid;
3293 /* do node specific alloc */
3294 if (nid != NUMA_NO_NODE) {
3295 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3296 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3301 /* allocate from next node when distributing huge pages */
3302 for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, &node_states[N_MEMORY]) {
3303 m = memblock_alloc_try_nid_raw(
3304 huge_page_size(h), huge_page_size(h),
3305 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3307 * Use the beginning of the huge page to store the
3308 * huge_bootmem_page struct (until gather_bootmem
3309 * puts them into the mem_map).
3319 * Only initialize the head struct page in memmap_init_reserved_pages,
3320 * rest of the struct pages will be initialized by the HugeTLB
3322 * The head struct page is used to get folio information by the HugeTLB
3323 * subsystem like zone id and node id.
3325 memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE),
3326 huge_page_size(h) - PAGE_SIZE);
3327 /* Put them into a private list first because mem_map is not up yet */
3328 INIT_LIST_HEAD(&m->list);
3329 list_add(&m->list, &huge_boot_pages[node]);
3334 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
3335 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
3336 unsigned long start_page_number,
3337 unsigned long end_page_number)
3339 enum zone_type zone = zone_idx(folio_zone(folio));
3340 int nid = folio_nid(folio);
3341 unsigned long head_pfn = folio_pfn(folio);
3342 unsigned long pfn, end_pfn = head_pfn + end_page_number;
3345 for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) {
3346 struct page *page = pfn_to_page(pfn);
3348 __init_single_page(page, pfn, zone, nid);
3349 prep_compound_tail((struct page *)folio, pfn - head_pfn);
3350 ret = page_ref_freeze(page, 1);
3355 static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
3357 unsigned long nr_pages)
3361 /* Prepare folio head */
3362 __folio_clear_reserved(folio);
3363 __folio_set_head(folio);
3364 ret = folio_ref_freeze(folio, 1);
3366 /* Initialize the necessary tail struct pages */
3367 hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);
3368 prep_compound_head((struct page *)folio, huge_page_order(h));
3371 static void __init prep_and_add_bootmem_folios(struct hstate *h,
3372 struct list_head *folio_list)
3374 unsigned long flags;
3375 struct folio *folio, *tmp_f;
3377 /* Send list for bulk vmemmap optimization processing */
3378 hugetlb_vmemmap_optimize_folios(h, folio_list);
3380 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
3381 if (!folio_test_hugetlb_vmemmap_optimized(folio)) {
3383 * If HVO fails, initialize all tail struct pages
3384 * We do not worry about potential long lock hold
3385 * time as this is early in boot and there should
3388 hugetlb_folio_init_tail_vmemmap(folio,
3389 HUGETLB_VMEMMAP_RESERVE_PAGES,
3390 pages_per_huge_page(h));
3392 /* Subdivide locks to achieve better parallel performance */
3393 spin_lock_irqsave(&hugetlb_lock, flags);
3394 __prep_account_new_huge_page(h, folio_nid(folio));
3395 enqueue_hugetlb_folio(h, folio);
3396 spin_unlock_irqrestore(&hugetlb_lock, flags);
3401 * Put bootmem huge pages into the standard lists after mem_map is up.
3402 * Note: This only applies to gigantic (order > MAX_PAGE_ORDER) pages.
3404 static void __init gather_bootmem_prealloc_node(unsigned long nid)
3406 LIST_HEAD(folio_list);
3407 struct huge_bootmem_page *m;
3408 struct hstate *h = NULL, *prev_h = NULL;
3410 list_for_each_entry(m, &huge_boot_pages[nid], list) {
3411 struct page *page = virt_to_page(m);
3412 struct folio *folio = (void *)page;
3416 * It is possible to have multiple huge page sizes (hstates)
3417 * in this list. If so, process each size separately.
3419 if (h != prev_h && prev_h != NULL)
3420 prep_and_add_bootmem_folios(prev_h, &folio_list);
3423 VM_BUG_ON(!hstate_is_gigantic(h));
3424 WARN_ON(folio_ref_count(folio) != 1);
3426 hugetlb_folio_init_vmemmap(folio, h,
3427 HUGETLB_VMEMMAP_RESERVE_PAGES);
3428 init_new_hugetlb_folio(h, folio);
3429 list_add(&folio->lru, &folio_list);
3432 * We need to restore the 'stolen' pages to totalram_pages
3433 * in order to fix confusing memory reports from free(1) and
3434 * other side-effects, like CommitLimit going negative.
3436 adjust_managed_page_count(page, pages_per_huge_page(h));
3440 prep_and_add_bootmem_folios(h, &folio_list);
3443 static void __init gather_bootmem_prealloc_parallel(unsigned long start,
3444 unsigned long end, void *arg)
3448 for (nid = start; nid < end; nid++)
3449 gather_bootmem_prealloc_node(nid);
3452 static void __init gather_bootmem_prealloc(void)
3454 struct padata_mt_job job = {
3455 .thread_fn = gather_bootmem_prealloc_parallel,
3458 .size = num_node_state(N_MEMORY),
3461 .max_threads = num_node_state(N_MEMORY),
3465 padata_do_multithreaded(&job);
3468 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3473 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3474 if (hstate_is_gigantic(h)) {
3475 if (!alloc_bootmem_huge_page(h, nid))
3478 struct folio *folio;
3479 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3481 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3482 &node_states[N_MEMORY], NULL);
3485 free_huge_folio(folio); /* free it into the hugepage allocator */
3489 if (i == h->max_huge_pages_node[nid])
3492 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3493 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3494 h->max_huge_pages_node[nid], buf, nid, i);
3495 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3496 h->max_huge_pages_node[nid] = i;
3499 static bool __init hugetlb_hstate_alloc_pages_specific_nodes(struct hstate *h)
3502 bool node_specific_alloc = false;
3504 for_each_online_node(i) {
3505 if (h->max_huge_pages_node[i] > 0) {
3506 hugetlb_hstate_alloc_pages_onenode(h, i);
3507 node_specific_alloc = true;
3511 return node_specific_alloc;
3514 static void __init hugetlb_hstate_alloc_pages_errcheck(unsigned long allocated, struct hstate *h)
3516 if (allocated < h->max_huge_pages) {
3519 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3520 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3521 h->max_huge_pages, buf, allocated);
3522 h->max_huge_pages = allocated;
3526 static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned long end, void *arg)
3528 struct hstate *h = (struct hstate *)arg;
3529 int i, num = end - start;
3530 nodemask_t node_alloc_noretry;
3531 LIST_HEAD(folio_list);
3532 int next_node = first_online_node;
3534 /* Bit mask controlling how hard we retry per-node allocations.*/
3535 nodes_clear(node_alloc_noretry);
3537 for (i = 0; i < num; ++i) {
3538 struct folio *folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
3539 &node_alloc_noretry, &next_node);
3543 list_move(&folio->lru, &folio_list);
3547 prep_and_add_allocated_folios(h, &folio_list);
3550 static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h)
3554 for (i = 0; i < h->max_huge_pages; ++i) {
3555 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3563 static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h)
3565 struct padata_mt_job job = {
3571 job.thread_fn = hugetlb_pages_alloc_boot_node;
3573 job.size = h->max_huge_pages;
3576 * job.max_threads is twice the num_node_state(N_MEMORY),
3578 * Tests below indicate that a multiplier of 2 significantly improves
3579 * performance, and although larger values also provide improvements,
3580 * the gains are marginal.
3582 * Therefore, choosing 2 as the multiplier strikes a good balance between
3583 * enhancing parallel processing capabilities and maintaining efficient
3584 * resource management.
3586 * +------------+-------+-------+-------+-------+-------+
3587 * | multiplier | 1 | 2 | 3 | 4 | 5 |
3588 * +------------+-------+-------+-------+-------+-------+
3589 * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms |
3590 * | 2T 4node | 979ms | 679ms | 543ms | 489ms | 481ms |
3591 * | 50G 2node | 71ms | 44ms | 37ms | 30ms | 31ms |
3592 * +------------+-------+-------+-------+-------+-------+
3594 job.max_threads = num_node_state(N_MEMORY) * 2;
3595 job.min_chunk = h->max_huge_pages / num_node_state(N_MEMORY) / 2;
3596 padata_do_multithreaded(&job);
3598 return h->nr_huge_pages;
3602 * NOTE: this routine is called in different contexts for gigantic and
3603 * non-gigantic pages.
3604 * - For gigantic pages, this is called early in the boot process and
3605 * pages are allocated from memblock allocated or something similar.
3606 * Gigantic pages are actually added to pools later with the routine
3607 * gather_bootmem_prealloc.
3608 * - For non-gigantic pages, this is called later in the boot process after
3609 * all of mm is up and functional. Pages are allocated from buddy and
3610 * then added to hugetlb pools.
3612 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3614 unsigned long allocated;
3615 static bool initialized __initdata;
3617 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3618 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3619 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3623 /* hugetlb_hstate_alloc_pages will be called many times, initialize huge_boot_pages once */
3627 for (i = 0; i < MAX_NUMNODES; i++)
3628 INIT_LIST_HEAD(&huge_boot_pages[i]);
3632 /* do node specific alloc */
3633 if (hugetlb_hstate_alloc_pages_specific_nodes(h))
3636 /* below will do all node balanced alloc */
3637 if (hstate_is_gigantic(h))
3638 allocated = hugetlb_gigantic_pages_alloc_boot(h);
3640 allocated = hugetlb_pages_alloc_boot(h);
3642 hugetlb_hstate_alloc_pages_errcheck(allocated, h);
3645 static void __init hugetlb_init_hstates(void)
3647 struct hstate *h, *h2;
3649 for_each_hstate(h) {
3650 /* oversize hugepages were init'ed in early boot */
3651 if (!hstate_is_gigantic(h))
3652 hugetlb_hstate_alloc_pages(h);
3655 * Set demote order for each hstate. Note that
3656 * h->demote_order is initially 0.
3657 * - We can not demote gigantic pages if runtime freeing
3658 * is not supported, so skip this.
3659 * - If CMA allocation is possible, we can not demote
3660 * HUGETLB_PAGE_ORDER or smaller size pages.
3662 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3664 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3666 for_each_hstate(h2) {
3669 if (h2->order < h->order &&
3670 h2->order > h->demote_order)
3671 h->demote_order = h2->order;
3676 static void __init report_hugepages(void)
3680 for_each_hstate(h) {
3683 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3684 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3685 buf, h->free_huge_pages);
3686 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3687 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3691 #ifdef CONFIG_HIGHMEM
3692 static void try_to_free_low(struct hstate *h, unsigned long count,
3693 nodemask_t *nodes_allowed)
3696 LIST_HEAD(page_list);
3698 lockdep_assert_held(&hugetlb_lock);
3699 if (hstate_is_gigantic(h))
3703 * Collect pages to be freed on a list, and free after dropping lock
3705 for_each_node_mask(i, *nodes_allowed) {
3706 struct folio *folio, *next;
3707 struct list_head *freel = &h->hugepage_freelists[i];
3708 list_for_each_entry_safe(folio, next, freel, lru) {
3709 if (count >= h->nr_huge_pages)
3711 if (folio_test_highmem(folio))
3713 remove_hugetlb_folio(h, folio, false);
3714 list_add(&folio->lru, &page_list);
3719 spin_unlock_irq(&hugetlb_lock);
3720 update_and_free_pages_bulk(h, &page_list);
3721 spin_lock_irq(&hugetlb_lock);
3724 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3725 nodemask_t *nodes_allowed)
3731 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3732 * balanced by operating on them in a round-robin fashion.
3733 * Returns 1 if an adjustment was made.
3735 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3740 lockdep_assert_held(&hugetlb_lock);
3741 VM_BUG_ON(delta != -1 && delta != 1);
3744 for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, nodes_allowed) {
3745 if (h->surplus_huge_pages_node[node])
3749 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3750 if (h->surplus_huge_pages_node[node] <
3751 h->nr_huge_pages_node[node])
3758 h->surplus_huge_pages += delta;
3759 h->surplus_huge_pages_node[node] += delta;
3763 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3764 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3765 nodemask_t *nodes_allowed)
3767 unsigned long min_count;
3768 unsigned long allocated;
3769 struct folio *folio;
3770 LIST_HEAD(page_list);
3771 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3774 * Bit mask controlling how hard we retry per-node allocations.
3775 * If we can not allocate the bit mask, do not attempt to allocate
3776 * the requested huge pages.
3778 if (node_alloc_noretry)
3779 nodes_clear(*node_alloc_noretry);
3784 * resize_lock mutex prevents concurrent adjustments to number of
3785 * pages in hstate via the proc/sysfs interfaces.
3787 mutex_lock(&h->resize_lock);
3788 flush_free_hpage_work(h);
3789 spin_lock_irq(&hugetlb_lock);
3792 * Check for a node specific request.
3793 * Changing node specific huge page count may require a corresponding
3794 * change to the global count. In any case, the passed node mask
3795 * (nodes_allowed) will restrict alloc/free to the specified node.
3797 if (nid != NUMA_NO_NODE) {
3798 unsigned long old_count = count;
3800 count += persistent_huge_pages(h) -
3801 (h->nr_huge_pages_node[nid] -
3802 h->surplus_huge_pages_node[nid]);
3804 * User may have specified a large count value which caused the
3805 * above calculation to overflow. In this case, they wanted
3806 * to allocate as many huge pages as possible. Set count to
3807 * largest possible value to align with their intention.
3809 if (count < old_count)
3814 * Gigantic pages runtime allocation depend on the capability for large
3815 * page range allocation.
3816 * If the system does not provide this feature, return an error when
3817 * the user tries to allocate gigantic pages but let the user free the
3818 * boottime allocated gigantic pages.
3820 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3821 if (count > persistent_huge_pages(h)) {
3822 spin_unlock_irq(&hugetlb_lock);
3823 mutex_unlock(&h->resize_lock);
3824 NODEMASK_FREE(node_alloc_noretry);
3827 /* Fall through to decrease pool */
3831 * Increase the pool size
3832 * First take pages out of surplus state. Then make up the
3833 * remaining difference by allocating fresh huge pages.
3835 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3836 * to convert a surplus huge page to a normal huge page. That is
3837 * not critical, though, it just means the overall size of the
3838 * pool might be one hugepage larger than it needs to be, but
3839 * within all the constraints specified by the sysctls.
3841 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3842 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3847 while (count > (persistent_huge_pages(h) + allocated)) {
3849 * If this allocation races such that we no longer need the
3850 * page, free_huge_folio will handle it by freeing the page
3851 * and reducing the surplus.
3853 spin_unlock_irq(&hugetlb_lock);
3855 /* yield cpu to avoid soft lockup */
3858 folio = alloc_pool_huge_folio(h, nodes_allowed,
3860 &h->next_nid_to_alloc);
3862 prep_and_add_allocated_folios(h, &page_list);
3863 spin_lock_irq(&hugetlb_lock);
3867 list_add(&folio->lru, &page_list);
3870 /* Bail for signals. Probably ctrl-c from user */
3871 if (signal_pending(current)) {
3872 prep_and_add_allocated_folios(h, &page_list);
3873 spin_lock_irq(&hugetlb_lock);
3877 spin_lock_irq(&hugetlb_lock);
3880 /* Add allocated pages to the pool */
3881 if (!list_empty(&page_list)) {
3882 spin_unlock_irq(&hugetlb_lock);
3883 prep_and_add_allocated_folios(h, &page_list);
3884 spin_lock_irq(&hugetlb_lock);
3888 * Decrease the pool size
3889 * First return free pages to the buddy allocator (being careful
3890 * to keep enough around to satisfy reservations). Then place
3891 * pages into surplus state as needed so the pool will shrink
3892 * to the desired size as pages become free.
3894 * By placing pages into the surplus state independent of the
3895 * overcommit value, we are allowing the surplus pool size to
3896 * exceed overcommit. There are few sane options here. Since
3897 * alloc_surplus_hugetlb_folio() is checking the global counter,
3898 * though, we'll note that we're not allowed to exceed surplus
3899 * and won't grow the pool anywhere else. Not until one of the
3900 * sysctls are changed, or the surplus pages go out of use.
3902 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3903 min_count = max(count, min_count);
3904 try_to_free_low(h, min_count, nodes_allowed);
3907 * Collect pages to be removed on list without dropping lock
3909 while (min_count < persistent_huge_pages(h)) {
3910 folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0);
3914 list_add(&folio->lru, &page_list);
3916 /* free the pages after dropping lock */
3917 spin_unlock_irq(&hugetlb_lock);
3918 update_and_free_pages_bulk(h, &page_list);
3919 flush_free_hpage_work(h);
3920 spin_lock_irq(&hugetlb_lock);
3922 while (count < persistent_huge_pages(h)) {
3923 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3927 h->max_huge_pages = persistent_huge_pages(h);
3928 spin_unlock_irq(&hugetlb_lock);
3929 mutex_unlock(&h->resize_lock);
3931 NODEMASK_FREE(node_alloc_noretry);
3936 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3938 int i, nid = folio_nid(folio);
3939 struct hstate *target_hstate;
3940 struct page *subpage;
3941 struct folio *inner_folio;
3944 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3946 remove_hugetlb_folio_for_demote(h, folio, false);
3947 spin_unlock_irq(&hugetlb_lock);
3950 * If vmemmap already existed for folio, the remove routine above would
3951 * have cleared the hugetlb folio flag. Hence the folio is technically
3952 * no longer a hugetlb folio. hugetlb_vmemmap_restore_folio can only be
3953 * passed hugetlb folios and will BUG otherwise.
3955 if (folio_test_hugetlb(folio)) {
3956 rc = hugetlb_vmemmap_restore_folio(h, folio);
3958 /* Allocation of vmemmmap failed, we can not demote folio */
3959 spin_lock_irq(&hugetlb_lock);
3960 folio_ref_unfreeze(folio, 1);
3961 add_hugetlb_folio(h, folio, false);
3967 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3968 * sizes as it will not ref count folios.
3970 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3973 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3974 * Without the mutex, pages added to target hstate could be marked
3977 * Note that we already hold h->resize_lock. To prevent deadlock,
3978 * use the convention of always taking larger size hstate mutex first.
3980 mutex_lock(&target_hstate->resize_lock);
3981 for (i = 0; i < pages_per_huge_page(h);
3982 i += pages_per_huge_page(target_hstate)) {
3983 subpage = folio_page(folio, i);
3984 inner_folio = page_folio(subpage);
3985 if (hstate_is_gigantic(target_hstate))
3986 prep_compound_gigantic_folio_for_demote(inner_folio,
3987 target_hstate->order);
3989 prep_compound_page(subpage, target_hstate->order);
3990 folio_change_private(inner_folio, NULL);
3991 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3992 free_huge_folio(inner_folio);
3994 mutex_unlock(&target_hstate->resize_lock);
3996 spin_lock_irq(&hugetlb_lock);
3999 * Not absolutely necessary, but for consistency update max_huge_pages
4000 * based on pool changes for the demoted page.
4002 h->max_huge_pages--;
4003 target_hstate->max_huge_pages +=
4004 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
4009 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
4010 __must_hold(&hugetlb_lock)
4013 struct folio *folio;
4015 lockdep_assert_held(&hugetlb_lock);
4017 /* We should never get here if no demote order */
4018 if (!h->demote_order) {
4019 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
4020 return -EINVAL; /* internal error */
4023 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
4024 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
4025 if (folio_test_hwpoison(folio))
4027 return demote_free_hugetlb_folio(h, folio);
4032 * Only way to get here is if all pages on free lists are poisoned.
4033 * Return -EBUSY so that caller will not retry.
4038 #define HSTATE_ATTR_RO(_name) \
4039 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
4041 #define HSTATE_ATTR_WO(_name) \
4042 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
4044 #define HSTATE_ATTR(_name) \
4045 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
4047 static struct kobject *hugepages_kobj;
4048 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4050 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
4052 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
4056 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4057 if (hstate_kobjs[i] == kobj) {
4059 *nidp = NUMA_NO_NODE;
4063 return kobj_to_node_hstate(kobj, nidp);
4066 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
4067 struct kobj_attribute *attr, char *buf)
4070 unsigned long nr_huge_pages;
4073 h = kobj_to_hstate(kobj, &nid);
4074 if (nid == NUMA_NO_NODE)
4075 nr_huge_pages = h->nr_huge_pages;
4077 nr_huge_pages = h->nr_huge_pages_node[nid];
4079 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
4082 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
4083 struct hstate *h, int nid,
4084 unsigned long count, size_t len)
4087 nodemask_t nodes_allowed, *n_mask;
4089 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
4092 if (nid == NUMA_NO_NODE) {
4094 * global hstate attribute
4096 if (!(obey_mempolicy &&
4097 init_nodemask_of_mempolicy(&nodes_allowed)))
4098 n_mask = &node_states[N_MEMORY];
4100 n_mask = &nodes_allowed;
4103 * Node specific request. count adjustment happens in
4104 * set_max_huge_pages() after acquiring hugetlb_lock.
4106 init_nodemask_of_node(&nodes_allowed, nid);
4107 n_mask = &nodes_allowed;
4110 err = set_max_huge_pages(h, count, nid, n_mask);
4112 return err ? err : len;
4115 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
4116 struct kobject *kobj, const char *buf,
4120 unsigned long count;
4124 err = kstrtoul(buf, 10, &count);
4128 h = kobj_to_hstate(kobj, &nid);
4129 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
4132 static ssize_t nr_hugepages_show(struct kobject *kobj,
4133 struct kobj_attribute *attr, char *buf)
4135 return nr_hugepages_show_common(kobj, attr, buf);
4138 static ssize_t nr_hugepages_store(struct kobject *kobj,
4139 struct kobj_attribute *attr, const char *buf, size_t len)
4141 return nr_hugepages_store_common(false, kobj, buf, len);
4143 HSTATE_ATTR(nr_hugepages);
4148 * hstate attribute for optionally mempolicy-based constraint on persistent
4149 * huge page alloc/free.
4151 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
4152 struct kobj_attribute *attr,
4155 return nr_hugepages_show_common(kobj, attr, buf);
4158 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
4159 struct kobj_attribute *attr, const char *buf, size_t len)
4161 return nr_hugepages_store_common(true, kobj, buf, len);
4163 HSTATE_ATTR(nr_hugepages_mempolicy);
4167 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
4168 struct kobj_attribute *attr, char *buf)
4170 struct hstate *h = kobj_to_hstate(kobj, NULL);
4171 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
4174 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
4175 struct kobj_attribute *attr, const char *buf, size_t count)
4178 unsigned long input;
4179 struct hstate *h = kobj_to_hstate(kobj, NULL);
4181 if (hstate_is_gigantic(h))
4184 err = kstrtoul(buf, 10, &input);
4188 spin_lock_irq(&hugetlb_lock);
4189 h->nr_overcommit_huge_pages = input;
4190 spin_unlock_irq(&hugetlb_lock);
4194 HSTATE_ATTR(nr_overcommit_hugepages);
4196 static ssize_t free_hugepages_show(struct kobject *kobj,
4197 struct kobj_attribute *attr, char *buf)
4200 unsigned long free_huge_pages;
4203 h = kobj_to_hstate(kobj, &nid);
4204 if (nid == NUMA_NO_NODE)
4205 free_huge_pages = h->free_huge_pages;
4207 free_huge_pages = h->free_huge_pages_node[nid];
4209 return sysfs_emit(buf, "%lu\n", free_huge_pages);
4211 HSTATE_ATTR_RO(free_hugepages);
4213 static ssize_t resv_hugepages_show(struct kobject *kobj,
4214 struct kobj_attribute *attr, char *buf)
4216 struct hstate *h = kobj_to_hstate(kobj, NULL);
4217 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
4219 HSTATE_ATTR_RO(resv_hugepages);
4221 static ssize_t surplus_hugepages_show(struct kobject *kobj,
4222 struct kobj_attribute *attr, char *buf)
4225 unsigned long surplus_huge_pages;
4228 h = kobj_to_hstate(kobj, &nid);
4229 if (nid == NUMA_NO_NODE)
4230 surplus_huge_pages = h->surplus_huge_pages;
4232 surplus_huge_pages = h->surplus_huge_pages_node[nid];
4234 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
4236 HSTATE_ATTR_RO(surplus_hugepages);
4238 static ssize_t demote_store(struct kobject *kobj,
4239 struct kobj_attribute *attr, const char *buf, size_t len)
4241 unsigned long nr_demote;
4242 unsigned long nr_available;
4243 nodemask_t nodes_allowed, *n_mask;
4248 err = kstrtoul(buf, 10, &nr_demote);
4251 h = kobj_to_hstate(kobj, &nid);
4253 if (nid != NUMA_NO_NODE) {
4254 init_nodemask_of_node(&nodes_allowed, nid);
4255 n_mask = &nodes_allowed;
4257 n_mask = &node_states[N_MEMORY];
4260 /* Synchronize with other sysfs operations modifying huge pages */
4261 mutex_lock(&h->resize_lock);
4262 spin_lock_irq(&hugetlb_lock);
4266 * Check for available pages to demote each time thorough the
4267 * loop as demote_pool_huge_page will drop hugetlb_lock.
4269 if (nid != NUMA_NO_NODE)
4270 nr_available = h->free_huge_pages_node[nid];
4272 nr_available = h->free_huge_pages;
4273 nr_available -= h->resv_huge_pages;
4277 err = demote_pool_huge_page(h, n_mask);
4284 spin_unlock_irq(&hugetlb_lock);
4285 mutex_unlock(&h->resize_lock);
4291 HSTATE_ATTR_WO(demote);
4293 static ssize_t demote_size_show(struct kobject *kobj,
4294 struct kobj_attribute *attr, char *buf)
4296 struct hstate *h = kobj_to_hstate(kobj, NULL);
4297 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
4299 return sysfs_emit(buf, "%lukB\n", demote_size);
4302 static ssize_t demote_size_store(struct kobject *kobj,
4303 struct kobj_attribute *attr,
4304 const char *buf, size_t count)
4306 struct hstate *h, *demote_hstate;
4307 unsigned long demote_size;
4308 unsigned int demote_order;
4310 demote_size = (unsigned long)memparse(buf, NULL);
4312 demote_hstate = size_to_hstate(demote_size);
4315 demote_order = demote_hstate->order;
4316 if (demote_order < HUGETLB_PAGE_ORDER)
4319 /* demote order must be smaller than hstate order */
4320 h = kobj_to_hstate(kobj, NULL);
4321 if (demote_order >= h->order)
4324 /* resize_lock synchronizes access to demote size and writes */
4325 mutex_lock(&h->resize_lock);
4326 h->demote_order = demote_order;
4327 mutex_unlock(&h->resize_lock);
4331 HSTATE_ATTR(demote_size);
4333 static struct attribute *hstate_attrs[] = {
4334 &nr_hugepages_attr.attr,
4335 &nr_overcommit_hugepages_attr.attr,
4336 &free_hugepages_attr.attr,
4337 &resv_hugepages_attr.attr,
4338 &surplus_hugepages_attr.attr,
4340 &nr_hugepages_mempolicy_attr.attr,
4345 static const struct attribute_group hstate_attr_group = {
4346 .attrs = hstate_attrs,
4349 static struct attribute *hstate_demote_attrs[] = {
4350 &demote_size_attr.attr,
4355 static const struct attribute_group hstate_demote_attr_group = {
4356 .attrs = hstate_demote_attrs,
4359 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4360 struct kobject **hstate_kobjs,
4361 const struct attribute_group *hstate_attr_group)
4364 int hi = hstate_index(h);
4366 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4367 if (!hstate_kobjs[hi])
4370 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4372 kobject_put(hstate_kobjs[hi]);
4373 hstate_kobjs[hi] = NULL;
4377 if (h->demote_order) {
4378 retval = sysfs_create_group(hstate_kobjs[hi],
4379 &hstate_demote_attr_group);
4381 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4382 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4383 kobject_put(hstate_kobjs[hi]);
4384 hstate_kobjs[hi] = NULL;
4393 static bool hugetlb_sysfs_initialized __ro_after_init;
4396 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4397 * with node devices in node_devices[] using a parallel array. The array
4398 * index of a node device or _hstate == node id.
4399 * This is here to avoid any static dependency of the node device driver, in
4400 * the base kernel, on the hugetlb module.
4402 struct node_hstate {
4403 struct kobject *hugepages_kobj;
4404 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4406 static struct node_hstate node_hstates[MAX_NUMNODES];
4409 * A subset of global hstate attributes for node devices
4411 static struct attribute *per_node_hstate_attrs[] = {
4412 &nr_hugepages_attr.attr,
4413 &free_hugepages_attr.attr,
4414 &surplus_hugepages_attr.attr,
4418 static const struct attribute_group per_node_hstate_attr_group = {
4419 .attrs = per_node_hstate_attrs,
4423 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4424 * Returns node id via non-NULL nidp.
4426 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4430 for (nid = 0; nid < nr_node_ids; nid++) {
4431 struct node_hstate *nhs = &node_hstates[nid];
4433 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4434 if (nhs->hstate_kobjs[i] == kobj) {
4446 * Unregister hstate attributes from a single node device.
4447 * No-op if no hstate attributes attached.
4449 void hugetlb_unregister_node(struct node *node)
4452 struct node_hstate *nhs = &node_hstates[node->dev.id];
4454 if (!nhs->hugepages_kobj)
4455 return; /* no hstate attributes */
4457 for_each_hstate(h) {
4458 int idx = hstate_index(h);
4459 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4463 if (h->demote_order)
4464 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4465 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4466 kobject_put(hstate_kobj);
4467 nhs->hstate_kobjs[idx] = NULL;
4470 kobject_put(nhs->hugepages_kobj);
4471 nhs->hugepages_kobj = NULL;
4476 * Register hstate attributes for a single node device.
4477 * No-op if attributes already registered.
4479 void hugetlb_register_node(struct node *node)
4482 struct node_hstate *nhs = &node_hstates[node->dev.id];
4485 if (!hugetlb_sysfs_initialized)
4488 if (nhs->hugepages_kobj)
4489 return; /* already allocated */
4491 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4493 if (!nhs->hugepages_kobj)
4496 for_each_hstate(h) {
4497 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4499 &per_node_hstate_attr_group);
4501 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4502 h->name, node->dev.id);
4503 hugetlb_unregister_node(node);
4510 * hugetlb init time: register hstate attributes for all registered node
4511 * devices of nodes that have memory. All on-line nodes should have
4512 * registered their associated device by this time.
4514 static void __init hugetlb_register_all_nodes(void)
4518 for_each_online_node(nid)
4519 hugetlb_register_node(node_devices[nid]);
4521 #else /* !CONFIG_NUMA */
4523 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4531 static void hugetlb_register_all_nodes(void) { }
4536 static void __init hugetlb_cma_check(void);
4538 static inline __init void hugetlb_cma_check(void)
4543 static void __init hugetlb_sysfs_init(void)
4548 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4549 if (!hugepages_kobj)
4552 for_each_hstate(h) {
4553 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4554 hstate_kobjs, &hstate_attr_group);
4556 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4560 hugetlb_sysfs_initialized = true;
4562 hugetlb_register_all_nodes();
4565 #ifdef CONFIG_SYSCTL
4566 static void hugetlb_sysctl_init(void);
4568 static inline void hugetlb_sysctl_init(void) { }
4571 static int __init hugetlb_init(void)
4575 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4578 if (!hugepages_supported()) {
4579 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4580 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4585 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4586 * architectures depend on setup being done here.
4588 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4589 if (!parsed_default_hugepagesz) {
4591 * If we did not parse a default huge page size, set
4592 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4593 * number of huge pages for this default size was implicitly
4594 * specified, set that here as well.
4595 * Note that the implicit setting will overwrite an explicit
4596 * setting. A warning will be printed in this case.
4598 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4599 if (default_hstate_max_huge_pages) {
4600 if (default_hstate.max_huge_pages) {
4603 string_get_size(huge_page_size(&default_hstate),
4604 1, STRING_UNITS_2, buf, 32);
4605 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4606 default_hstate.max_huge_pages, buf);
4607 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4608 default_hstate_max_huge_pages);
4610 default_hstate.max_huge_pages =
4611 default_hstate_max_huge_pages;
4613 for_each_online_node(i)
4614 default_hstate.max_huge_pages_node[i] =
4615 default_hugepages_in_node[i];
4619 hugetlb_cma_check();
4620 hugetlb_init_hstates();
4621 gather_bootmem_prealloc();
4624 hugetlb_sysfs_init();
4625 hugetlb_cgroup_file_init();
4626 hugetlb_sysctl_init();
4629 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4631 num_fault_mutexes = 1;
4633 hugetlb_fault_mutex_table =
4634 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4636 BUG_ON(!hugetlb_fault_mutex_table);
4638 for (i = 0; i < num_fault_mutexes; i++)
4639 mutex_init(&hugetlb_fault_mutex_table[i]);
4642 subsys_initcall(hugetlb_init);
4644 /* Overwritten by architectures with more huge page sizes */
4645 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4647 return size == HPAGE_SIZE;
4650 void __init hugetlb_add_hstate(unsigned int order)
4655 if (size_to_hstate(PAGE_SIZE << order)) {
4658 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4659 BUG_ON(order < order_base_2(__NR_USED_SUBPAGE));
4660 h = &hstates[hugetlb_max_hstate++];
4661 mutex_init(&h->resize_lock);
4663 h->mask = ~(huge_page_size(h) - 1);
4664 for (i = 0; i < MAX_NUMNODES; ++i)
4665 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4666 INIT_LIST_HEAD(&h->hugepage_activelist);
4667 h->next_nid_to_alloc = first_memory_node;
4668 h->next_nid_to_free = first_memory_node;
4669 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4670 huge_page_size(h)/SZ_1K);
4675 bool __init __weak hugetlb_node_alloc_supported(void)
4680 static void __init hugepages_clear_pages_in_node(void)
4682 if (!hugetlb_max_hstate) {
4683 default_hstate_max_huge_pages = 0;
4684 memset(default_hugepages_in_node, 0,
4685 sizeof(default_hugepages_in_node));
4687 parsed_hstate->max_huge_pages = 0;
4688 memset(parsed_hstate->max_huge_pages_node, 0,
4689 sizeof(parsed_hstate->max_huge_pages_node));
4694 * hugepages command line processing
4695 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4696 * specification. If not, ignore the hugepages value. hugepages can also
4697 * be the first huge page command line option in which case it implicitly
4698 * specifies the number of huge pages for the default size.
4700 static int __init hugepages_setup(char *s)
4703 static unsigned long *last_mhp;
4704 int node = NUMA_NO_NODE;
4709 if (!parsed_valid_hugepagesz) {
4710 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4711 parsed_valid_hugepagesz = true;
4716 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4717 * yet, so this hugepages= parameter goes to the "default hstate".
4718 * Otherwise, it goes with the previously parsed hugepagesz or
4719 * default_hugepagesz.
4721 else if (!hugetlb_max_hstate)
4722 mhp = &default_hstate_max_huge_pages;
4724 mhp = &parsed_hstate->max_huge_pages;
4726 if (mhp == last_mhp) {
4727 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4733 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4735 /* Parameter is node format */
4736 if (p[count] == ':') {
4737 if (!hugetlb_node_alloc_supported()) {
4738 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4741 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4743 node = array_index_nospec(tmp, MAX_NUMNODES);
4745 /* Parse hugepages */
4746 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4748 if (!hugetlb_max_hstate)
4749 default_hugepages_in_node[node] = tmp;
4751 parsed_hstate->max_huge_pages_node[node] = tmp;
4753 /* Go to parse next node*/
4754 if (p[count] == ',')
4767 * Global state is always initialized later in hugetlb_init.
4768 * But we need to allocate gigantic hstates here early to still
4769 * use the bootmem allocator.
4771 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4772 hugetlb_hstate_alloc_pages(parsed_hstate);
4779 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4780 hugepages_clear_pages_in_node();
4783 __setup("hugepages=", hugepages_setup);
4786 * hugepagesz command line processing
4787 * A specific huge page size can only be specified once with hugepagesz.
4788 * hugepagesz is followed by hugepages on the command line. The global
4789 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4790 * hugepagesz argument was valid.
4792 static int __init hugepagesz_setup(char *s)
4797 parsed_valid_hugepagesz = false;
4798 size = (unsigned long)memparse(s, NULL);
4800 if (!arch_hugetlb_valid_size(size)) {
4801 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4805 h = size_to_hstate(size);
4808 * hstate for this size already exists. This is normally
4809 * an error, but is allowed if the existing hstate is the
4810 * default hstate. More specifically, it is only allowed if
4811 * the number of huge pages for the default hstate was not
4812 * previously specified.
4814 if (!parsed_default_hugepagesz || h != &default_hstate ||
4815 default_hstate.max_huge_pages) {
4816 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4821 * No need to call hugetlb_add_hstate() as hstate already
4822 * exists. But, do set parsed_hstate so that a following
4823 * hugepages= parameter will be applied to this hstate.
4826 parsed_valid_hugepagesz = true;
4830 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4831 parsed_valid_hugepagesz = true;
4834 __setup("hugepagesz=", hugepagesz_setup);
4837 * default_hugepagesz command line input
4838 * Only one instance of default_hugepagesz allowed on command line.
4840 static int __init default_hugepagesz_setup(char *s)
4845 parsed_valid_hugepagesz = false;
4846 if (parsed_default_hugepagesz) {
4847 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4851 size = (unsigned long)memparse(s, NULL);
4853 if (!arch_hugetlb_valid_size(size)) {
4854 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4858 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4859 parsed_valid_hugepagesz = true;
4860 parsed_default_hugepagesz = true;
4861 default_hstate_idx = hstate_index(size_to_hstate(size));
4864 * The number of default huge pages (for this size) could have been
4865 * specified as the first hugetlb parameter: hugepages=X. If so,
4866 * then default_hstate_max_huge_pages is set. If the default huge
4867 * page size is gigantic (> MAX_PAGE_ORDER), then the pages must be
4868 * allocated here from bootmem allocator.
4870 if (default_hstate_max_huge_pages) {
4871 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4872 for_each_online_node(i)
4873 default_hstate.max_huge_pages_node[i] =
4874 default_hugepages_in_node[i];
4875 if (hstate_is_gigantic(&default_hstate))
4876 hugetlb_hstate_alloc_pages(&default_hstate);
4877 default_hstate_max_huge_pages = 0;
4882 __setup("default_hugepagesz=", default_hugepagesz_setup);
4884 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4887 struct mempolicy *mpol = get_task_policy(current);
4890 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4891 * (from policy_nodemask) specifically for hugetlb case
4893 if (mpol->mode == MPOL_BIND &&
4894 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4895 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4896 return &mpol->nodes;
4901 static unsigned int allowed_mems_nr(struct hstate *h)
4904 unsigned int nr = 0;
4905 nodemask_t *mbind_nodemask;
4906 unsigned int *array = h->free_huge_pages_node;
4907 gfp_t gfp_mask = htlb_alloc_mask(h);
4909 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4910 for_each_node_mask(node, cpuset_current_mems_allowed) {
4911 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4918 #ifdef CONFIG_SYSCTL
4919 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4920 void *buffer, size_t *length,
4921 loff_t *ppos, unsigned long *out)
4923 struct ctl_table dup_table;
4926 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4927 * can duplicate the @table and alter the duplicate of it.
4930 dup_table.data = out;
4932 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4935 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4936 struct ctl_table *table, int write,
4937 void *buffer, size_t *length, loff_t *ppos)
4939 struct hstate *h = &default_hstate;
4940 unsigned long tmp = h->max_huge_pages;
4943 if (!hugepages_supported())
4946 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4952 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4953 NUMA_NO_NODE, tmp, *length);
4958 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4959 void *buffer, size_t *length, loff_t *ppos)
4962 return hugetlb_sysctl_handler_common(false, table, write,
4963 buffer, length, ppos);
4967 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4968 void *buffer, size_t *length, loff_t *ppos)
4970 return hugetlb_sysctl_handler_common(true, table, write,
4971 buffer, length, ppos);
4973 #endif /* CONFIG_NUMA */
4975 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4976 void *buffer, size_t *length, loff_t *ppos)
4978 struct hstate *h = &default_hstate;
4982 if (!hugepages_supported())
4985 tmp = h->nr_overcommit_huge_pages;
4987 if (write && hstate_is_gigantic(h))
4990 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4996 spin_lock_irq(&hugetlb_lock);
4997 h->nr_overcommit_huge_pages = tmp;
4998 spin_unlock_irq(&hugetlb_lock);
5004 static struct ctl_table hugetlb_table[] = {
5006 .procname = "nr_hugepages",
5008 .maxlen = sizeof(unsigned long),
5010 .proc_handler = hugetlb_sysctl_handler,
5014 .procname = "nr_hugepages_mempolicy",
5016 .maxlen = sizeof(unsigned long),
5018 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
5022 .procname = "hugetlb_shm_group",
5023 .data = &sysctl_hugetlb_shm_group,
5024 .maxlen = sizeof(gid_t),
5026 .proc_handler = proc_dointvec,
5029 .procname = "nr_overcommit_hugepages",
5031 .maxlen = sizeof(unsigned long),
5033 .proc_handler = hugetlb_overcommit_handler,
5038 static void hugetlb_sysctl_init(void)
5040 register_sysctl_init("vm", hugetlb_table);
5042 #endif /* CONFIG_SYSCTL */
5044 void hugetlb_report_meminfo(struct seq_file *m)
5047 unsigned long total = 0;
5049 if (!hugepages_supported())
5052 for_each_hstate(h) {
5053 unsigned long count = h->nr_huge_pages;
5055 total += huge_page_size(h) * count;
5057 if (h == &default_hstate)
5059 "HugePages_Total: %5lu\n"
5060 "HugePages_Free: %5lu\n"
5061 "HugePages_Rsvd: %5lu\n"
5062 "HugePages_Surp: %5lu\n"
5063 "Hugepagesize: %8lu kB\n",
5067 h->surplus_huge_pages,
5068 huge_page_size(h) / SZ_1K);
5071 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
5074 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
5076 struct hstate *h = &default_hstate;
5078 if (!hugepages_supported())
5081 return sysfs_emit_at(buf, len,
5082 "Node %d HugePages_Total: %5u\n"
5083 "Node %d HugePages_Free: %5u\n"
5084 "Node %d HugePages_Surp: %5u\n",
5085 nid, h->nr_huge_pages_node[nid],
5086 nid, h->free_huge_pages_node[nid],
5087 nid, h->surplus_huge_pages_node[nid]);
5090 void hugetlb_show_meminfo_node(int nid)
5094 if (!hugepages_supported())
5098 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
5100 h->nr_huge_pages_node[nid],
5101 h->free_huge_pages_node[nid],
5102 h->surplus_huge_pages_node[nid],
5103 huge_page_size(h) / SZ_1K);
5106 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
5108 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
5109 K(atomic_long_read(&mm->hugetlb_usage)));
5112 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
5113 unsigned long hugetlb_total_pages(void)
5116 unsigned long nr_total_pages = 0;
5119 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
5120 return nr_total_pages;
5123 static int hugetlb_acct_memory(struct hstate *h, long delta)
5130 spin_lock_irq(&hugetlb_lock);
5132 * When cpuset is configured, it breaks the strict hugetlb page
5133 * reservation as the accounting is done on a global variable. Such
5134 * reservation is completely rubbish in the presence of cpuset because
5135 * the reservation is not checked against page availability for the
5136 * current cpuset. Application can still potentially OOM'ed by kernel
5137 * with lack of free htlb page in cpuset that the task is in.
5138 * Attempt to enforce strict accounting with cpuset is almost
5139 * impossible (or too ugly) because cpuset is too fluid that
5140 * task or memory node can be dynamically moved between cpusets.
5142 * The change of semantics for shared hugetlb mapping with cpuset is
5143 * undesirable. However, in order to preserve some of the semantics,
5144 * we fall back to check against current free page availability as
5145 * a best attempt and hopefully to minimize the impact of changing
5146 * semantics that cpuset has.
5148 * Apart from cpuset, we also have memory policy mechanism that
5149 * also determines from which node the kernel will allocate memory
5150 * in a NUMA system. So similar to cpuset, we also should consider
5151 * the memory policy of the current task. Similar to the description
5155 if (gather_surplus_pages(h, delta) < 0)
5158 if (delta > allowed_mems_nr(h)) {
5159 return_unused_surplus_pages(h, delta);
5166 return_unused_surplus_pages(h, (unsigned long) -delta);
5169 spin_unlock_irq(&hugetlb_lock);
5173 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
5175 struct resv_map *resv = vma_resv_map(vma);
5178 * HPAGE_RESV_OWNER indicates a private mapping.
5179 * This new VMA should share its siblings reservation map if present.
5180 * The VMA will only ever have a valid reservation map pointer where
5181 * it is being copied for another still existing VMA. As that VMA
5182 * has a reference to the reservation map it cannot disappear until
5183 * after this open call completes. It is therefore safe to take a
5184 * new reference here without additional locking.
5186 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
5187 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
5188 kref_get(&resv->refs);
5192 * vma_lock structure for sharable mappings is vma specific.
5193 * Clear old pointer (if copied via vm_area_dup) and allocate
5194 * new structure. Before clearing, make sure vma_lock is not
5197 if (vma->vm_flags & VM_MAYSHARE) {
5198 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
5201 if (vma_lock->vma != vma) {
5202 vma->vm_private_data = NULL;
5203 hugetlb_vma_lock_alloc(vma);
5205 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
5207 hugetlb_vma_lock_alloc(vma);
5211 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
5213 struct hstate *h = hstate_vma(vma);
5214 struct resv_map *resv;
5215 struct hugepage_subpool *spool = subpool_vma(vma);
5216 unsigned long reserve, start, end;
5219 hugetlb_vma_lock_free(vma);
5221 resv = vma_resv_map(vma);
5222 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5225 start = vma_hugecache_offset(h, vma, vma->vm_start);
5226 end = vma_hugecache_offset(h, vma, vma->vm_end);
5228 reserve = (end - start) - region_count(resv, start, end);
5229 hugetlb_cgroup_uncharge_counter(resv, start, end);
5232 * Decrement reserve counts. The global reserve count may be
5233 * adjusted if the subpool has a minimum size.
5235 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
5236 hugetlb_acct_memory(h, -gbl_reserve);
5239 kref_put(&resv->refs, resv_map_release);
5242 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
5244 if (addr & ~(huge_page_mask(hstate_vma(vma))))
5248 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
5249 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
5250 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
5252 if (addr & ~PUD_MASK) {
5254 * hugetlb_vm_op_split is called right before we attempt to
5255 * split the VMA. We will need to unshare PMDs in the old and
5256 * new VMAs, so let's unshare before we split.
5258 unsigned long floor = addr & PUD_MASK;
5259 unsigned long ceil = floor + PUD_SIZE;
5261 if (floor >= vma->vm_start && ceil <= vma->vm_end)
5262 hugetlb_unshare_pmds(vma, floor, ceil);
5268 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
5270 return huge_page_size(hstate_vma(vma));
5274 * We cannot handle pagefaults against hugetlb pages at all. They cause
5275 * handle_mm_fault() to try to instantiate regular-sized pages in the
5276 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
5279 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
5286 * When a new function is introduced to vm_operations_struct and added
5287 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
5288 * This is because under System V memory model, mappings created via
5289 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
5290 * their original vm_ops are overwritten with shm_vm_ops.
5292 const struct vm_operations_struct hugetlb_vm_ops = {
5293 .fault = hugetlb_vm_op_fault,
5294 .open = hugetlb_vm_op_open,
5295 .close = hugetlb_vm_op_close,
5296 .may_split = hugetlb_vm_op_split,
5297 .pagesize = hugetlb_vm_op_pagesize,
5300 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
5304 unsigned int shift = huge_page_shift(hstate_vma(vma));
5307 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
5308 vma->vm_page_prot)));
5310 entry = huge_pte_wrprotect(mk_huge_pte(page,
5311 vma->vm_page_prot));
5313 entry = pte_mkyoung(entry);
5314 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
5319 static void set_huge_ptep_writable(struct vm_area_struct *vma,
5320 unsigned long address, pte_t *ptep)
5324 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
5325 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
5326 update_mmu_cache(vma, address, ptep);
5329 bool is_hugetlb_entry_migration(pte_t pte)
5333 if (huge_pte_none(pte) || pte_present(pte))
5335 swp = pte_to_swp_entry(pte);
5336 if (is_migration_entry(swp))
5342 bool is_hugetlb_entry_hwpoisoned(pte_t pte)
5346 if (huge_pte_none(pte) || pte_present(pte))
5348 swp = pte_to_swp_entry(pte);
5349 if (is_hwpoison_entry(swp))
5356 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5357 struct folio *new_folio, pte_t old, unsigned long sz)
5359 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5361 __folio_mark_uptodate(new_folio);
5362 hugetlb_add_new_anon_rmap(new_folio, vma, addr);
5363 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5364 newpte = huge_pte_mkuffd_wp(newpte);
5365 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5366 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5367 folio_set_hugetlb_migratable(new_folio);
5370 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5371 struct vm_area_struct *dst_vma,
5372 struct vm_area_struct *src_vma)
5374 pte_t *src_pte, *dst_pte, entry;
5375 struct folio *pte_folio;
5377 bool cow = is_cow_mapping(src_vma->vm_flags);
5378 struct hstate *h = hstate_vma(src_vma);
5379 unsigned long sz = huge_page_size(h);
5380 unsigned long npages = pages_per_huge_page(h);
5381 struct mmu_notifier_range range;
5382 unsigned long last_addr_mask;
5386 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5389 mmu_notifier_invalidate_range_start(&range);
5390 vma_assert_write_locked(src_vma);
5391 raw_write_seqcount_begin(&src->write_protect_seq);
5394 * For shared mappings the vma lock must be held before
5395 * calling hugetlb_walk() in the src vma. Otherwise, the
5396 * returned ptep could go away if part of a shared pmd and
5397 * another thread calls huge_pmd_unshare.
5399 hugetlb_vma_lock_read(src_vma);
5402 last_addr_mask = hugetlb_mask_last_page(h);
5403 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5404 spinlock_t *src_ptl, *dst_ptl;
5405 src_pte = hugetlb_walk(src_vma, addr, sz);
5407 addr |= last_addr_mask;
5410 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5417 * If the pagetables are shared don't copy or take references.
5419 * dst_pte == src_pte is the common case of src/dest sharing.
5420 * However, src could have 'unshared' and dst shares with
5421 * another vma. So page_count of ptep page is checked instead
5422 * to reliably determine whether pte is shared.
5424 if (page_count(virt_to_page(dst_pte)) > 1) {
5425 addr |= last_addr_mask;
5429 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5430 src_ptl = huge_pte_lockptr(h, src, src_pte);
5431 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5432 entry = huge_ptep_get(src_pte);
5434 if (huge_pte_none(entry)) {
5436 * Skip if src entry none.
5439 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5440 if (!userfaultfd_wp(dst_vma))
5441 entry = huge_pte_clear_uffd_wp(entry);
5442 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5443 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5444 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5445 bool uffd_wp = pte_swp_uffd_wp(entry);
5447 if (!is_readable_migration_entry(swp_entry) && cow) {
5449 * COW mappings require pages in both
5450 * parent and child to be set to read.
5452 swp_entry = make_readable_migration_entry(
5453 swp_offset(swp_entry));
5454 entry = swp_entry_to_pte(swp_entry);
5455 if (userfaultfd_wp(src_vma) && uffd_wp)
5456 entry = pte_swp_mkuffd_wp(entry);
5457 set_huge_pte_at(src, addr, src_pte, entry, sz);
5459 if (!userfaultfd_wp(dst_vma))
5460 entry = huge_pte_clear_uffd_wp(entry);
5461 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5462 } else if (unlikely(is_pte_marker(entry))) {
5463 pte_marker marker = copy_pte_marker(
5464 pte_to_swp_entry(entry), dst_vma);
5467 set_huge_pte_at(dst, addr, dst_pte,
5468 make_pte_marker(marker), sz);
5470 entry = huge_ptep_get(src_pte);
5471 pte_folio = page_folio(pte_page(entry));
5472 folio_get(pte_folio);
5475 * Failing to duplicate the anon rmap is a rare case
5476 * where we see pinned hugetlb pages while they're
5477 * prone to COW. We need to do the COW earlier during
5480 * When pre-allocating the page or copying data, we
5481 * need to be without the pgtable locks since we could
5482 * sleep during the process.
5484 if (!folio_test_anon(pte_folio)) {
5485 hugetlb_add_file_rmap(pte_folio);
5486 } else if (hugetlb_try_dup_anon_rmap(pte_folio, src_vma)) {
5487 pte_t src_pte_old = entry;
5488 struct folio *new_folio;
5490 spin_unlock(src_ptl);
5491 spin_unlock(dst_ptl);
5492 /* Do not use reserve as it's private owned */
5493 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5494 if (IS_ERR(new_folio)) {
5495 folio_put(pte_folio);
5496 ret = PTR_ERR(new_folio);
5499 ret = copy_user_large_folio(new_folio,
5502 folio_put(pte_folio);
5504 folio_put(new_folio);
5508 /* Install the new hugetlb folio if src pte stable */
5509 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5510 src_ptl = huge_pte_lockptr(h, src, src_pte);
5511 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5512 entry = huge_ptep_get(src_pte);
5513 if (!pte_same(src_pte_old, entry)) {
5514 restore_reserve_on_error(h, dst_vma, addr,
5516 folio_put(new_folio);
5517 /* huge_ptep of dst_pte won't change as in child */
5520 hugetlb_install_folio(dst_vma, dst_pte, addr,
5521 new_folio, src_pte_old, sz);
5522 spin_unlock(src_ptl);
5523 spin_unlock(dst_ptl);
5529 * No need to notify as we are downgrading page
5530 * table protection not changing it to point
5533 * See Documentation/mm/mmu_notifier.rst
5535 huge_ptep_set_wrprotect(src, addr, src_pte);
5536 entry = huge_pte_wrprotect(entry);
5539 if (!userfaultfd_wp(dst_vma))
5540 entry = huge_pte_clear_uffd_wp(entry);
5542 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5543 hugetlb_count_add(npages, dst);
5545 spin_unlock(src_ptl);
5546 spin_unlock(dst_ptl);
5550 raw_write_seqcount_end(&src->write_protect_seq);
5551 mmu_notifier_invalidate_range_end(&range);
5553 hugetlb_vma_unlock_read(src_vma);
5559 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5560 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5563 struct hstate *h = hstate_vma(vma);
5564 struct mm_struct *mm = vma->vm_mm;
5565 spinlock_t *src_ptl, *dst_ptl;
5568 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5569 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5572 * We don't have to worry about the ordering of src and dst ptlocks
5573 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5575 if (src_ptl != dst_ptl)
5576 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5578 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5579 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5581 if (src_ptl != dst_ptl)
5582 spin_unlock(src_ptl);
5583 spin_unlock(dst_ptl);
5586 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5587 struct vm_area_struct *new_vma,
5588 unsigned long old_addr, unsigned long new_addr,
5591 struct hstate *h = hstate_vma(vma);
5592 struct address_space *mapping = vma->vm_file->f_mapping;
5593 unsigned long sz = huge_page_size(h);
5594 struct mm_struct *mm = vma->vm_mm;
5595 unsigned long old_end = old_addr + len;
5596 unsigned long last_addr_mask;
5597 pte_t *src_pte, *dst_pte;
5598 struct mmu_notifier_range range;
5599 bool shared_pmd = false;
5601 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5603 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5605 * In case of shared PMDs, we should cover the maximum possible
5608 flush_cache_range(vma, range.start, range.end);
5610 mmu_notifier_invalidate_range_start(&range);
5611 last_addr_mask = hugetlb_mask_last_page(h);
5612 /* Prevent race with file truncation */
5613 hugetlb_vma_lock_write(vma);
5614 i_mmap_lock_write(mapping);
5615 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5616 src_pte = hugetlb_walk(vma, old_addr, sz);
5618 old_addr |= last_addr_mask;
5619 new_addr |= last_addr_mask;
5622 if (huge_pte_none(huge_ptep_get(src_pte)))
5625 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5627 old_addr |= last_addr_mask;
5628 new_addr |= last_addr_mask;
5632 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5636 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5640 flush_hugetlb_tlb_range(vma, range.start, range.end);
5642 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5643 mmu_notifier_invalidate_range_end(&range);
5644 i_mmap_unlock_write(mapping);
5645 hugetlb_vma_unlock_write(vma);
5647 return len + old_addr - old_end;
5650 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5651 unsigned long start, unsigned long end,
5652 struct page *ref_page, zap_flags_t zap_flags)
5654 struct mm_struct *mm = vma->vm_mm;
5655 unsigned long address;
5660 struct hstate *h = hstate_vma(vma);
5661 unsigned long sz = huge_page_size(h);
5662 bool adjust_reservation = false;
5663 unsigned long last_addr_mask;
5664 bool force_flush = false;
5666 WARN_ON(!is_vm_hugetlb_page(vma));
5667 BUG_ON(start & ~huge_page_mask(h));
5668 BUG_ON(end & ~huge_page_mask(h));
5671 * This is a hugetlb vma, all the pte entries should point
5674 tlb_change_page_size(tlb, sz);
5675 tlb_start_vma(tlb, vma);
5677 last_addr_mask = hugetlb_mask_last_page(h);
5679 for (; address < end; address += sz) {
5680 ptep = hugetlb_walk(vma, address, sz);
5682 address |= last_addr_mask;
5686 ptl = huge_pte_lock(h, mm, ptep);
5687 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5689 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5691 address |= last_addr_mask;
5695 pte = huge_ptep_get(ptep);
5696 if (huge_pte_none(pte)) {
5702 * Migrating hugepage or HWPoisoned hugepage is already
5703 * unmapped and its refcount is dropped, so just clear pte here.
5705 if (unlikely(!pte_present(pte))) {
5707 * If the pte was wr-protected by uffd-wp in any of the
5708 * swap forms, meanwhile the caller does not want to
5709 * drop the uffd-wp bit in this zap, then replace the
5710 * pte with a marker.
5712 if (pte_swp_uffd_wp_any(pte) &&
5713 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5714 set_huge_pte_at(mm, address, ptep,
5715 make_pte_marker(PTE_MARKER_UFFD_WP),
5718 huge_pte_clear(mm, address, ptep, sz);
5723 page = pte_page(pte);
5725 * If a reference page is supplied, it is because a specific
5726 * page is being unmapped, not a range. Ensure the page we
5727 * are about to unmap is the actual page of interest.
5730 if (page != ref_page) {
5735 * Mark the VMA as having unmapped its page so that
5736 * future faults in this VMA will fail rather than
5737 * looking like data was lost
5739 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5742 pte = huge_ptep_get_and_clear(mm, address, ptep);
5743 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5744 if (huge_pte_dirty(pte))
5745 set_page_dirty(page);
5746 /* Leave a uffd-wp pte marker if needed */
5747 if (huge_pte_uffd_wp(pte) &&
5748 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5749 set_huge_pte_at(mm, address, ptep,
5750 make_pte_marker(PTE_MARKER_UFFD_WP),
5752 hugetlb_count_sub(pages_per_huge_page(h), mm);
5753 hugetlb_remove_rmap(page_folio(page));
5756 * Restore the reservation for anonymous page, otherwise the
5757 * backing page could be stolen by someone.
5758 * If there we are freeing a surplus, do not set the restore
5761 if (!h->surplus_huge_pages && __vma_private_lock(vma) &&
5762 folio_test_anon(page_folio(page))) {
5763 folio_set_hugetlb_restore_reserve(page_folio(page));
5764 /* Reservation to be adjusted after the spin lock */
5765 adjust_reservation = true;
5771 * Adjust the reservation for the region that will have the
5772 * reserve restored. Keep in mind that vma_needs_reservation() changes
5773 * resv->adds_in_progress if it succeeds. If this is not done,
5774 * do_exit() will not see it, and will keep the reservation
5777 if (adjust_reservation && vma_needs_reservation(h, vma, address))
5778 vma_add_reservation(h, vma, address);
5780 tlb_remove_page_size(tlb, page, huge_page_size(h));
5782 * Bail out after unmapping reference page if supplied
5787 tlb_end_vma(tlb, vma);
5790 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5791 * could defer the flush until now, since by holding i_mmap_rwsem we
5792 * guaranteed that the last refernece would not be dropped. But we must
5793 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5794 * dropped and the last reference to the shared PMDs page might be
5797 * In theory we could defer the freeing of the PMD pages as well, but
5798 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5799 * detect sharing, so we cannot defer the release of the page either.
5800 * Instead, do flush now.
5803 tlb_flush_mmu_tlbonly(tlb);
5806 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5807 unsigned long *start, unsigned long *end)
5809 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5812 adjust_range_if_pmd_sharing_possible(vma, start, end);
5813 hugetlb_vma_lock_write(vma);
5815 i_mmap_lock_write(vma->vm_file->f_mapping);
5818 void __hugetlb_zap_end(struct vm_area_struct *vma,
5819 struct zap_details *details)
5821 zap_flags_t zap_flags = details ? details->zap_flags : 0;
5823 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5826 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5828 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5829 * When the vma_lock is freed, this makes the vma ineligible
5830 * for pmd sharing. And, i_mmap_rwsem is required to set up
5831 * pmd sharing. This is important as page tables for this
5832 * unmapped range will be asynchrously deleted. If the page
5833 * tables are shared, there will be issues when accessed by
5836 __hugetlb_vma_unlock_write_free(vma);
5838 hugetlb_vma_unlock_write(vma);
5842 i_mmap_unlock_write(vma->vm_file->f_mapping);
5845 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5846 unsigned long end, struct page *ref_page,
5847 zap_flags_t zap_flags)
5849 struct mmu_notifier_range range;
5850 struct mmu_gather tlb;
5852 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5854 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5855 mmu_notifier_invalidate_range_start(&range);
5856 tlb_gather_mmu(&tlb, vma->vm_mm);
5858 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5860 mmu_notifier_invalidate_range_end(&range);
5861 tlb_finish_mmu(&tlb);
5865 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5866 * mapping it owns the reserve page for. The intention is to unmap the page
5867 * from other VMAs and let the children be SIGKILLed if they are faulting the
5870 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5871 struct page *page, unsigned long address)
5873 struct hstate *h = hstate_vma(vma);
5874 struct vm_area_struct *iter_vma;
5875 struct address_space *mapping;
5879 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5880 * from page cache lookup which is in HPAGE_SIZE units.
5882 address = address & huge_page_mask(h);
5883 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5885 mapping = vma->vm_file->f_mapping;
5888 * Take the mapping lock for the duration of the table walk. As
5889 * this mapping should be shared between all the VMAs,
5890 * __unmap_hugepage_range() is called as the lock is already held
5892 i_mmap_lock_write(mapping);
5893 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5894 /* Do not unmap the current VMA */
5895 if (iter_vma == vma)
5899 * Shared VMAs have their own reserves and do not affect
5900 * MAP_PRIVATE accounting but it is possible that a shared
5901 * VMA is using the same page so check and skip such VMAs.
5903 if (iter_vma->vm_flags & VM_MAYSHARE)
5907 * Unmap the page from other VMAs without their own reserves.
5908 * They get marked to be SIGKILLed if they fault in these
5909 * areas. This is because a future no-page fault on this VMA
5910 * could insert a zeroed page instead of the data existing
5911 * from the time of fork. This would look like data corruption
5913 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5914 unmap_hugepage_range(iter_vma, address,
5915 address + huge_page_size(h), page, 0);
5917 i_mmap_unlock_write(mapping);
5921 * hugetlb_wp() should be called with page lock of the original hugepage held.
5922 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5923 * cannot race with other handlers or page migration.
5924 * Keep the pte_same checks anyway to make transition from the mutex easier.
5926 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5927 unsigned long address, pte_t *ptep, unsigned int flags,
5928 struct folio *pagecache_folio, spinlock_t *ptl,
5929 struct vm_fault *vmf)
5931 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5932 pte_t pte = huge_ptep_get(ptep);
5933 struct hstate *h = hstate_vma(vma);
5934 struct folio *old_folio;
5935 struct folio *new_folio;
5936 int outside_reserve = 0;
5938 unsigned long haddr = address & huge_page_mask(h);
5939 struct mmu_notifier_range range;
5942 * Never handle CoW for uffd-wp protected pages. It should be only
5943 * handled when the uffd-wp protection is removed.
5945 * Note that only the CoW optimization path (in hugetlb_no_page())
5946 * can trigger this, because hugetlb_fault() will always resolve
5947 * uffd-wp bit first.
5949 if (!unshare && huge_pte_uffd_wp(pte))
5953 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5954 * PTE mapped R/O such as maybe_mkwrite() would do.
5956 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5957 return VM_FAULT_SIGSEGV;
5959 /* Let's take out MAP_SHARED mappings first. */
5960 if (vma->vm_flags & VM_MAYSHARE) {
5961 set_huge_ptep_writable(vma, haddr, ptep);
5965 old_folio = page_folio(pte_page(pte));
5967 delayacct_wpcopy_start();
5971 * If no-one else is actually using this page, we're the exclusive
5972 * owner and can reuse this page.
5974 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5975 if (!PageAnonExclusive(&old_folio->page)) {
5976 folio_move_anon_rmap(old_folio, vma);
5977 SetPageAnonExclusive(&old_folio->page);
5979 if (likely(!unshare))
5980 set_huge_ptep_writable(vma, haddr, ptep);
5982 delayacct_wpcopy_end();
5985 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5986 PageAnonExclusive(&old_folio->page), &old_folio->page);
5989 * If the process that created a MAP_PRIVATE mapping is about to
5990 * perform a COW due to a shared page count, attempt to satisfy
5991 * the allocation without using the existing reserves. The pagecache
5992 * page is used to determine if the reserve at this address was
5993 * consumed or not. If reserves were used, a partial faulted mapping
5994 * at the time of fork() could consume its reserves on COW instead
5995 * of the full address range.
5997 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5998 old_folio != pagecache_folio)
5999 outside_reserve = 1;
6001 folio_get(old_folio);
6004 * Drop page table lock as buddy allocator may be called. It will
6005 * be acquired again before returning to the caller, as expected.
6008 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
6010 if (IS_ERR(new_folio)) {
6012 * If a process owning a MAP_PRIVATE mapping fails to COW,
6013 * it is due to references held by a child and an insufficient
6014 * huge page pool. To guarantee the original mappers
6015 * reliability, unmap the page from child processes. The child
6016 * may get SIGKILLed if it later faults.
6018 if (outside_reserve) {
6019 struct address_space *mapping = vma->vm_file->f_mapping;
6023 folio_put(old_folio);
6025 * Drop hugetlb_fault_mutex and vma_lock before
6026 * unmapping. unmapping needs to hold vma_lock
6027 * in write mode. Dropping vma_lock in read mode
6028 * here is OK as COW mappings do not interact with
6031 * Reacquire both after unmap operation.
6033 idx = vma_hugecache_offset(h, vma, haddr);
6034 hash = hugetlb_fault_mutex_hash(mapping, idx);
6035 hugetlb_vma_unlock_read(vma);
6036 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6038 unmap_ref_private(mm, vma, &old_folio->page, haddr);
6040 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6041 hugetlb_vma_lock_read(vma);
6043 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
6045 pte_same(huge_ptep_get(ptep), pte)))
6046 goto retry_avoidcopy;
6048 * race occurs while re-acquiring page table
6049 * lock, and our job is done.
6051 delayacct_wpcopy_end();
6055 ret = vmf_error(PTR_ERR(new_folio));
6056 goto out_release_old;
6060 * When the original hugepage is shared one, it does not have
6061 * anon_vma prepared.
6063 ret = vmf_anon_prepare(vmf);
6065 goto out_release_all;
6067 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
6068 ret = VM_FAULT_HWPOISON_LARGE;
6069 goto out_release_all;
6071 __folio_mark_uptodate(new_folio);
6073 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
6074 haddr + huge_page_size(h));
6075 mmu_notifier_invalidate_range_start(&range);
6078 * Retake the page table lock to check for racing updates
6079 * before the page tables are altered
6082 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
6083 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
6084 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
6086 /* Break COW or unshare */
6087 huge_ptep_clear_flush(vma, haddr, ptep);
6088 hugetlb_remove_rmap(old_folio);
6089 hugetlb_add_new_anon_rmap(new_folio, vma, haddr);
6090 if (huge_pte_uffd_wp(pte))
6091 newpte = huge_pte_mkuffd_wp(newpte);
6092 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
6093 folio_set_hugetlb_migratable(new_folio);
6094 /* Make the old page be freed below */
6095 new_folio = old_folio;
6098 mmu_notifier_invalidate_range_end(&range);
6101 * No restore in case of successful pagetable update (Break COW or
6104 if (new_folio != old_folio)
6105 restore_reserve_on_error(h, vma, haddr, new_folio);
6106 folio_put(new_folio);
6108 folio_put(old_folio);
6110 spin_lock(ptl); /* Caller expects lock to be held */
6112 delayacct_wpcopy_end();
6117 * Return whether there is a pagecache page to back given address within VMA.
6119 static bool hugetlbfs_pagecache_present(struct hstate *h,
6120 struct vm_area_struct *vma, unsigned long address)
6122 struct address_space *mapping = vma->vm_file->f_mapping;
6123 pgoff_t idx = linear_page_index(vma, address);
6124 struct folio *folio;
6126 folio = filemap_get_folio(mapping, idx);
6133 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
6136 struct inode *inode = mapping->host;
6137 struct hstate *h = hstate_inode(inode);
6140 idx <<= huge_page_order(h);
6141 __folio_set_locked(folio);
6142 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
6144 if (unlikely(err)) {
6145 __folio_clear_locked(folio);
6148 folio_clear_hugetlb_restore_reserve(folio);
6151 * mark folio dirty so that it will not be removed from cache/file
6152 * by non-hugetlbfs specific code paths.
6154 folio_mark_dirty(folio);
6156 spin_lock(&inode->i_lock);
6157 inode->i_blocks += blocks_per_huge_page(h);
6158 spin_unlock(&inode->i_lock);
6162 static inline vm_fault_t hugetlb_handle_userfault(struct vm_fault *vmf,
6163 struct address_space *mapping,
6164 unsigned long reason)
6169 * vma_lock and hugetlb_fault_mutex must be dropped before handling
6170 * userfault. Also mmap_lock could be dropped due to handling
6171 * userfault, any vma operation should be careful from here.
6173 hugetlb_vma_unlock_read(vmf->vma);
6174 hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff);
6175 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6176 return handle_userfault(vmf, reason);
6180 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
6181 * false if pte changed or is changing.
6183 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
6184 pte_t *ptep, pte_t old_pte)
6189 ptl = huge_pte_lock(h, mm, ptep);
6190 same = pte_same(huge_ptep_get(ptep), old_pte);
6196 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
6197 struct vm_area_struct *vma,
6198 struct address_space *mapping, pgoff_t idx,
6199 unsigned long address, pte_t *ptep,
6200 pte_t old_pte, unsigned int flags,
6201 struct vm_fault *vmf)
6203 struct hstate *h = hstate_vma(vma);
6204 vm_fault_t ret = VM_FAULT_SIGBUS;
6207 struct folio *folio;
6210 unsigned long haddr = address & huge_page_mask(h);
6211 bool new_folio, new_pagecache_folio = false;
6212 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
6215 * Currently, we are forced to kill the process in the event the
6216 * original mapper has unmapped pages from the child due to a failed
6217 * COW/unsharing. Warn that such a situation has occurred as it may not
6220 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
6221 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
6227 * Use page lock to guard against racing truncation
6228 * before we get page_table_lock.
6231 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6232 if (IS_ERR(folio)) {
6233 size = i_size_read(mapping->host) >> huge_page_shift(h);
6236 /* Check for page in userfault range */
6237 if (userfaultfd_missing(vma)) {
6239 * Since hugetlb_no_page() was examining pte
6240 * without pgtable lock, we need to re-test under
6241 * lock because the pte may not be stable and could
6242 * have changed from under us. Try to detect
6243 * either changed or during-changing ptes and retry
6244 * properly when needed.
6246 * Note that userfaultfd is actually fine with
6247 * false positives (e.g. caused by pte changed),
6248 * but not wrong logical events (e.g. caused by
6249 * reading a pte during changing). The latter can
6250 * confuse the userspace, so the strictness is very
6251 * much preferred. E.g., MISSING event should
6252 * never happen on the page after UFFDIO_COPY has
6253 * correctly installed the page and returned.
6255 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6260 return hugetlb_handle_userfault(vmf, mapping,
6264 if (!(vma->vm_flags & VM_MAYSHARE)) {
6265 ret = vmf_anon_prepare(vmf);
6270 folio = alloc_hugetlb_folio(vma, haddr, 0);
6271 if (IS_ERR(folio)) {
6273 * Returning error will result in faulting task being
6274 * sent SIGBUS. The hugetlb fault mutex prevents two
6275 * tasks from racing to fault in the same page which
6276 * could result in false unable to allocate errors.
6277 * Page migration does not take the fault mutex, but
6278 * does a clear then write of pte's under page table
6279 * lock. Page fault code could race with migration,
6280 * notice the clear pte and try to allocate a page
6281 * here. Before returning error, get ptl and make
6282 * sure there really is no pte entry.
6284 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
6285 ret = vmf_error(PTR_ERR(folio));
6290 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
6291 __folio_mark_uptodate(folio);
6294 if (vma->vm_flags & VM_MAYSHARE) {
6295 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
6298 * err can't be -EEXIST which implies someone
6299 * else consumed the reservation since hugetlb
6300 * fault mutex is held when add a hugetlb page
6301 * to the page cache. So it's safe to call
6302 * restore_reserve_on_error() here.
6304 restore_reserve_on_error(h, vma, haddr, folio);
6306 ret = VM_FAULT_SIGBUS;
6309 new_pagecache_folio = true;
6316 * If memory error occurs between mmap() and fault, some process
6317 * don't have hwpoisoned swap entry for errored virtual address.
6318 * So we need to block hugepage fault by PG_hwpoison bit check.
6320 if (unlikely(folio_test_hwpoison(folio))) {
6321 ret = VM_FAULT_HWPOISON_LARGE |
6322 VM_FAULT_SET_HINDEX(hstate_index(h));
6323 goto backout_unlocked;
6326 /* Check for page in userfault range. */
6327 if (userfaultfd_minor(vma)) {
6328 folio_unlock(folio);
6330 /* See comment in userfaultfd_missing() block above */
6331 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6335 return hugetlb_handle_userfault(vmf, mapping,
6341 * If we are going to COW a private mapping later, we examine the
6342 * pending reservations for this page now. This will ensure that
6343 * any allocations necessary to record that reservation occur outside
6346 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6347 if (vma_needs_reservation(h, vma, haddr) < 0) {
6349 goto backout_unlocked;
6351 /* Just decrements count, does not deallocate */
6352 vma_end_reservation(h, vma, haddr);
6355 ptl = huge_pte_lock(h, mm, ptep);
6357 /* If pte changed from under us, retry */
6358 if (!pte_same(huge_ptep_get(ptep), old_pte))
6362 hugetlb_add_new_anon_rmap(folio, vma, haddr);
6364 hugetlb_add_file_rmap(folio);
6365 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6366 && (vma->vm_flags & VM_SHARED)));
6368 * If this pte was previously wr-protected, keep it wr-protected even
6371 if (unlikely(pte_marker_uffd_wp(old_pte)))
6372 new_pte = huge_pte_mkuffd_wp(new_pte);
6373 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6375 hugetlb_count_add(pages_per_huge_page(h), mm);
6376 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6377 /* Optimization, do the COW without a second fault */
6378 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl, vmf);
6384 * Only set hugetlb_migratable in newly allocated pages. Existing pages
6385 * found in the pagecache may not have hugetlb_migratable if they have
6386 * been isolated for migration.
6389 folio_set_hugetlb_migratable(folio);
6391 folio_unlock(folio);
6393 hugetlb_vma_unlock_read(vma);
6394 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6400 if (new_folio && !new_pagecache_folio)
6401 restore_reserve_on_error(h, vma, haddr, folio);
6403 folio_unlock(folio);
6409 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6411 unsigned long key[2];
6414 key[0] = (unsigned long) mapping;
6417 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6419 return hash & (num_fault_mutexes - 1);
6423 * For uniprocessor systems we always use a single mutex, so just
6424 * return 0 and avoid the hashing overhead.
6426 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6432 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6433 unsigned long address, unsigned int flags)
6439 struct folio *folio = NULL;
6440 struct folio *pagecache_folio = NULL;
6441 struct hstate *h = hstate_vma(vma);
6442 struct address_space *mapping;
6443 int need_wait_lock = 0;
6444 unsigned long haddr = address & huge_page_mask(h);
6445 struct vm_fault vmf = {
6448 .real_address = address,
6450 .pgoff = vma_hugecache_offset(h, vma, haddr),
6451 /* TODO: Track hugetlb faults using vm_fault */
6454 * Some fields may not be initialized, be careful as it may
6455 * be hard to debug if called functions make assumptions
6460 * Serialize hugepage allocation and instantiation, so that we don't
6461 * get spurious allocation failures if two CPUs race to instantiate
6462 * the same page in the page cache.
6464 mapping = vma->vm_file->f_mapping;
6465 hash = hugetlb_fault_mutex_hash(mapping, vmf.pgoff);
6466 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6469 * Acquire vma lock before calling huge_pte_alloc and hold
6470 * until finished with ptep. This prevents huge_pmd_unshare from
6471 * being called elsewhere and making the ptep no longer valid.
6473 hugetlb_vma_lock_read(vma);
6474 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6476 hugetlb_vma_unlock_read(vma);
6477 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6478 return VM_FAULT_OOM;
6481 entry = huge_ptep_get(ptep);
6482 if (huge_pte_none_mostly(entry)) {
6483 if (is_pte_marker(entry)) {
6485 pte_marker_get(pte_to_swp_entry(entry));
6487 if (marker & PTE_MARKER_POISONED) {
6488 ret = VM_FAULT_HWPOISON_LARGE;
6494 * Other PTE markers should be handled the same way as none PTE.
6496 * hugetlb_no_page will drop vma lock and hugetlb fault
6497 * mutex internally, which make us return immediately.
6499 return hugetlb_no_page(mm, vma, mapping, vmf.pgoff, address,
6500 ptep, entry, flags, &vmf);
6506 * entry could be a migration/hwpoison entry at this point, so this
6507 * check prevents the kernel from going below assuming that we have
6508 * an active hugepage in pagecache. This goto expects the 2nd page
6509 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6510 * properly handle it.
6512 if (!pte_present(entry)) {
6513 if (unlikely(is_hugetlb_entry_migration(entry))) {
6515 * Release the hugetlb fault lock now, but retain
6516 * the vma lock, because it is needed to guard the
6517 * huge_pte_lockptr() later in
6518 * migration_entry_wait_huge(). The vma lock will
6519 * be released there.
6521 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6522 migration_entry_wait_huge(vma, ptep);
6524 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6525 ret = VM_FAULT_HWPOISON_LARGE |
6526 VM_FAULT_SET_HINDEX(hstate_index(h));
6531 * If we are going to COW/unshare the mapping later, we examine the
6532 * pending reservations for this page now. This will ensure that any
6533 * allocations necessary to record that reservation occur outside the
6534 * spinlock. Also lookup the pagecache page now as it is used to
6535 * determine if a reservation has been consumed.
6537 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6538 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6539 if (vma_needs_reservation(h, vma, haddr) < 0) {
6543 /* Just decrements count, does not deallocate */
6544 vma_end_reservation(h, vma, haddr);
6546 pagecache_folio = filemap_lock_hugetlb_folio(h, mapping,
6548 if (IS_ERR(pagecache_folio))
6549 pagecache_folio = NULL;
6552 ptl = huge_pte_lock(h, mm, ptep);
6554 /* Check for a racing update before calling hugetlb_wp() */
6555 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6558 /* Handle userfault-wp first, before trying to lock more pages */
6559 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6560 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6561 if (!userfaultfd_wp_async(vma)) {
6563 if (pagecache_folio) {
6564 folio_unlock(pagecache_folio);
6565 folio_put(pagecache_folio);
6567 hugetlb_vma_unlock_read(vma);
6568 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6569 return handle_userfault(&vmf, VM_UFFD_WP);
6572 entry = huge_pte_clear_uffd_wp(entry);
6573 set_huge_pte_at(mm, haddr, ptep, entry,
6574 huge_page_size(hstate_vma(vma)));
6575 /* Fallthrough to CoW */
6579 * hugetlb_wp() requires page locks of pte_page(entry) and
6580 * pagecache_folio, so here we need take the former one
6581 * when folio != pagecache_folio or !pagecache_folio.
6583 folio = page_folio(pte_page(entry));
6584 if (folio != pagecache_folio)
6585 if (!folio_trylock(folio)) {
6592 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6593 if (!huge_pte_write(entry)) {
6594 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6595 pagecache_folio, ptl, &vmf);
6597 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6598 entry = huge_pte_mkdirty(entry);
6601 entry = pte_mkyoung(entry);
6602 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6603 flags & FAULT_FLAG_WRITE))
6604 update_mmu_cache(vma, haddr, ptep);
6606 if (folio != pagecache_folio)
6607 folio_unlock(folio);
6612 if (pagecache_folio) {
6613 folio_unlock(pagecache_folio);
6614 folio_put(pagecache_folio);
6617 hugetlb_vma_unlock_read(vma);
6618 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6620 * Generally it's safe to hold refcount during waiting page lock. But
6621 * here we just wait to defer the next page fault to avoid busy loop and
6622 * the page is not used after unlocked before returning from the current
6623 * page fault. So we are safe from accessing freed page, even if we wait
6624 * here without taking refcount.
6627 folio_wait_locked(folio);
6631 #ifdef CONFIG_USERFAULTFD
6633 * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte().
6635 static struct folio *alloc_hugetlb_folio_vma(struct hstate *h,
6636 struct vm_area_struct *vma, unsigned long address)
6638 struct mempolicy *mpol;
6639 nodemask_t *nodemask;
6640 struct folio *folio;
6644 gfp_mask = htlb_alloc_mask(h);
6645 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
6646 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
6647 mpol_cond_put(mpol);
6653 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6654 * with modifications for hugetlb pages.
6656 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6657 struct vm_area_struct *dst_vma,
6658 unsigned long dst_addr,
6659 unsigned long src_addr,
6661 struct folio **foliop)
6663 struct mm_struct *dst_mm = dst_vma->vm_mm;
6664 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6665 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6666 struct hstate *h = hstate_vma(dst_vma);
6667 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6668 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6670 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6674 struct folio *folio;
6676 bool folio_in_pagecache = false;
6678 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6679 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6681 /* Don't overwrite any existing PTEs (even markers) */
6682 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6687 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6688 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6691 /* No need to invalidate - it was non-present before */
6692 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6700 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6703 folio_in_pagecache = true;
6704 } else if (!*foliop) {
6705 /* If a folio already exists, then it's UFFDIO_COPY for
6706 * a non-missing case. Return -EEXIST.
6709 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6714 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6715 if (IS_ERR(folio)) {
6720 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6723 /* fallback to copy_from_user outside mmap_lock */
6724 if (unlikely(ret)) {
6726 /* Free the allocated folio which may have
6727 * consumed a reservation.
6729 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6732 /* Allocate a temporary folio to hold the copied
6735 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6741 /* Set the outparam foliop and return to the caller to
6742 * copy the contents outside the lock. Don't free the
6749 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6756 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6757 if (IS_ERR(folio)) {
6763 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6773 * If we just allocated a new page, we need a memory barrier to ensure
6774 * that preceding stores to the page become visible before the
6775 * set_pte_at() write. The memory barrier inside __folio_mark_uptodate
6778 * In the case where we have not allocated a new page (is_continue),
6779 * the page must already be uptodate. UFFDIO_CONTINUE already includes
6780 * an earlier smp_wmb() to ensure that prior stores will be visible
6781 * before the set_pte_at() write.
6784 __folio_mark_uptodate(folio);
6786 WARN_ON_ONCE(!folio_test_uptodate(folio));
6788 /* Add shared, newly allocated pages to the page cache. */
6789 if (vm_shared && !is_continue) {
6790 size = i_size_read(mapping->host) >> huge_page_shift(h);
6793 goto out_release_nounlock;
6796 * Serialization between remove_inode_hugepages() and
6797 * hugetlb_add_to_page_cache() below happens through the
6798 * hugetlb_fault_mutex_table that here must be hold by
6801 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6803 goto out_release_nounlock;
6804 folio_in_pagecache = true;
6807 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6810 if (folio_test_hwpoison(folio))
6811 goto out_release_unlock;
6814 * We allow to overwrite a pte marker: consider when both MISSING|WP
6815 * registered, we firstly wr-protect a none pte which has no page cache
6816 * page backing it, then access the page.
6819 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6820 goto out_release_unlock;
6822 if (folio_in_pagecache)
6823 hugetlb_add_file_rmap(folio);
6825 hugetlb_add_new_anon_rmap(folio, dst_vma, dst_addr);
6828 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6829 * with wp flag set, don't set pte write bit.
6831 if (wp_enabled || (is_continue && !vm_shared))
6834 writable = dst_vma->vm_flags & VM_WRITE;
6836 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6838 * Always mark UFFDIO_COPY page dirty; note that this may not be
6839 * extremely important for hugetlbfs for now since swapping is not
6840 * supported, but we should still be clear in that this page cannot be
6841 * thrown away at will, even if write bit not set.
6843 _dst_pte = huge_pte_mkdirty(_dst_pte);
6844 _dst_pte = pte_mkyoung(_dst_pte);
6847 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6849 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6851 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6853 /* No need to invalidate - it was non-present before */
6854 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6858 folio_set_hugetlb_migratable(folio);
6859 if (vm_shared || is_continue)
6860 folio_unlock(folio);
6866 if (vm_shared || is_continue)
6867 folio_unlock(folio);
6868 out_release_nounlock:
6869 if (!folio_in_pagecache)
6870 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6874 #endif /* CONFIG_USERFAULTFD */
6876 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6877 unsigned long address, unsigned int flags,
6878 unsigned int *page_mask)
6880 struct hstate *h = hstate_vma(vma);
6881 struct mm_struct *mm = vma->vm_mm;
6882 unsigned long haddr = address & huge_page_mask(h);
6883 struct page *page = NULL;
6888 hugetlb_vma_lock_read(vma);
6889 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6893 ptl = huge_pte_lock(h, mm, pte);
6894 entry = huge_ptep_get(pte);
6895 if (pte_present(entry)) {
6896 page = pte_page(entry);
6898 if (!huge_pte_write(entry)) {
6899 if (flags & FOLL_WRITE) {
6904 if (gup_must_unshare(vma, flags, page)) {
6905 /* Tell the caller to do unsharing */
6906 page = ERR_PTR(-EMLINK);
6911 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6914 * Note that page may be a sub-page, and with vmemmap
6915 * optimizations the page struct may be read only.
6916 * try_grab_page() will increase the ref count on the
6917 * head page, so this will be OK.
6919 * try_grab_page() should always be able to get the page here,
6920 * because we hold the ptl lock and have verified pte_present().
6922 ret = try_grab_page(page, flags);
6924 if (WARN_ON_ONCE(ret)) {
6925 page = ERR_PTR(ret);
6929 *page_mask = (1U << huge_page_order(h)) - 1;
6934 hugetlb_vma_unlock_read(vma);
6937 * Fixup retval for dump requests: if pagecache doesn't exist,
6938 * don't try to allocate a new page but just skip it.
6940 if (!page && (flags & FOLL_DUMP) &&
6941 !hugetlbfs_pagecache_present(h, vma, address))
6942 page = ERR_PTR(-EFAULT);
6947 long hugetlb_change_protection(struct vm_area_struct *vma,
6948 unsigned long address, unsigned long end,
6949 pgprot_t newprot, unsigned long cp_flags)
6951 struct mm_struct *mm = vma->vm_mm;
6952 unsigned long start = address;
6955 struct hstate *h = hstate_vma(vma);
6956 long pages = 0, psize = huge_page_size(h);
6957 bool shared_pmd = false;
6958 struct mmu_notifier_range range;
6959 unsigned long last_addr_mask;
6960 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6961 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6964 * In the case of shared PMDs, the area to flush could be beyond
6965 * start/end. Set range.start/range.end to cover the maximum possible
6966 * range if PMD sharing is possible.
6968 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6970 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6972 BUG_ON(address >= end);
6973 flush_cache_range(vma, range.start, range.end);
6975 mmu_notifier_invalidate_range_start(&range);
6976 hugetlb_vma_lock_write(vma);
6977 i_mmap_lock_write(vma->vm_file->f_mapping);
6978 last_addr_mask = hugetlb_mask_last_page(h);
6979 for (; address < end; address += psize) {
6981 ptep = hugetlb_walk(vma, address, psize);
6984 address |= last_addr_mask;
6988 * Userfaultfd wr-protect requires pgtable
6989 * pre-allocations to install pte markers.
6991 ptep = huge_pte_alloc(mm, vma, address, psize);
6997 ptl = huge_pte_lock(h, mm, ptep);
6998 if (huge_pmd_unshare(mm, vma, address, ptep)) {
7000 * When uffd-wp is enabled on the vma, unshare
7001 * shouldn't happen at all. Warn about it if it
7002 * happened due to some reason.
7004 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
7008 address |= last_addr_mask;
7011 pte = huge_ptep_get(ptep);
7012 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
7013 /* Nothing to do. */
7014 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
7015 swp_entry_t entry = pte_to_swp_entry(pte);
7016 struct page *page = pfn_swap_entry_to_page(entry);
7019 if (is_writable_migration_entry(entry)) {
7021 entry = make_readable_exclusive_migration_entry(
7024 entry = make_readable_migration_entry(
7026 newpte = swp_entry_to_pte(entry);
7031 newpte = pte_swp_mkuffd_wp(newpte);
7032 else if (uffd_wp_resolve)
7033 newpte = pte_swp_clear_uffd_wp(newpte);
7034 if (!pte_same(pte, newpte))
7035 set_huge_pte_at(mm, address, ptep, newpte, psize);
7036 } else if (unlikely(is_pte_marker(pte))) {
7038 * Do nothing on a poison marker; page is
7039 * corrupted, permissons do not apply. Here
7040 * pte_marker_uffd_wp()==true implies !poison
7041 * because they're mutual exclusive.
7043 if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)
7044 /* Safe to modify directly (non-present->none). */
7045 huge_pte_clear(mm, address, ptep, psize);
7046 } else if (!huge_pte_none(pte)) {
7048 unsigned int shift = huge_page_shift(hstate_vma(vma));
7050 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
7051 pte = huge_pte_modify(old_pte, newprot);
7052 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
7054 pte = huge_pte_mkuffd_wp(pte);
7055 else if (uffd_wp_resolve)
7056 pte = huge_pte_clear_uffd_wp(pte);
7057 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
7061 if (unlikely(uffd_wp))
7062 /* Safe to modify directly (none->non-present). */
7063 set_huge_pte_at(mm, address, ptep,
7064 make_pte_marker(PTE_MARKER_UFFD_WP),
7070 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
7071 * may have cleared our pud entry and done put_page on the page table:
7072 * once we release i_mmap_rwsem, another task can do the final put_page
7073 * and that page table be reused and filled with junk. If we actually
7074 * did unshare a page of pmds, flush the range corresponding to the pud.
7077 flush_hugetlb_tlb_range(vma, range.start, range.end);
7079 flush_hugetlb_tlb_range(vma, start, end);
7081 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
7082 * downgrading page table protection not changing it to point to a new
7085 * See Documentation/mm/mmu_notifier.rst
7087 i_mmap_unlock_write(vma->vm_file->f_mapping);
7088 hugetlb_vma_unlock_write(vma);
7089 mmu_notifier_invalidate_range_end(&range);
7091 return pages > 0 ? (pages << h->order) : pages;
7094 /* Return true if reservation was successful, false otherwise. */
7095 bool hugetlb_reserve_pages(struct inode *inode,
7097 struct vm_area_struct *vma,
7098 vm_flags_t vm_flags)
7100 long chg = -1, add = -1;
7101 struct hstate *h = hstate_inode(inode);
7102 struct hugepage_subpool *spool = subpool_inode(inode);
7103 struct resv_map *resv_map;
7104 struct hugetlb_cgroup *h_cg = NULL;
7105 long gbl_reserve, regions_needed = 0;
7107 /* This should never happen */
7109 VM_WARN(1, "%s called with a negative range\n", __func__);
7114 * vma specific semaphore used for pmd sharing and fault/truncation
7117 hugetlb_vma_lock_alloc(vma);
7120 * Only apply hugepage reservation if asked. At fault time, an
7121 * attempt will be made for VM_NORESERVE to allocate a page
7122 * without using reserves
7124 if (vm_flags & VM_NORESERVE)
7128 * Shared mappings base their reservation on the number of pages that
7129 * are already allocated on behalf of the file. Private mappings need
7130 * to reserve the full area even if read-only as mprotect() may be
7131 * called to make the mapping read-write. Assume !vma is a shm mapping
7133 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7135 * resv_map can not be NULL as hugetlb_reserve_pages is only
7136 * called for inodes for which resv_maps were created (see
7137 * hugetlbfs_get_inode).
7139 resv_map = inode_resv_map(inode);
7141 chg = region_chg(resv_map, from, to, ®ions_needed);
7143 /* Private mapping. */
7144 resv_map = resv_map_alloc();
7150 set_vma_resv_map(vma, resv_map);
7151 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
7157 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
7158 chg * pages_per_huge_page(h), &h_cg) < 0)
7161 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
7162 /* For private mappings, the hugetlb_cgroup uncharge info hangs
7165 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
7169 * There must be enough pages in the subpool for the mapping. If
7170 * the subpool has a minimum size, there may be some global
7171 * reservations already in place (gbl_reserve).
7173 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
7174 if (gbl_reserve < 0)
7175 goto out_uncharge_cgroup;
7178 * Check enough hugepages are available for the reservation.
7179 * Hand the pages back to the subpool if there are not
7181 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
7185 * Account for the reservations made. Shared mappings record regions
7186 * that have reservations as they are shared by multiple VMAs.
7187 * When the last VMA disappears, the region map says how much
7188 * the reservation was and the page cache tells how much of
7189 * the reservation was consumed. Private mappings are per-VMA and
7190 * only the consumed reservations are tracked. When the VMA
7191 * disappears, the original reservation is the VMA size and the
7192 * consumed reservations are stored in the map. Hence, nothing
7193 * else has to be done for private mappings here
7195 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7196 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
7198 if (unlikely(add < 0)) {
7199 hugetlb_acct_memory(h, -gbl_reserve);
7201 } else if (unlikely(chg > add)) {
7203 * pages in this range were added to the reserve
7204 * map between region_chg and region_add. This
7205 * indicates a race with alloc_hugetlb_folio. Adjust
7206 * the subpool and reserve counts modified above
7207 * based on the difference.
7212 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
7213 * reference to h_cg->css. See comment below for detail.
7215 hugetlb_cgroup_uncharge_cgroup_rsvd(
7217 (chg - add) * pages_per_huge_page(h), h_cg);
7219 rsv_adjust = hugepage_subpool_put_pages(spool,
7221 hugetlb_acct_memory(h, -rsv_adjust);
7224 * The file_regions will hold their own reference to
7225 * h_cg->css. So we should release the reference held
7226 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
7229 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
7235 /* put back original number of pages, chg */
7236 (void)hugepage_subpool_put_pages(spool, chg);
7237 out_uncharge_cgroup:
7238 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
7239 chg * pages_per_huge_page(h), h_cg);
7241 hugetlb_vma_lock_free(vma);
7242 if (!vma || vma->vm_flags & VM_MAYSHARE)
7243 /* Only call region_abort if the region_chg succeeded but the
7244 * region_add failed or didn't run.
7246 if (chg >= 0 && add < 0)
7247 region_abort(resv_map, from, to, regions_needed);
7248 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
7249 kref_put(&resv_map->refs, resv_map_release);
7250 set_vma_resv_map(vma, NULL);
7255 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
7258 struct hstate *h = hstate_inode(inode);
7259 struct resv_map *resv_map = inode_resv_map(inode);
7261 struct hugepage_subpool *spool = subpool_inode(inode);
7265 * Since this routine can be called in the evict inode path for all
7266 * hugetlbfs inodes, resv_map could be NULL.
7269 chg = region_del(resv_map, start, end);
7271 * region_del() can fail in the rare case where a region
7272 * must be split and another region descriptor can not be
7273 * allocated. If end == LONG_MAX, it will not fail.
7279 spin_lock(&inode->i_lock);
7280 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
7281 spin_unlock(&inode->i_lock);
7284 * If the subpool has a minimum size, the number of global
7285 * reservations to be released may be adjusted.
7287 * Note that !resv_map implies freed == 0. So (chg - freed)
7288 * won't go negative.
7290 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
7291 hugetlb_acct_memory(h, -gbl_reserve);
7296 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7297 static unsigned long page_table_shareable(struct vm_area_struct *svma,
7298 struct vm_area_struct *vma,
7299 unsigned long addr, pgoff_t idx)
7301 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
7303 unsigned long sbase = saddr & PUD_MASK;
7304 unsigned long s_end = sbase + PUD_SIZE;
7306 /* Allow segments to share if only one is marked locked */
7307 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
7308 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
7311 * match the virtual addresses, permission and the alignment of the
7314 * Also, vma_lock (vm_private_data) is required for sharing.
7316 if (pmd_index(addr) != pmd_index(saddr) ||
7317 vm_flags != svm_flags ||
7318 !range_in_vma(svma, sbase, s_end) ||
7319 !svma->vm_private_data)
7325 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7327 unsigned long start = addr & PUD_MASK;
7328 unsigned long end = start + PUD_SIZE;
7330 #ifdef CONFIG_USERFAULTFD
7331 if (uffd_disable_huge_pmd_share(vma))
7335 * check on proper vm_flags and page table alignment
7337 if (!(vma->vm_flags & VM_MAYSHARE))
7339 if (!vma->vm_private_data) /* vma lock required for sharing */
7341 if (!range_in_vma(vma, start, end))
7347 * Determine if start,end range within vma could be mapped by shared pmd.
7348 * If yes, adjust start and end to cover range associated with possible
7349 * shared pmd mappings.
7351 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7352 unsigned long *start, unsigned long *end)
7354 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7355 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7358 * vma needs to span at least one aligned PUD size, and the range
7359 * must be at least partially within in.
7361 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7362 (*end <= v_start) || (*start >= v_end))
7365 /* Extend the range to be PUD aligned for a worst case scenario */
7366 if (*start > v_start)
7367 *start = ALIGN_DOWN(*start, PUD_SIZE);
7370 *end = ALIGN(*end, PUD_SIZE);
7374 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7375 * and returns the corresponding pte. While this is not necessary for the
7376 * !shared pmd case because we can allocate the pmd later as well, it makes the
7377 * code much cleaner. pmd allocation is essential for the shared case because
7378 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7379 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7380 * bad pmd for sharing.
7382 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7383 unsigned long addr, pud_t *pud)
7385 struct address_space *mapping = vma->vm_file->f_mapping;
7386 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7388 struct vm_area_struct *svma;
7389 unsigned long saddr;
7393 i_mmap_lock_read(mapping);
7394 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7398 saddr = page_table_shareable(svma, vma, addr, idx);
7400 spte = hugetlb_walk(svma, saddr,
7401 vma_mmu_pagesize(svma));
7403 get_page(virt_to_page(spte));
7412 spin_lock(&mm->page_table_lock);
7413 if (pud_none(*pud)) {
7414 pud_populate(mm, pud,
7415 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7418 put_page(virt_to_page(spte));
7420 spin_unlock(&mm->page_table_lock);
7422 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7423 i_mmap_unlock_read(mapping);
7428 * unmap huge page backed by shared pte.
7430 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7431 * indicated by page_count > 1, unmap is achieved by clearing pud and
7432 * decrementing the ref count. If count == 1, the pte page is not shared.
7434 * Called with page table lock held.
7436 * returns: 1 successfully unmapped a shared pte page
7437 * 0 the underlying pte page is not shared, or it is the last user
7439 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7440 unsigned long addr, pte_t *ptep)
7442 pgd_t *pgd = pgd_offset(mm, addr);
7443 p4d_t *p4d = p4d_offset(pgd, addr);
7444 pud_t *pud = pud_offset(p4d, addr);
7446 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7447 hugetlb_vma_assert_locked(vma);
7448 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7449 if (page_count(virt_to_page(ptep)) == 1)
7453 put_page(virt_to_page(ptep));
7458 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7460 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7461 unsigned long addr, pud_t *pud)
7466 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7467 unsigned long addr, pte_t *ptep)
7472 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7473 unsigned long *start, unsigned long *end)
7477 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7481 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7483 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7484 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7485 unsigned long addr, unsigned long sz)
7492 pgd = pgd_offset(mm, addr);
7493 p4d = p4d_alloc(mm, pgd, addr);
7496 pud = pud_alloc(mm, p4d, addr);
7498 if (sz == PUD_SIZE) {
7501 BUG_ON(sz != PMD_SIZE);
7502 if (want_pmd_share(vma, addr) && pud_none(*pud))
7503 pte = huge_pmd_share(mm, vma, addr, pud);
7505 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7510 pte_t pteval = ptep_get_lockless(pte);
7512 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7519 * huge_pte_offset() - Walk the page table to resolve the hugepage
7520 * entry at address @addr
7522 * Return: Pointer to page table entry (PUD or PMD) for
7523 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7524 * size @sz doesn't match the hugepage size at this level of the page
7527 pte_t *huge_pte_offset(struct mm_struct *mm,
7528 unsigned long addr, unsigned long sz)
7535 pgd = pgd_offset(mm, addr);
7536 if (!pgd_present(*pgd))
7538 p4d = p4d_offset(pgd, addr);
7539 if (!p4d_present(*p4d))
7542 pud = pud_offset(p4d, addr);
7544 /* must be pud huge, non-present or none */
7545 return (pte_t *)pud;
7546 if (!pud_present(*pud))
7548 /* must have a valid entry and size to go further */
7550 pmd = pmd_offset(pud, addr);
7551 /* must be pmd huge, non-present or none */
7552 return (pte_t *)pmd;
7556 * Return a mask that can be used to update an address to the last huge
7557 * page in a page table page mapping size. Used to skip non-present
7558 * page table entries when linearly scanning address ranges. Architectures
7559 * with unique huge page to page table relationships can define their own
7560 * version of this routine.
7562 unsigned long hugetlb_mask_last_page(struct hstate *h)
7564 unsigned long hp_size = huge_page_size(h);
7566 if (hp_size == PUD_SIZE)
7567 return P4D_SIZE - PUD_SIZE;
7568 else if (hp_size == PMD_SIZE)
7569 return PUD_SIZE - PMD_SIZE;
7576 /* See description above. Architectures can provide their own version. */
7577 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7579 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7580 if (huge_page_size(h) == PMD_SIZE)
7581 return PUD_SIZE - PMD_SIZE;
7586 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7589 * These functions are overwritable if your architecture needs its own
7592 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7596 spin_lock_irq(&hugetlb_lock);
7597 if (!folio_test_hugetlb(folio) ||
7598 !folio_test_hugetlb_migratable(folio) ||
7599 !folio_try_get(folio)) {
7603 folio_clear_hugetlb_migratable(folio);
7604 list_move_tail(&folio->lru, list);
7606 spin_unlock_irq(&hugetlb_lock);
7610 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7615 spin_lock_irq(&hugetlb_lock);
7616 if (folio_test_hugetlb(folio)) {
7618 if (folio_test_hugetlb_freed(folio))
7620 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7621 ret = folio_try_get(folio);
7625 spin_unlock_irq(&hugetlb_lock);
7629 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7630 bool *migratable_cleared)
7634 spin_lock_irq(&hugetlb_lock);
7635 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7636 spin_unlock_irq(&hugetlb_lock);
7640 void folio_putback_active_hugetlb(struct folio *folio)
7642 spin_lock_irq(&hugetlb_lock);
7643 folio_set_hugetlb_migratable(folio);
7644 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7645 spin_unlock_irq(&hugetlb_lock);
7649 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7651 struct hstate *h = folio_hstate(old_folio);
7653 hugetlb_cgroup_migrate(old_folio, new_folio);
7654 set_page_owner_migrate_reason(&new_folio->page, reason);
7657 * transfer temporary state of the new hugetlb folio. This is
7658 * reverse to other transitions because the newpage is going to
7659 * be final while the old one will be freed so it takes over
7660 * the temporary status.
7662 * Also note that we have to transfer the per-node surplus state
7663 * here as well otherwise the global surplus count will not match
7666 if (folio_test_hugetlb_temporary(new_folio)) {
7667 int old_nid = folio_nid(old_folio);
7668 int new_nid = folio_nid(new_folio);
7670 folio_set_hugetlb_temporary(old_folio);
7671 folio_clear_hugetlb_temporary(new_folio);
7675 * There is no need to transfer the per-node surplus state
7676 * when we do not cross the node.
7678 if (new_nid == old_nid)
7680 spin_lock_irq(&hugetlb_lock);
7681 if (h->surplus_huge_pages_node[old_nid]) {
7682 h->surplus_huge_pages_node[old_nid]--;
7683 h->surplus_huge_pages_node[new_nid]++;
7685 spin_unlock_irq(&hugetlb_lock);
7689 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7690 unsigned long start,
7693 struct hstate *h = hstate_vma(vma);
7694 unsigned long sz = huge_page_size(h);
7695 struct mm_struct *mm = vma->vm_mm;
7696 struct mmu_notifier_range range;
7697 unsigned long address;
7701 if (!(vma->vm_flags & VM_MAYSHARE))
7707 flush_cache_range(vma, start, end);
7709 * No need to call adjust_range_if_pmd_sharing_possible(), because
7710 * we have already done the PUD_SIZE alignment.
7712 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7714 mmu_notifier_invalidate_range_start(&range);
7715 hugetlb_vma_lock_write(vma);
7716 i_mmap_lock_write(vma->vm_file->f_mapping);
7717 for (address = start; address < end; address += PUD_SIZE) {
7718 ptep = hugetlb_walk(vma, address, sz);
7721 ptl = huge_pte_lock(h, mm, ptep);
7722 huge_pmd_unshare(mm, vma, address, ptep);
7725 flush_hugetlb_tlb_range(vma, start, end);
7726 i_mmap_unlock_write(vma->vm_file->f_mapping);
7727 hugetlb_vma_unlock_write(vma);
7729 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7730 * Documentation/mm/mmu_notifier.rst.
7732 mmu_notifier_invalidate_range_end(&range);
7736 * This function will unconditionally remove all the shared pmd pgtable entries
7737 * within the specific vma for a hugetlbfs memory range.
7739 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7741 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7742 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7746 static bool cma_reserve_called __initdata;
7748 static int __init cmdline_parse_hugetlb_cma(char *p)
7755 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7758 if (s[count] == ':') {
7759 if (tmp >= MAX_NUMNODES)
7761 nid = array_index_nospec(tmp, MAX_NUMNODES);
7764 tmp = memparse(s, &s);
7765 hugetlb_cma_size_in_node[nid] = tmp;
7766 hugetlb_cma_size += tmp;
7769 * Skip the separator if have one, otherwise
7770 * break the parsing.
7777 hugetlb_cma_size = memparse(p, &p);
7785 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7787 void __init hugetlb_cma_reserve(int order)
7789 unsigned long size, reserved, per_node;
7790 bool node_specific_cma_alloc = false;
7794 * HugeTLB CMA reservation is required for gigantic
7795 * huge pages which could not be allocated via the
7796 * page allocator. Just warn if there is any change
7797 * breaking this assumption.
7799 VM_WARN_ON(order <= MAX_PAGE_ORDER);
7800 cma_reserve_called = true;
7802 if (!hugetlb_cma_size)
7805 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7806 if (hugetlb_cma_size_in_node[nid] == 0)
7809 if (!node_online(nid)) {
7810 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7811 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7812 hugetlb_cma_size_in_node[nid] = 0;
7816 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7817 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7818 nid, (PAGE_SIZE << order) / SZ_1M);
7819 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7820 hugetlb_cma_size_in_node[nid] = 0;
7822 node_specific_cma_alloc = true;
7826 /* Validate the CMA size again in case some invalid nodes specified. */
7827 if (!hugetlb_cma_size)
7830 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7831 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7832 (PAGE_SIZE << order) / SZ_1M);
7833 hugetlb_cma_size = 0;
7837 if (!node_specific_cma_alloc) {
7839 * If 3 GB area is requested on a machine with 4 numa nodes,
7840 * let's allocate 1 GB on first three nodes and ignore the last one.
7842 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7843 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7844 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7848 for_each_online_node(nid) {
7850 char name[CMA_MAX_NAME];
7852 if (node_specific_cma_alloc) {
7853 if (hugetlb_cma_size_in_node[nid] == 0)
7856 size = hugetlb_cma_size_in_node[nid];
7858 size = min(per_node, hugetlb_cma_size - reserved);
7861 size = round_up(size, PAGE_SIZE << order);
7863 snprintf(name, sizeof(name), "hugetlb%d", nid);
7865 * Note that 'order per bit' is based on smallest size that
7866 * may be returned to CMA allocator in the case of
7867 * huge page demotion.
7869 res = cma_declare_contiguous_nid(0, size, 0,
7870 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7872 &hugetlb_cma[nid], nid);
7874 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7880 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7883 if (reserved >= hugetlb_cma_size)
7889 * hugetlb_cma_size is used to determine if allocations from
7890 * cma are possible. Set to zero if no cma regions are set up.
7892 hugetlb_cma_size = 0;
7895 static void __init hugetlb_cma_check(void)
7897 if (!hugetlb_cma_size || cma_reserve_called)
7900 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7903 #endif /* CONFIG_CMA */