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>
39 #include <asm/pgalloc.h>
43 #include <linux/hugetlb.h>
44 #include <linux/hugetlb_cgroup.h>
45 #include <linux/node.h>
46 #include <linux/page_owner.h>
48 #include "hugetlb_vmemmap.h"
50 int hugetlb_max_hstate __read_mostly;
51 unsigned int default_hstate_idx;
52 struct hstate hstates[HUGE_MAX_HSTATE];
55 static struct cma *hugetlb_cma[MAX_NUMNODES];
56 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
57 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
63 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
68 static unsigned long hugetlb_cma_size __initdata;
70 __initdata LIST_HEAD(huge_boot_pages);
72 /* for command line parsing */
73 static struct hstate * __initdata parsed_hstate;
74 static unsigned long __initdata default_hstate_max_huge_pages;
75 static bool __initdata parsed_valid_hugepagesz = true;
76 static bool __initdata parsed_default_hugepagesz;
77 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
80 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
81 * free_huge_pages, and surplus_huge_pages.
83 DEFINE_SPINLOCK(hugetlb_lock);
86 * Serializes faults on the same logical page. This is used to
87 * prevent spurious OOMs when the hugepage pool is fully utilized.
89 static int num_fault_mutexes;
90 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92 /* Forward declaration */
93 static int hugetlb_acct_memory(struct hstate *h, long delta);
94 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
95 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
96 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
97 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
98 unsigned long start, unsigned long end);
100 static inline bool subpool_is_free(struct hugepage_subpool *spool)
104 if (spool->max_hpages != -1)
105 return spool->used_hpages == 0;
106 if (spool->min_hpages != -1)
107 return spool->rsv_hpages == spool->min_hpages;
112 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
113 unsigned long irq_flags)
115 spin_unlock_irqrestore(&spool->lock, irq_flags);
117 /* If no pages are used, and no other handles to the subpool
118 * remain, give up any reservations based on minimum size and
119 * free the subpool */
120 if (subpool_is_free(spool)) {
121 if (spool->min_hpages != -1)
122 hugetlb_acct_memory(spool->hstate,
128 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 struct hugepage_subpool *spool;
133 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
137 spin_lock_init(&spool->lock);
139 spool->max_hpages = max_hpages;
141 spool->min_hpages = min_hpages;
143 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
147 spool->rsv_hpages = min_hpages;
152 void hugepage_put_subpool(struct hugepage_subpool *spool)
156 spin_lock_irqsave(&spool->lock, flags);
157 BUG_ON(!spool->count);
159 unlock_or_release_subpool(spool, flags);
163 * Subpool accounting for allocating and reserving pages.
164 * Return -ENOMEM if there are not enough resources to satisfy the
165 * request. Otherwise, return the number of pages by which the
166 * global pools must be adjusted (upward). The returned value may
167 * only be different than the passed value (delta) in the case where
168 * a subpool minimum size must be maintained.
170 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
178 spin_lock_irq(&spool->lock);
180 if (spool->max_hpages != -1) { /* maximum size accounting */
181 if ((spool->used_hpages + delta) <= spool->max_hpages)
182 spool->used_hpages += delta;
189 /* minimum size accounting */
190 if (spool->min_hpages != -1 && spool->rsv_hpages) {
191 if (delta > spool->rsv_hpages) {
193 * Asking for more reserves than those already taken on
194 * behalf of subpool. Return difference.
196 ret = delta - spool->rsv_hpages;
197 spool->rsv_hpages = 0;
199 ret = 0; /* reserves already accounted for */
200 spool->rsv_hpages -= delta;
205 spin_unlock_irq(&spool->lock);
210 * Subpool accounting for freeing and unreserving pages.
211 * Return the number of global page reservations that must be dropped.
212 * The return value may only be different than the passed value (delta)
213 * in the case where a subpool minimum size must be maintained.
215 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
224 spin_lock_irqsave(&spool->lock, flags);
226 if (spool->max_hpages != -1) /* maximum size accounting */
227 spool->used_hpages -= delta;
229 /* minimum size accounting */
230 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
231 if (spool->rsv_hpages + delta <= spool->min_hpages)
234 ret = spool->rsv_hpages + delta - spool->min_hpages;
236 spool->rsv_hpages += delta;
237 if (spool->rsv_hpages > spool->min_hpages)
238 spool->rsv_hpages = spool->min_hpages;
242 * If hugetlbfs_put_super couldn't free spool due to an outstanding
243 * quota reference, free it now.
245 unlock_or_release_subpool(spool, flags);
250 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
252 return HUGETLBFS_SB(inode->i_sb)->spool;
255 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
257 return subpool_inode(file_inode(vma->vm_file));
261 * hugetlb vma_lock helper routines
263 static bool __vma_shareable_lock(struct vm_area_struct *vma)
265 return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) &&
266 vma->vm_private_data;
269 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
271 if (__vma_shareable_lock(vma)) {
272 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
274 down_read(&vma_lock->rw_sema);
278 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
280 if (__vma_shareable_lock(vma)) {
281 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
283 up_read(&vma_lock->rw_sema);
287 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
289 if (__vma_shareable_lock(vma)) {
290 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
292 down_write(&vma_lock->rw_sema);
296 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
298 if (__vma_shareable_lock(vma)) {
299 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
301 up_write(&vma_lock->rw_sema);
305 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
307 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
309 if (!__vma_shareable_lock(vma))
312 return down_write_trylock(&vma_lock->rw_sema);
315 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
317 if (__vma_shareable_lock(vma)) {
318 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
320 lockdep_assert_held(&vma_lock->rw_sema);
324 void hugetlb_vma_lock_release(struct kref *kref)
326 struct hugetlb_vma_lock *vma_lock = container_of(kref,
327 struct hugetlb_vma_lock, refs);
332 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
334 struct vm_area_struct *vma = vma_lock->vma;
337 * vma_lock structure may or not be released as a result of put,
338 * it certainly will no longer be attached to vma so clear pointer.
339 * Semaphore synchronizes access to vma_lock->vma field.
341 vma_lock->vma = NULL;
342 vma->vm_private_data = NULL;
343 up_write(&vma_lock->rw_sema);
344 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
347 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
349 if (__vma_shareable_lock(vma)) {
350 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
352 __hugetlb_vma_unlock_write_put(vma_lock);
356 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
359 * Only present in sharable vmas.
361 if (!vma || !__vma_shareable_lock(vma))
364 if (vma->vm_private_data) {
365 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
367 down_write(&vma_lock->rw_sema);
368 __hugetlb_vma_unlock_write_put(vma_lock);
372 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
374 struct hugetlb_vma_lock *vma_lock;
376 /* Only establish in (flags) sharable vmas */
377 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
380 /* Should never get here with non-NULL vm_private_data */
381 if (vma->vm_private_data)
384 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
387 * If we can not allocate structure, then vma can not
388 * participate in pmd sharing. This is only a possible
389 * performance enhancement and memory saving issue.
390 * However, the lock is also used to synchronize page
391 * faults with truncation. If the lock is not present,
392 * unlikely races could leave pages in a file past i_size
393 * until the file is removed. Warn in the unlikely case of
394 * allocation failure.
396 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
400 kref_init(&vma_lock->refs);
401 init_rwsem(&vma_lock->rw_sema);
403 vma->vm_private_data = vma_lock;
406 /* Helper that removes a struct file_region from the resv_map cache and returns
409 static struct file_region *
410 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
412 struct file_region *nrg;
414 VM_BUG_ON(resv->region_cache_count <= 0);
416 resv->region_cache_count--;
417 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
418 list_del(&nrg->link);
426 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
427 struct file_region *rg)
429 #ifdef CONFIG_CGROUP_HUGETLB
430 nrg->reservation_counter = rg->reservation_counter;
437 /* Helper that records hugetlb_cgroup uncharge info. */
438 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
440 struct resv_map *resv,
441 struct file_region *nrg)
443 #ifdef CONFIG_CGROUP_HUGETLB
445 nrg->reservation_counter =
446 &h_cg->rsvd_hugepage[hstate_index(h)];
447 nrg->css = &h_cg->css;
449 * The caller will hold exactly one h_cg->css reference for the
450 * whole contiguous reservation region. But this area might be
451 * scattered when there are already some file_regions reside in
452 * it. As a result, many file_regions may share only one css
453 * reference. In order to ensure that one file_region must hold
454 * exactly one h_cg->css reference, we should do css_get for
455 * each file_region and leave the reference held by caller
459 if (!resv->pages_per_hpage)
460 resv->pages_per_hpage = pages_per_huge_page(h);
461 /* pages_per_hpage should be the same for all entries in
464 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
466 nrg->reservation_counter = NULL;
472 static void put_uncharge_info(struct file_region *rg)
474 #ifdef CONFIG_CGROUP_HUGETLB
480 static bool has_same_uncharge_info(struct file_region *rg,
481 struct file_region *org)
483 #ifdef CONFIG_CGROUP_HUGETLB
484 return rg->reservation_counter == org->reservation_counter &&
492 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
494 struct file_region *nrg, *prg;
496 prg = list_prev_entry(rg, link);
497 if (&prg->link != &resv->regions && prg->to == rg->from &&
498 has_same_uncharge_info(prg, rg)) {
502 put_uncharge_info(rg);
508 nrg = list_next_entry(rg, link);
509 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
510 has_same_uncharge_info(nrg, rg)) {
511 nrg->from = rg->from;
514 put_uncharge_info(rg);
520 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
521 long to, struct hstate *h, struct hugetlb_cgroup *cg,
522 long *regions_needed)
524 struct file_region *nrg;
526 if (!regions_needed) {
527 nrg = get_file_region_entry_from_cache(map, from, to);
528 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
529 list_add(&nrg->link, rg);
530 coalesce_file_region(map, nrg);
532 *regions_needed += 1;
538 * Must be called with resv->lock held.
540 * Calling this with regions_needed != NULL will count the number of pages
541 * to be added but will not modify the linked list. And regions_needed will
542 * indicate the number of file_regions needed in the cache to carry out to add
543 * the regions for this range.
545 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
546 struct hugetlb_cgroup *h_cg,
547 struct hstate *h, long *regions_needed)
550 struct list_head *head = &resv->regions;
551 long last_accounted_offset = f;
552 struct file_region *iter, *trg = NULL;
553 struct list_head *rg = NULL;
558 /* In this loop, we essentially handle an entry for the range
559 * [last_accounted_offset, iter->from), at every iteration, with some
562 list_for_each_entry_safe(iter, trg, head, link) {
563 /* Skip irrelevant regions that start before our range. */
564 if (iter->from < f) {
565 /* If this region ends after the last accounted offset,
566 * then we need to update last_accounted_offset.
568 if (iter->to > last_accounted_offset)
569 last_accounted_offset = iter->to;
573 /* When we find a region that starts beyond our range, we've
576 if (iter->from >= t) {
577 rg = iter->link.prev;
581 /* Add an entry for last_accounted_offset -> iter->from, and
582 * update last_accounted_offset.
584 if (iter->from > last_accounted_offset)
585 add += hugetlb_resv_map_add(resv, iter->link.prev,
586 last_accounted_offset,
590 last_accounted_offset = iter->to;
593 /* Handle the case where our range extends beyond
594 * last_accounted_offset.
598 if (last_accounted_offset < t)
599 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
600 t, h, h_cg, regions_needed);
605 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
607 static int allocate_file_region_entries(struct resv_map *resv,
609 __must_hold(&resv->lock)
611 LIST_HEAD(allocated_regions);
612 int to_allocate = 0, i = 0;
613 struct file_region *trg = NULL, *rg = NULL;
615 VM_BUG_ON(regions_needed < 0);
618 * Check for sufficient descriptors in the cache to accommodate
619 * the number of in progress add operations plus regions_needed.
621 * This is a while loop because when we drop the lock, some other call
622 * to region_add or region_del may have consumed some region_entries,
623 * so we keep looping here until we finally have enough entries for
624 * (adds_in_progress + regions_needed).
626 while (resv->region_cache_count <
627 (resv->adds_in_progress + regions_needed)) {
628 to_allocate = resv->adds_in_progress + regions_needed -
629 resv->region_cache_count;
631 /* At this point, we should have enough entries in the cache
632 * for all the existing adds_in_progress. We should only be
633 * needing to allocate for regions_needed.
635 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
637 spin_unlock(&resv->lock);
638 for (i = 0; i < to_allocate; i++) {
639 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
642 list_add(&trg->link, &allocated_regions);
645 spin_lock(&resv->lock);
647 list_splice(&allocated_regions, &resv->region_cache);
648 resv->region_cache_count += to_allocate;
654 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
662 * Add the huge page range represented by [f, t) to the reserve
663 * map. Regions will be taken from the cache to fill in this range.
664 * Sufficient regions should exist in the cache due to the previous
665 * call to region_chg with the same range, but in some cases the cache will not
666 * have sufficient entries due to races with other code doing region_add or
667 * region_del. The extra needed entries will be allocated.
669 * regions_needed is the out value provided by a previous call to region_chg.
671 * Return the number of new huge pages added to the map. This number is greater
672 * than or equal to zero. If file_region entries needed to be allocated for
673 * this operation and we were not able to allocate, it returns -ENOMEM.
674 * region_add of regions of length 1 never allocate file_regions and cannot
675 * fail; region_chg will always allocate at least 1 entry and a region_add for
676 * 1 page will only require at most 1 entry.
678 static long region_add(struct resv_map *resv, long f, long t,
679 long in_regions_needed, struct hstate *h,
680 struct hugetlb_cgroup *h_cg)
682 long add = 0, actual_regions_needed = 0;
684 spin_lock(&resv->lock);
687 /* Count how many regions are actually needed to execute this add. */
688 add_reservation_in_range(resv, f, t, NULL, NULL,
689 &actual_regions_needed);
692 * Check for sufficient descriptors in the cache to accommodate
693 * this add operation. Note that actual_regions_needed may be greater
694 * than in_regions_needed, as the resv_map may have been modified since
695 * the region_chg call. In this case, we need to make sure that we
696 * allocate extra entries, such that we have enough for all the
697 * existing adds_in_progress, plus the excess needed for this
700 if (actual_regions_needed > in_regions_needed &&
701 resv->region_cache_count <
702 resv->adds_in_progress +
703 (actual_regions_needed - in_regions_needed)) {
704 /* region_add operation of range 1 should never need to
705 * allocate file_region entries.
707 VM_BUG_ON(t - f <= 1);
709 if (allocate_file_region_entries(
710 resv, actual_regions_needed - in_regions_needed)) {
717 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
719 resv->adds_in_progress -= in_regions_needed;
721 spin_unlock(&resv->lock);
726 * Examine the existing reserve map and determine how many
727 * huge pages in the specified range [f, t) are NOT currently
728 * represented. This routine is called before a subsequent
729 * call to region_add that will actually modify the reserve
730 * map to add the specified range [f, t). region_chg does
731 * not change the number of huge pages represented by the
732 * map. A number of new file_region structures is added to the cache as a
733 * placeholder, for the subsequent region_add call to use. At least 1
734 * file_region structure is added.
736 * out_regions_needed is the number of regions added to the
737 * resv->adds_in_progress. This value needs to be provided to a follow up call
738 * to region_add or region_abort for proper accounting.
740 * Returns the number of huge pages that need to be added to the existing
741 * reservation map for the range [f, t). This number is greater or equal to
742 * zero. -ENOMEM is returned if a new file_region structure or cache entry
743 * is needed and can not be allocated.
745 static long region_chg(struct resv_map *resv, long f, long t,
746 long *out_regions_needed)
750 spin_lock(&resv->lock);
752 /* Count how many hugepages in this range are NOT represented. */
753 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
756 if (*out_regions_needed == 0)
757 *out_regions_needed = 1;
759 if (allocate_file_region_entries(resv, *out_regions_needed))
762 resv->adds_in_progress += *out_regions_needed;
764 spin_unlock(&resv->lock);
769 * Abort the in progress add operation. The adds_in_progress field
770 * of the resv_map keeps track of the operations in progress between
771 * calls to region_chg and region_add. Operations are sometimes
772 * aborted after the call to region_chg. In such cases, region_abort
773 * is called to decrement the adds_in_progress counter. regions_needed
774 * is the value returned by the region_chg call, it is used to decrement
775 * the adds_in_progress counter.
777 * NOTE: The range arguments [f, t) are not needed or used in this
778 * routine. They are kept to make reading the calling code easier as
779 * arguments will match the associated region_chg call.
781 static void region_abort(struct resv_map *resv, long f, long t,
784 spin_lock(&resv->lock);
785 VM_BUG_ON(!resv->region_cache_count);
786 resv->adds_in_progress -= regions_needed;
787 spin_unlock(&resv->lock);
791 * Delete the specified range [f, t) from the reserve map. If the
792 * t parameter is LONG_MAX, this indicates that ALL regions after f
793 * should be deleted. Locate the regions which intersect [f, t)
794 * and either trim, delete or split the existing regions.
796 * Returns the number of huge pages deleted from the reserve map.
797 * In the normal case, the return value is zero or more. In the
798 * case where a region must be split, a new region descriptor must
799 * be allocated. If the allocation fails, -ENOMEM will be returned.
800 * NOTE: If the parameter t == LONG_MAX, then we will never split
801 * a region and possibly return -ENOMEM. Callers specifying
802 * t == LONG_MAX do not need to check for -ENOMEM error.
804 static long region_del(struct resv_map *resv, long f, long t)
806 struct list_head *head = &resv->regions;
807 struct file_region *rg, *trg;
808 struct file_region *nrg = NULL;
812 spin_lock(&resv->lock);
813 list_for_each_entry_safe(rg, trg, head, link) {
815 * Skip regions before the range to be deleted. file_region
816 * ranges are normally of the form [from, to). However, there
817 * may be a "placeholder" entry in the map which is of the form
818 * (from, to) with from == to. Check for placeholder entries
819 * at the beginning of the range to be deleted.
821 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
827 if (f > rg->from && t < rg->to) { /* Must split region */
829 * Check for an entry in the cache before dropping
830 * lock and attempting allocation.
833 resv->region_cache_count > resv->adds_in_progress) {
834 nrg = list_first_entry(&resv->region_cache,
837 list_del(&nrg->link);
838 resv->region_cache_count--;
842 spin_unlock(&resv->lock);
843 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
850 hugetlb_cgroup_uncharge_file_region(
851 resv, rg, t - f, false);
853 /* New entry for end of split region */
857 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
859 INIT_LIST_HEAD(&nrg->link);
861 /* Original entry is trimmed */
864 list_add(&nrg->link, &rg->link);
869 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
870 del += rg->to - rg->from;
871 hugetlb_cgroup_uncharge_file_region(resv, rg,
872 rg->to - rg->from, true);
878 if (f <= rg->from) { /* Trim beginning of region */
879 hugetlb_cgroup_uncharge_file_region(resv, rg,
880 t - rg->from, false);
884 } else { /* Trim end of region */
885 hugetlb_cgroup_uncharge_file_region(resv, rg,
893 spin_unlock(&resv->lock);
899 * A rare out of memory error was encountered which prevented removal of
900 * the reserve map region for a page. The huge page itself was free'ed
901 * and removed from the page cache. This routine will adjust the subpool
902 * usage count, and the global reserve count if needed. By incrementing
903 * these counts, the reserve map entry which could not be deleted will
904 * appear as a "reserved" entry instead of simply dangling with incorrect
907 void hugetlb_fix_reserve_counts(struct inode *inode)
909 struct hugepage_subpool *spool = subpool_inode(inode);
911 bool reserved = false;
913 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
914 if (rsv_adjust > 0) {
915 struct hstate *h = hstate_inode(inode);
917 if (!hugetlb_acct_memory(h, 1))
919 } else if (!rsv_adjust) {
924 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
928 * Count and return the number of huge pages in the reserve map
929 * that intersect with the range [f, t).
931 static long region_count(struct resv_map *resv, long f, long t)
933 struct list_head *head = &resv->regions;
934 struct file_region *rg;
937 spin_lock(&resv->lock);
938 /* Locate each segment we overlap with, and count that overlap. */
939 list_for_each_entry(rg, head, link) {
948 seg_from = max(rg->from, f);
949 seg_to = min(rg->to, t);
951 chg += seg_to - seg_from;
953 spin_unlock(&resv->lock);
959 * Convert the address within this vma to the page offset within
960 * the mapping, in pagecache page units; huge pages here.
962 static pgoff_t vma_hugecache_offset(struct hstate *h,
963 struct vm_area_struct *vma, unsigned long address)
965 return ((address - vma->vm_start) >> huge_page_shift(h)) +
966 (vma->vm_pgoff >> huge_page_order(h));
969 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
970 unsigned long address)
972 return vma_hugecache_offset(hstate_vma(vma), vma, address);
974 EXPORT_SYMBOL_GPL(linear_hugepage_index);
977 * Return the size of the pages allocated when backing a VMA. In the majority
978 * cases this will be same size as used by the page table entries.
980 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
982 if (vma->vm_ops && vma->vm_ops->pagesize)
983 return vma->vm_ops->pagesize(vma);
986 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
989 * Return the page size being used by the MMU to back a VMA. In the majority
990 * of cases, the page size used by the kernel matches the MMU size. On
991 * architectures where it differs, an architecture-specific 'strong'
992 * version of this symbol is required.
994 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
996 return vma_kernel_pagesize(vma);
1000 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1001 * bits of the reservation map pointer, which are always clear due to
1004 #define HPAGE_RESV_OWNER (1UL << 0)
1005 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1006 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1009 * These helpers are used to track how many pages are reserved for
1010 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1011 * is guaranteed to have their future faults succeed.
1013 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1014 * the reserve counters are updated with the hugetlb_lock held. It is safe
1015 * to reset the VMA at fork() time as it is not in use yet and there is no
1016 * chance of the global counters getting corrupted as a result of the values.
1018 * The private mapping reservation is represented in a subtly different
1019 * manner to a shared mapping. A shared mapping has a region map associated
1020 * with the underlying file, this region map represents the backing file
1021 * pages which have ever had a reservation assigned which this persists even
1022 * after the page is instantiated. A private mapping has a region map
1023 * associated with the original mmap which is attached to all VMAs which
1024 * reference it, this region map represents those offsets which have consumed
1025 * reservation ie. where pages have been instantiated.
1027 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1029 return (unsigned long)vma->vm_private_data;
1032 static void set_vma_private_data(struct vm_area_struct *vma,
1033 unsigned long value)
1035 vma->vm_private_data = (void *)value;
1039 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1040 struct hugetlb_cgroup *h_cg,
1043 #ifdef CONFIG_CGROUP_HUGETLB
1045 resv_map->reservation_counter = NULL;
1046 resv_map->pages_per_hpage = 0;
1047 resv_map->css = NULL;
1049 resv_map->reservation_counter =
1050 &h_cg->rsvd_hugepage[hstate_index(h)];
1051 resv_map->pages_per_hpage = pages_per_huge_page(h);
1052 resv_map->css = &h_cg->css;
1057 struct resv_map *resv_map_alloc(void)
1059 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1060 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1062 if (!resv_map || !rg) {
1068 kref_init(&resv_map->refs);
1069 spin_lock_init(&resv_map->lock);
1070 INIT_LIST_HEAD(&resv_map->regions);
1072 resv_map->adds_in_progress = 0;
1074 * Initialize these to 0. On shared mappings, 0's here indicate these
1075 * fields don't do cgroup accounting. On private mappings, these will be
1076 * re-initialized to the proper values, to indicate that hugetlb cgroup
1077 * reservations are to be un-charged from here.
1079 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1081 INIT_LIST_HEAD(&resv_map->region_cache);
1082 list_add(&rg->link, &resv_map->region_cache);
1083 resv_map->region_cache_count = 1;
1088 void resv_map_release(struct kref *ref)
1090 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1091 struct list_head *head = &resv_map->region_cache;
1092 struct file_region *rg, *trg;
1094 /* Clear out any active regions before we release the map. */
1095 region_del(resv_map, 0, LONG_MAX);
1097 /* ... and any entries left in the cache */
1098 list_for_each_entry_safe(rg, trg, head, link) {
1099 list_del(&rg->link);
1103 VM_BUG_ON(resv_map->adds_in_progress);
1108 static inline struct resv_map *inode_resv_map(struct inode *inode)
1111 * At inode evict time, i_mapping may not point to the original
1112 * address space within the inode. This original address space
1113 * contains the pointer to the resv_map. So, always use the
1114 * address space embedded within the inode.
1115 * The VERY common case is inode->mapping == &inode->i_data but,
1116 * this may not be true for device special inodes.
1118 return (struct resv_map *)(&inode->i_data)->private_data;
1121 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1123 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1124 if (vma->vm_flags & VM_MAYSHARE) {
1125 struct address_space *mapping = vma->vm_file->f_mapping;
1126 struct inode *inode = mapping->host;
1128 return inode_resv_map(inode);
1131 return (struct resv_map *)(get_vma_private_data(vma) &
1136 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1138 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1139 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1141 set_vma_private_data(vma, (get_vma_private_data(vma) &
1142 HPAGE_RESV_MASK) | (unsigned long)map);
1145 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1147 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1148 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1150 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1153 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1155 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1157 return (get_vma_private_data(vma) & flag) != 0;
1160 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1162 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1164 * Clear vm_private_data
1165 * - For shared mappings this is a per-vma semaphore that may be
1166 * allocated in a subsequent call to hugetlb_vm_op_open.
1167 * Before clearing, make sure pointer is not associated with vma
1168 * as this will leak the structure. This is the case when called
1169 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1170 * been called to allocate a new structure.
1171 * - For MAP_PRIVATE mappings, this is the reserve map which does
1172 * not apply to children. Faults generated by the children are
1173 * not guaranteed to succeed, even if read-only.
1175 if (vma->vm_flags & VM_MAYSHARE) {
1176 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1178 if (vma_lock && vma_lock->vma != vma)
1179 vma->vm_private_data = NULL;
1181 vma->vm_private_data = NULL;
1185 * Reset and decrement one ref on hugepage private reservation.
1186 * Called with mm->mmap_sem writer semaphore held.
1187 * This function should be only used by move_vma() and operate on
1188 * same sized vma. It should never come here with last ref on the
1191 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1194 * Clear the old hugetlb private page reservation.
1195 * It has already been transferred to new_vma.
1197 * During a mremap() operation of a hugetlb vma we call move_vma()
1198 * which copies vma into new_vma and unmaps vma. After the copy
1199 * operation both new_vma and vma share a reference to the resv_map
1200 * struct, and at that point vma is about to be unmapped. We don't
1201 * want to return the reservation to the pool at unmap of vma because
1202 * the reservation still lives on in new_vma, so simply decrement the
1203 * ref here and remove the resv_map reference from this vma.
1205 struct resv_map *reservations = vma_resv_map(vma);
1207 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1208 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1209 kref_put(&reservations->refs, resv_map_release);
1212 hugetlb_dup_vma_private(vma);
1215 /* Returns true if the VMA has associated reserve pages */
1216 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1218 if (vma->vm_flags & VM_NORESERVE) {
1220 * This address is already reserved by other process(chg == 0),
1221 * so, we should decrement reserved count. Without decrementing,
1222 * reserve count remains after releasing inode, because this
1223 * allocated page will go into page cache and is regarded as
1224 * coming from reserved pool in releasing step. Currently, we
1225 * don't have any other solution to deal with this situation
1226 * properly, so add work-around here.
1228 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1234 /* Shared mappings always use reserves */
1235 if (vma->vm_flags & VM_MAYSHARE) {
1237 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1238 * be a region map for all pages. The only situation where
1239 * there is no region map is if a hole was punched via
1240 * fallocate. In this case, there really are no reserves to
1241 * use. This situation is indicated if chg != 0.
1250 * Only the process that called mmap() has reserves for
1253 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1255 * Like the shared case above, a hole punch or truncate
1256 * could have been performed on the private mapping.
1257 * Examine the value of chg to determine if reserves
1258 * actually exist or were previously consumed.
1259 * Very Subtle - The value of chg comes from a previous
1260 * call to vma_needs_reserves(). The reserve map for
1261 * private mappings has different (opposite) semantics
1262 * than that of shared mappings. vma_needs_reserves()
1263 * has already taken this difference in semantics into
1264 * account. Therefore, the meaning of chg is the same
1265 * as in the shared case above. Code could easily be
1266 * combined, but keeping it separate draws attention to
1267 * subtle differences.
1278 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1280 int nid = folio_nid(folio);
1282 lockdep_assert_held(&hugetlb_lock);
1283 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1285 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1286 h->free_huge_pages++;
1287 h->free_huge_pages_node[nid]++;
1288 folio_set_hugetlb_freed(folio);
1291 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1294 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1296 lockdep_assert_held(&hugetlb_lock);
1297 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1298 if (pin && !is_longterm_pinnable_page(page))
1301 if (PageHWPoison(page))
1304 list_move(&page->lru, &h->hugepage_activelist);
1305 set_page_refcounted(page);
1306 ClearHPageFreed(page);
1307 h->free_huge_pages--;
1308 h->free_huge_pages_node[nid]--;
1315 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1318 unsigned int cpuset_mems_cookie;
1319 struct zonelist *zonelist;
1322 int node = NUMA_NO_NODE;
1324 zonelist = node_zonelist(nid, gfp_mask);
1327 cpuset_mems_cookie = read_mems_allowed_begin();
1328 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1331 if (!cpuset_zone_allowed(zone, gfp_mask))
1334 * no need to ask again on the same node. Pool is node rather than
1337 if (zone_to_nid(zone) == node)
1339 node = zone_to_nid(zone);
1341 page = dequeue_huge_page_node_exact(h, node);
1345 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1351 static unsigned long available_huge_pages(struct hstate *h)
1353 return h->free_huge_pages - h->resv_huge_pages;
1356 static struct page *dequeue_huge_page_vma(struct hstate *h,
1357 struct vm_area_struct *vma,
1358 unsigned long address, int avoid_reserve,
1361 struct page *page = NULL;
1362 struct mempolicy *mpol;
1364 nodemask_t *nodemask;
1368 * A child process with MAP_PRIVATE mappings created by their parent
1369 * have no page reserves. This check ensures that reservations are
1370 * not "stolen". The child may still get SIGKILLed
1372 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1375 /* If reserves cannot be used, ensure enough pages are in the pool */
1376 if (avoid_reserve && !available_huge_pages(h))
1379 gfp_mask = htlb_alloc_mask(h);
1380 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1382 if (mpol_is_preferred_many(mpol)) {
1383 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1385 /* Fallback to all nodes if page==NULL */
1390 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1392 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1393 SetHPageRestoreReserve(page);
1394 h->resv_huge_pages--;
1397 mpol_cond_put(mpol);
1405 * common helper functions for hstate_next_node_to_{alloc|free}.
1406 * We may have allocated or freed a huge page based on a different
1407 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1408 * be outside of *nodes_allowed. Ensure that we use an allowed
1409 * node for alloc or free.
1411 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1413 nid = next_node_in(nid, *nodes_allowed);
1414 VM_BUG_ON(nid >= MAX_NUMNODES);
1419 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1421 if (!node_isset(nid, *nodes_allowed))
1422 nid = next_node_allowed(nid, nodes_allowed);
1427 * returns the previously saved node ["this node"] from which to
1428 * allocate a persistent huge page for the pool and advance the
1429 * next node from which to allocate, handling wrap at end of node
1432 static int hstate_next_node_to_alloc(struct hstate *h,
1433 nodemask_t *nodes_allowed)
1437 VM_BUG_ON(!nodes_allowed);
1439 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1440 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1446 * helper for remove_pool_huge_page() - return the previously saved
1447 * node ["this node"] from which to free a huge page. Advance the
1448 * next node id whether or not we find a free huge page to free so
1449 * that the next attempt to free addresses the next node.
1451 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1455 VM_BUG_ON(!nodes_allowed);
1457 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1458 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1463 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1464 for (nr_nodes = nodes_weight(*mask); \
1466 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1469 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1470 for (nr_nodes = nodes_weight(*mask); \
1472 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1475 /* used to demote non-gigantic_huge pages as well */
1476 static void __destroy_compound_gigantic_folio(struct folio *folio,
1477 unsigned int order, bool demote)
1480 int nr_pages = 1 << order;
1483 atomic_set(folio_mapcount_ptr(folio), 0);
1484 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
1485 atomic_set(folio_pincount_ptr(folio), 0);
1487 for (i = 1; i < nr_pages; i++) {
1488 p = folio_page(folio, i);
1490 clear_compound_head(p);
1492 set_page_refcounted(p);
1495 folio_set_compound_order(folio, 0);
1496 __folio_clear_head(folio);
1499 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1502 __destroy_compound_gigantic_folio(folio, order, true);
1505 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1506 static void destroy_compound_gigantic_folio(struct folio *folio,
1509 __destroy_compound_gigantic_folio(folio, order, false);
1512 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1515 * If the page isn't allocated using the cma allocator,
1516 * cma_release() returns false.
1519 int nid = folio_nid(folio);
1521 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1525 free_contig_range(folio_pfn(folio), 1 << order);
1528 #ifdef CONFIG_CONTIG_ALLOC
1529 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1530 int nid, nodemask_t *nodemask)
1533 unsigned long nr_pages = pages_per_huge_page(h);
1534 if (nid == NUMA_NO_NODE)
1535 nid = numa_mem_id();
1541 if (hugetlb_cma[nid]) {
1542 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1543 huge_page_order(h), true);
1545 return page_folio(page);
1548 if (!(gfp_mask & __GFP_THISNODE)) {
1549 for_each_node_mask(node, *nodemask) {
1550 if (node == nid || !hugetlb_cma[node])
1553 page = cma_alloc(hugetlb_cma[node], nr_pages,
1554 huge_page_order(h), true);
1556 return page_folio(page);
1562 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1563 return page ? page_folio(page) : NULL;
1566 #else /* !CONFIG_CONTIG_ALLOC */
1567 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1568 int nid, nodemask_t *nodemask)
1572 #endif /* CONFIG_CONTIG_ALLOC */
1574 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1575 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1576 int nid, nodemask_t *nodemask)
1580 static inline void free_gigantic_folio(struct folio *folio,
1581 unsigned int order) { }
1582 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1583 unsigned int order) { }
1587 * Remove hugetlb folio from lists, and update dtor so that the folio appears
1588 * as just a compound page.
1590 * A reference is held on the folio, except in the case of demote.
1592 * Must be called with hugetlb lock held.
1594 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1595 bool adjust_surplus,
1598 int nid = folio_nid(folio);
1600 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1601 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1603 lockdep_assert_held(&hugetlb_lock);
1604 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1607 list_del(&folio->lru);
1609 if (folio_test_hugetlb_freed(folio)) {
1610 h->free_huge_pages--;
1611 h->free_huge_pages_node[nid]--;
1613 if (adjust_surplus) {
1614 h->surplus_huge_pages--;
1615 h->surplus_huge_pages_node[nid]--;
1621 * For non-gigantic pages set the destructor to the normal compound
1622 * page dtor. This is needed in case someone takes an additional
1623 * temporary ref to the page, and freeing is delayed until they drop
1626 * For gigantic pages set the destructor to the null dtor. This
1627 * destructor will never be called. Before freeing the gigantic
1628 * page destroy_compound_gigantic_folio will turn the folio into a
1629 * simple group of pages. After this the destructor does not
1632 * This handles the case where more than one ref is held when and
1633 * after update_and_free_hugetlb_folio is called.
1635 * In the case of demote we do not ref count the page as it will soon
1636 * be turned into a page of smaller size.
1639 folio_ref_unfreeze(folio, 1);
1640 if (hstate_is_gigantic(h))
1641 folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR);
1643 folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR);
1646 h->nr_huge_pages_node[nid]--;
1649 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1650 bool adjust_surplus)
1652 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1655 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1656 bool adjust_surplus)
1658 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1661 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1662 bool adjust_surplus)
1665 int nid = folio_nid(folio);
1667 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1669 lockdep_assert_held(&hugetlb_lock);
1671 INIT_LIST_HEAD(&folio->lru);
1673 h->nr_huge_pages_node[nid]++;
1675 if (adjust_surplus) {
1676 h->surplus_huge_pages++;
1677 h->surplus_huge_pages_node[nid]++;
1680 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1681 folio_change_private(folio, NULL);
1683 * We have to set hugetlb_vmemmap_optimized again as above
1684 * folio_change_private(folio, NULL) cleared it.
1686 folio_set_hugetlb_vmemmap_optimized(folio);
1689 * This folio is about to be managed by the hugetlb allocator and
1690 * should have no users. Drop our reference, and check for others
1693 zeroed = folio_put_testzero(folio);
1694 if (unlikely(!zeroed))
1696 * It is VERY unlikely soneone else has taken a ref on
1697 * the page. In this case, we simply return as the
1698 * hugetlb destructor (free_huge_page) will be called
1699 * when this other ref is dropped.
1703 arch_clear_hugepage_flags(&folio->page);
1704 enqueue_hugetlb_folio(h, folio);
1707 static void __update_and_free_page(struct hstate *h, struct page *page)
1710 struct folio *folio = page_folio(page);
1711 struct page *subpage;
1713 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1717 * If we don't know which subpages are hwpoisoned, we can't free
1718 * the hugepage, so it's leaked intentionally.
1720 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1723 if (hugetlb_vmemmap_restore(h, page)) {
1724 spin_lock_irq(&hugetlb_lock);
1726 * If we cannot allocate vmemmap pages, just refuse to free the
1727 * page and put the page back on the hugetlb free list and treat
1728 * as a surplus page.
1730 add_hugetlb_folio(h, folio, true);
1731 spin_unlock_irq(&hugetlb_lock);
1736 * Move PageHWPoison flag from head page to the raw error pages,
1737 * which makes any healthy subpages reusable.
1739 if (unlikely(folio_test_hwpoison(folio)))
1740 hugetlb_clear_page_hwpoison(&folio->page);
1742 for (i = 0; i < pages_per_huge_page(h); i++) {
1743 subpage = folio_page(folio, i);
1744 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1745 1 << PG_referenced | 1 << PG_dirty |
1746 1 << PG_active | 1 << PG_private |
1751 * Non-gigantic pages demoted from CMA allocated gigantic pages
1752 * need to be given back to CMA in free_gigantic_folio.
1754 if (hstate_is_gigantic(h) ||
1755 hugetlb_cma_folio(folio, huge_page_order(h))) {
1756 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1757 free_gigantic_folio(folio, huge_page_order(h));
1759 __free_pages(page, huge_page_order(h));
1764 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1765 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1766 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1767 * the vmemmap pages.
1769 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1770 * freed and frees them one-by-one. As the page->mapping pointer is going
1771 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1772 * structure of a lockless linked list of huge pages to be freed.
1774 static LLIST_HEAD(hpage_freelist);
1776 static void free_hpage_workfn(struct work_struct *work)
1778 struct llist_node *node;
1780 node = llist_del_all(&hpage_freelist);
1786 page = container_of((struct address_space **)node,
1787 struct page, mapping);
1789 page->mapping = NULL;
1791 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1792 * is going to trigger because a previous call to
1793 * remove_hugetlb_folio() will call folio_set_compound_dtor
1794 * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate()
1797 h = size_to_hstate(page_size(page));
1799 __update_and_free_page(h, page);
1804 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1806 static inline void flush_free_hpage_work(struct hstate *h)
1808 if (hugetlb_vmemmap_optimizable(h))
1809 flush_work(&free_hpage_work);
1812 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1815 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1816 __update_and_free_page(h, &folio->page);
1821 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1823 * Only call schedule_work() if hpage_freelist is previously
1824 * empty. Otherwise, schedule_work() had been called but the workfn
1825 * hasn't retrieved the list yet.
1827 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1828 schedule_work(&free_hpage_work);
1831 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1833 struct page *page, *t_page;
1834 struct folio *folio;
1836 list_for_each_entry_safe(page, t_page, list, lru) {
1837 folio = page_folio(page);
1838 update_and_free_hugetlb_folio(h, folio, false);
1843 struct hstate *size_to_hstate(unsigned long size)
1847 for_each_hstate(h) {
1848 if (huge_page_size(h) == size)
1854 void free_huge_page(struct page *page)
1857 * Can't pass hstate in here because it is called from the
1858 * compound page destructor.
1860 struct folio *folio = page_folio(page);
1861 struct hstate *h = folio_hstate(folio);
1862 int nid = folio_nid(folio);
1863 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1864 bool restore_reserve;
1865 unsigned long flags;
1867 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1868 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1870 hugetlb_set_folio_subpool(folio, NULL);
1871 if (folio_test_anon(folio))
1872 __ClearPageAnonExclusive(&folio->page);
1873 folio->mapping = NULL;
1874 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1875 folio_clear_hugetlb_restore_reserve(folio);
1878 * If HPageRestoreReserve was set on page, page allocation consumed a
1879 * reservation. If the page was associated with a subpool, there
1880 * would have been a page reserved in the subpool before allocation
1881 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1882 * reservation, do not call hugepage_subpool_put_pages() as this will
1883 * remove the reserved page from the subpool.
1885 if (!restore_reserve) {
1887 * A return code of zero implies that the subpool will be
1888 * under its minimum size if the reservation is not restored
1889 * after page is free. Therefore, force restore_reserve
1892 if (hugepage_subpool_put_pages(spool, 1) == 0)
1893 restore_reserve = true;
1896 spin_lock_irqsave(&hugetlb_lock, flags);
1897 folio_clear_hugetlb_migratable(folio);
1898 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1899 pages_per_huge_page(h), folio);
1900 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1901 pages_per_huge_page(h), folio);
1902 if (restore_reserve)
1903 h->resv_huge_pages++;
1905 if (folio_test_hugetlb_temporary(folio)) {
1906 remove_hugetlb_folio(h, folio, false);
1907 spin_unlock_irqrestore(&hugetlb_lock, flags);
1908 update_and_free_hugetlb_folio(h, folio, true);
1909 } else if (h->surplus_huge_pages_node[nid]) {
1910 /* remove the page from active list */
1911 remove_hugetlb_folio(h, folio, true);
1912 spin_unlock_irqrestore(&hugetlb_lock, flags);
1913 update_and_free_hugetlb_folio(h, folio, true);
1915 arch_clear_hugepage_flags(page);
1916 enqueue_hugetlb_folio(h, folio);
1917 spin_unlock_irqrestore(&hugetlb_lock, flags);
1922 * Must be called with the hugetlb lock held
1924 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1926 lockdep_assert_held(&hugetlb_lock);
1928 h->nr_huge_pages_node[nid]++;
1931 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1933 hugetlb_vmemmap_optimize(h, &folio->page);
1934 INIT_LIST_HEAD(&folio->lru);
1935 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1936 hugetlb_set_folio_subpool(folio, NULL);
1937 set_hugetlb_cgroup(folio, NULL);
1938 set_hugetlb_cgroup_rsvd(folio, NULL);
1941 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1943 __prep_new_hugetlb_folio(h, folio);
1944 spin_lock_irq(&hugetlb_lock);
1945 __prep_account_new_huge_page(h, nid);
1946 spin_unlock_irq(&hugetlb_lock);
1949 static bool __prep_compound_gigantic_folio(struct folio *folio,
1950 unsigned int order, bool demote)
1953 int nr_pages = 1 << order;
1956 __folio_clear_reserved(folio);
1957 __folio_set_head(folio);
1958 /* we rely on prep_new_hugetlb_folio to set the destructor */
1959 folio_set_compound_order(folio, order);
1960 for (i = 0; i < nr_pages; i++) {
1961 p = folio_page(folio, i);
1964 * For gigantic hugepages allocated through bootmem at
1965 * boot, it's safer to be consistent with the not-gigantic
1966 * hugepages and clear the PG_reserved bit from all tail pages
1967 * too. Otherwise drivers using get_user_pages() to access tail
1968 * pages may get the reference counting wrong if they see
1969 * PG_reserved set on a tail page (despite the head page not
1970 * having PG_reserved set). Enforcing this consistency between
1971 * head and tail pages allows drivers to optimize away a check
1972 * on the head page when they need know if put_page() is needed
1973 * after get_user_pages().
1975 if (i != 0) /* head page cleared above */
1976 __ClearPageReserved(p);
1978 * Subtle and very unlikely
1980 * Gigantic 'page allocators' such as memblock or cma will
1981 * return a set of pages with each page ref counted. We need
1982 * to turn this set of pages into a compound page with tail
1983 * page ref counts set to zero. Code such as speculative page
1984 * cache adding could take a ref on a 'to be' tail page.
1985 * We need to respect any increased ref count, and only set
1986 * the ref count to zero if count is currently 1. If count
1987 * is not 1, we return an error. An error return indicates
1988 * the set of pages can not be converted to a gigantic page.
1989 * The caller who allocated the pages should then discard the
1990 * pages using the appropriate free interface.
1992 * In the case of demote, the ref count will be zero.
1995 if (!page_ref_freeze(p, 1)) {
1996 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2000 VM_BUG_ON_PAGE(page_count(p), p);
2003 set_compound_head(p, &folio->page);
2005 atomic_set(folio_mapcount_ptr(folio), -1);
2006 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
2007 atomic_set(folio_pincount_ptr(folio), 0);
2011 /* undo page modifications made above */
2012 for (j = 0; j < i; j++) {
2013 p = folio_page(folio, j);
2015 clear_compound_head(p);
2016 set_page_refcounted(p);
2018 /* need to clear PG_reserved on remaining tail pages */
2019 for (; j < nr_pages; j++) {
2020 p = folio_page(folio, j);
2021 __ClearPageReserved(p);
2023 folio_set_compound_order(folio, 0);
2024 __folio_clear_head(folio);
2028 static bool prep_compound_gigantic_folio(struct folio *folio,
2031 return __prep_compound_gigantic_folio(folio, order, false);
2034 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2037 return __prep_compound_gigantic_folio(folio, order, true);
2041 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2042 * transparent huge pages. See the PageTransHuge() documentation for more
2045 int PageHuge(struct page *page)
2047 if (!PageCompound(page))
2050 page = compound_head(page);
2051 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
2053 EXPORT_SYMBOL_GPL(PageHuge);
2056 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
2057 * normal or transparent huge pages.
2059 int PageHeadHuge(struct page *page_head)
2061 if (!PageHead(page_head))
2064 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
2066 EXPORT_SYMBOL_GPL(PageHeadHuge);
2069 * Find and lock address space (mapping) in write mode.
2071 * Upon entry, the page is locked which means that page_mapping() is
2072 * stable. Due to locking order, we can only trylock_write. If we can
2073 * not get the lock, simply return NULL to caller.
2075 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2077 struct address_space *mapping = page_mapping(hpage);
2082 if (i_mmap_trylock_write(mapping))
2088 pgoff_t hugetlb_basepage_index(struct page *page)
2090 struct page *page_head = compound_head(page);
2091 pgoff_t index = page_index(page_head);
2092 unsigned long compound_idx;
2094 if (compound_order(page_head) >= MAX_ORDER)
2095 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2097 compound_idx = page - page_head;
2099 return (index << compound_order(page_head)) + compound_idx;
2102 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2103 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2104 nodemask_t *node_alloc_noretry)
2106 int order = huge_page_order(h);
2108 bool alloc_try_hard = true;
2112 * By default we always try hard to allocate the page with
2113 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2114 * a loop (to adjust global huge page counts) and previous allocation
2115 * failed, do not continue to try hard on the same node. Use the
2116 * node_alloc_noretry bitmap to manage this state information.
2118 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2119 alloc_try_hard = false;
2120 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2122 gfp_mask |= __GFP_RETRY_MAYFAIL;
2123 if (nid == NUMA_NO_NODE)
2124 nid = numa_mem_id();
2126 page = __alloc_pages(gfp_mask, order, nid, nmask);
2128 /* Freeze head page */
2129 if (page && !page_ref_freeze(page, 1)) {
2130 __free_pages(page, order);
2131 if (retry) { /* retry once */
2135 /* WOW! twice in a row. */
2136 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2141 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2142 * indicates an overall state change. Clear bit so that we resume
2143 * normal 'try hard' allocations.
2145 if (node_alloc_noretry && page && !alloc_try_hard)
2146 node_clear(nid, *node_alloc_noretry);
2149 * If we tried hard to get a page but failed, set bit so that
2150 * subsequent attempts will not try as hard until there is an
2151 * overall state change.
2153 if (node_alloc_noretry && !page && alloc_try_hard)
2154 node_set(nid, *node_alloc_noretry);
2157 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2161 __count_vm_event(HTLB_BUDDY_PGALLOC);
2162 return page_folio(page);
2166 * Common helper to allocate a fresh hugetlb page. All specific allocators
2167 * should use this function to get new hugetlb pages
2169 * Note that returned page is 'frozen': ref count of head page and all tail
2172 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2173 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2174 nodemask_t *node_alloc_noretry)
2176 struct folio *folio;
2180 if (hstate_is_gigantic(h))
2181 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2183 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2184 nid, nmask, node_alloc_noretry);
2187 if (hstate_is_gigantic(h)) {
2188 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2190 * Rare failure to convert pages to compound page.
2191 * Free pages and try again - ONCE!
2193 free_gigantic_folio(folio, huge_page_order(h));
2201 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2207 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2210 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2211 nodemask_t *node_alloc_noretry)
2213 struct folio *folio;
2215 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2217 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2218 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2219 nodes_allowed, node_alloc_noretry);
2221 free_huge_page(&folio->page); /* free it into the hugepage allocator */
2230 * Remove huge page from pool from next node to free. Attempt to keep
2231 * persistent huge pages more or less balanced over allowed nodes.
2232 * This routine only 'removes' the hugetlb page. The caller must make
2233 * an additional call to free the page to low level allocators.
2234 * Called with hugetlb_lock locked.
2236 static struct page *remove_pool_huge_page(struct hstate *h,
2237 nodemask_t *nodes_allowed,
2241 struct page *page = NULL;
2242 struct folio *folio;
2244 lockdep_assert_held(&hugetlb_lock);
2245 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2247 * If we're returning unused surplus pages, only examine
2248 * nodes with surplus pages.
2250 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2251 !list_empty(&h->hugepage_freelists[node])) {
2252 page = list_entry(h->hugepage_freelists[node].next,
2254 folio = page_folio(page);
2255 remove_hugetlb_folio(h, folio, acct_surplus);
2264 * Dissolve a given free hugepage into free buddy pages. This function does
2265 * nothing for in-use hugepages and non-hugepages.
2266 * This function returns values like below:
2268 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2269 * when the system is under memory pressure and the feature of
2270 * freeing unused vmemmap pages associated with each hugetlb page
2272 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2273 * (allocated or reserved.)
2274 * 0: successfully dissolved free hugepages or the page is not a
2275 * hugepage (considered as already dissolved)
2277 int dissolve_free_huge_page(struct page *page)
2280 struct folio *folio = page_folio(page);
2283 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2284 if (!folio_test_hugetlb(folio))
2287 spin_lock_irq(&hugetlb_lock);
2288 if (!folio_test_hugetlb(folio)) {
2293 if (!folio_ref_count(folio)) {
2294 struct hstate *h = folio_hstate(folio);
2295 if (!available_huge_pages(h))
2299 * We should make sure that the page is already on the free list
2300 * when it is dissolved.
2302 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2303 spin_unlock_irq(&hugetlb_lock);
2307 * Theoretically, we should return -EBUSY when we
2308 * encounter this race. In fact, we have a chance
2309 * to successfully dissolve the page if we do a
2310 * retry. Because the race window is quite small.
2311 * If we seize this opportunity, it is an optimization
2312 * for increasing the success rate of dissolving page.
2317 remove_hugetlb_folio(h, folio, false);
2318 h->max_huge_pages--;
2319 spin_unlock_irq(&hugetlb_lock);
2322 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2323 * before freeing the page. update_and_free_hugtlb_folio will fail to
2324 * free the page if it can not allocate required vmemmap. We
2325 * need to adjust max_huge_pages if the page is not freed.
2326 * Attempt to allocate vmemmmap here so that we can take
2327 * appropriate action on failure.
2329 rc = hugetlb_vmemmap_restore(h, &folio->page);
2331 update_and_free_hugetlb_folio(h, folio, false);
2333 spin_lock_irq(&hugetlb_lock);
2334 add_hugetlb_folio(h, folio, false);
2335 h->max_huge_pages++;
2336 spin_unlock_irq(&hugetlb_lock);
2342 spin_unlock_irq(&hugetlb_lock);
2347 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2348 * make specified memory blocks removable from the system.
2349 * Note that this will dissolve a free gigantic hugepage completely, if any
2350 * part of it lies within the given range.
2351 * Also note that if dissolve_free_huge_page() returns with an error, all
2352 * free hugepages that were dissolved before that error are lost.
2354 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2362 if (!hugepages_supported())
2365 order = huge_page_order(&default_hstate);
2367 order = min(order, huge_page_order(h));
2369 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2370 page = pfn_to_page(pfn);
2371 rc = dissolve_free_huge_page(page);
2380 * Allocates a fresh surplus page from the page allocator.
2382 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2383 int nid, nodemask_t *nmask)
2385 struct folio *folio = NULL;
2387 if (hstate_is_gigantic(h))
2390 spin_lock_irq(&hugetlb_lock);
2391 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2393 spin_unlock_irq(&hugetlb_lock);
2395 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2399 spin_lock_irq(&hugetlb_lock);
2401 * We could have raced with the pool size change.
2402 * Double check that and simply deallocate the new page
2403 * if we would end up overcommiting the surpluses. Abuse
2404 * temporary page to workaround the nasty free_huge_page
2407 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2408 folio_set_hugetlb_temporary(folio);
2409 spin_unlock_irq(&hugetlb_lock);
2410 free_huge_page(&folio->page);
2414 h->surplus_huge_pages++;
2415 h->surplus_huge_pages_node[folio_nid(folio)]++;
2418 spin_unlock_irq(&hugetlb_lock);
2420 return &folio->page;
2423 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2424 int nid, nodemask_t *nmask)
2426 struct folio *folio;
2428 if (hstate_is_gigantic(h))
2431 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2435 /* fresh huge pages are frozen */
2436 folio_ref_unfreeze(folio, 1);
2438 * We do not account these pages as surplus because they are only
2439 * temporary and will be released properly on the last reference
2441 folio_set_hugetlb_temporary(folio);
2443 return &folio->page;
2447 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2450 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2451 struct vm_area_struct *vma, unsigned long addr)
2453 struct page *page = NULL;
2454 struct mempolicy *mpol;
2455 gfp_t gfp_mask = htlb_alloc_mask(h);
2457 nodemask_t *nodemask;
2459 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2460 if (mpol_is_preferred_many(mpol)) {
2461 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2463 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2464 page = alloc_surplus_huge_page(h, gfp, nid, nodemask);
2466 /* Fallback to all nodes if page==NULL */
2471 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2472 mpol_cond_put(mpol);
2476 /* page migration callback function */
2477 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2478 nodemask_t *nmask, gfp_t gfp_mask)
2480 spin_lock_irq(&hugetlb_lock);
2481 if (available_huge_pages(h)) {
2484 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2486 spin_unlock_irq(&hugetlb_lock);
2490 spin_unlock_irq(&hugetlb_lock);
2492 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2495 /* mempolicy aware migration callback */
2496 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2497 unsigned long address)
2499 struct mempolicy *mpol;
2500 nodemask_t *nodemask;
2505 gfp_mask = htlb_alloc_mask(h);
2506 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2507 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2508 mpol_cond_put(mpol);
2514 * Increase the hugetlb pool such that it can accommodate a reservation
2517 static int gather_surplus_pages(struct hstate *h, long delta)
2518 __must_hold(&hugetlb_lock)
2520 LIST_HEAD(surplus_list);
2521 struct page *page, *tmp;
2524 long needed, allocated;
2525 bool alloc_ok = true;
2527 lockdep_assert_held(&hugetlb_lock);
2528 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2530 h->resv_huge_pages += delta;
2538 spin_unlock_irq(&hugetlb_lock);
2539 for (i = 0; i < needed; i++) {
2540 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2541 NUMA_NO_NODE, NULL);
2546 list_add(&page->lru, &surplus_list);
2552 * After retaking hugetlb_lock, we need to recalculate 'needed'
2553 * because either resv_huge_pages or free_huge_pages may have changed.
2555 spin_lock_irq(&hugetlb_lock);
2556 needed = (h->resv_huge_pages + delta) -
2557 (h->free_huge_pages + allocated);
2562 * We were not able to allocate enough pages to
2563 * satisfy the entire reservation so we free what
2564 * we've allocated so far.
2569 * The surplus_list now contains _at_least_ the number of extra pages
2570 * needed to accommodate the reservation. Add the appropriate number
2571 * of pages to the hugetlb pool and free the extras back to the buddy
2572 * allocator. Commit the entire reservation here to prevent another
2573 * process from stealing the pages as they are added to the pool but
2574 * before they are reserved.
2576 needed += allocated;
2577 h->resv_huge_pages += delta;
2580 /* Free the needed pages to the hugetlb pool */
2581 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2584 /* Add the page to the hugetlb allocator */
2585 enqueue_hugetlb_folio(h, page_folio(page));
2588 spin_unlock_irq(&hugetlb_lock);
2591 * Free unnecessary surplus pages to the buddy allocator.
2592 * Pages have no ref count, call free_huge_page directly.
2594 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2595 free_huge_page(page);
2596 spin_lock_irq(&hugetlb_lock);
2602 * This routine has two main purposes:
2603 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2604 * in unused_resv_pages. This corresponds to the prior adjustments made
2605 * to the associated reservation map.
2606 * 2) Free any unused surplus pages that may have been allocated to satisfy
2607 * the reservation. As many as unused_resv_pages may be freed.
2609 static void return_unused_surplus_pages(struct hstate *h,
2610 unsigned long unused_resv_pages)
2612 unsigned long nr_pages;
2614 LIST_HEAD(page_list);
2616 lockdep_assert_held(&hugetlb_lock);
2617 /* Uncommit the reservation */
2618 h->resv_huge_pages -= unused_resv_pages;
2620 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2624 * Part (or even all) of the reservation could have been backed
2625 * by pre-allocated pages. Only free surplus pages.
2627 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2630 * We want to release as many surplus pages as possible, spread
2631 * evenly across all nodes with memory. Iterate across these nodes
2632 * until we can no longer free unreserved surplus pages. This occurs
2633 * when the nodes with surplus pages have no free pages.
2634 * remove_pool_huge_page() will balance the freed pages across the
2635 * on-line nodes with memory and will handle the hstate accounting.
2637 while (nr_pages--) {
2638 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2642 list_add(&page->lru, &page_list);
2646 spin_unlock_irq(&hugetlb_lock);
2647 update_and_free_pages_bulk(h, &page_list);
2648 spin_lock_irq(&hugetlb_lock);
2653 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2654 * are used by the huge page allocation routines to manage reservations.
2656 * vma_needs_reservation is called to determine if the huge page at addr
2657 * within the vma has an associated reservation. If a reservation is
2658 * needed, the value 1 is returned. The caller is then responsible for
2659 * managing the global reservation and subpool usage counts. After
2660 * the huge page has been allocated, vma_commit_reservation is called
2661 * to add the page to the reservation map. If the page allocation fails,
2662 * the reservation must be ended instead of committed. vma_end_reservation
2663 * is called in such cases.
2665 * In the normal case, vma_commit_reservation returns the same value
2666 * as the preceding vma_needs_reservation call. The only time this
2667 * is not the case is if a reserve map was changed between calls. It
2668 * is the responsibility of the caller to notice the difference and
2669 * take appropriate action.
2671 * vma_add_reservation is used in error paths where a reservation must
2672 * be restored when a newly allocated huge page must be freed. It is
2673 * to be called after calling vma_needs_reservation to determine if a
2674 * reservation exists.
2676 * vma_del_reservation is used in error paths where an entry in the reserve
2677 * map was created during huge page allocation and must be removed. It is to
2678 * be called after calling vma_needs_reservation to determine if a reservation
2681 enum vma_resv_mode {
2688 static long __vma_reservation_common(struct hstate *h,
2689 struct vm_area_struct *vma, unsigned long addr,
2690 enum vma_resv_mode mode)
2692 struct resv_map *resv;
2695 long dummy_out_regions_needed;
2697 resv = vma_resv_map(vma);
2701 idx = vma_hugecache_offset(h, vma, addr);
2703 case VMA_NEEDS_RESV:
2704 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2705 /* We assume that vma_reservation_* routines always operate on
2706 * 1 page, and that adding to resv map a 1 page entry can only
2707 * ever require 1 region.
2709 VM_BUG_ON(dummy_out_regions_needed != 1);
2711 case VMA_COMMIT_RESV:
2712 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2713 /* region_add calls of range 1 should never fail. */
2717 region_abort(resv, idx, idx + 1, 1);
2721 if (vma->vm_flags & VM_MAYSHARE) {
2722 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2723 /* region_add calls of range 1 should never fail. */
2726 region_abort(resv, idx, idx + 1, 1);
2727 ret = region_del(resv, idx, idx + 1);
2731 if (vma->vm_flags & VM_MAYSHARE) {
2732 region_abort(resv, idx, idx + 1, 1);
2733 ret = region_del(resv, idx, idx + 1);
2735 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2736 /* region_add calls of range 1 should never fail. */
2744 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2747 * We know private mapping must have HPAGE_RESV_OWNER set.
2749 * In most cases, reserves always exist for private mappings.
2750 * However, a file associated with mapping could have been
2751 * hole punched or truncated after reserves were consumed.
2752 * As subsequent fault on such a range will not use reserves.
2753 * Subtle - The reserve map for private mappings has the
2754 * opposite meaning than that of shared mappings. If NO
2755 * entry is in the reserve map, it means a reservation exists.
2756 * If an entry exists in the reserve map, it means the
2757 * reservation has already been consumed. As a result, the
2758 * return value of this routine is the opposite of the
2759 * value returned from reserve map manipulation routines above.
2768 static long vma_needs_reservation(struct hstate *h,
2769 struct vm_area_struct *vma, unsigned long addr)
2771 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2774 static long vma_commit_reservation(struct hstate *h,
2775 struct vm_area_struct *vma, unsigned long addr)
2777 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2780 static void vma_end_reservation(struct hstate *h,
2781 struct vm_area_struct *vma, unsigned long addr)
2783 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2786 static long vma_add_reservation(struct hstate *h,
2787 struct vm_area_struct *vma, unsigned long addr)
2789 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2792 static long vma_del_reservation(struct hstate *h,
2793 struct vm_area_struct *vma, unsigned long addr)
2795 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2799 * This routine is called to restore reservation information on error paths.
2800 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2801 * the hugetlb mutex should remain held when calling this routine.
2803 * It handles two specific cases:
2804 * 1) A reservation was in place and the page consumed the reservation.
2805 * HPageRestoreReserve is set in the page.
2806 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2807 * not set. However, alloc_huge_page always updates the reserve map.
2809 * In case 1, free_huge_page later in the error path will increment the
2810 * global reserve count. But, free_huge_page does not have enough context
2811 * to adjust the reservation map. This case deals primarily with private
2812 * mappings. Adjust the reserve map here to be consistent with global
2813 * reserve count adjustments to be made by free_huge_page. Make sure the
2814 * reserve map indicates there is a reservation present.
2816 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2818 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2819 unsigned long address, struct page *page)
2821 long rc = vma_needs_reservation(h, vma, address);
2823 if (HPageRestoreReserve(page)) {
2824 if (unlikely(rc < 0))
2826 * Rare out of memory condition in reserve map
2827 * manipulation. Clear HPageRestoreReserve so that
2828 * global reserve count will not be incremented
2829 * by free_huge_page. This will make it appear
2830 * as though the reservation for this page was
2831 * consumed. This may prevent the task from
2832 * faulting in the page at a later time. This
2833 * is better than inconsistent global huge page
2834 * accounting of reserve counts.
2836 ClearHPageRestoreReserve(page);
2838 (void)vma_add_reservation(h, vma, address);
2840 vma_end_reservation(h, vma, address);
2844 * This indicates there is an entry in the reserve map
2845 * not added by alloc_huge_page. We know it was added
2846 * before the alloc_huge_page call, otherwise
2847 * HPageRestoreReserve would be set on the page.
2848 * Remove the entry so that a subsequent allocation
2849 * does not consume a reservation.
2851 rc = vma_del_reservation(h, vma, address);
2854 * VERY rare out of memory condition. Since
2855 * we can not delete the entry, set
2856 * HPageRestoreReserve so that the reserve
2857 * count will be incremented when the page
2858 * is freed. This reserve will be consumed
2859 * on a subsequent allocation.
2861 SetHPageRestoreReserve(page);
2862 } else if (rc < 0) {
2864 * Rare out of memory condition from
2865 * vma_needs_reservation call. Memory allocation is
2866 * only attempted if a new entry is needed. Therefore,
2867 * this implies there is not an entry in the
2870 * For shared mappings, no entry in the map indicates
2871 * no reservation. We are done.
2873 if (!(vma->vm_flags & VM_MAYSHARE))
2875 * For private mappings, no entry indicates
2876 * a reservation is present. Since we can
2877 * not add an entry, set SetHPageRestoreReserve
2878 * on the page so reserve count will be
2879 * incremented when freed. This reserve will
2880 * be consumed on a subsequent allocation.
2882 SetHPageRestoreReserve(page);
2885 * No reservation present, do nothing
2887 vma_end_reservation(h, vma, address);
2892 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2894 * @h: struct hstate old page belongs to
2895 * @old_folio: Old folio to dissolve
2896 * @list: List to isolate the page in case we need to
2897 * Returns 0 on success, otherwise negated error.
2899 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2900 struct folio *old_folio, struct list_head *list)
2902 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2903 int nid = folio_nid(old_folio);
2904 struct folio *new_folio;
2908 * Before dissolving the folio, we need to allocate a new one for the
2909 * pool to remain stable. Here, we allocate the folio and 'prep' it
2910 * by doing everything but actually updating counters and adding to
2911 * the pool. This simplifies and let us do most of the processing
2914 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2917 __prep_new_hugetlb_folio(h, new_folio);
2920 spin_lock_irq(&hugetlb_lock);
2921 if (!folio_test_hugetlb(old_folio)) {
2923 * Freed from under us. Drop new_folio too.
2926 } else if (folio_ref_count(old_folio)) {
2928 * Someone has grabbed the folio, try to isolate it here.
2929 * Fail with -EBUSY if not possible.
2931 spin_unlock_irq(&hugetlb_lock);
2932 ret = isolate_hugetlb(&old_folio->page, list);
2933 spin_lock_irq(&hugetlb_lock);
2935 } else if (!folio_test_hugetlb_freed(old_folio)) {
2937 * Folio's refcount is 0 but it has not been enqueued in the
2938 * freelist yet. Race window is small, so we can succeed here if
2941 spin_unlock_irq(&hugetlb_lock);
2946 * Ok, old_folio is still a genuine free hugepage. Remove it from
2947 * the freelist and decrease the counters. These will be
2948 * incremented again when calling __prep_account_new_huge_page()
2949 * and enqueue_hugetlb_folio() for new_folio. The counters will
2950 * remain stable since this happens under the lock.
2952 remove_hugetlb_folio(h, old_folio, false);
2955 * Ref count on new_folio is already zero as it was dropped
2956 * earlier. It can be directly added to the pool free list.
2958 __prep_account_new_huge_page(h, nid);
2959 enqueue_hugetlb_folio(h, new_folio);
2962 * Folio has been replaced, we can safely free the old one.
2964 spin_unlock_irq(&hugetlb_lock);
2965 update_and_free_hugetlb_folio(h, old_folio, false);
2971 spin_unlock_irq(&hugetlb_lock);
2972 /* Folio has a zero ref count, but needs a ref to be freed */
2973 folio_ref_unfreeze(new_folio, 1);
2974 update_and_free_hugetlb_folio(h, new_folio, false);
2979 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2982 struct folio *folio = page_folio(page);
2986 * The page might have been dissolved from under our feet, so make sure
2987 * to carefully check the state under the lock.
2988 * Return success when racing as if we dissolved the page ourselves.
2990 spin_lock_irq(&hugetlb_lock);
2991 if (folio_test_hugetlb(folio)) {
2992 h = folio_hstate(folio);
2994 spin_unlock_irq(&hugetlb_lock);
2997 spin_unlock_irq(&hugetlb_lock);
3000 * Fence off gigantic pages as there is a cyclic dependency between
3001 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3002 * of bailing out right away without further retrying.
3004 if (hstate_is_gigantic(h))
3007 if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list))
3009 else if (!folio_ref_count(folio))
3010 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3015 struct page *alloc_huge_page(struct vm_area_struct *vma,
3016 unsigned long addr, int avoid_reserve)
3018 struct hugepage_subpool *spool = subpool_vma(vma);
3019 struct hstate *h = hstate_vma(vma);
3021 struct folio *folio;
3022 long map_chg, map_commit;
3025 struct hugetlb_cgroup *h_cg;
3026 bool deferred_reserve;
3028 idx = hstate_index(h);
3030 * Examine the region/reserve map to determine if the process
3031 * has a reservation for the page to be allocated. A return
3032 * code of zero indicates a reservation exists (no change).
3034 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3036 return ERR_PTR(-ENOMEM);
3039 * Processes that did not create the mapping will have no
3040 * reserves as indicated by the region/reserve map. Check
3041 * that the allocation will not exceed the subpool limit.
3042 * Allocations for MAP_NORESERVE mappings also need to be
3043 * checked against any subpool limit.
3045 if (map_chg || avoid_reserve) {
3046 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3048 vma_end_reservation(h, vma, addr);
3049 return ERR_PTR(-ENOSPC);
3053 * Even though there was no reservation in the region/reserve
3054 * map, there could be reservations associated with the
3055 * subpool that can be used. This would be indicated if the
3056 * return value of hugepage_subpool_get_pages() is zero.
3057 * However, if avoid_reserve is specified we still avoid even
3058 * the subpool reservations.
3064 /* If this allocation is not consuming a reservation, charge it now.
3066 deferred_reserve = map_chg || avoid_reserve;
3067 if (deferred_reserve) {
3068 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3069 idx, pages_per_huge_page(h), &h_cg);
3071 goto out_subpool_put;
3074 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3076 goto out_uncharge_cgroup_reservation;
3078 spin_lock_irq(&hugetlb_lock);
3080 * glb_chg is passed to indicate whether or not a page must be taken
3081 * from the global free pool (global change). gbl_chg == 0 indicates
3082 * a reservation exists for the allocation.
3084 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
3086 spin_unlock_irq(&hugetlb_lock);
3087 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
3089 goto out_uncharge_cgroup;
3090 spin_lock_irq(&hugetlb_lock);
3091 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3092 SetHPageRestoreReserve(page);
3093 h->resv_huge_pages--;
3095 list_add(&page->lru, &h->hugepage_activelist);
3096 set_page_refcounted(page);
3099 folio = page_folio(page);
3100 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
3101 /* If allocation is not consuming a reservation, also store the
3102 * hugetlb_cgroup pointer on the page.
3104 if (deferred_reserve) {
3105 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3109 spin_unlock_irq(&hugetlb_lock);
3111 hugetlb_set_page_subpool(page, spool);
3113 map_commit = vma_commit_reservation(h, vma, addr);
3114 if (unlikely(map_chg > map_commit)) {
3116 * The page was added to the reservation map between
3117 * vma_needs_reservation and vma_commit_reservation.
3118 * This indicates a race with hugetlb_reserve_pages.
3119 * Adjust for the subpool count incremented above AND
3120 * in hugetlb_reserve_pages for the same page. Also,
3121 * the reservation count added in hugetlb_reserve_pages
3122 * no longer applies.
3126 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3127 hugetlb_acct_memory(h, -rsv_adjust);
3128 if (deferred_reserve)
3129 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3130 pages_per_huge_page(h), folio);
3134 out_uncharge_cgroup:
3135 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3136 out_uncharge_cgroup_reservation:
3137 if (deferred_reserve)
3138 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3141 if (map_chg || avoid_reserve)
3142 hugepage_subpool_put_pages(spool, 1);
3143 vma_end_reservation(h, vma, addr);
3144 return ERR_PTR(-ENOSPC);
3147 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3148 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3149 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3151 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3154 /* do node specific alloc */
3155 if (nid != NUMA_NO_NODE) {
3156 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3157 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3162 /* allocate from next node when distributing huge pages */
3163 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3164 m = memblock_alloc_try_nid_raw(
3165 huge_page_size(h), huge_page_size(h),
3166 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3168 * Use the beginning of the huge page to store the
3169 * huge_bootmem_page struct (until gather_bootmem
3170 * puts them into the mem_map).
3178 /* Put them into a private list first because mem_map is not up yet */
3179 INIT_LIST_HEAD(&m->list);
3180 list_add(&m->list, &huge_boot_pages);
3186 * Put bootmem huge pages into the standard lists after mem_map is up.
3187 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3189 static void __init gather_bootmem_prealloc(void)
3191 struct huge_bootmem_page *m;
3193 list_for_each_entry(m, &huge_boot_pages, list) {
3194 struct page *page = virt_to_page(m);
3195 struct folio *folio = page_folio(page);
3196 struct hstate *h = m->hstate;
3198 VM_BUG_ON(!hstate_is_gigantic(h));
3199 WARN_ON(folio_ref_count(folio) != 1);
3200 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3201 WARN_ON(folio_test_reserved(folio));
3202 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3203 free_huge_page(page); /* add to the hugepage allocator */
3205 /* VERY unlikely inflated ref count on a tail page */
3206 free_gigantic_folio(folio, huge_page_order(h));
3210 * We need to restore the 'stolen' pages to totalram_pages
3211 * in order to fix confusing memory reports from free(1) and
3212 * other side-effects, like CommitLimit going negative.
3214 adjust_managed_page_count(page, pages_per_huge_page(h));
3218 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3223 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3224 if (hstate_is_gigantic(h)) {
3225 if (!alloc_bootmem_huge_page(h, nid))
3228 struct folio *folio;
3229 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3231 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3232 &node_states[N_MEMORY], NULL);
3235 free_huge_page(&folio->page); /* free it into the hugepage allocator */
3239 if (i == h->max_huge_pages_node[nid])
3242 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3243 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3244 h->max_huge_pages_node[nid], buf, nid, i);
3245 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3246 h->max_huge_pages_node[nid] = i;
3249 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3252 nodemask_t *node_alloc_noretry;
3253 bool node_specific_alloc = false;
3255 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3256 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3257 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3261 /* do node specific alloc */
3262 for_each_online_node(i) {
3263 if (h->max_huge_pages_node[i] > 0) {
3264 hugetlb_hstate_alloc_pages_onenode(h, i);
3265 node_specific_alloc = true;
3269 if (node_specific_alloc)
3272 /* below will do all node balanced alloc */
3273 if (!hstate_is_gigantic(h)) {
3275 * Bit mask controlling how hard we retry per-node allocations.
3276 * Ignore errors as lower level routines can deal with
3277 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3278 * time, we are likely in bigger trouble.
3280 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3283 /* allocations done at boot time */
3284 node_alloc_noretry = NULL;
3287 /* bit mask controlling how hard we retry per-node allocations */
3288 if (node_alloc_noretry)
3289 nodes_clear(*node_alloc_noretry);
3291 for (i = 0; i < h->max_huge_pages; ++i) {
3292 if (hstate_is_gigantic(h)) {
3293 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3295 } else if (!alloc_pool_huge_page(h,
3296 &node_states[N_MEMORY],
3297 node_alloc_noretry))
3301 if (i < h->max_huge_pages) {
3304 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3305 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3306 h->max_huge_pages, buf, i);
3307 h->max_huge_pages = i;
3309 kfree(node_alloc_noretry);
3312 static void __init hugetlb_init_hstates(void)
3314 struct hstate *h, *h2;
3316 for_each_hstate(h) {
3317 /* oversize hugepages were init'ed in early boot */
3318 if (!hstate_is_gigantic(h))
3319 hugetlb_hstate_alloc_pages(h);
3322 * Set demote order for each hstate. Note that
3323 * h->demote_order is initially 0.
3324 * - We can not demote gigantic pages if runtime freeing
3325 * is not supported, so skip this.
3326 * - If CMA allocation is possible, we can not demote
3327 * HUGETLB_PAGE_ORDER or smaller size pages.
3329 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3331 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3333 for_each_hstate(h2) {
3336 if (h2->order < h->order &&
3337 h2->order > h->demote_order)
3338 h->demote_order = h2->order;
3343 static void __init report_hugepages(void)
3347 for_each_hstate(h) {
3350 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3351 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3352 buf, h->free_huge_pages);
3353 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3354 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3358 #ifdef CONFIG_HIGHMEM
3359 static void try_to_free_low(struct hstate *h, unsigned long count,
3360 nodemask_t *nodes_allowed)
3363 LIST_HEAD(page_list);
3365 lockdep_assert_held(&hugetlb_lock);
3366 if (hstate_is_gigantic(h))
3370 * Collect pages to be freed on a list, and free after dropping lock
3372 for_each_node_mask(i, *nodes_allowed) {
3373 struct page *page, *next;
3374 struct list_head *freel = &h->hugepage_freelists[i];
3375 list_for_each_entry_safe(page, next, freel, lru) {
3376 if (count >= h->nr_huge_pages)
3378 if (PageHighMem(page))
3380 remove_hugetlb_folio(h, page_folio(page), false);
3381 list_add(&page->lru, &page_list);
3386 spin_unlock_irq(&hugetlb_lock);
3387 update_and_free_pages_bulk(h, &page_list);
3388 spin_lock_irq(&hugetlb_lock);
3391 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3392 nodemask_t *nodes_allowed)
3398 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3399 * balanced by operating on them in a round-robin fashion.
3400 * Returns 1 if an adjustment was made.
3402 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3407 lockdep_assert_held(&hugetlb_lock);
3408 VM_BUG_ON(delta != -1 && delta != 1);
3411 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3412 if (h->surplus_huge_pages_node[node])
3416 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3417 if (h->surplus_huge_pages_node[node] <
3418 h->nr_huge_pages_node[node])
3425 h->surplus_huge_pages += delta;
3426 h->surplus_huge_pages_node[node] += delta;
3430 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3431 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3432 nodemask_t *nodes_allowed)
3434 unsigned long min_count, ret;
3436 LIST_HEAD(page_list);
3437 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3440 * Bit mask controlling how hard we retry per-node allocations.
3441 * If we can not allocate the bit mask, do not attempt to allocate
3442 * the requested huge pages.
3444 if (node_alloc_noretry)
3445 nodes_clear(*node_alloc_noretry);
3450 * resize_lock mutex prevents concurrent adjustments to number of
3451 * pages in hstate via the proc/sysfs interfaces.
3453 mutex_lock(&h->resize_lock);
3454 flush_free_hpage_work(h);
3455 spin_lock_irq(&hugetlb_lock);
3458 * Check for a node specific request.
3459 * Changing node specific huge page count may require a corresponding
3460 * change to the global count. In any case, the passed node mask
3461 * (nodes_allowed) will restrict alloc/free to the specified node.
3463 if (nid != NUMA_NO_NODE) {
3464 unsigned long old_count = count;
3466 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3468 * User may have specified a large count value which caused the
3469 * above calculation to overflow. In this case, they wanted
3470 * to allocate as many huge pages as possible. Set count to
3471 * largest possible value to align with their intention.
3473 if (count < old_count)
3478 * Gigantic pages runtime allocation depend on the capability for large
3479 * page range allocation.
3480 * If the system does not provide this feature, return an error when
3481 * the user tries to allocate gigantic pages but let the user free the
3482 * boottime allocated gigantic pages.
3484 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3485 if (count > persistent_huge_pages(h)) {
3486 spin_unlock_irq(&hugetlb_lock);
3487 mutex_unlock(&h->resize_lock);
3488 NODEMASK_FREE(node_alloc_noretry);
3491 /* Fall through to decrease pool */
3495 * Increase the pool size
3496 * First take pages out of surplus state. Then make up the
3497 * remaining difference by allocating fresh huge pages.
3499 * We might race with alloc_surplus_huge_page() here and be unable
3500 * to convert a surplus huge page to a normal huge page. That is
3501 * not critical, though, it just means the overall size of the
3502 * pool might be one hugepage larger than it needs to be, but
3503 * within all the constraints specified by the sysctls.
3505 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3506 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3510 while (count > persistent_huge_pages(h)) {
3512 * If this allocation races such that we no longer need the
3513 * page, free_huge_page will handle it by freeing the page
3514 * and reducing the surplus.
3516 spin_unlock_irq(&hugetlb_lock);
3518 /* yield cpu to avoid soft lockup */
3521 ret = alloc_pool_huge_page(h, nodes_allowed,
3522 node_alloc_noretry);
3523 spin_lock_irq(&hugetlb_lock);
3527 /* Bail for signals. Probably ctrl-c from user */
3528 if (signal_pending(current))
3533 * Decrease the pool size
3534 * First return free pages to the buddy allocator (being careful
3535 * to keep enough around to satisfy reservations). Then place
3536 * pages into surplus state as needed so the pool will shrink
3537 * to the desired size as pages become free.
3539 * By placing pages into the surplus state independent of the
3540 * overcommit value, we are allowing the surplus pool size to
3541 * exceed overcommit. There are few sane options here. Since
3542 * alloc_surplus_huge_page() is checking the global counter,
3543 * though, we'll note that we're not allowed to exceed surplus
3544 * and won't grow the pool anywhere else. Not until one of the
3545 * sysctls are changed, or the surplus pages go out of use.
3547 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3548 min_count = max(count, min_count);
3549 try_to_free_low(h, min_count, nodes_allowed);
3552 * Collect pages to be removed on list without dropping lock
3554 while (min_count < persistent_huge_pages(h)) {
3555 page = remove_pool_huge_page(h, nodes_allowed, 0);
3559 list_add(&page->lru, &page_list);
3561 /* free the pages after dropping lock */
3562 spin_unlock_irq(&hugetlb_lock);
3563 update_and_free_pages_bulk(h, &page_list);
3564 flush_free_hpage_work(h);
3565 spin_lock_irq(&hugetlb_lock);
3567 while (count < persistent_huge_pages(h)) {
3568 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3572 h->max_huge_pages = persistent_huge_pages(h);
3573 spin_unlock_irq(&hugetlb_lock);
3574 mutex_unlock(&h->resize_lock);
3576 NODEMASK_FREE(node_alloc_noretry);
3581 static int demote_free_huge_page(struct hstate *h, struct page *page)
3583 int i, nid = page_to_nid(page);
3584 struct hstate *target_hstate;
3585 struct folio *folio = page_folio(page);
3586 struct page *subpage;
3589 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3591 remove_hugetlb_folio_for_demote(h, folio, false);
3592 spin_unlock_irq(&hugetlb_lock);
3594 rc = hugetlb_vmemmap_restore(h, page);
3596 /* Allocation of vmemmmap failed, we can not demote page */
3597 spin_lock_irq(&hugetlb_lock);
3598 set_page_refcounted(page);
3599 add_hugetlb_folio(h, page_folio(page), false);
3604 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3605 * sizes as it will not ref count pages.
3607 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3610 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3611 * Without the mutex, pages added to target hstate could be marked
3614 * Note that we already hold h->resize_lock. To prevent deadlock,
3615 * use the convention of always taking larger size hstate mutex first.
3617 mutex_lock(&target_hstate->resize_lock);
3618 for (i = 0; i < pages_per_huge_page(h);
3619 i += pages_per_huge_page(target_hstate)) {
3620 subpage = nth_page(page, i);
3621 folio = page_folio(subpage);
3622 if (hstate_is_gigantic(target_hstate))
3623 prep_compound_gigantic_folio_for_demote(folio,
3624 target_hstate->order);
3626 prep_compound_page(subpage, target_hstate->order);
3627 set_page_private(subpage, 0);
3628 prep_new_hugetlb_folio(target_hstate, folio, nid);
3629 free_huge_page(subpage);
3631 mutex_unlock(&target_hstate->resize_lock);
3633 spin_lock_irq(&hugetlb_lock);
3636 * Not absolutely necessary, but for consistency update max_huge_pages
3637 * based on pool changes for the demoted page.
3639 h->max_huge_pages--;
3640 target_hstate->max_huge_pages +=
3641 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3646 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3647 __must_hold(&hugetlb_lock)
3652 lockdep_assert_held(&hugetlb_lock);
3654 /* We should never get here if no demote order */
3655 if (!h->demote_order) {
3656 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3657 return -EINVAL; /* internal error */
3660 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3661 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3662 if (PageHWPoison(page))
3665 return demote_free_huge_page(h, page);
3670 * Only way to get here is if all pages on free lists are poisoned.
3671 * Return -EBUSY so that caller will not retry.
3676 #define HSTATE_ATTR_RO(_name) \
3677 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3679 #define HSTATE_ATTR_WO(_name) \
3680 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3682 #define HSTATE_ATTR(_name) \
3683 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3685 static struct kobject *hugepages_kobj;
3686 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3688 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3690 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3694 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3695 if (hstate_kobjs[i] == kobj) {
3697 *nidp = NUMA_NO_NODE;
3701 return kobj_to_node_hstate(kobj, nidp);
3704 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3705 struct kobj_attribute *attr, char *buf)
3708 unsigned long nr_huge_pages;
3711 h = kobj_to_hstate(kobj, &nid);
3712 if (nid == NUMA_NO_NODE)
3713 nr_huge_pages = h->nr_huge_pages;
3715 nr_huge_pages = h->nr_huge_pages_node[nid];
3717 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3720 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3721 struct hstate *h, int nid,
3722 unsigned long count, size_t len)
3725 nodemask_t nodes_allowed, *n_mask;
3727 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3730 if (nid == NUMA_NO_NODE) {
3732 * global hstate attribute
3734 if (!(obey_mempolicy &&
3735 init_nodemask_of_mempolicy(&nodes_allowed)))
3736 n_mask = &node_states[N_MEMORY];
3738 n_mask = &nodes_allowed;
3741 * Node specific request. count adjustment happens in
3742 * set_max_huge_pages() after acquiring hugetlb_lock.
3744 init_nodemask_of_node(&nodes_allowed, nid);
3745 n_mask = &nodes_allowed;
3748 err = set_max_huge_pages(h, count, nid, n_mask);
3750 return err ? err : len;
3753 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3754 struct kobject *kobj, const char *buf,
3758 unsigned long count;
3762 err = kstrtoul(buf, 10, &count);
3766 h = kobj_to_hstate(kobj, &nid);
3767 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3770 static ssize_t nr_hugepages_show(struct kobject *kobj,
3771 struct kobj_attribute *attr, char *buf)
3773 return nr_hugepages_show_common(kobj, attr, buf);
3776 static ssize_t nr_hugepages_store(struct kobject *kobj,
3777 struct kobj_attribute *attr, const char *buf, size_t len)
3779 return nr_hugepages_store_common(false, kobj, buf, len);
3781 HSTATE_ATTR(nr_hugepages);
3786 * hstate attribute for optionally mempolicy-based constraint on persistent
3787 * huge page alloc/free.
3789 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3790 struct kobj_attribute *attr,
3793 return nr_hugepages_show_common(kobj, attr, buf);
3796 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3797 struct kobj_attribute *attr, const char *buf, size_t len)
3799 return nr_hugepages_store_common(true, kobj, buf, len);
3801 HSTATE_ATTR(nr_hugepages_mempolicy);
3805 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3806 struct kobj_attribute *attr, char *buf)
3808 struct hstate *h = kobj_to_hstate(kobj, NULL);
3809 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3812 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3813 struct kobj_attribute *attr, const char *buf, size_t count)
3816 unsigned long input;
3817 struct hstate *h = kobj_to_hstate(kobj, NULL);
3819 if (hstate_is_gigantic(h))
3822 err = kstrtoul(buf, 10, &input);
3826 spin_lock_irq(&hugetlb_lock);
3827 h->nr_overcommit_huge_pages = input;
3828 spin_unlock_irq(&hugetlb_lock);
3832 HSTATE_ATTR(nr_overcommit_hugepages);
3834 static ssize_t free_hugepages_show(struct kobject *kobj,
3835 struct kobj_attribute *attr, char *buf)
3838 unsigned long free_huge_pages;
3841 h = kobj_to_hstate(kobj, &nid);
3842 if (nid == NUMA_NO_NODE)
3843 free_huge_pages = h->free_huge_pages;
3845 free_huge_pages = h->free_huge_pages_node[nid];
3847 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3849 HSTATE_ATTR_RO(free_hugepages);
3851 static ssize_t resv_hugepages_show(struct kobject *kobj,
3852 struct kobj_attribute *attr, char *buf)
3854 struct hstate *h = kobj_to_hstate(kobj, NULL);
3855 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3857 HSTATE_ATTR_RO(resv_hugepages);
3859 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3860 struct kobj_attribute *attr, char *buf)
3863 unsigned long surplus_huge_pages;
3866 h = kobj_to_hstate(kobj, &nid);
3867 if (nid == NUMA_NO_NODE)
3868 surplus_huge_pages = h->surplus_huge_pages;
3870 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3872 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3874 HSTATE_ATTR_RO(surplus_hugepages);
3876 static ssize_t demote_store(struct kobject *kobj,
3877 struct kobj_attribute *attr, const char *buf, size_t len)
3879 unsigned long nr_demote;
3880 unsigned long nr_available;
3881 nodemask_t nodes_allowed, *n_mask;
3886 err = kstrtoul(buf, 10, &nr_demote);
3889 h = kobj_to_hstate(kobj, &nid);
3891 if (nid != NUMA_NO_NODE) {
3892 init_nodemask_of_node(&nodes_allowed, nid);
3893 n_mask = &nodes_allowed;
3895 n_mask = &node_states[N_MEMORY];
3898 /* Synchronize with other sysfs operations modifying huge pages */
3899 mutex_lock(&h->resize_lock);
3900 spin_lock_irq(&hugetlb_lock);
3904 * Check for available pages to demote each time thorough the
3905 * loop as demote_pool_huge_page will drop hugetlb_lock.
3907 if (nid != NUMA_NO_NODE)
3908 nr_available = h->free_huge_pages_node[nid];
3910 nr_available = h->free_huge_pages;
3911 nr_available -= h->resv_huge_pages;
3915 err = demote_pool_huge_page(h, n_mask);
3922 spin_unlock_irq(&hugetlb_lock);
3923 mutex_unlock(&h->resize_lock);
3929 HSTATE_ATTR_WO(demote);
3931 static ssize_t demote_size_show(struct kobject *kobj,
3932 struct kobj_attribute *attr, char *buf)
3934 struct hstate *h = kobj_to_hstate(kobj, NULL);
3935 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3937 return sysfs_emit(buf, "%lukB\n", demote_size);
3940 static ssize_t demote_size_store(struct kobject *kobj,
3941 struct kobj_attribute *attr,
3942 const char *buf, size_t count)
3944 struct hstate *h, *demote_hstate;
3945 unsigned long demote_size;
3946 unsigned int demote_order;
3948 demote_size = (unsigned long)memparse(buf, NULL);
3950 demote_hstate = size_to_hstate(demote_size);
3953 demote_order = demote_hstate->order;
3954 if (demote_order < HUGETLB_PAGE_ORDER)
3957 /* demote order must be smaller than hstate order */
3958 h = kobj_to_hstate(kobj, NULL);
3959 if (demote_order >= h->order)
3962 /* resize_lock synchronizes access to demote size and writes */
3963 mutex_lock(&h->resize_lock);
3964 h->demote_order = demote_order;
3965 mutex_unlock(&h->resize_lock);
3969 HSTATE_ATTR(demote_size);
3971 static struct attribute *hstate_attrs[] = {
3972 &nr_hugepages_attr.attr,
3973 &nr_overcommit_hugepages_attr.attr,
3974 &free_hugepages_attr.attr,
3975 &resv_hugepages_attr.attr,
3976 &surplus_hugepages_attr.attr,
3978 &nr_hugepages_mempolicy_attr.attr,
3983 static const struct attribute_group hstate_attr_group = {
3984 .attrs = hstate_attrs,
3987 static struct attribute *hstate_demote_attrs[] = {
3988 &demote_size_attr.attr,
3993 static const struct attribute_group hstate_demote_attr_group = {
3994 .attrs = hstate_demote_attrs,
3997 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3998 struct kobject **hstate_kobjs,
3999 const struct attribute_group *hstate_attr_group)
4002 int hi = hstate_index(h);
4004 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4005 if (!hstate_kobjs[hi])
4008 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4010 kobject_put(hstate_kobjs[hi]);
4011 hstate_kobjs[hi] = NULL;
4015 if (h->demote_order) {
4016 retval = sysfs_create_group(hstate_kobjs[hi],
4017 &hstate_demote_attr_group);
4019 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4020 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4021 kobject_put(hstate_kobjs[hi]);
4022 hstate_kobjs[hi] = NULL;
4031 static bool hugetlb_sysfs_initialized __ro_after_init;
4034 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4035 * with node devices in node_devices[] using a parallel array. The array
4036 * index of a node device or _hstate == node id.
4037 * This is here to avoid any static dependency of the node device driver, in
4038 * the base kernel, on the hugetlb module.
4040 struct node_hstate {
4041 struct kobject *hugepages_kobj;
4042 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4044 static struct node_hstate node_hstates[MAX_NUMNODES];
4047 * A subset of global hstate attributes for node devices
4049 static struct attribute *per_node_hstate_attrs[] = {
4050 &nr_hugepages_attr.attr,
4051 &free_hugepages_attr.attr,
4052 &surplus_hugepages_attr.attr,
4056 static const struct attribute_group per_node_hstate_attr_group = {
4057 .attrs = per_node_hstate_attrs,
4061 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4062 * Returns node id via non-NULL nidp.
4064 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4068 for (nid = 0; nid < nr_node_ids; nid++) {
4069 struct node_hstate *nhs = &node_hstates[nid];
4071 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4072 if (nhs->hstate_kobjs[i] == kobj) {
4084 * Unregister hstate attributes from a single node device.
4085 * No-op if no hstate attributes attached.
4087 void hugetlb_unregister_node(struct node *node)
4090 struct node_hstate *nhs = &node_hstates[node->dev.id];
4092 if (!nhs->hugepages_kobj)
4093 return; /* no hstate attributes */
4095 for_each_hstate(h) {
4096 int idx = hstate_index(h);
4097 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4101 if (h->demote_order)
4102 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4103 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4104 kobject_put(hstate_kobj);
4105 nhs->hstate_kobjs[idx] = NULL;
4108 kobject_put(nhs->hugepages_kobj);
4109 nhs->hugepages_kobj = NULL;
4114 * Register hstate attributes for a single node device.
4115 * No-op if attributes already registered.
4117 void hugetlb_register_node(struct node *node)
4120 struct node_hstate *nhs = &node_hstates[node->dev.id];
4123 if (!hugetlb_sysfs_initialized)
4126 if (nhs->hugepages_kobj)
4127 return; /* already allocated */
4129 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4131 if (!nhs->hugepages_kobj)
4134 for_each_hstate(h) {
4135 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4137 &per_node_hstate_attr_group);
4139 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4140 h->name, node->dev.id);
4141 hugetlb_unregister_node(node);
4148 * hugetlb init time: register hstate attributes for all registered node
4149 * devices of nodes that have memory. All on-line nodes should have
4150 * registered their associated device by this time.
4152 static void __init hugetlb_register_all_nodes(void)
4156 for_each_online_node(nid)
4157 hugetlb_register_node(node_devices[nid]);
4159 #else /* !CONFIG_NUMA */
4161 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4169 static void hugetlb_register_all_nodes(void) { }
4174 static void __init hugetlb_cma_check(void);
4176 static inline __init void hugetlb_cma_check(void)
4181 static void __init hugetlb_sysfs_init(void)
4186 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4187 if (!hugepages_kobj)
4190 for_each_hstate(h) {
4191 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4192 hstate_kobjs, &hstate_attr_group);
4194 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4198 hugetlb_sysfs_initialized = true;
4200 hugetlb_register_all_nodes();
4203 static int __init hugetlb_init(void)
4207 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4210 if (!hugepages_supported()) {
4211 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4212 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4217 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4218 * architectures depend on setup being done here.
4220 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4221 if (!parsed_default_hugepagesz) {
4223 * If we did not parse a default huge page size, set
4224 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4225 * number of huge pages for this default size was implicitly
4226 * specified, set that here as well.
4227 * Note that the implicit setting will overwrite an explicit
4228 * setting. A warning will be printed in this case.
4230 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4231 if (default_hstate_max_huge_pages) {
4232 if (default_hstate.max_huge_pages) {
4235 string_get_size(huge_page_size(&default_hstate),
4236 1, STRING_UNITS_2, buf, 32);
4237 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4238 default_hstate.max_huge_pages, buf);
4239 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4240 default_hstate_max_huge_pages);
4242 default_hstate.max_huge_pages =
4243 default_hstate_max_huge_pages;
4245 for_each_online_node(i)
4246 default_hstate.max_huge_pages_node[i] =
4247 default_hugepages_in_node[i];
4251 hugetlb_cma_check();
4252 hugetlb_init_hstates();
4253 gather_bootmem_prealloc();
4256 hugetlb_sysfs_init();
4257 hugetlb_cgroup_file_init();
4260 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4262 num_fault_mutexes = 1;
4264 hugetlb_fault_mutex_table =
4265 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4267 BUG_ON(!hugetlb_fault_mutex_table);
4269 for (i = 0; i < num_fault_mutexes; i++)
4270 mutex_init(&hugetlb_fault_mutex_table[i]);
4273 subsys_initcall(hugetlb_init);
4275 /* Overwritten by architectures with more huge page sizes */
4276 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4278 return size == HPAGE_SIZE;
4281 void __init hugetlb_add_hstate(unsigned int order)
4286 if (size_to_hstate(PAGE_SIZE << order)) {
4289 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4291 h = &hstates[hugetlb_max_hstate++];
4292 mutex_init(&h->resize_lock);
4294 h->mask = ~(huge_page_size(h) - 1);
4295 for (i = 0; i < MAX_NUMNODES; ++i)
4296 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4297 INIT_LIST_HEAD(&h->hugepage_activelist);
4298 h->next_nid_to_alloc = first_memory_node;
4299 h->next_nid_to_free = first_memory_node;
4300 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4301 huge_page_size(h)/SZ_1K);
4306 bool __init __weak hugetlb_node_alloc_supported(void)
4311 static void __init hugepages_clear_pages_in_node(void)
4313 if (!hugetlb_max_hstate) {
4314 default_hstate_max_huge_pages = 0;
4315 memset(default_hugepages_in_node, 0,
4316 sizeof(default_hugepages_in_node));
4318 parsed_hstate->max_huge_pages = 0;
4319 memset(parsed_hstate->max_huge_pages_node, 0,
4320 sizeof(parsed_hstate->max_huge_pages_node));
4325 * hugepages command line processing
4326 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4327 * specification. If not, ignore the hugepages value. hugepages can also
4328 * be the first huge page command line option in which case it implicitly
4329 * specifies the number of huge pages for the default size.
4331 static int __init hugepages_setup(char *s)
4334 static unsigned long *last_mhp;
4335 int node = NUMA_NO_NODE;
4340 if (!parsed_valid_hugepagesz) {
4341 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4342 parsed_valid_hugepagesz = true;
4347 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4348 * yet, so this hugepages= parameter goes to the "default hstate".
4349 * Otherwise, it goes with the previously parsed hugepagesz or
4350 * default_hugepagesz.
4352 else if (!hugetlb_max_hstate)
4353 mhp = &default_hstate_max_huge_pages;
4355 mhp = &parsed_hstate->max_huge_pages;
4357 if (mhp == last_mhp) {
4358 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4364 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4366 /* Parameter is node format */
4367 if (p[count] == ':') {
4368 if (!hugetlb_node_alloc_supported()) {
4369 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4372 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4374 node = array_index_nospec(tmp, MAX_NUMNODES);
4376 /* Parse hugepages */
4377 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4379 if (!hugetlb_max_hstate)
4380 default_hugepages_in_node[node] = tmp;
4382 parsed_hstate->max_huge_pages_node[node] = tmp;
4384 /* Go to parse next node*/
4385 if (p[count] == ',')
4398 * Global state is always initialized later in hugetlb_init.
4399 * But we need to allocate gigantic hstates here early to still
4400 * use the bootmem allocator.
4402 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4403 hugetlb_hstate_alloc_pages(parsed_hstate);
4410 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4411 hugepages_clear_pages_in_node();
4414 __setup("hugepages=", hugepages_setup);
4417 * hugepagesz command line processing
4418 * A specific huge page size can only be specified once with hugepagesz.
4419 * hugepagesz is followed by hugepages on the command line. The global
4420 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4421 * hugepagesz argument was valid.
4423 static int __init hugepagesz_setup(char *s)
4428 parsed_valid_hugepagesz = false;
4429 size = (unsigned long)memparse(s, NULL);
4431 if (!arch_hugetlb_valid_size(size)) {
4432 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4436 h = size_to_hstate(size);
4439 * hstate for this size already exists. This is normally
4440 * an error, but is allowed if the existing hstate is the
4441 * default hstate. More specifically, it is only allowed if
4442 * the number of huge pages for the default hstate was not
4443 * previously specified.
4445 if (!parsed_default_hugepagesz || h != &default_hstate ||
4446 default_hstate.max_huge_pages) {
4447 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4452 * No need to call hugetlb_add_hstate() as hstate already
4453 * exists. But, do set parsed_hstate so that a following
4454 * hugepages= parameter will be applied to this hstate.
4457 parsed_valid_hugepagesz = true;
4461 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4462 parsed_valid_hugepagesz = true;
4465 __setup("hugepagesz=", hugepagesz_setup);
4468 * default_hugepagesz command line input
4469 * Only one instance of default_hugepagesz allowed on command line.
4471 static int __init default_hugepagesz_setup(char *s)
4476 parsed_valid_hugepagesz = false;
4477 if (parsed_default_hugepagesz) {
4478 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4482 size = (unsigned long)memparse(s, NULL);
4484 if (!arch_hugetlb_valid_size(size)) {
4485 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4489 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4490 parsed_valid_hugepagesz = true;
4491 parsed_default_hugepagesz = true;
4492 default_hstate_idx = hstate_index(size_to_hstate(size));
4495 * The number of default huge pages (for this size) could have been
4496 * specified as the first hugetlb parameter: hugepages=X. If so,
4497 * then default_hstate_max_huge_pages is set. If the default huge
4498 * page size is gigantic (>= MAX_ORDER), then the pages must be
4499 * allocated here from bootmem allocator.
4501 if (default_hstate_max_huge_pages) {
4502 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4503 for_each_online_node(i)
4504 default_hstate.max_huge_pages_node[i] =
4505 default_hugepages_in_node[i];
4506 if (hstate_is_gigantic(&default_hstate))
4507 hugetlb_hstate_alloc_pages(&default_hstate);
4508 default_hstate_max_huge_pages = 0;
4513 __setup("default_hugepagesz=", default_hugepagesz_setup);
4515 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4518 struct mempolicy *mpol = get_task_policy(current);
4521 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4522 * (from policy_nodemask) specifically for hugetlb case
4524 if (mpol->mode == MPOL_BIND &&
4525 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4526 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4527 return &mpol->nodes;
4532 static unsigned int allowed_mems_nr(struct hstate *h)
4535 unsigned int nr = 0;
4536 nodemask_t *mbind_nodemask;
4537 unsigned int *array = h->free_huge_pages_node;
4538 gfp_t gfp_mask = htlb_alloc_mask(h);
4540 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4541 for_each_node_mask(node, cpuset_current_mems_allowed) {
4542 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4549 #ifdef CONFIG_SYSCTL
4550 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4551 void *buffer, size_t *length,
4552 loff_t *ppos, unsigned long *out)
4554 struct ctl_table dup_table;
4557 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4558 * can duplicate the @table and alter the duplicate of it.
4561 dup_table.data = out;
4563 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4566 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4567 struct ctl_table *table, int write,
4568 void *buffer, size_t *length, loff_t *ppos)
4570 struct hstate *h = &default_hstate;
4571 unsigned long tmp = h->max_huge_pages;
4574 if (!hugepages_supported())
4577 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4583 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4584 NUMA_NO_NODE, tmp, *length);
4589 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4590 void *buffer, size_t *length, loff_t *ppos)
4593 return hugetlb_sysctl_handler_common(false, table, write,
4594 buffer, length, ppos);
4598 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4599 void *buffer, size_t *length, loff_t *ppos)
4601 return hugetlb_sysctl_handler_common(true, table, write,
4602 buffer, length, ppos);
4604 #endif /* CONFIG_NUMA */
4606 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4607 void *buffer, size_t *length, loff_t *ppos)
4609 struct hstate *h = &default_hstate;
4613 if (!hugepages_supported())
4616 tmp = h->nr_overcommit_huge_pages;
4618 if (write && hstate_is_gigantic(h))
4621 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4627 spin_lock_irq(&hugetlb_lock);
4628 h->nr_overcommit_huge_pages = tmp;
4629 spin_unlock_irq(&hugetlb_lock);
4635 #endif /* CONFIG_SYSCTL */
4637 void hugetlb_report_meminfo(struct seq_file *m)
4640 unsigned long total = 0;
4642 if (!hugepages_supported())
4645 for_each_hstate(h) {
4646 unsigned long count = h->nr_huge_pages;
4648 total += huge_page_size(h) * count;
4650 if (h == &default_hstate)
4652 "HugePages_Total: %5lu\n"
4653 "HugePages_Free: %5lu\n"
4654 "HugePages_Rsvd: %5lu\n"
4655 "HugePages_Surp: %5lu\n"
4656 "Hugepagesize: %8lu kB\n",
4660 h->surplus_huge_pages,
4661 huge_page_size(h) / SZ_1K);
4664 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4667 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4669 struct hstate *h = &default_hstate;
4671 if (!hugepages_supported())
4674 return sysfs_emit_at(buf, len,
4675 "Node %d HugePages_Total: %5u\n"
4676 "Node %d HugePages_Free: %5u\n"
4677 "Node %d HugePages_Surp: %5u\n",
4678 nid, h->nr_huge_pages_node[nid],
4679 nid, h->free_huge_pages_node[nid],
4680 nid, h->surplus_huge_pages_node[nid]);
4683 void hugetlb_show_meminfo_node(int nid)
4687 if (!hugepages_supported())
4691 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4693 h->nr_huge_pages_node[nid],
4694 h->free_huge_pages_node[nid],
4695 h->surplus_huge_pages_node[nid],
4696 huge_page_size(h) / SZ_1K);
4699 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4701 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4702 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4705 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4706 unsigned long hugetlb_total_pages(void)
4709 unsigned long nr_total_pages = 0;
4712 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4713 return nr_total_pages;
4716 static int hugetlb_acct_memory(struct hstate *h, long delta)
4723 spin_lock_irq(&hugetlb_lock);
4725 * When cpuset is configured, it breaks the strict hugetlb page
4726 * reservation as the accounting is done on a global variable. Such
4727 * reservation is completely rubbish in the presence of cpuset because
4728 * the reservation is not checked against page availability for the
4729 * current cpuset. Application can still potentially OOM'ed by kernel
4730 * with lack of free htlb page in cpuset that the task is in.
4731 * Attempt to enforce strict accounting with cpuset is almost
4732 * impossible (or too ugly) because cpuset is too fluid that
4733 * task or memory node can be dynamically moved between cpusets.
4735 * The change of semantics for shared hugetlb mapping with cpuset is
4736 * undesirable. However, in order to preserve some of the semantics,
4737 * we fall back to check against current free page availability as
4738 * a best attempt and hopefully to minimize the impact of changing
4739 * semantics that cpuset has.
4741 * Apart from cpuset, we also have memory policy mechanism that
4742 * also determines from which node the kernel will allocate memory
4743 * in a NUMA system. So similar to cpuset, we also should consider
4744 * the memory policy of the current task. Similar to the description
4748 if (gather_surplus_pages(h, delta) < 0)
4751 if (delta > allowed_mems_nr(h)) {
4752 return_unused_surplus_pages(h, delta);
4759 return_unused_surplus_pages(h, (unsigned long) -delta);
4762 spin_unlock_irq(&hugetlb_lock);
4766 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4768 struct resv_map *resv = vma_resv_map(vma);
4771 * HPAGE_RESV_OWNER indicates a private mapping.
4772 * This new VMA should share its siblings reservation map if present.
4773 * The VMA will only ever have a valid reservation map pointer where
4774 * it is being copied for another still existing VMA. As that VMA
4775 * has a reference to the reservation map it cannot disappear until
4776 * after this open call completes. It is therefore safe to take a
4777 * new reference here without additional locking.
4779 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4780 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4781 kref_get(&resv->refs);
4785 * vma_lock structure for sharable mappings is vma specific.
4786 * Clear old pointer (if copied via vm_area_dup) and allocate
4787 * new structure. Before clearing, make sure vma_lock is not
4790 if (vma->vm_flags & VM_MAYSHARE) {
4791 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4794 if (vma_lock->vma != vma) {
4795 vma->vm_private_data = NULL;
4796 hugetlb_vma_lock_alloc(vma);
4798 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4800 hugetlb_vma_lock_alloc(vma);
4804 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4806 struct hstate *h = hstate_vma(vma);
4807 struct resv_map *resv;
4808 struct hugepage_subpool *spool = subpool_vma(vma);
4809 unsigned long reserve, start, end;
4812 hugetlb_vma_lock_free(vma);
4814 resv = vma_resv_map(vma);
4815 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4818 start = vma_hugecache_offset(h, vma, vma->vm_start);
4819 end = vma_hugecache_offset(h, vma, vma->vm_end);
4821 reserve = (end - start) - region_count(resv, start, end);
4822 hugetlb_cgroup_uncharge_counter(resv, start, end);
4825 * Decrement reserve counts. The global reserve count may be
4826 * adjusted if the subpool has a minimum size.
4828 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4829 hugetlb_acct_memory(h, -gbl_reserve);
4832 kref_put(&resv->refs, resv_map_release);
4835 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4837 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4841 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4842 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4843 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4845 if (addr & ~PUD_MASK) {
4847 * hugetlb_vm_op_split is called right before we attempt to
4848 * split the VMA. We will need to unshare PMDs in the old and
4849 * new VMAs, so let's unshare before we split.
4851 unsigned long floor = addr & PUD_MASK;
4852 unsigned long ceil = floor + PUD_SIZE;
4854 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4855 hugetlb_unshare_pmds(vma, floor, ceil);
4861 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4863 return huge_page_size(hstate_vma(vma));
4867 * We cannot handle pagefaults against hugetlb pages at all. They cause
4868 * handle_mm_fault() to try to instantiate regular-sized pages in the
4869 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4872 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4879 * When a new function is introduced to vm_operations_struct and added
4880 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4881 * This is because under System V memory model, mappings created via
4882 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4883 * their original vm_ops are overwritten with shm_vm_ops.
4885 const struct vm_operations_struct hugetlb_vm_ops = {
4886 .fault = hugetlb_vm_op_fault,
4887 .open = hugetlb_vm_op_open,
4888 .close = hugetlb_vm_op_close,
4889 .may_split = hugetlb_vm_op_split,
4890 .pagesize = hugetlb_vm_op_pagesize,
4893 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4897 unsigned int shift = huge_page_shift(hstate_vma(vma));
4900 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4901 vma->vm_page_prot)));
4903 entry = huge_pte_wrprotect(mk_huge_pte(page,
4904 vma->vm_page_prot));
4906 entry = pte_mkyoung(entry);
4907 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4912 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4913 unsigned long address, pte_t *ptep)
4917 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4918 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4919 update_mmu_cache(vma, address, ptep);
4922 bool is_hugetlb_entry_migration(pte_t pte)
4926 if (huge_pte_none(pte) || pte_present(pte))
4928 swp = pte_to_swp_entry(pte);
4929 if (is_migration_entry(swp))
4935 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4939 if (huge_pte_none(pte) || pte_present(pte))
4941 swp = pte_to_swp_entry(pte);
4942 if (is_hwpoison_entry(swp))
4949 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4950 struct page *new_page)
4952 __SetPageUptodate(new_page);
4953 hugepage_add_new_anon_rmap(new_page, vma, addr);
4954 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4955 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4956 SetHPageMigratable(new_page);
4959 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4960 struct vm_area_struct *dst_vma,
4961 struct vm_area_struct *src_vma)
4963 pte_t *src_pte, *dst_pte, entry;
4964 struct page *ptepage;
4966 bool cow = is_cow_mapping(src_vma->vm_flags);
4967 struct hstate *h = hstate_vma(src_vma);
4968 unsigned long sz = huge_page_size(h);
4969 unsigned long npages = pages_per_huge_page(h);
4970 struct mmu_notifier_range range;
4971 unsigned long last_addr_mask;
4975 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4978 mmu_notifier_invalidate_range_start(&range);
4979 mmap_assert_write_locked(src);
4980 raw_write_seqcount_begin(&src->write_protect_seq);
4983 * For shared mappings the vma lock must be held before
4984 * calling huge_pte_offset in the src vma. Otherwise, the
4985 * returned ptep could go away if part of a shared pmd and
4986 * another thread calls huge_pmd_unshare.
4988 hugetlb_vma_lock_read(src_vma);
4991 last_addr_mask = hugetlb_mask_last_page(h);
4992 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4993 spinlock_t *src_ptl, *dst_ptl;
4994 src_pte = huge_pte_offset(src, addr, sz);
4996 addr |= last_addr_mask;
4999 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5006 * If the pagetables are shared don't copy or take references.
5008 * dst_pte == src_pte is the common case of src/dest sharing.
5009 * However, src could have 'unshared' and dst shares with
5010 * another vma. So page_count of ptep page is checked instead
5011 * to reliably determine whether pte is shared.
5013 if (page_count(virt_to_page(dst_pte)) > 1) {
5014 addr |= last_addr_mask;
5018 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5019 src_ptl = huge_pte_lockptr(h, src, src_pte);
5020 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5021 entry = huge_ptep_get(src_pte);
5023 if (huge_pte_none(entry)) {
5025 * Skip if src entry none.
5028 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5029 bool uffd_wp = huge_pte_uffd_wp(entry);
5031 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5032 entry = huge_pte_clear_uffd_wp(entry);
5033 set_huge_pte_at(dst, addr, dst_pte, entry);
5034 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5035 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5036 bool uffd_wp = huge_pte_uffd_wp(entry);
5038 if (!is_readable_migration_entry(swp_entry) && cow) {
5040 * COW mappings require pages in both
5041 * parent and child to be set to read.
5043 swp_entry = make_readable_migration_entry(
5044 swp_offset(swp_entry));
5045 entry = swp_entry_to_pte(swp_entry);
5046 if (userfaultfd_wp(src_vma) && uffd_wp)
5047 entry = huge_pte_mkuffd_wp(entry);
5048 set_huge_pte_at(src, addr, src_pte, entry);
5050 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5051 entry = huge_pte_clear_uffd_wp(entry);
5052 set_huge_pte_at(dst, addr, dst_pte, entry);
5053 } else if (unlikely(is_pte_marker(entry))) {
5055 * We copy the pte marker only if the dst vma has
5058 if (userfaultfd_wp(dst_vma))
5059 set_huge_pte_at(dst, addr, dst_pte, entry);
5061 entry = huge_ptep_get(src_pte);
5062 ptepage = pte_page(entry);
5066 * Failing to duplicate the anon rmap is a rare case
5067 * where we see pinned hugetlb pages while they're
5068 * prone to COW. We need to do the COW earlier during
5071 * When pre-allocating the page or copying data, we
5072 * need to be without the pgtable locks since we could
5073 * sleep during the process.
5075 if (!PageAnon(ptepage)) {
5076 page_dup_file_rmap(ptepage, true);
5077 } else if (page_try_dup_anon_rmap(ptepage, true,
5079 pte_t src_pte_old = entry;
5082 spin_unlock(src_ptl);
5083 spin_unlock(dst_ptl);
5084 /* Do not use reserve as it's private owned */
5085 new = alloc_huge_page(dst_vma, addr, 1);
5091 copy_user_huge_page(new, ptepage, addr, dst_vma,
5095 /* Install the new huge page if src pte stable */
5096 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5097 src_ptl = huge_pte_lockptr(h, src, src_pte);
5098 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5099 entry = huge_ptep_get(src_pte);
5100 if (!pte_same(src_pte_old, entry)) {
5101 restore_reserve_on_error(h, dst_vma, addr,
5104 /* huge_ptep of dst_pte won't change as in child */
5107 hugetlb_install_page(dst_vma, dst_pte, addr, new);
5108 spin_unlock(src_ptl);
5109 spin_unlock(dst_ptl);
5115 * No need to notify as we are downgrading page
5116 * table protection not changing it to point
5119 * See Documentation/mm/mmu_notifier.rst
5121 huge_ptep_set_wrprotect(src, addr, src_pte);
5122 entry = huge_pte_wrprotect(entry);
5125 set_huge_pte_at(dst, addr, dst_pte, entry);
5126 hugetlb_count_add(npages, dst);
5128 spin_unlock(src_ptl);
5129 spin_unlock(dst_ptl);
5133 raw_write_seqcount_end(&src->write_protect_seq);
5134 mmu_notifier_invalidate_range_end(&range);
5136 hugetlb_vma_unlock_read(src_vma);
5142 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5143 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5145 struct hstate *h = hstate_vma(vma);
5146 struct mm_struct *mm = vma->vm_mm;
5147 spinlock_t *src_ptl, *dst_ptl;
5150 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5151 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5154 * We don't have to worry about the ordering of src and dst ptlocks
5155 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
5157 if (src_ptl != dst_ptl)
5158 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5160 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5161 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5163 if (src_ptl != dst_ptl)
5164 spin_unlock(src_ptl);
5165 spin_unlock(dst_ptl);
5168 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5169 struct vm_area_struct *new_vma,
5170 unsigned long old_addr, unsigned long new_addr,
5173 struct hstate *h = hstate_vma(vma);
5174 struct address_space *mapping = vma->vm_file->f_mapping;
5175 unsigned long sz = huge_page_size(h);
5176 struct mm_struct *mm = vma->vm_mm;
5177 unsigned long old_end = old_addr + len;
5178 unsigned long last_addr_mask;
5179 pte_t *src_pte, *dst_pte;
5180 struct mmu_notifier_range range;
5181 bool shared_pmd = false;
5183 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
5185 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5187 * In case of shared PMDs, we should cover the maximum possible
5190 flush_cache_range(vma, range.start, range.end);
5192 mmu_notifier_invalidate_range_start(&range);
5193 last_addr_mask = hugetlb_mask_last_page(h);
5194 /* Prevent race with file truncation */
5195 hugetlb_vma_lock_write(vma);
5196 i_mmap_lock_write(mapping);
5197 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5198 src_pte = huge_pte_offset(mm, old_addr, sz);
5200 old_addr |= last_addr_mask;
5201 new_addr |= last_addr_mask;
5204 if (huge_pte_none(huge_ptep_get(src_pte)))
5207 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5209 old_addr |= last_addr_mask;
5210 new_addr |= last_addr_mask;
5214 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5218 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5222 flush_tlb_range(vma, range.start, range.end);
5224 flush_tlb_range(vma, old_end - len, old_end);
5225 mmu_notifier_invalidate_range_end(&range);
5226 i_mmap_unlock_write(mapping);
5227 hugetlb_vma_unlock_write(vma);
5229 return len + old_addr - old_end;
5232 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5233 unsigned long start, unsigned long end,
5234 struct page *ref_page, zap_flags_t zap_flags)
5236 struct mm_struct *mm = vma->vm_mm;
5237 unsigned long address;
5242 struct hstate *h = hstate_vma(vma);
5243 unsigned long sz = huge_page_size(h);
5244 unsigned long last_addr_mask;
5245 bool force_flush = false;
5247 WARN_ON(!is_vm_hugetlb_page(vma));
5248 BUG_ON(start & ~huge_page_mask(h));
5249 BUG_ON(end & ~huge_page_mask(h));
5252 * This is a hugetlb vma, all the pte entries should point
5255 tlb_change_page_size(tlb, sz);
5256 tlb_start_vma(tlb, vma);
5258 last_addr_mask = hugetlb_mask_last_page(h);
5260 for (; address < end; address += sz) {
5261 ptep = huge_pte_offset(mm, address, sz);
5263 address |= last_addr_mask;
5267 ptl = huge_pte_lock(h, mm, ptep);
5268 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5270 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5272 address |= last_addr_mask;
5276 pte = huge_ptep_get(ptep);
5277 if (huge_pte_none(pte)) {
5283 * Migrating hugepage or HWPoisoned hugepage is already
5284 * unmapped and its refcount is dropped, so just clear pte here.
5286 if (unlikely(!pte_present(pte))) {
5288 * If the pte was wr-protected by uffd-wp in any of the
5289 * swap forms, meanwhile the caller does not want to
5290 * drop the uffd-wp bit in this zap, then replace the
5291 * pte with a marker.
5293 if (pte_swp_uffd_wp_any(pte) &&
5294 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5295 set_huge_pte_at(mm, address, ptep,
5296 make_pte_marker(PTE_MARKER_UFFD_WP));
5298 huge_pte_clear(mm, address, ptep, sz);
5303 page = pte_page(pte);
5305 * If a reference page is supplied, it is because a specific
5306 * page is being unmapped, not a range. Ensure the page we
5307 * are about to unmap is the actual page of interest.
5310 if (page != ref_page) {
5315 * Mark the VMA as having unmapped its page so that
5316 * future faults in this VMA will fail rather than
5317 * looking like data was lost
5319 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5322 pte = huge_ptep_get_and_clear(mm, address, ptep);
5323 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5324 if (huge_pte_dirty(pte))
5325 set_page_dirty(page);
5326 /* Leave a uffd-wp pte marker if needed */
5327 if (huge_pte_uffd_wp(pte) &&
5328 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5329 set_huge_pte_at(mm, address, ptep,
5330 make_pte_marker(PTE_MARKER_UFFD_WP));
5331 hugetlb_count_sub(pages_per_huge_page(h), mm);
5332 page_remove_rmap(page, vma, true);
5335 tlb_remove_page_size(tlb, page, huge_page_size(h));
5337 * Bail out after unmapping reference page if supplied
5342 tlb_end_vma(tlb, vma);
5345 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5346 * could defer the flush until now, since by holding i_mmap_rwsem we
5347 * guaranteed that the last refernece would not be dropped. But we must
5348 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5349 * dropped and the last reference to the shared PMDs page might be
5352 * In theory we could defer the freeing of the PMD pages as well, but
5353 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5354 * detect sharing, so we cannot defer the release of the page either.
5355 * Instead, do flush now.
5358 tlb_flush_mmu_tlbonly(tlb);
5361 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5362 struct vm_area_struct *vma, unsigned long start,
5363 unsigned long end, struct page *ref_page,
5364 zap_flags_t zap_flags)
5366 hugetlb_vma_lock_write(vma);
5367 i_mmap_lock_write(vma->vm_file->f_mapping);
5369 /* mmu notification performed in caller */
5370 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5372 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5374 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5375 * When the vma_lock is freed, this makes the vma ineligible
5376 * for pmd sharing. And, i_mmap_rwsem is required to set up
5377 * pmd sharing. This is important as page tables for this
5378 * unmapped range will be asynchrously deleted. If the page
5379 * tables are shared, there will be issues when accessed by
5382 __hugetlb_vma_unlock_write_free(vma);
5383 i_mmap_unlock_write(vma->vm_file->f_mapping);
5385 i_mmap_unlock_write(vma->vm_file->f_mapping);
5386 hugetlb_vma_unlock_write(vma);
5390 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5391 unsigned long end, struct page *ref_page,
5392 zap_flags_t zap_flags)
5394 struct mmu_notifier_range range;
5395 struct mmu_gather tlb;
5397 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
5399 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5400 mmu_notifier_invalidate_range_start(&range);
5401 tlb_gather_mmu(&tlb, vma->vm_mm);
5403 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5405 mmu_notifier_invalidate_range_end(&range);
5406 tlb_finish_mmu(&tlb);
5410 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5411 * mapping it owns the reserve page for. The intention is to unmap the page
5412 * from other VMAs and let the children be SIGKILLed if they are faulting the
5415 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5416 struct page *page, unsigned long address)
5418 struct hstate *h = hstate_vma(vma);
5419 struct vm_area_struct *iter_vma;
5420 struct address_space *mapping;
5424 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5425 * from page cache lookup which is in HPAGE_SIZE units.
5427 address = address & huge_page_mask(h);
5428 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5430 mapping = vma->vm_file->f_mapping;
5433 * Take the mapping lock for the duration of the table walk. As
5434 * this mapping should be shared between all the VMAs,
5435 * __unmap_hugepage_range() is called as the lock is already held
5437 i_mmap_lock_write(mapping);
5438 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5439 /* Do not unmap the current VMA */
5440 if (iter_vma == vma)
5444 * Shared VMAs have their own reserves and do not affect
5445 * MAP_PRIVATE accounting but it is possible that a shared
5446 * VMA is using the same page so check and skip such VMAs.
5448 if (iter_vma->vm_flags & VM_MAYSHARE)
5452 * Unmap the page from other VMAs without their own reserves.
5453 * They get marked to be SIGKILLed if they fault in these
5454 * areas. This is because a future no-page fault on this VMA
5455 * could insert a zeroed page instead of the data existing
5456 * from the time of fork. This would look like data corruption
5458 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5459 unmap_hugepage_range(iter_vma, address,
5460 address + huge_page_size(h), page, 0);
5462 i_mmap_unlock_write(mapping);
5466 * hugetlb_wp() should be called with page lock of the original hugepage held.
5467 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5468 * cannot race with other handlers or page migration.
5469 * Keep the pte_same checks anyway to make transition from the mutex easier.
5471 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5472 unsigned long address, pte_t *ptep, unsigned int flags,
5473 struct page *pagecache_page, spinlock_t *ptl)
5475 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5477 struct hstate *h = hstate_vma(vma);
5478 struct page *old_page, *new_page;
5479 int outside_reserve = 0;
5481 unsigned long haddr = address & huge_page_mask(h);
5482 struct mmu_notifier_range range;
5485 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5486 * PTE mapped R/O such as maybe_mkwrite() would do.
5488 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5489 return VM_FAULT_SIGSEGV;
5491 /* Let's take out MAP_SHARED mappings first. */
5492 if (vma->vm_flags & VM_MAYSHARE) {
5493 set_huge_ptep_writable(vma, haddr, ptep);
5497 pte = huge_ptep_get(ptep);
5498 old_page = pte_page(pte);
5500 delayacct_wpcopy_start();
5504 * If no-one else is actually using this page, we're the exclusive
5505 * owner and can reuse this page.
5507 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5508 if (!PageAnonExclusive(old_page))
5509 page_move_anon_rmap(old_page, vma);
5510 if (likely(!unshare))
5511 set_huge_ptep_writable(vma, haddr, ptep);
5513 delayacct_wpcopy_end();
5516 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5520 * If the process that created a MAP_PRIVATE mapping is about to
5521 * perform a COW due to a shared page count, attempt to satisfy
5522 * the allocation without using the existing reserves. The pagecache
5523 * page is used to determine if the reserve at this address was
5524 * consumed or not. If reserves were used, a partial faulted mapping
5525 * at the time of fork() could consume its reserves on COW instead
5526 * of the full address range.
5528 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5529 old_page != pagecache_page)
5530 outside_reserve = 1;
5535 * Drop page table lock as buddy allocator may be called. It will
5536 * be acquired again before returning to the caller, as expected.
5539 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5541 if (IS_ERR(new_page)) {
5543 * If a process owning a MAP_PRIVATE mapping fails to COW,
5544 * it is due to references held by a child and an insufficient
5545 * huge page pool. To guarantee the original mappers
5546 * reliability, unmap the page from child processes. The child
5547 * may get SIGKILLed if it later faults.
5549 if (outside_reserve) {
5550 struct address_space *mapping = vma->vm_file->f_mapping;
5556 * Drop hugetlb_fault_mutex and vma_lock before
5557 * unmapping. unmapping needs to hold vma_lock
5558 * in write mode. Dropping vma_lock in read mode
5559 * here is OK as COW mappings do not interact with
5562 * Reacquire both after unmap operation.
5564 idx = vma_hugecache_offset(h, vma, haddr);
5565 hash = hugetlb_fault_mutex_hash(mapping, idx);
5566 hugetlb_vma_unlock_read(vma);
5567 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5569 unmap_ref_private(mm, vma, old_page, haddr);
5571 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5572 hugetlb_vma_lock_read(vma);
5574 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5576 pte_same(huge_ptep_get(ptep), pte)))
5577 goto retry_avoidcopy;
5579 * race occurs while re-acquiring page table
5580 * lock, and our job is done.
5582 delayacct_wpcopy_end();
5586 ret = vmf_error(PTR_ERR(new_page));
5587 goto out_release_old;
5591 * When the original hugepage is shared one, it does not have
5592 * anon_vma prepared.
5594 if (unlikely(anon_vma_prepare(vma))) {
5596 goto out_release_all;
5599 copy_user_huge_page(new_page, old_page, address, vma,
5600 pages_per_huge_page(h));
5601 __SetPageUptodate(new_page);
5603 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5604 haddr + huge_page_size(h));
5605 mmu_notifier_invalidate_range_start(&range);
5608 * Retake the page table lock to check for racing updates
5609 * before the page tables are altered
5612 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5613 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5614 /* Break COW or unshare */
5615 huge_ptep_clear_flush(vma, haddr, ptep);
5616 mmu_notifier_invalidate_range(mm, range.start, range.end);
5617 page_remove_rmap(old_page, vma, true);
5618 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5619 set_huge_pte_at(mm, haddr, ptep,
5620 make_huge_pte(vma, new_page, !unshare));
5621 SetHPageMigratable(new_page);
5622 /* Make the old page be freed below */
5623 new_page = old_page;
5626 mmu_notifier_invalidate_range_end(&range);
5629 * No restore in case of successful pagetable update (Break COW or
5632 if (new_page != old_page)
5633 restore_reserve_on_error(h, vma, haddr, new_page);
5638 spin_lock(ptl); /* Caller expects lock to be held */
5640 delayacct_wpcopy_end();
5645 * Return whether there is a pagecache page to back given address within VMA.
5646 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5648 static bool hugetlbfs_pagecache_present(struct hstate *h,
5649 struct vm_area_struct *vma, unsigned long address)
5651 struct address_space *mapping;
5655 mapping = vma->vm_file->f_mapping;
5656 idx = vma_hugecache_offset(h, vma, address);
5658 page = find_get_page(mapping, idx);
5661 return page != NULL;
5664 int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping,
5667 struct folio *folio = page_folio(page);
5668 struct inode *inode = mapping->host;
5669 struct hstate *h = hstate_inode(inode);
5672 __folio_set_locked(folio);
5673 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5675 if (unlikely(err)) {
5676 __folio_clear_locked(folio);
5679 ClearHPageRestoreReserve(page);
5682 * mark folio dirty so that it will not be removed from cache/file
5683 * by non-hugetlbfs specific code paths.
5685 folio_mark_dirty(folio);
5687 spin_lock(&inode->i_lock);
5688 inode->i_blocks += blocks_per_huge_page(h);
5689 spin_unlock(&inode->i_lock);
5693 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5694 struct address_space *mapping,
5697 unsigned long haddr,
5699 unsigned long reason)
5702 struct vm_fault vmf = {
5705 .real_address = addr,
5709 * Hard to debug if it ends up being
5710 * used by a callee that assumes
5711 * something about the other
5712 * uninitialized fields... same as in
5718 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5719 * userfault. Also mmap_lock could be dropped due to handling
5720 * userfault, any vma operation should be careful from here.
5722 hugetlb_vma_unlock_read(vma);
5723 hash = hugetlb_fault_mutex_hash(mapping, idx);
5724 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5725 return handle_userfault(&vmf, reason);
5729 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5730 * false if pte changed or is changing.
5732 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5733 pte_t *ptep, pte_t old_pte)
5738 ptl = huge_pte_lock(h, mm, ptep);
5739 same = pte_same(huge_ptep_get(ptep), old_pte);
5745 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5746 struct vm_area_struct *vma,
5747 struct address_space *mapping, pgoff_t idx,
5748 unsigned long address, pte_t *ptep,
5749 pte_t old_pte, unsigned int flags)
5751 struct hstate *h = hstate_vma(vma);
5752 vm_fault_t ret = VM_FAULT_SIGBUS;
5758 unsigned long haddr = address & huge_page_mask(h);
5759 bool new_page, new_pagecache_page = false;
5760 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5763 * Currently, we are forced to kill the process in the event the
5764 * original mapper has unmapped pages from the child due to a failed
5765 * COW/unsharing. Warn that such a situation has occurred as it may not
5768 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5769 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5775 * Use page lock to guard against racing truncation
5776 * before we get page_table_lock.
5779 page = find_lock_page(mapping, idx);
5781 size = i_size_read(mapping->host) >> huge_page_shift(h);
5784 /* Check for page in userfault range */
5785 if (userfaultfd_missing(vma)) {
5787 * Since hugetlb_no_page() was examining pte
5788 * without pgtable lock, we need to re-test under
5789 * lock because the pte may not be stable and could
5790 * have changed from under us. Try to detect
5791 * either changed or during-changing ptes and retry
5792 * properly when needed.
5794 * Note that userfaultfd is actually fine with
5795 * false positives (e.g. caused by pte changed),
5796 * but not wrong logical events (e.g. caused by
5797 * reading a pte during changing). The latter can
5798 * confuse the userspace, so the strictness is very
5799 * much preferred. E.g., MISSING event should
5800 * never happen on the page after UFFDIO_COPY has
5801 * correctly installed the page and returned.
5803 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5808 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5813 page = alloc_huge_page(vma, haddr, 0);
5816 * Returning error will result in faulting task being
5817 * sent SIGBUS. The hugetlb fault mutex prevents two
5818 * tasks from racing to fault in the same page which
5819 * could result in false unable to allocate errors.
5820 * Page migration does not take the fault mutex, but
5821 * does a clear then write of pte's under page table
5822 * lock. Page fault code could race with migration,
5823 * notice the clear pte and try to allocate a page
5824 * here. Before returning error, get ptl and make
5825 * sure there really is no pte entry.
5827 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5828 ret = vmf_error(PTR_ERR(page));
5833 clear_huge_page(page, address, pages_per_huge_page(h));
5834 __SetPageUptodate(page);
5837 if (vma->vm_flags & VM_MAYSHARE) {
5838 int err = hugetlb_add_to_page_cache(page, mapping, idx);
5841 * err can't be -EEXIST which implies someone
5842 * else consumed the reservation since hugetlb
5843 * fault mutex is held when add a hugetlb page
5844 * to the page cache. So it's safe to call
5845 * restore_reserve_on_error() here.
5847 restore_reserve_on_error(h, vma, haddr, page);
5851 new_pagecache_page = true;
5854 if (unlikely(anon_vma_prepare(vma))) {
5856 goto backout_unlocked;
5862 * If memory error occurs between mmap() and fault, some process
5863 * don't have hwpoisoned swap entry for errored virtual address.
5864 * So we need to block hugepage fault by PG_hwpoison bit check.
5866 if (unlikely(PageHWPoison(page))) {
5867 ret = VM_FAULT_HWPOISON_LARGE |
5868 VM_FAULT_SET_HINDEX(hstate_index(h));
5869 goto backout_unlocked;
5872 /* Check for page in userfault range. */
5873 if (userfaultfd_minor(vma)) {
5876 /* See comment in userfaultfd_missing() block above */
5877 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5881 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5888 * If we are going to COW a private mapping later, we examine the
5889 * pending reservations for this page now. This will ensure that
5890 * any allocations necessary to record that reservation occur outside
5893 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5894 if (vma_needs_reservation(h, vma, haddr) < 0) {
5896 goto backout_unlocked;
5898 /* Just decrements count, does not deallocate */
5899 vma_end_reservation(h, vma, haddr);
5902 ptl = huge_pte_lock(h, mm, ptep);
5904 /* If pte changed from under us, retry */
5905 if (!pte_same(huge_ptep_get(ptep), old_pte))
5909 hugepage_add_new_anon_rmap(page, vma, haddr);
5911 page_dup_file_rmap(page, true);
5912 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5913 && (vma->vm_flags & VM_SHARED)));
5915 * If this pte was previously wr-protected, keep it wr-protected even
5918 if (unlikely(pte_marker_uffd_wp(old_pte)))
5919 new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte));
5920 set_huge_pte_at(mm, haddr, ptep, new_pte);
5922 hugetlb_count_add(pages_per_huge_page(h), mm);
5923 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5924 /* Optimization, do the COW without a second fault */
5925 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5931 * Only set HPageMigratable in newly allocated pages. Existing pages
5932 * found in the pagecache may not have HPageMigratableset if they have
5933 * been isolated for migration.
5936 SetHPageMigratable(page);
5940 hugetlb_vma_unlock_read(vma);
5941 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5947 if (new_page && !new_pagecache_page)
5948 restore_reserve_on_error(h, vma, haddr, page);
5956 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5958 unsigned long key[2];
5961 key[0] = (unsigned long) mapping;
5964 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5966 return hash & (num_fault_mutexes - 1);
5970 * For uniprocessor systems we always use a single mutex, so just
5971 * return 0 and avoid the hashing overhead.
5973 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5979 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5980 unsigned long address, unsigned int flags)
5987 struct page *page = NULL;
5988 struct page *pagecache_page = NULL;
5989 struct hstate *h = hstate_vma(vma);
5990 struct address_space *mapping;
5991 int need_wait_lock = 0;
5992 unsigned long haddr = address & huge_page_mask(h);
5994 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5997 * Since we hold no locks, ptep could be stale. That is
5998 * OK as we are only making decisions based on content and
5999 * not actually modifying content here.
6001 entry = huge_ptep_get(ptep);
6002 if (unlikely(is_hugetlb_entry_migration(entry))) {
6003 migration_entry_wait_huge(vma, ptep);
6005 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6006 return VM_FAULT_HWPOISON_LARGE |
6007 VM_FAULT_SET_HINDEX(hstate_index(h));
6011 * Serialize hugepage allocation and instantiation, so that we don't
6012 * get spurious allocation failures if two CPUs race to instantiate
6013 * the same page in the page cache.
6015 mapping = vma->vm_file->f_mapping;
6016 idx = vma_hugecache_offset(h, vma, haddr);
6017 hash = hugetlb_fault_mutex_hash(mapping, idx);
6018 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6021 * Acquire vma lock before calling huge_pte_alloc and hold
6022 * until finished with ptep. This prevents huge_pmd_unshare from
6023 * being called elsewhere and making the ptep no longer valid.
6025 * ptep could have already be assigned via huge_pte_offset. That
6026 * is OK, as huge_pte_alloc will return the same value unless
6027 * something has changed.
6029 hugetlb_vma_lock_read(vma);
6030 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6032 hugetlb_vma_unlock_read(vma);
6033 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6034 return VM_FAULT_OOM;
6037 entry = huge_ptep_get(ptep);
6038 /* PTE markers should be handled the same way as none pte */
6039 if (huge_pte_none_mostly(entry))
6041 * hugetlb_no_page will drop vma lock and hugetlb fault
6042 * mutex internally, which make us return immediately.
6044 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6050 * entry could be a migration/hwpoison entry at this point, so this
6051 * check prevents the kernel from going below assuming that we have
6052 * an active hugepage in pagecache. This goto expects the 2nd page
6053 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6054 * properly handle it.
6056 if (!pte_present(entry))
6060 * If we are going to COW/unshare the mapping later, we examine the
6061 * pending reservations for this page now. This will ensure that any
6062 * allocations necessary to record that reservation occur outside the
6063 * spinlock. Also lookup the pagecache page now as it is used to
6064 * determine if a reservation has been consumed.
6066 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6067 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6068 if (vma_needs_reservation(h, vma, haddr) < 0) {
6072 /* Just decrements count, does not deallocate */
6073 vma_end_reservation(h, vma, haddr);
6075 pagecache_page = find_lock_page(mapping, idx);
6078 ptl = huge_pte_lock(h, mm, ptep);
6080 /* Check for a racing update before calling hugetlb_wp() */
6081 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6084 /* Handle userfault-wp first, before trying to lock more pages */
6085 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6086 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6087 struct vm_fault vmf = {
6090 .real_address = address,
6095 if (pagecache_page) {
6096 unlock_page(pagecache_page);
6097 put_page(pagecache_page);
6099 hugetlb_vma_unlock_read(vma);
6100 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6101 return handle_userfault(&vmf, VM_UFFD_WP);
6105 * hugetlb_wp() requires page locks of pte_page(entry) and
6106 * pagecache_page, so here we need take the former one
6107 * when page != pagecache_page or !pagecache_page.
6109 page = pte_page(entry);
6110 if (page != pagecache_page)
6111 if (!trylock_page(page)) {
6118 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6119 if (!huge_pte_write(entry)) {
6120 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6121 pagecache_page, ptl);
6123 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6124 entry = huge_pte_mkdirty(entry);
6127 entry = pte_mkyoung(entry);
6128 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6129 flags & FAULT_FLAG_WRITE))
6130 update_mmu_cache(vma, haddr, ptep);
6132 if (page != pagecache_page)
6138 if (pagecache_page) {
6139 unlock_page(pagecache_page);
6140 put_page(pagecache_page);
6143 hugetlb_vma_unlock_read(vma);
6144 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6146 * Generally it's safe to hold refcount during waiting page lock. But
6147 * here we just wait to defer the next page fault to avoid busy loop and
6148 * the page is not used after unlocked before returning from the current
6149 * page fault. So we are safe from accessing freed page, even if we wait
6150 * here without taking refcount.
6153 wait_on_page_locked(page);
6157 #ifdef CONFIG_USERFAULTFD
6159 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
6160 * modifications for huge pages.
6162 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
6164 struct vm_area_struct *dst_vma,
6165 unsigned long dst_addr,
6166 unsigned long src_addr,
6167 enum mcopy_atomic_mode mode,
6168 struct page **pagep,
6171 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
6172 struct hstate *h = hstate_vma(dst_vma);
6173 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6174 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6176 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6182 bool page_in_pagecache = false;
6186 page = find_lock_page(mapping, idx);
6189 page_in_pagecache = true;
6190 } else if (!*pagep) {
6191 /* If a page already exists, then it's UFFDIO_COPY for
6192 * a non-missing case. Return -EEXIST.
6195 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6200 page = alloc_huge_page(dst_vma, dst_addr, 0);
6206 ret = copy_huge_page_from_user(page,
6207 (const void __user *) src_addr,
6208 pages_per_huge_page(h), false);
6210 /* fallback to copy_from_user outside mmap_lock */
6211 if (unlikely(ret)) {
6213 /* Free the allocated page which may have
6214 * consumed a reservation.
6216 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6219 /* Allocate a temporary page to hold the copied
6222 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
6228 /* Set the outparam pagep and return to the caller to
6229 * copy the contents outside the lock. Don't free the
6236 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6243 page = alloc_huge_page(dst_vma, dst_addr, 0);
6250 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
6251 pages_per_huge_page(h));
6257 * The memory barrier inside __SetPageUptodate makes sure that
6258 * preceding stores to the page contents become visible before
6259 * the set_pte_at() write.
6261 __SetPageUptodate(page);
6263 /* Add shared, newly allocated pages to the page cache. */
6264 if (vm_shared && !is_continue) {
6265 size = i_size_read(mapping->host) >> huge_page_shift(h);
6268 goto out_release_nounlock;
6271 * Serialization between remove_inode_hugepages() and
6272 * hugetlb_add_to_page_cache() below happens through the
6273 * hugetlb_fault_mutex_table that here must be hold by
6276 ret = hugetlb_add_to_page_cache(page, mapping, idx);
6278 goto out_release_nounlock;
6279 page_in_pagecache = true;
6282 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6285 if (PageHWPoison(page))
6286 goto out_release_unlock;
6289 * We allow to overwrite a pte marker: consider when both MISSING|WP
6290 * registered, we firstly wr-protect a none pte which has no page cache
6291 * page backing it, then access the page.
6294 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6295 goto out_release_unlock;
6297 if (page_in_pagecache)
6298 page_dup_file_rmap(page, true);
6300 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6303 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6304 * with wp flag set, don't set pte write bit.
6306 if (wp_copy || (is_continue && !vm_shared))
6309 writable = dst_vma->vm_flags & VM_WRITE;
6311 _dst_pte = make_huge_pte(dst_vma, page, writable);
6313 * Always mark UFFDIO_COPY page dirty; note that this may not be
6314 * extremely important for hugetlbfs for now since swapping is not
6315 * supported, but we should still be clear in that this page cannot be
6316 * thrown away at will, even if write bit not set.
6318 _dst_pte = huge_pte_mkdirty(_dst_pte);
6319 _dst_pte = pte_mkyoung(_dst_pte);
6322 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6324 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6326 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6328 /* No need to invalidate - it was non-present before */
6329 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6333 SetHPageMigratable(page);
6334 if (vm_shared || is_continue)
6341 if (vm_shared || is_continue)
6343 out_release_nounlock:
6344 if (!page_in_pagecache)
6345 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6349 #endif /* CONFIG_USERFAULTFD */
6351 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6352 int refs, struct page **pages,
6353 struct vm_area_struct **vmas)
6357 for (nr = 0; nr < refs; nr++) {
6359 pages[nr] = nth_page(page, nr);
6365 static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma,
6366 unsigned int flags, pte_t *pte,
6369 pte_t pteval = huge_ptep_get(pte);
6372 if (is_swap_pte(pteval))
6374 if (huge_pte_write(pteval))
6376 if (flags & FOLL_WRITE)
6378 if (gup_must_unshare(vma, flags, pte_page(pteval))) {
6385 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6386 unsigned long address, unsigned int flags)
6388 struct hstate *h = hstate_vma(vma);
6389 struct mm_struct *mm = vma->vm_mm;
6390 unsigned long haddr = address & huge_page_mask(h);
6391 struct page *page = NULL;
6396 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
6397 * follow_hugetlb_page().
6399 if (WARN_ON_ONCE(flags & FOLL_PIN))
6403 pte = huge_pte_offset(mm, haddr, huge_page_size(h));
6407 ptl = huge_pte_lock(h, mm, pte);
6408 entry = huge_ptep_get(pte);
6409 if (pte_present(entry)) {
6410 page = pte_page(entry) +
6411 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6413 * Note that page may be a sub-page, and with vmemmap
6414 * optimizations the page struct may be read only.
6415 * try_grab_page() will increase the ref count on the
6416 * head page, so this will be OK.
6418 * try_grab_page() should always be able to get the page here,
6419 * because we hold the ptl lock and have verified pte_present().
6421 if (try_grab_page(page, flags)) {
6426 if (is_hugetlb_entry_migration(entry)) {
6428 __migration_entry_wait_huge(pte, ptl);
6432 * hwpoisoned entry is treated as no_page_table in
6433 * follow_page_mask().
6441 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6442 struct page **pages, struct vm_area_struct **vmas,
6443 unsigned long *position, unsigned long *nr_pages,
6444 long i, unsigned int flags, int *locked)
6446 unsigned long pfn_offset;
6447 unsigned long vaddr = *position;
6448 unsigned long remainder = *nr_pages;
6449 struct hstate *h = hstate_vma(vma);
6450 int err = -EFAULT, refs;
6452 while (vaddr < vma->vm_end && remainder) {
6454 spinlock_t *ptl = NULL;
6455 bool unshare = false;
6460 * If we have a pending SIGKILL, don't keep faulting pages and
6461 * potentially allocating memory.
6463 if (fatal_signal_pending(current)) {
6469 * Some archs (sparc64, sh*) have multiple pte_ts to
6470 * each hugepage. We have to make sure we get the
6471 * first, for the page indexing below to work.
6473 * Note that page table lock is not held when pte is null.
6475 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6478 ptl = huge_pte_lock(h, mm, pte);
6479 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6482 * When coredumping, it suits get_dump_page if we just return
6483 * an error where there's an empty slot with no huge pagecache
6484 * to back it. This way, we avoid allocating a hugepage, and
6485 * the sparse dumpfile avoids allocating disk blocks, but its
6486 * huge holes still show up with zeroes where they need to be.
6488 if (absent && (flags & FOLL_DUMP) &&
6489 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6497 * We need call hugetlb_fault for both hugepages under migration
6498 * (in which case hugetlb_fault waits for the migration,) and
6499 * hwpoisoned hugepages (in which case we need to prevent the
6500 * caller from accessing to them.) In order to do this, we use
6501 * here is_swap_pte instead of is_hugetlb_entry_migration and
6502 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6503 * both cases, and because we can't follow correct pages
6504 * directly from any kind of swap entries.
6507 __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) {
6509 unsigned int fault_flags = 0;
6513 if (flags & FOLL_WRITE)
6514 fault_flags |= FAULT_FLAG_WRITE;
6516 fault_flags |= FAULT_FLAG_UNSHARE;
6518 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6519 FAULT_FLAG_KILLABLE;
6520 if (flags & FOLL_INTERRUPTIBLE)
6521 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
6523 if (flags & FOLL_NOWAIT)
6524 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6525 FAULT_FLAG_RETRY_NOWAIT;
6526 if (flags & FOLL_TRIED) {
6528 * Note: FAULT_FLAG_ALLOW_RETRY and
6529 * FAULT_FLAG_TRIED can co-exist
6531 fault_flags |= FAULT_FLAG_TRIED;
6533 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6534 if (ret & VM_FAULT_ERROR) {
6535 err = vm_fault_to_errno(ret, flags);
6539 if (ret & VM_FAULT_RETRY) {
6541 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6545 * VM_FAULT_RETRY must not return an
6546 * error, it will return zero
6549 * No need to update "position" as the
6550 * caller will not check it after
6551 * *nr_pages is set to 0.
6558 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6559 page = pte_page(huge_ptep_get(pte));
6561 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6562 !PageAnonExclusive(page), page);
6565 * If subpage information not requested, update counters
6566 * and skip the same_page loop below.
6568 if (!pages && !vmas && !pfn_offset &&
6569 (vaddr + huge_page_size(h) < vma->vm_end) &&
6570 (remainder >= pages_per_huge_page(h))) {
6571 vaddr += huge_page_size(h);
6572 remainder -= pages_per_huge_page(h);
6573 i += pages_per_huge_page(h);
6578 /* vaddr may not be aligned to PAGE_SIZE */
6579 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6580 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6583 record_subpages_vmas(nth_page(page, pfn_offset),
6585 likely(pages) ? pages + i : NULL,
6586 vmas ? vmas + i : NULL);
6590 * try_grab_folio() should always succeed here,
6591 * because: a) we hold the ptl lock, and b) we've just
6592 * checked that the huge page is present in the page
6593 * tables. If the huge page is present, then the tail
6594 * pages must also be present. The ptl prevents the
6595 * head page and tail pages from being rearranged in
6596 * any way. As this is hugetlb, the pages will never
6597 * be p2pdma or not longterm pinable. So this page
6598 * must be available at this point, unless the page
6599 * refcount overflowed:
6601 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6610 vaddr += (refs << PAGE_SHIFT);
6616 *nr_pages = remainder;
6618 * setting position is actually required only if remainder is
6619 * not zero but it's faster not to add a "if (remainder)"
6627 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6628 unsigned long address, unsigned long end,
6629 pgprot_t newprot, unsigned long cp_flags)
6631 struct mm_struct *mm = vma->vm_mm;
6632 unsigned long start = address;
6635 struct hstate *h = hstate_vma(vma);
6636 unsigned long pages = 0, psize = huge_page_size(h);
6637 bool shared_pmd = false;
6638 struct mmu_notifier_range range;
6639 unsigned long last_addr_mask;
6640 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6641 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6644 * In the case of shared PMDs, the area to flush could be beyond
6645 * start/end. Set range.start/range.end to cover the maximum possible
6646 * range if PMD sharing is possible.
6648 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6649 0, vma, mm, start, end);
6650 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6652 BUG_ON(address >= end);
6653 flush_cache_range(vma, range.start, range.end);
6655 mmu_notifier_invalidate_range_start(&range);
6656 hugetlb_vma_lock_write(vma);
6657 i_mmap_lock_write(vma->vm_file->f_mapping);
6658 last_addr_mask = hugetlb_mask_last_page(h);
6659 for (; address < end; address += psize) {
6661 ptep = huge_pte_offset(mm, address, psize);
6663 address |= last_addr_mask;
6666 ptl = huge_pte_lock(h, mm, ptep);
6667 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6669 * When uffd-wp is enabled on the vma, unshare
6670 * shouldn't happen at all. Warn about it if it
6671 * happened due to some reason.
6673 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6677 address |= last_addr_mask;
6680 pte = huge_ptep_get(ptep);
6681 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6682 /* Nothing to do. */
6683 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6684 swp_entry_t entry = pte_to_swp_entry(pte);
6685 struct page *page = pfn_swap_entry_to_page(entry);
6688 if (is_writable_migration_entry(entry)) {
6690 entry = make_readable_exclusive_migration_entry(
6693 entry = make_readable_migration_entry(
6695 newpte = swp_entry_to_pte(entry);
6700 newpte = pte_swp_mkuffd_wp(newpte);
6701 else if (uffd_wp_resolve)
6702 newpte = pte_swp_clear_uffd_wp(newpte);
6703 if (!pte_same(pte, newpte))
6704 set_huge_pte_at(mm, address, ptep, newpte);
6705 } else if (unlikely(is_pte_marker(pte))) {
6706 /* No other markers apply for now. */
6707 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6708 if (uffd_wp_resolve)
6709 /* Safe to modify directly (non-present->none). */
6710 huge_pte_clear(mm, address, ptep, psize);
6711 } else if (!huge_pte_none(pte)) {
6713 unsigned int shift = huge_page_shift(hstate_vma(vma));
6715 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6716 pte = huge_pte_modify(old_pte, newprot);
6717 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6719 pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte));
6720 else if (uffd_wp_resolve)
6721 pte = huge_pte_clear_uffd_wp(pte);
6722 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6726 if (unlikely(uffd_wp))
6727 /* Safe to modify directly (none->non-present). */
6728 set_huge_pte_at(mm, address, ptep,
6729 make_pte_marker(PTE_MARKER_UFFD_WP));
6734 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6735 * may have cleared our pud entry and done put_page on the page table:
6736 * once we release i_mmap_rwsem, another task can do the final put_page
6737 * and that page table be reused and filled with junk. If we actually
6738 * did unshare a page of pmds, flush the range corresponding to the pud.
6741 flush_hugetlb_tlb_range(vma, range.start, range.end);
6743 flush_hugetlb_tlb_range(vma, start, end);
6745 * No need to call mmu_notifier_invalidate_range() we are downgrading
6746 * page table protection not changing it to point to a new page.
6748 * See Documentation/mm/mmu_notifier.rst
6750 i_mmap_unlock_write(vma->vm_file->f_mapping);
6751 hugetlb_vma_unlock_write(vma);
6752 mmu_notifier_invalidate_range_end(&range);
6754 return pages << h->order;
6757 /* Return true if reservation was successful, false otherwise. */
6758 bool hugetlb_reserve_pages(struct inode *inode,
6760 struct vm_area_struct *vma,
6761 vm_flags_t vm_flags)
6764 struct hstate *h = hstate_inode(inode);
6765 struct hugepage_subpool *spool = subpool_inode(inode);
6766 struct resv_map *resv_map;
6767 struct hugetlb_cgroup *h_cg = NULL;
6768 long gbl_reserve, regions_needed = 0;
6770 /* This should never happen */
6772 VM_WARN(1, "%s called with a negative range\n", __func__);
6777 * vma specific semaphore used for pmd sharing and fault/truncation
6780 hugetlb_vma_lock_alloc(vma);
6783 * Only apply hugepage reservation if asked. At fault time, an
6784 * attempt will be made for VM_NORESERVE to allocate a page
6785 * without using reserves
6787 if (vm_flags & VM_NORESERVE)
6791 * Shared mappings base their reservation on the number of pages that
6792 * are already allocated on behalf of the file. Private mappings need
6793 * to reserve the full area even if read-only as mprotect() may be
6794 * called to make the mapping read-write. Assume !vma is a shm mapping
6796 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6798 * resv_map can not be NULL as hugetlb_reserve_pages is only
6799 * called for inodes for which resv_maps were created (see
6800 * hugetlbfs_get_inode).
6802 resv_map = inode_resv_map(inode);
6804 chg = region_chg(resv_map, from, to, ®ions_needed);
6806 /* Private mapping. */
6807 resv_map = resv_map_alloc();
6813 set_vma_resv_map(vma, resv_map);
6814 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6820 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6821 chg * pages_per_huge_page(h), &h_cg) < 0)
6824 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6825 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6828 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6832 * There must be enough pages in the subpool for the mapping. If
6833 * the subpool has a minimum size, there may be some global
6834 * reservations already in place (gbl_reserve).
6836 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6837 if (gbl_reserve < 0)
6838 goto out_uncharge_cgroup;
6841 * Check enough hugepages are available for the reservation.
6842 * Hand the pages back to the subpool if there are not
6844 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6848 * Account for the reservations made. Shared mappings record regions
6849 * that have reservations as they are shared by multiple VMAs.
6850 * When the last VMA disappears, the region map says how much
6851 * the reservation was and the page cache tells how much of
6852 * the reservation was consumed. Private mappings are per-VMA and
6853 * only the consumed reservations are tracked. When the VMA
6854 * disappears, the original reservation is the VMA size and the
6855 * consumed reservations are stored in the map. Hence, nothing
6856 * else has to be done for private mappings here
6858 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6859 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6861 if (unlikely(add < 0)) {
6862 hugetlb_acct_memory(h, -gbl_reserve);
6864 } else if (unlikely(chg > add)) {
6866 * pages in this range were added to the reserve
6867 * map between region_chg and region_add. This
6868 * indicates a race with alloc_huge_page. Adjust
6869 * the subpool and reserve counts modified above
6870 * based on the difference.
6875 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6876 * reference to h_cg->css. See comment below for detail.
6878 hugetlb_cgroup_uncharge_cgroup_rsvd(
6880 (chg - add) * pages_per_huge_page(h), h_cg);
6882 rsv_adjust = hugepage_subpool_put_pages(spool,
6884 hugetlb_acct_memory(h, -rsv_adjust);
6887 * The file_regions will hold their own reference to
6888 * h_cg->css. So we should release the reference held
6889 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6892 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6898 /* put back original number of pages, chg */
6899 (void)hugepage_subpool_put_pages(spool, chg);
6900 out_uncharge_cgroup:
6901 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6902 chg * pages_per_huge_page(h), h_cg);
6904 hugetlb_vma_lock_free(vma);
6905 if (!vma || vma->vm_flags & VM_MAYSHARE)
6906 /* Only call region_abort if the region_chg succeeded but the
6907 * region_add failed or didn't run.
6909 if (chg >= 0 && add < 0)
6910 region_abort(resv_map, from, to, regions_needed);
6911 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6912 kref_put(&resv_map->refs, resv_map_release);
6916 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6919 struct hstate *h = hstate_inode(inode);
6920 struct resv_map *resv_map = inode_resv_map(inode);
6922 struct hugepage_subpool *spool = subpool_inode(inode);
6926 * Since this routine can be called in the evict inode path for all
6927 * hugetlbfs inodes, resv_map could be NULL.
6930 chg = region_del(resv_map, start, end);
6932 * region_del() can fail in the rare case where a region
6933 * must be split and another region descriptor can not be
6934 * allocated. If end == LONG_MAX, it will not fail.
6940 spin_lock(&inode->i_lock);
6941 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6942 spin_unlock(&inode->i_lock);
6945 * If the subpool has a minimum size, the number of global
6946 * reservations to be released may be adjusted.
6948 * Note that !resv_map implies freed == 0. So (chg - freed)
6949 * won't go negative.
6951 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6952 hugetlb_acct_memory(h, -gbl_reserve);
6957 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6958 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6959 struct vm_area_struct *vma,
6960 unsigned long addr, pgoff_t idx)
6962 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6964 unsigned long sbase = saddr & PUD_MASK;
6965 unsigned long s_end = sbase + PUD_SIZE;
6967 /* Allow segments to share if only one is marked locked */
6968 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6969 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6972 * match the virtual addresses, permission and the alignment of the
6975 * Also, vma_lock (vm_private_data) is required for sharing.
6977 if (pmd_index(addr) != pmd_index(saddr) ||
6978 vm_flags != svm_flags ||
6979 !range_in_vma(svma, sbase, s_end) ||
6980 !svma->vm_private_data)
6986 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6988 unsigned long start = addr & PUD_MASK;
6989 unsigned long end = start + PUD_SIZE;
6991 #ifdef CONFIG_USERFAULTFD
6992 if (uffd_disable_huge_pmd_share(vma))
6996 * check on proper vm_flags and page table alignment
6998 if (!(vma->vm_flags & VM_MAYSHARE))
7000 if (!vma->vm_private_data) /* vma lock required for sharing */
7002 if (!range_in_vma(vma, start, end))
7008 * Determine if start,end range within vma could be mapped by shared pmd.
7009 * If yes, adjust start and end to cover range associated with possible
7010 * shared pmd mappings.
7012 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7013 unsigned long *start, unsigned long *end)
7015 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7016 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7019 * vma needs to span at least one aligned PUD size, and the range
7020 * must be at least partially within in.
7022 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7023 (*end <= v_start) || (*start >= v_end))
7026 /* Extend the range to be PUD aligned for a worst case scenario */
7027 if (*start > v_start)
7028 *start = ALIGN_DOWN(*start, PUD_SIZE);
7031 *end = ALIGN(*end, PUD_SIZE);
7035 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7036 * and returns the corresponding pte. While this is not necessary for the
7037 * !shared pmd case because we can allocate the pmd later as well, it makes the
7038 * code much cleaner. pmd allocation is essential for the shared case because
7039 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7040 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7041 * bad pmd for sharing.
7043 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7044 unsigned long addr, pud_t *pud)
7046 struct address_space *mapping = vma->vm_file->f_mapping;
7047 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7049 struct vm_area_struct *svma;
7050 unsigned long saddr;
7055 i_mmap_lock_read(mapping);
7056 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7060 saddr = page_table_shareable(svma, vma, addr, idx);
7062 spte = huge_pte_offset(svma->vm_mm, saddr,
7063 vma_mmu_pagesize(svma));
7065 get_page(virt_to_page(spte));
7074 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
7075 if (pud_none(*pud)) {
7076 pud_populate(mm, pud,
7077 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7080 put_page(virt_to_page(spte));
7084 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7085 i_mmap_unlock_read(mapping);
7090 * unmap huge page backed by shared pte.
7092 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7093 * indicated by page_count > 1, unmap is achieved by clearing pud and
7094 * decrementing the ref count. If count == 1, the pte page is not shared.
7096 * Called with page table lock held.
7098 * returns: 1 successfully unmapped a shared pte page
7099 * 0 the underlying pte page is not shared, or it is the last user
7101 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7102 unsigned long addr, pte_t *ptep)
7104 pgd_t *pgd = pgd_offset(mm, addr);
7105 p4d_t *p4d = p4d_offset(pgd, addr);
7106 pud_t *pud = pud_offset(p4d, addr);
7108 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7109 hugetlb_vma_assert_locked(vma);
7110 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7111 if (page_count(virt_to_page(ptep)) == 1)
7115 put_page(virt_to_page(ptep));
7120 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7122 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7123 unsigned long addr, pud_t *pud)
7128 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7129 unsigned long addr, pte_t *ptep)
7134 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7135 unsigned long *start, unsigned long *end)
7139 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7143 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7145 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7146 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7147 unsigned long addr, unsigned long sz)
7154 pgd = pgd_offset(mm, addr);
7155 p4d = p4d_alloc(mm, pgd, addr);
7158 pud = pud_alloc(mm, p4d, addr);
7160 if (sz == PUD_SIZE) {
7163 BUG_ON(sz != PMD_SIZE);
7164 if (want_pmd_share(vma, addr) && pud_none(*pud))
7165 pte = huge_pmd_share(mm, vma, addr, pud);
7167 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7170 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
7176 * huge_pte_offset() - Walk the page table to resolve the hugepage
7177 * entry at address @addr
7179 * Return: Pointer to page table entry (PUD or PMD) for
7180 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7181 * size @sz doesn't match the hugepage size at this level of the page
7184 pte_t *huge_pte_offset(struct mm_struct *mm,
7185 unsigned long addr, unsigned long sz)
7192 pgd = pgd_offset(mm, addr);
7193 if (!pgd_present(*pgd))
7195 p4d = p4d_offset(pgd, addr);
7196 if (!p4d_present(*p4d))
7199 pud = pud_offset(p4d, addr);
7201 /* must be pud huge, non-present or none */
7202 return (pte_t *)pud;
7203 if (!pud_present(*pud))
7205 /* must have a valid entry and size to go further */
7207 pmd = pmd_offset(pud, addr);
7208 /* must be pmd huge, non-present or none */
7209 return (pte_t *)pmd;
7213 * Return a mask that can be used to update an address to the last huge
7214 * page in a page table page mapping size. Used to skip non-present
7215 * page table entries when linearly scanning address ranges. Architectures
7216 * with unique huge page to page table relationships can define their own
7217 * version of this routine.
7219 unsigned long hugetlb_mask_last_page(struct hstate *h)
7221 unsigned long hp_size = huge_page_size(h);
7223 if (hp_size == PUD_SIZE)
7224 return P4D_SIZE - PUD_SIZE;
7225 else if (hp_size == PMD_SIZE)
7226 return PUD_SIZE - PMD_SIZE;
7233 /* See description above. Architectures can provide their own version. */
7234 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7236 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7237 if (huge_page_size(h) == PMD_SIZE)
7238 return PUD_SIZE - PMD_SIZE;
7243 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7246 * These functions are overwritable if your architecture needs its own
7249 int isolate_hugetlb(struct page *page, struct list_head *list)
7253 spin_lock_irq(&hugetlb_lock);
7254 if (!PageHeadHuge(page) ||
7255 !HPageMigratable(page) ||
7256 !get_page_unless_zero(page)) {
7260 ClearHPageMigratable(page);
7261 list_move_tail(&page->lru, list);
7263 spin_unlock_irq(&hugetlb_lock);
7267 int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison)
7272 spin_lock_irq(&hugetlb_lock);
7273 if (PageHeadHuge(page)) {
7275 if (HPageFreed(page))
7277 else if (HPageMigratable(page) || unpoison)
7278 ret = get_page_unless_zero(page);
7282 spin_unlock_irq(&hugetlb_lock);
7286 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7287 bool *migratable_cleared)
7291 spin_lock_irq(&hugetlb_lock);
7292 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7293 spin_unlock_irq(&hugetlb_lock);
7297 void putback_active_hugepage(struct page *page)
7299 spin_lock_irq(&hugetlb_lock);
7300 SetHPageMigratable(page);
7301 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7302 spin_unlock_irq(&hugetlb_lock);
7306 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7308 struct hstate *h = folio_hstate(old_folio);
7310 hugetlb_cgroup_migrate(old_folio, new_folio);
7311 set_page_owner_migrate_reason(&new_folio->page, reason);
7314 * transfer temporary state of the new hugetlb folio. This is
7315 * reverse to other transitions because the newpage is going to
7316 * be final while the old one will be freed so it takes over
7317 * the temporary status.
7319 * Also note that we have to transfer the per-node surplus state
7320 * here as well otherwise the global surplus count will not match
7323 if (folio_test_hugetlb_temporary(new_folio)) {
7324 int old_nid = folio_nid(old_folio);
7325 int new_nid = folio_nid(new_folio);
7327 folio_set_hugetlb_temporary(old_folio);
7328 folio_clear_hugetlb_temporary(new_folio);
7332 * There is no need to transfer the per-node surplus state
7333 * when we do not cross the node.
7335 if (new_nid == old_nid)
7337 spin_lock_irq(&hugetlb_lock);
7338 if (h->surplus_huge_pages_node[old_nid]) {
7339 h->surplus_huge_pages_node[old_nid]--;
7340 h->surplus_huge_pages_node[new_nid]++;
7342 spin_unlock_irq(&hugetlb_lock);
7346 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7347 unsigned long start,
7350 struct hstate *h = hstate_vma(vma);
7351 unsigned long sz = huge_page_size(h);
7352 struct mm_struct *mm = vma->vm_mm;
7353 struct mmu_notifier_range range;
7354 unsigned long address;
7358 if (!(vma->vm_flags & VM_MAYSHARE))
7364 flush_cache_range(vma, start, end);
7366 * No need to call adjust_range_if_pmd_sharing_possible(), because
7367 * we have already done the PUD_SIZE alignment.
7369 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7371 mmu_notifier_invalidate_range_start(&range);
7372 hugetlb_vma_lock_write(vma);
7373 i_mmap_lock_write(vma->vm_file->f_mapping);
7374 for (address = start; address < end; address += PUD_SIZE) {
7375 ptep = huge_pte_offset(mm, address, sz);
7378 ptl = huge_pte_lock(h, mm, ptep);
7379 huge_pmd_unshare(mm, vma, address, ptep);
7382 flush_hugetlb_tlb_range(vma, start, end);
7383 i_mmap_unlock_write(vma->vm_file->f_mapping);
7384 hugetlb_vma_unlock_write(vma);
7386 * No need to call mmu_notifier_invalidate_range(), see
7387 * Documentation/mm/mmu_notifier.rst.
7389 mmu_notifier_invalidate_range_end(&range);
7393 * This function will unconditionally remove all the shared pmd pgtable entries
7394 * within the specific vma for a hugetlbfs memory range.
7396 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7398 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7399 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7403 static bool cma_reserve_called __initdata;
7405 static int __init cmdline_parse_hugetlb_cma(char *p)
7412 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7415 if (s[count] == ':') {
7416 if (tmp >= MAX_NUMNODES)
7418 nid = array_index_nospec(tmp, MAX_NUMNODES);
7421 tmp = memparse(s, &s);
7422 hugetlb_cma_size_in_node[nid] = tmp;
7423 hugetlb_cma_size += tmp;
7426 * Skip the separator if have one, otherwise
7427 * break the parsing.
7434 hugetlb_cma_size = memparse(p, &p);
7442 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7444 void __init hugetlb_cma_reserve(int order)
7446 unsigned long size, reserved, per_node;
7447 bool node_specific_cma_alloc = false;
7450 cma_reserve_called = true;
7452 if (!hugetlb_cma_size)
7455 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7456 if (hugetlb_cma_size_in_node[nid] == 0)
7459 if (!node_online(nid)) {
7460 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7461 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7462 hugetlb_cma_size_in_node[nid] = 0;
7466 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7467 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7468 nid, (PAGE_SIZE << order) / SZ_1M);
7469 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7470 hugetlb_cma_size_in_node[nid] = 0;
7472 node_specific_cma_alloc = true;
7476 /* Validate the CMA size again in case some invalid nodes specified. */
7477 if (!hugetlb_cma_size)
7480 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7481 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7482 (PAGE_SIZE << order) / SZ_1M);
7483 hugetlb_cma_size = 0;
7487 if (!node_specific_cma_alloc) {
7489 * If 3 GB area is requested on a machine with 4 numa nodes,
7490 * let's allocate 1 GB on first three nodes and ignore the last one.
7492 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7493 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7494 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7498 for_each_online_node(nid) {
7500 char name[CMA_MAX_NAME];
7502 if (node_specific_cma_alloc) {
7503 if (hugetlb_cma_size_in_node[nid] == 0)
7506 size = hugetlb_cma_size_in_node[nid];
7508 size = min(per_node, hugetlb_cma_size - reserved);
7511 size = round_up(size, PAGE_SIZE << order);
7513 snprintf(name, sizeof(name), "hugetlb%d", nid);
7515 * Note that 'order per bit' is based on smallest size that
7516 * may be returned to CMA allocator in the case of
7517 * huge page demotion.
7519 res = cma_declare_contiguous_nid(0, size, 0,
7520 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7522 &hugetlb_cma[nid], nid);
7524 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7530 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7533 if (reserved >= hugetlb_cma_size)
7539 * hugetlb_cma_size is used to determine if allocations from
7540 * cma are possible. Set to zero if no cma regions are set up.
7542 hugetlb_cma_size = 0;
7545 static void __init hugetlb_cma_check(void)
7547 if (!hugetlb_cma_size || cma_reserve_called)
7550 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7553 #endif /* CONFIG_CMA */