1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
40 #include <asm/pgalloc.h>
44 #include <linux/hugetlb.h>
45 #include <linux/hugetlb_cgroup.h>
46 #include <linux/node.h>
47 #include <linux/page_owner.h>
49 #include "hugetlb_vmemmap.h"
51 int hugetlb_max_hstate __read_mostly;
52 unsigned int default_hstate_idx;
53 struct hstate hstates[HUGE_MAX_HSTATE];
56 static struct cma *hugetlb_cma[MAX_NUMNODES];
57 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
58 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
60 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
64 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
69 static unsigned long hugetlb_cma_size __initdata;
71 __initdata LIST_HEAD(huge_boot_pages);
73 /* for command line parsing */
74 static struct hstate * __initdata parsed_hstate;
75 static unsigned long __initdata default_hstate_max_huge_pages;
76 static bool __initdata parsed_valid_hugepagesz = true;
77 static bool __initdata parsed_default_hugepagesz;
78 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
81 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
82 * free_huge_pages, and surplus_huge_pages.
84 DEFINE_SPINLOCK(hugetlb_lock);
87 * Serializes faults on the same logical page. This is used to
88 * prevent spurious OOMs when the hugepage pool is fully utilized.
90 static int num_fault_mutexes;
91 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
93 /* Forward declaration */
94 static int hugetlb_acct_memory(struct hstate *h, long delta);
95 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
96 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
97 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
98 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
99 unsigned long start, unsigned long end);
101 static inline bool subpool_is_free(struct hugepage_subpool *spool)
105 if (spool->max_hpages != -1)
106 return spool->used_hpages == 0;
107 if (spool->min_hpages != -1)
108 return spool->rsv_hpages == spool->min_hpages;
113 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
114 unsigned long irq_flags)
116 spin_unlock_irqrestore(&spool->lock, irq_flags);
118 /* If no pages are used, and no other handles to the subpool
119 * remain, give up any reservations based on minimum size and
120 * free the subpool */
121 if (subpool_is_free(spool)) {
122 if (spool->min_hpages != -1)
123 hugetlb_acct_memory(spool->hstate,
129 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
132 struct hugepage_subpool *spool;
134 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
138 spin_lock_init(&spool->lock);
140 spool->max_hpages = max_hpages;
142 spool->min_hpages = min_hpages;
144 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
148 spool->rsv_hpages = min_hpages;
153 void hugepage_put_subpool(struct hugepage_subpool *spool)
157 spin_lock_irqsave(&spool->lock, flags);
158 BUG_ON(!spool->count);
160 unlock_or_release_subpool(spool, flags);
164 * Subpool accounting for allocating and reserving pages.
165 * Return -ENOMEM if there are not enough resources to satisfy the
166 * request. Otherwise, return the number of pages by which the
167 * global pools must be adjusted (upward). The returned value may
168 * only be different than the passed value (delta) in the case where
169 * a subpool minimum size must be maintained.
171 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
179 spin_lock_irq(&spool->lock);
181 if (spool->max_hpages != -1) { /* maximum size accounting */
182 if ((spool->used_hpages + delta) <= spool->max_hpages)
183 spool->used_hpages += delta;
190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->rsv_hpages) {
192 if (delta > spool->rsv_hpages) {
194 * Asking for more reserves than those already taken on
195 * behalf of subpool. Return difference.
197 ret = delta - spool->rsv_hpages;
198 spool->rsv_hpages = 0;
200 ret = 0; /* reserves already accounted for */
201 spool->rsv_hpages -= delta;
206 spin_unlock_irq(&spool->lock);
211 * Subpool accounting for freeing and unreserving pages.
212 * Return the number of global page reservations that must be dropped.
213 * The return value may only be different than the passed value (delta)
214 * in the case where a subpool minimum size must be maintained.
216 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
225 spin_lock_irqsave(&spool->lock, flags);
227 if (spool->max_hpages != -1) /* maximum size accounting */
228 spool->used_hpages -= delta;
230 /* minimum size accounting */
231 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
232 if (spool->rsv_hpages + delta <= spool->min_hpages)
235 ret = spool->rsv_hpages + delta - spool->min_hpages;
237 spool->rsv_hpages += delta;
238 if (spool->rsv_hpages > spool->min_hpages)
239 spool->rsv_hpages = spool->min_hpages;
243 * If hugetlbfs_put_super couldn't free spool due to an outstanding
244 * quota reference, free it now.
246 unlock_or_release_subpool(spool, flags);
251 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
253 return HUGETLBFS_SB(inode->i_sb)->spool;
256 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
258 return subpool_inode(file_inode(vma->vm_file));
262 * hugetlb vma_lock helper routines
264 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
266 if (__vma_shareable_lock(vma)) {
267 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
269 down_read(&vma_lock->rw_sema);
273 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
275 if (__vma_shareable_lock(vma)) {
276 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
278 up_read(&vma_lock->rw_sema);
282 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
284 if (__vma_shareable_lock(vma)) {
285 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
287 down_write(&vma_lock->rw_sema);
291 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
293 if (__vma_shareable_lock(vma)) {
294 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
296 up_write(&vma_lock->rw_sema);
300 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
302 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
304 if (!__vma_shareable_lock(vma))
307 return down_write_trylock(&vma_lock->rw_sema);
310 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
312 if (__vma_shareable_lock(vma)) {
313 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
315 lockdep_assert_held(&vma_lock->rw_sema);
319 void hugetlb_vma_lock_release(struct kref *kref)
321 struct hugetlb_vma_lock *vma_lock = container_of(kref,
322 struct hugetlb_vma_lock, refs);
327 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
329 struct vm_area_struct *vma = vma_lock->vma;
332 * vma_lock structure may or not be released as a result of put,
333 * it certainly will no longer be attached to vma so clear pointer.
334 * Semaphore synchronizes access to vma_lock->vma field.
336 vma_lock->vma = NULL;
337 vma->vm_private_data = NULL;
338 up_write(&vma_lock->rw_sema);
339 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
342 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
344 if (__vma_shareable_lock(vma)) {
345 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
347 __hugetlb_vma_unlock_write_put(vma_lock);
351 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
354 * Only present in sharable vmas.
356 if (!vma || !__vma_shareable_lock(vma))
359 if (vma->vm_private_data) {
360 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
362 down_write(&vma_lock->rw_sema);
363 __hugetlb_vma_unlock_write_put(vma_lock);
367 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
369 struct hugetlb_vma_lock *vma_lock;
371 /* Only establish in (flags) sharable vmas */
372 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
375 /* Should never get here with non-NULL vm_private_data */
376 if (vma->vm_private_data)
379 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
382 * If we can not allocate structure, then vma can not
383 * participate in pmd sharing. This is only a possible
384 * performance enhancement and memory saving issue.
385 * However, the lock is also used to synchronize page
386 * faults with truncation. If the lock is not present,
387 * unlikely races could leave pages in a file past i_size
388 * until the file is removed. Warn in the unlikely case of
389 * allocation failure.
391 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
395 kref_init(&vma_lock->refs);
396 init_rwsem(&vma_lock->rw_sema);
398 vma->vm_private_data = vma_lock;
401 /* Helper that removes a struct file_region from the resv_map cache and returns
404 static struct file_region *
405 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
407 struct file_region *nrg;
409 VM_BUG_ON(resv->region_cache_count <= 0);
411 resv->region_cache_count--;
412 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
413 list_del(&nrg->link);
421 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
422 struct file_region *rg)
424 #ifdef CONFIG_CGROUP_HUGETLB
425 nrg->reservation_counter = rg->reservation_counter;
432 /* Helper that records hugetlb_cgroup uncharge info. */
433 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
435 struct resv_map *resv,
436 struct file_region *nrg)
438 #ifdef CONFIG_CGROUP_HUGETLB
440 nrg->reservation_counter =
441 &h_cg->rsvd_hugepage[hstate_index(h)];
442 nrg->css = &h_cg->css;
444 * The caller will hold exactly one h_cg->css reference for the
445 * whole contiguous reservation region. But this area might be
446 * scattered when there are already some file_regions reside in
447 * it. As a result, many file_regions may share only one css
448 * reference. In order to ensure that one file_region must hold
449 * exactly one h_cg->css reference, we should do css_get for
450 * each file_region and leave the reference held by caller
454 if (!resv->pages_per_hpage)
455 resv->pages_per_hpage = pages_per_huge_page(h);
456 /* pages_per_hpage should be the same for all entries in
459 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
461 nrg->reservation_counter = NULL;
467 static void put_uncharge_info(struct file_region *rg)
469 #ifdef CONFIG_CGROUP_HUGETLB
475 static bool has_same_uncharge_info(struct file_region *rg,
476 struct file_region *org)
478 #ifdef CONFIG_CGROUP_HUGETLB
479 return rg->reservation_counter == org->reservation_counter &&
487 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
489 struct file_region *nrg, *prg;
491 prg = list_prev_entry(rg, link);
492 if (&prg->link != &resv->regions && prg->to == rg->from &&
493 has_same_uncharge_info(prg, rg)) {
497 put_uncharge_info(rg);
503 nrg = list_next_entry(rg, link);
504 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
505 has_same_uncharge_info(nrg, rg)) {
506 nrg->from = rg->from;
509 put_uncharge_info(rg);
515 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
516 long to, struct hstate *h, struct hugetlb_cgroup *cg,
517 long *regions_needed)
519 struct file_region *nrg;
521 if (!regions_needed) {
522 nrg = get_file_region_entry_from_cache(map, from, to);
523 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
524 list_add(&nrg->link, rg);
525 coalesce_file_region(map, nrg);
527 *regions_needed += 1;
533 * Must be called with resv->lock held.
535 * Calling this with regions_needed != NULL will count the number of pages
536 * to be added but will not modify the linked list. And regions_needed will
537 * indicate the number of file_regions needed in the cache to carry out to add
538 * the regions for this range.
540 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
541 struct hugetlb_cgroup *h_cg,
542 struct hstate *h, long *regions_needed)
545 struct list_head *head = &resv->regions;
546 long last_accounted_offset = f;
547 struct file_region *iter, *trg = NULL;
548 struct list_head *rg = NULL;
553 /* In this loop, we essentially handle an entry for the range
554 * [last_accounted_offset, iter->from), at every iteration, with some
557 list_for_each_entry_safe(iter, trg, head, link) {
558 /* Skip irrelevant regions that start before our range. */
559 if (iter->from < f) {
560 /* If this region ends after the last accounted offset,
561 * then we need to update last_accounted_offset.
563 if (iter->to > last_accounted_offset)
564 last_accounted_offset = iter->to;
568 /* When we find a region that starts beyond our range, we've
571 if (iter->from >= t) {
572 rg = iter->link.prev;
576 /* Add an entry for last_accounted_offset -> iter->from, and
577 * update last_accounted_offset.
579 if (iter->from > last_accounted_offset)
580 add += hugetlb_resv_map_add(resv, iter->link.prev,
581 last_accounted_offset,
585 last_accounted_offset = iter->to;
588 /* Handle the case where our range extends beyond
589 * last_accounted_offset.
593 if (last_accounted_offset < t)
594 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
595 t, h, h_cg, regions_needed);
600 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
602 static int allocate_file_region_entries(struct resv_map *resv,
604 __must_hold(&resv->lock)
606 LIST_HEAD(allocated_regions);
607 int to_allocate = 0, i = 0;
608 struct file_region *trg = NULL, *rg = NULL;
610 VM_BUG_ON(regions_needed < 0);
613 * Check for sufficient descriptors in the cache to accommodate
614 * the number of in progress add operations plus regions_needed.
616 * This is a while loop because when we drop the lock, some other call
617 * to region_add or region_del may have consumed some region_entries,
618 * so we keep looping here until we finally have enough entries for
619 * (adds_in_progress + regions_needed).
621 while (resv->region_cache_count <
622 (resv->adds_in_progress + regions_needed)) {
623 to_allocate = resv->adds_in_progress + regions_needed -
624 resv->region_cache_count;
626 /* At this point, we should have enough entries in the cache
627 * for all the existing adds_in_progress. We should only be
628 * needing to allocate for regions_needed.
630 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
632 spin_unlock(&resv->lock);
633 for (i = 0; i < to_allocate; i++) {
634 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
637 list_add(&trg->link, &allocated_regions);
640 spin_lock(&resv->lock);
642 list_splice(&allocated_regions, &resv->region_cache);
643 resv->region_cache_count += to_allocate;
649 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
657 * Add the huge page range represented by [f, t) to the reserve
658 * map. Regions will be taken from the cache to fill in this range.
659 * Sufficient regions should exist in the cache due to the previous
660 * call to region_chg with the same range, but in some cases the cache will not
661 * have sufficient entries due to races with other code doing region_add or
662 * region_del. The extra needed entries will be allocated.
664 * regions_needed is the out value provided by a previous call to region_chg.
666 * Return the number of new huge pages added to the map. This number is greater
667 * than or equal to zero. If file_region entries needed to be allocated for
668 * this operation and we were not able to allocate, it returns -ENOMEM.
669 * region_add of regions of length 1 never allocate file_regions and cannot
670 * fail; region_chg will always allocate at least 1 entry and a region_add for
671 * 1 page will only require at most 1 entry.
673 static long region_add(struct resv_map *resv, long f, long t,
674 long in_regions_needed, struct hstate *h,
675 struct hugetlb_cgroup *h_cg)
677 long add = 0, actual_regions_needed = 0;
679 spin_lock(&resv->lock);
682 /* Count how many regions are actually needed to execute this add. */
683 add_reservation_in_range(resv, f, t, NULL, NULL,
684 &actual_regions_needed);
687 * Check for sufficient descriptors in the cache to accommodate
688 * this add operation. Note that actual_regions_needed may be greater
689 * than in_regions_needed, as the resv_map may have been modified since
690 * the region_chg call. In this case, we need to make sure that we
691 * allocate extra entries, such that we have enough for all the
692 * existing adds_in_progress, plus the excess needed for this
695 if (actual_regions_needed > in_regions_needed &&
696 resv->region_cache_count <
697 resv->adds_in_progress +
698 (actual_regions_needed - in_regions_needed)) {
699 /* region_add operation of range 1 should never need to
700 * allocate file_region entries.
702 VM_BUG_ON(t - f <= 1);
704 if (allocate_file_region_entries(
705 resv, actual_regions_needed - in_regions_needed)) {
712 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
714 resv->adds_in_progress -= in_regions_needed;
716 spin_unlock(&resv->lock);
721 * Examine the existing reserve map and determine how many
722 * huge pages in the specified range [f, t) are NOT currently
723 * represented. This routine is called before a subsequent
724 * call to region_add that will actually modify the reserve
725 * map to add the specified range [f, t). region_chg does
726 * not change the number of huge pages represented by the
727 * map. A number of new file_region structures is added to the cache as a
728 * placeholder, for the subsequent region_add call to use. At least 1
729 * file_region structure is added.
731 * out_regions_needed is the number of regions added to the
732 * resv->adds_in_progress. This value needs to be provided to a follow up call
733 * to region_add or region_abort for proper accounting.
735 * Returns the number of huge pages that need to be added to the existing
736 * reservation map for the range [f, t). This number is greater or equal to
737 * zero. -ENOMEM is returned if a new file_region structure or cache entry
738 * is needed and can not be allocated.
740 static long region_chg(struct resv_map *resv, long f, long t,
741 long *out_regions_needed)
745 spin_lock(&resv->lock);
747 /* Count how many hugepages in this range are NOT represented. */
748 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
751 if (*out_regions_needed == 0)
752 *out_regions_needed = 1;
754 if (allocate_file_region_entries(resv, *out_regions_needed))
757 resv->adds_in_progress += *out_regions_needed;
759 spin_unlock(&resv->lock);
764 * Abort the in progress add operation. The adds_in_progress field
765 * of the resv_map keeps track of the operations in progress between
766 * calls to region_chg and region_add. Operations are sometimes
767 * aborted after the call to region_chg. In such cases, region_abort
768 * is called to decrement the adds_in_progress counter. regions_needed
769 * is the value returned by the region_chg call, it is used to decrement
770 * the adds_in_progress counter.
772 * NOTE: The range arguments [f, t) are not needed or used in this
773 * routine. They are kept to make reading the calling code easier as
774 * arguments will match the associated region_chg call.
776 static void region_abort(struct resv_map *resv, long f, long t,
779 spin_lock(&resv->lock);
780 VM_BUG_ON(!resv->region_cache_count);
781 resv->adds_in_progress -= regions_needed;
782 spin_unlock(&resv->lock);
786 * Delete the specified range [f, t) from the reserve map. If the
787 * t parameter is LONG_MAX, this indicates that ALL regions after f
788 * should be deleted. Locate the regions which intersect [f, t)
789 * and either trim, delete or split the existing regions.
791 * Returns the number of huge pages deleted from the reserve map.
792 * In the normal case, the return value is zero or more. In the
793 * case where a region must be split, a new region descriptor must
794 * be allocated. If the allocation fails, -ENOMEM will be returned.
795 * NOTE: If the parameter t == LONG_MAX, then we will never split
796 * a region and possibly return -ENOMEM. Callers specifying
797 * t == LONG_MAX do not need to check for -ENOMEM error.
799 static long region_del(struct resv_map *resv, long f, long t)
801 struct list_head *head = &resv->regions;
802 struct file_region *rg, *trg;
803 struct file_region *nrg = NULL;
807 spin_lock(&resv->lock);
808 list_for_each_entry_safe(rg, trg, head, link) {
810 * Skip regions before the range to be deleted. file_region
811 * ranges are normally of the form [from, to). However, there
812 * may be a "placeholder" entry in the map which is of the form
813 * (from, to) with from == to. Check for placeholder entries
814 * at the beginning of the range to be deleted.
816 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
822 if (f > rg->from && t < rg->to) { /* Must split region */
824 * Check for an entry in the cache before dropping
825 * lock and attempting allocation.
828 resv->region_cache_count > resv->adds_in_progress) {
829 nrg = list_first_entry(&resv->region_cache,
832 list_del(&nrg->link);
833 resv->region_cache_count--;
837 spin_unlock(&resv->lock);
838 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
845 hugetlb_cgroup_uncharge_file_region(
846 resv, rg, t - f, false);
848 /* New entry for end of split region */
852 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
854 INIT_LIST_HEAD(&nrg->link);
856 /* Original entry is trimmed */
859 list_add(&nrg->link, &rg->link);
864 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
865 del += rg->to - rg->from;
866 hugetlb_cgroup_uncharge_file_region(resv, rg,
867 rg->to - rg->from, true);
873 if (f <= rg->from) { /* Trim beginning of region */
874 hugetlb_cgroup_uncharge_file_region(resv, rg,
875 t - rg->from, false);
879 } else { /* Trim end of region */
880 hugetlb_cgroup_uncharge_file_region(resv, rg,
888 spin_unlock(&resv->lock);
894 * A rare out of memory error was encountered which prevented removal of
895 * the reserve map region for a page. The huge page itself was free'ed
896 * and removed from the page cache. This routine will adjust the subpool
897 * usage count, and the global reserve count if needed. By incrementing
898 * these counts, the reserve map entry which could not be deleted will
899 * appear as a "reserved" entry instead of simply dangling with incorrect
902 void hugetlb_fix_reserve_counts(struct inode *inode)
904 struct hugepage_subpool *spool = subpool_inode(inode);
906 bool reserved = false;
908 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
909 if (rsv_adjust > 0) {
910 struct hstate *h = hstate_inode(inode);
912 if (!hugetlb_acct_memory(h, 1))
914 } else if (!rsv_adjust) {
919 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
923 * Count and return the number of huge pages in the reserve map
924 * that intersect with the range [f, t).
926 static long region_count(struct resv_map *resv, long f, long t)
928 struct list_head *head = &resv->regions;
929 struct file_region *rg;
932 spin_lock(&resv->lock);
933 /* Locate each segment we overlap with, and count that overlap. */
934 list_for_each_entry(rg, head, link) {
943 seg_from = max(rg->from, f);
944 seg_to = min(rg->to, t);
946 chg += seg_to - seg_from;
948 spin_unlock(&resv->lock);
954 * Convert the address within this vma to the page offset within
955 * the mapping, in pagecache page units; huge pages here.
957 static pgoff_t vma_hugecache_offset(struct hstate *h,
958 struct vm_area_struct *vma, unsigned long address)
960 return ((address - vma->vm_start) >> huge_page_shift(h)) +
961 (vma->vm_pgoff >> huge_page_order(h));
964 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
965 unsigned long address)
967 return vma_hugecache_offset(hstate_vma(vma), vma, address);
969 EXPORT_SYMBOL_GPL(linear_hugepage_index);
972 * Return the size of the pages allocated when backing a VMA. In the majority
973 * cases this will be same size as used by the page table entries.
975 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
977 if (vma->vm_ops && vma->vm_ops->pagesize)
978 return vma->vm_ops->pagesize(vma);
981 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
984 * Return the page size being used by the MMU to back a VMA. In the majority
985 * of cases, the page size used by the kernel matches the MMU size. On
986 * architectures where it differs, an architecture-specific 'strong'
987 * version of this symbol is required.
989 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
991 return vma_kernel_pagesize(vma);
995 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
996 * bits of the reservation map pointer, which are always clear due to
999 #define HPAGE_RESV_OWNER (1UL << 0)
1000 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1001 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1004 * These helpers are used to track how many pages are reserved for
1005 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1006 * is guaranteed to have their future faults succeed.
1008 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1009 * the reserve counters are updated with the hugetlb_lock held. It is safe
1010 * to reset the VMA at fork() time as it is not in use yet and there is no
1011 * chance of the global counters getting corrupted as a result of the values.
1013 * The private mapping reservation is represented in a subtly different
1014 * manner to a shared mapping. A shared mapping has a region map associated
1015 * with the underlying file, this region map represents the backing file
1016 * pages which have ever had a reservation assigned which this persists even
1017 * after the page is instantiated. A private mapping has a region map
1018 * associated with the original mmap which is attached to all VMAs which
1019 * reference it, this region map represents those offsets which have consumed
1020 * reservation ie. where pages have been instantiated.
1022 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1024 return (unsigned long)vma->vm_private_data;
1027 static void set_vma_private_data(struct vm_area_struct *vma,
1028 unsigned long value)
1030 vma->vm_private_data = (void *)value;
1034 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1035 struct hugetlb_cgroup *h_cg,
1038 #ifdef CONFIG_CGROUP_HUGETLB
1040 resv_map->reservation_counter = NULL;
1041 resv_map->pages_per_hpage = 0;
1042 resv_map->css = NULL;
1044 resv_map->reservation_counter =
1045 &h_cg->rsvd_hugepage[hstate_index(h)];
1046 resv_map->pages_per_hpage = pages_per_huge_page(h);
1047 resv_map->css = &h_cg->css;
1052 struct resv_map *resv_map_alloc(void)
1054 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1055 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1057 if (!resv_map || !rg) {
1063 kref_init(&resv_map->refs);
1064 spin_lock_init(&resv_map->lock);
1065 INIT_LIST_HEAD(&resv_map->regions);
1067 resv_map->adds_in_progress = 0;
1069 * Initialize these to 0. On shared mappings, 0's here indicate these
1070 * fields don't do cgroup accounting. On private mappings, these will be
1071 * re-initialized to the proper values, to indicate that hugetlb cgroup
1072 * reservations are to be un-charged from here.
1074 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1076 INIT_LIST_HEAD(&resv_map->region_cache);
1077 list_add(&rg->link, &resv_map->region_cache);
1078 resv_map->region_cache_count = 1;
1083 void resv_map_release(struct kref *ref)
1085 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1086 struct list_head *head = &resv_map->region_cache;
1087 struct file_region *rg, *trg;
1089 /* Clear out any active regions before we release the map. */
1090 region_del(resv_map, 0, LONG_MAX);
1092 /* ... and any entries left in the cache */
1093 list_for_each_entry_safe(rg, trg, head, link) {
1094 list_del(&rg->link);
1098 VM_BUG_ON(resv_map->adds_in_progress);
1103 static inline struct resv_map *inode_resv_map(struct inode *inode)
1106 * At inode evict time, i_mapping may not point to the original
1107 * address space within the inode. This original address space
1108 * contains the pointer to the resv_map. So, always use the
1109 * address space embedded within the inode.
1110 * The VERY common case is inode->mapping == &inode->i_data but,
1111 * this may not be true for device special inodes.
1113 return (struct resv_map *)(&inode->i_data)->private_data;
1116 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1118 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1119 if (vma->vm_flags & VM_MAYSHARE) {
1120 struct address_space *mapping = vma->vm_file->f_mapping;
1121 struct inode *inode = mapping->host;
1123 return inode_resv_map(inode);
1126 return (struct resv_map *)(get_vma_private_data(vma) &
1131 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1133 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1134 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1136 set_vma_private_data(vma, (get_vma_private_data(vma) &
1137 HPAGE_RESV_MASK) | (unsigned long)map);
1140 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1142 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1143 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1145 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1148 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1150 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1152 return (get_vma_private_data(vma) & flag) != 0;
1155 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1157 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1159 * Clear vm_private_data
1160 * - For shared mappings this is a per-vma semaphore that may be
1161 * allocated in a subsequent call to hugetlb_vm_op_open.
1162 * Before clearing, make sure pointer is not associated with vma
1163 * as this will leak the structure. This is the case when called
1164 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1165 * been called to allocate a new structure.
1166 * - For MAP_PRIVATE mappings, this is the reserve map which does
1167 * not apply to children. Faults generated by the children are
1168 * not guaranteed to succeed, even if read-only.
1170 if (vma->vm_flags & VM_MAYSHARE) {
1171 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1173 if (vma_lock && vma_lock->vma != vma)
1174 vma->vm_private_data = NULL;
1176 vma->vm_private_data = NULL;
1180 * Reset and decrement one ref on hugepage private reservation.
1181 * Called with mm->mmap_lock writer semaphore held.
1182 * This function should be only used by move_vma() and operate on
1183 * same sized vma. It should never come here with last ref on the
1186 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1189 * Clear the old hugetlb private page reservation.
1190 * It has already been transferred to new_vma.
1192 * During a mremap() operation of a hugetlb vma we call move_vma()
1193 * which copies vma into new_vma and unmaps vma. After the copy
1194 * operation both new_vma and vma share a reference to the resv_map
1195 * struct, and at that point vma is about to be unmapped. We don't
1196 * want to return the reservation to the pool at unmap of vma because
1197 * the reservation still lives on in new_vma, so simply decrement the
1198 * ref here and remove the resv_map reference from this vma.
1200 struct resv_map *reservations = vma_resv_map(vma);
1202 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1203 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1204 kref_put(&reservations->refs, resv_map_release);
1207 hugetlb_dup_vma_private(vma);
1210 /* Returns true if the VMA has associated reserve pages */
1211 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1213 if (vma->vm_flags & VM_NORESERVE) {
1215 * This address is already reserved by other process(chg == 0),
1216 * so, we should decrement reserved count. Without decrementing,
1217 * reserve count remains after releasing inode, because this
1218 * allocated page will go into page cache and is regarded as
1219 * coming from reserved pool in releasing step. Currently, we
1220 * don't have any other solution to deal with this situation
1221 * properly, so add work-around here.
1223 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1229 /* Shared mappings always use reserves */
1230 if (vma->vm_flags & VM_MAYSHARE) {
1232 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1233 * be a region map for all pages. The only situation where
1234 * there is no region map is if a hole was punched via
1235 * fallocate. In this case, there really are no reserves to
1236 * use. This situation is indicated if chg != 0.
1245 * Only the process that called mmap() has reserves for
1248 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1250 * Like the shared case above, a hole punch or truncate
1251 * could have been performed on the private mapping.
1252 * Examine the value of chg to determine if reserves
1253 * actually exist or were previously consumed.
1254 * Very Subtle - The value of chg comes from a previous
1255 * call to vma_needs_reserves(). The reserve map for
1256 * private mappings has different (opposite) semantics
1257 * than that of shared mappings. vma_needs_reserves()
1258 * has already taken this difference in semantics into
1259 * account. Therefore, the meaning of chg is the same
1260 * as in the shared case above. Code could easily be
1261 * combined, but keeping it separate draws attention to
1262 * subtle differences.
1273 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1275 int nid = folio_nid(folio);
1277 lockdep_assert_held(&hugetlb_lock);
1278 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1280 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1281 h->free_huge_pages++;
1282 h->free_huge_pages_node[nid]++;
1283 folio_set_hugetlb_freed(folio);
1286 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1289 struct folio *folio;
1290 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1292 lockdep_assert_held(&hugetlb_lock);
1293 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1294 if (pin && !folio_is_longterm_pinnable(folio))
1297 if (folio_test_hwpoison(folio))
1300 list_move(&folio->lru, &h->hugepage_activelist);
1301 folio_ref_unfreeze(folio, 1);
1302 folio_clear_hugetlb_freed(folio);
1303 h->free_huge_pages--;
1304 h->free_huge_pages_node[nid]--;
1311 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1312 int nid, nodemask_t *nmask)
1314 unsigned int cpuset_mems_cookie;
1315 struct zonelist *zonelist;
1318 int node = NUMA_NO_NODE;
1320 zonelist = node_zonelist(nid, gfp_mask);
1323 cpuset_mems_cookie = read_mems_allowed_begin();
1324 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1325 struct folio *folio;
1327 if (!cpuset_zone_allowed(zone, gfp_mask))
1330 * no need to ask again on the same node. Pool is node rather than
1333 if (zone_to_nid(zone) == node)
1335 node = zone_to_nid(zone);
1337 folio = dequeue_hugetlb_folio_node_exact(h, node);
1341 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1347 static unsigned long available_huge_pages(struct hstate *h)
1349 return h->free_huge_pages - h->resv_huge_pages;
1352 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1353 struct vm_area_struct *vma,
1354 unsigned long address, int avoid_reserve,
1357 struct folio *folio = NULL;
1358 struct mempolicy *mpol;
1360 nodemask_t *nodemask;
1364 * A child process with MAP_PRIVATE mappings created by their parent
1365 * have no page reserves. This check ensures that reservations are
1366 * not "stolen". The child may still get SIGKILLed
1368 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1371 /* If reserves cannot be used, ensure enough pages are in the pool */
1372 if (avoid_reserve && !available_huge_pages(h))
1375 gfp_mask = htlb_alloc_mask(h);
1376 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1378 if (mpol_is_preferred_many(mpol)) {
1379 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1382 /* Fallback to all nodes if page==NULL */
1387 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1390 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1391 folio_set_hugetlb_restore_reserve(folio);
1392 h->resv_huge_pages--;
1395 mpol_cond_put(mpol);
1403 * common helper functions for hstate_next_node_to_{alloc|free}.
1404 * We may have allocated or freed a huge page based on a different
1405 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1406 * be outside of *nodes_allowed. Ensure that we use an allowed
1407 * node for alloc or free.
1409 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1411 nid = next_node_in(nid, *nodes_allowed);
1412 VM_BUG_ON(nid >= MAX_NUMNODES);
1417 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1419 if (!node_isset(nid, *nodes_allowed))
1420 nid = next_node_allowed(nid, nodes_allowed);
1425 * returns the previously saved node ["this node"] from which to
1426 * allocate a persistent huge page for the pool and advance the
1427 * next node from which to allocate, handling wrap at end of node
1430 static int hstate_next_node_to_alloc(struct hstate *h,
1431 nodemask_t *nodes_allowed)
1435 VM_BUG_ON(!nodes_allowed);
1437 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1438 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1444 * helper for remove_pool_huge_page() - return the previously saved
1445 * node ["this node"] from which to free a huge page. Advance the
1446 * next node id whether or not we find a free huge page to free so
1447 * that the next attempt to free addresses the next node.
1449 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1453 VM_BUG_ON(!nodes_allowed);
1455 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1456 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1461 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1462 for (nr_nodes = nodes_weight(*mask); \
1464 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1467 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1468 for (nr_nodes = nodes_weight(*mask); \
1470 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1473 /* used to demote non-gigantic_huge pages as well */
1474 static void __destroy_compound_gigantic_folio(struct folio *folio,
1475 unsigned int order, bool demote)
1478 int nr_pages = 1 << order;
1481 atomic_set(&folio->_entire_mapcount, 0);
1482 atomic_set(&folio->_nr_pages_mapped, 0);
1483 atomic_set(&folio->_pincount, 0);
1485 for (i = 1; i < nr_pages; i++) {
1486 p = folio_page(folio, i);
1487 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1489 clear_compound_head(p);
1491 set_page_refcounted(p);
1494 __folio_clear_head(folio);
1497 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1500 __destroy_compound_gigantic_folio(folio, order, true);
1503 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1504 static void destroy_compound_gigantic_folio(struct folio *folio,
1507 __destroy_compound_gigantic_folio(folio, order, false);
1510 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1513 * If the page isn't allocated using the cma allocator,
1514 * cma_release() returns false.
1517 int nid = folio_nid(folio);
1519 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1523 free_contig_range(folio_pfn(folio), 1 << order);
1526 #ifdef CONFIG_CONTIG_ALLOC
1527 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1528 int nid, nodemask_t *nodemask)
1531 unsigned long nr_pages = pages_per_huge_page(h);
1532 if (nid == NUMA_NO_NODE)
1533 nid = numa_mem_id();
1539 if (hugetlb_cma[nid]) {
1540 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1541 huge_page_order(h), true);
1543 return page_folio(page);
1546 if (!(gfp_mask & __GFP_THISNODE)) {
1547 for_each_node_mask(node, *nodemask) {
1548 if (node == nid || !hugetlb_cma[node])
1551 page = cma_alloc(hugetlb_cma[node], nr_pages,
1552 huge_page_order(h), true);
1554 return page_folio(page);
1560 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1561 return page ? page_folio(page) : NULL;
1564 #else /* !CONFIG_CONTIG_ALLOC */
1565 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1566 int nid, nodemask_t *nodemask)
1570 #endif /* CONFIG_CONTIG_ALLOC */
1572 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1573 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1574 int nid, nodemask_t *nodemask)
1578 static inline void free_gigantic_folio(struct folio *folio,
1579 unsigned int order) { }
1580 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1581 unsigned int order) { }
1584 static inline void __clear_hugetlb_destructor(struct hstate *h,
1585 struct folio *folio)
1587 lockdep_assert_held(&hugetlb_lock);
1589 folio_clear_hugetlb(folio);
1593 * Remove hugetlb folio from lists.
1594 * If vmemmap exists for the folio, update dtor so that the folio appears
1595 * as just a compound page. Otherwise, wait until after allocating vmemmap
1598 * A reference is held on the folio, except in the case of demote.
1600 * Must be called with hugetlb lock held.
1602 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1603 bool adjust_surplus,
1606 int nid = folio_nid(folio);
1608 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1609 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1611 lockdep_assert_held(&hugetlb_lock);
1612 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1615 list_del(&folio->lru);
1617 if (folio_test_hugetlb_freed(folio)) {
1618 h->free_huge_pages--;
1619 h->free_huge_pages_node[nid]--;
1621 if (adjust_surplus) {
1622 h->surplus_huge_pages--;
1623 h->surplus_huge_pages_node[nid]--;
1627 * We can only clear the hugetlb destructor after allocating vmemmap
1628 * pages. Otherwise, someone (memory error handling) may try to write
1629 * to tail struct pages.
1631 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1632 __clear_hugetlb_destructor(h, folio);
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);
1642 h->nr_huge_pages_node[nid]--;
1645 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1646 bool adjust_surplus)
1648 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1651 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1652 bool adjust_surplus)
1654 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1657 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1658 bool adjust_surplus)
1661 int nid = folio_nid(folio);
1663 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1665 lockdep_assert_held(&hugetlb_lock);
1667 INIT_LIST_HEAD(&folio->lru);
1669 h->nr_huge_pages_node[nid]++;
1671 if (adjust_surplus) {
1672 h->surplus_huge_pages++;
1673 h->surplus_huge_pages_node[nid]++;
1676 folio_set_hugetlb(folio);
1677 folio_change_private(folio, NULL);
1679 * We have to set hugetlb_vmemmap_optimized again as above
1680 * folio_change_private(folio, NULL) cleared it.
1682 folio_set_hugetlb_vmemmap_optimized(folio);
1685 * This folio is about to be managed by the hugetlb allocator and
1686 * should have no users. Drop our reference, and check for others
1689 zeroed = folio_put_testzero(folio);
1690 if (unlikely(!zeroed))
1692 * It is VERY unlikely soneone else has taken a ref
1693 * on the folio. In this case, we simply return as
1694 * free_huge_folio() will be called when this other ref
1699 arch_clear_hugepage_flags(&folio->page);
1700 enqueue_hugetlb_folio(h, folio);
1703 static void __update_and_free_hugetlb_folio(struct hstate *h,
1704 struct folio *folio)
1706 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1708 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1712 * If we don't know which subpages are hwpoisoned, we can't free
1713 * the hugepage, so it's leaked intentionally.
1715 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1718 if (hugetlb_vmemmap_restore(h, &folio->page)) {
1719 spin_lock_irq(&hugetlb_lock);
1721 * If we cannot allocate vmemmap pages, just refuse to free the
1722 * page and put the page back on the hugetlb free list and treat
1723 * as a surplus page.
1725 add_hugetlb_folio(h, folio, true);
1726 spin_unlock_irq(&hugetlb_lock);
1731 * Move PageHWPoison flag from head page to the raw error pages,
1732 * which makes any healthy subpages reusable.
1734 if (unlikely(folio_test_hwpoison(folio)))
1735 folio_clear_hugetlb_hwpoison(folio);
1738 * If vmemmap pages were allocated above, then we need to clear the
1739 * hugetlb destructor under the hugetlb lock.
1742 spin_lock_irq(&hugetlb_lock);
1743 __clear_hugetlb_destructor(h, folio);
1744 spin_unlock_irq(&hugetlb_lock);
1748 * Non-gigantic pages demoted from CMA allocated gigantic pages
1749 * need to be given back to CMA in free_gigantic_folio.
1751 if (hstate_is_gigantic(h) ||
1752 hugetlb_cma_folio(folio, huge_page_order(h))) {
1753 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1754 free_gigantic_folio(folio, huge_page_order(h));
1756 __free_pages(&folio->page, huge_page_order(h));
1761 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1762 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1763 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1764 * the vmemmap pages.
1766 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1767 * freed and frees them one-by-one. As the page->mapping pointer is going
1768 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1769 * structure of a lockless linked list of huge pages to be freed.
1771 static LLIST_HEAD(hpage_freelist);
1773 static void free_hpage_workfn(struct work_struct *work)
1775 struct llist_node *node;
1777 node = llist_del_all(&hpage_freelist);
1783 page = container_of((struct address_space **)node,
1784 struct page, mapping);
1786 page->mapping = NULL;
1788 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1789 * folio_hstate() is going to trigger because a previous call to
1790 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1791 * not use folio_hstate() directly.
1793 h = size_to_hstate(page_size(page));
1795 __update_and_free_hugetlb_folio(h, page_folio(page));
1800 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1802 static inline void flush_free_hpage_work(struct hstate *h)
1804 if (hugetlb_vmemmap_optimizable(h))
1805 flush_work(&free_hpage_work);
1808 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1811 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1812 __update_and_free_hugetlb_folio(h, folio);
1817 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1819 * Only call schedule_work() if hpage_freelist is previously
1820 * empty. Otherwise, schedule_work() had been called but the workfn
1821 * hasn't retrieved the list yet.
1823 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1824 schedule_work(&free_hpage_work);
1827 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1829 struct page *page, *t_page;
1830 struct folio *folio;
1832 list_for_each_entry_safe(page, t_page, list, lru) {
1833 folio = page_folio(page);
1834 update_and_free_hugetlb_folio(h, folio, false);
1839 struct hstate *size_to_hstate(unsigned long size)
1843 for_each_hstate(h) {
1844 if (huge_page_size(h) == size)
1850 void free_huge_folio(struct folio *folio)
1853 * Can't pass hstate in here because it is called from the
1854 * compound page destructor.
1856 struct hstate *h = folio_hstate(folio);
1857 int nid = folio_nid(folio);
1858 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1859 bool restore_reserve;
1860 unsigned long flags;
1862 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1863 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1865 hugetlb_set_folio_subpool(folio, NULL);
1866 if (folio_test_anon(folio))
1867 __ClearPageAnonExclusive(&folio->page);
1868 folio->mapping = NULL;
1869 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1870 folio_clear_hugetlb_restore_reserve(folio);
1873 * If HPageRestoreReserve was set on page, page allocation consumed a
1874 * reservation. If the page was associated with a subpool, there
1875 * would have been a page reserved in the subpool before allocation
1876 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1877 * reservation, do not call hugepage_subpool_put_pages() as this will
1878 * remove the reserved page from the subpool.
1880 if (!restore_reserve) {
1882 * A return code of zero implies that the subpool will be
1883 * under its minimum size if the reservation is not restored
1884 * after page is free. Therefore, force restore_reserve
1887 if (hugepage_subpool_put_pages(spool, 1) == 0)
1888 restore_reserve = true;
1891 spin_lock_irqsave(&hugetlb_lock, flags);
1892 folio_clear_hugetlb_migratable(folio);
1893 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1894 pages_per_huge_page(h), folio);
1895 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1896 pages_per_huge_page(h), folio);
1897 if (restore_reserve)
1898 h->resv_huge_pages++;
1900 if (folio_test_hugetlb_temporary(folio)) {
1901 remove_hugetlb_folio(h, folio, false);
1902 spin_unlock_irqrestore(&hugetlb_lock, flags);
1903 update_and_free_hugetlb_folio(h, folio, true);
1904 } else if (h->surplus_huge_pages_node[nid]) {
1905 /* remove the page from active list */
1906 remove_hugetlb_folio(h, folio, true);
1907 spin_unlock_irqrestore(&hugetlb_lock, flags);
1908 update_and_free_hugetlb_folio(h, folio, true);
1910 arch_clear_hugepage_flags(&folio->page);
1911 enqueue_hugetlb_folio(h, folio);
1912 spin_unlock_irqrestore(&hugetlb_lock, flags);
1917 * Must be called with the hugetlb lock held
1919 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1921 lockdep_assert_held(&hugetlb_lock);
1923 h->nr_huge_pages_node[nid]++;
1926 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1928 hugetlb_vmemmap_optimize(h, &folio->page);
1929 INIT_LIST_HEAD(&folio->lru);
1930 folio_set_hugetlb(folio);
1931 hugetlb_set_folio_subpool(folio, NULL);
1932 set_hugetlb_cgroup(folio, NULL);
1933 set_hugetlb_cgroup_rsvd(folio, NULL);
1936 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1938 __prep_new_hugetlb_folio(h, folio);
1939 spin_lock_irq(&hugetlb_lock);
1940 __prep_account_new_huge_page(h, nid);
1941 spin_unlock_irq(&hugetlb_lock);
1944 static bool __prep_compound_gigantic_folio(struct folio *folio,
1945 unsigned int order, bool demote)
1948 int nr_pages = 1 << order;
1951 __folio_clear_reserved(folio);
1952 for (i = 0; i < nr_pages; i++) {
1953 p = folio_page(folio, i);
1956 * For gigantic hugepages allocated through bootmem at
1957 * boot, it's safer to be consistent with the not-gigantic
1958 * hugepages and clear the PG_reserved bit from all tail pages
1959 * too. Otherwise drivers using get_user_pages() to access tail
1960 * pages may get the reference counting wrong if they see
1961 * PG_reserved set on a tail page (despite the head page not
1962 * having PG_reserved set). Enforcing this consistency between
1963 * head and tail pages allows drivers to optimize away a check
1964 * on the head page when they need know if put_page() is needed
1965 * after get_user_pages().
1967 if (i != 0) /* head page cleared above */
1968 __ClearPageReserved(p);
1970 * Subtle and very unlikely
1972 * Gigantic 'page allocators' such as memblock or cma will
1973 * return a set of pages with each page ref counted. We need
1974 * to turn this set of pages into a compound page with tail
1975 * page ref counts set to zero. Code such as speculative page
1976 * cache adding could take a ref on a 'to be' tail page.
1977 * We need to respect any increased ref count, and only set
1978 * the ref count to zero if count is currently 1. If count
1979 * is not 1, we return an error. An error return indicates
1980 * the set of pages can not be converted to a gigantic page.
1981 * The caller who allocated the pages should then discard the
1982 * pages using the appropriate free interface.
1984 * In the case of demote, the ref count will be zero.
1987 if (!page_ref_freeze(p, 1)) {
1988 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1992 VM_BUG_ON_PAGE(page_count(p), p);
1995 set_compound_head(p, &folio->page);
1997 __folio_set_head(folio);
1998 /* we rely on prep_new_hugetlb_folio to set the destructor */
1999 folio_set_order(folio, order);
2000 atomic_set(&folio->_entire_mapcount, -1);
2001 atomic_set(&folio->_nr_pages_mapped, 0);
2002 atomic_set(&folio->_pincount, 0);
2006 /* undo page modifications made above */
2007 for (j = 0; j < i; j++) {
2008 p = folio_page(folio, j);
2010 clear_compound_head(p);
2011 set_page_refcounted(p);
2013 /* need to clear PG_reserved on remaining tail pages */
2014 for (; j < nr_pages; j++) {
2015 p = folio_page(folio, j);
2016 __ClearPageReserved(p);
2021 static bool prep_compound_gigantic_folio(struct folio *folio,
2024 return __prep_compound_gigantic_folio(folio, order, false);
2027 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2030 return __prep_compound_gigantic_folio(folio, order, true);
2034 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2035 * transparent huge pages. See the PageTransHuge() documentation for more
2038 int PageHuge(struct page *page)
2040 struct folio *folio;
2042 if (!PageCompound(page))
2044 folio = page_folio(page);
2045 return folio_test_hugetlb(folio);
2047 EXPORT_SYMBOL_GPL(PageHuge);
2050 * Find and lock address space (mapping) in write mode.
2052 * Upon entry, the page is locked which means that page_mapping() is
2053 * stable. Due to locking order, we can only trylock_write. If we can
2054 * not get the lock, simply return NULL to caller.
2056 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2058 struct address_space *mapping = page_mapping(hpage);
2063 if (i_mmap_trylock_write(mapping))
2069 pgoff_t hugetlb_basepage_index(struct page *page)
2071 struct page *page_head = compound_head(page);
2072 pgoff_t index = page_index(page_head);
2073 unsigned long compound_idx;
2075 if (compound_order(page_head) > MAX_ORDER)
2076 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2078 compound_idx = page - page_head;
2080 return (index << compound_order(page_head)) + compound_idx;
2083 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2084 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2085 nodemask_t *node_alloc_noretry)
2087 int order = huge_page_order(h);
2089 bool alloc_try_hard = true;
2093 * By default we always try hard to allocate the page with
2094 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2095 * a loop (to adjust global huge page counts) and previous allocation
2096 * failed, do not continue to try hard on the same node. Use the
2097 * node_alloc_noretry bitmap to manage this state information.
2099 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2100 alloc_try_hard = false;
2101 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2103 gfp_mask |= __GFP_RETRY_MAYFAIL;
2104 if (nid == NUMA_NO_NODE)
2105 nid = numa_mem_id();
2107 page = __alloc_pages(gfp_mask, order, nid, nmask);
2109 /* Freeze head page */
2110 if (page && !page_ref_freeze(page, 1)) {
2111 __free_pages(page, order);
2112 if (retry) { /* retry once */
2116 /* WOW! twice in a row. */
2117 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2122 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2123 * indicates an overall state change. Clear bit so that we resume
2124 * normal 'try hard' allocations.
2126 if (node_alloc_noretry && page && !alloc_try_hard)
2127 node_clear(nid, *node_alloc_noretry);
2130 * If we tried hard to get a page but failed, set bit so that
2131 * subsequent attempts will not try as hard until there is an
2132 * overall state change.
2134 if (node_alloc_noretry && !page && alloc_try_hard)
2135 node_set(nid, *node_alloc_noretry);
2138 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2142 __count_vm_event(HTLB_BUDDY_PGALLOC);
2143 return page_folio(page);
2147 * Common helper to allocate a fresh hugetlb page. All specific allocators
2148 * should use this function to get new hugetlb pages
2150 * Note that returned page is 'frozen': ref count of head page and all tail
2153 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2154 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2155 nodemask_t *node_alloc_noretry)
2157 struct folio *folio;
2161 if (hstate_is_gigantic(h))
2162 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2164 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2165 nid, nmask, node_alloc_noretry);
2168 if (hstate_is_gigantic(h)) {
2169 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2171 * Rare failure to convert pages to compound page.
2172 * Free pages and try again - ONCE!
2174 free_gigantic_folio(folio, huge_page_order(h));
2182 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2188 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2191 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2192 nodemask_t *node_alloc_noretry)
2194 struct folio *folio;
2196 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2198 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2199 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2200 nodes_allowed, node_alloc_noretry);
2202 free_huge_folio(folio); /* free it into the hugepage allocator */
2211 * Remove huge page from pool from next node to free. Attempt to keep
2212 * persistent huge pages more or less balanced over allowed nodes.
2213 * This routine only 'removes' the hugetlb page. The caller must make
2214 * an additional call to free the page to low level allocators.
2215 * Called with hugetlb_lock locked.
2217 static struct page *remove_pool_huge_page(struct hstate *h,
2218 nodemask_t *nodes_allowed,
2222 struct page *page = NULL;
2223 struct folio *folio;
2225 lockdep_assert_held(&hugetlb_lock);
2226 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2228 * If we're returning unused surplus pages, only examine
2229 * nodes with surplus pages.
2231 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2232 !list_empty(&h->hugepage_freelists[node])) {
2233 page = list_entry(h->hugepage_freelists[node].next,
2235 folio = page_folio(page);
2236 remove_hugetlb_folio(h, folio, acct_surplus);
2245 * Dissolve a given free hugepage into free buddy pages. This function does
2246 * nothing for in-use hugepages and non-hugepages.
2247 * This function returns values like below:
2249 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2250 * when the system is under memory pressure and the feature of
2251 * freeing unused vmemmap pages associated with each hugetlb page
2253 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2254 * (allocated or reserved.)
2255 * 0: successfully dissolved free hugepages or the page is not a
2256 * hugepage (considered as already dissolved)
2258 int dissolve_free_huge_page(struct page *page)
2261 struct folio *folio = page_folio(page);
2264 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2265 if (!folio_test_hugetlb(folio))
2268 spin_lock_irq(&hugetlb_lock);
2269 if (!folio_test_hugetlb(folio)) {
2274 if (!folio_ref_count(folio)) {
2275 struct hstate *h = folio_hstate(folio);
2276 if (!available_huge_pages(h))
2280 * We should make sure that the page is already on the free list
2281 * when it is dissolved.
2283 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2284 spin_unlock_irq(&hugetlb_lock);
2288 * Theoretically, we should return -EBUSY when we
2289 * encounter this race. In fact, we have a chance
2290 * to successfully dissolve the page if we do a
2291 * retry. Because the race window is quite small.
2292 * If we seize this opportunity, it is an optimization
2293 * for increasing the success rate of dissolving page.
2298 remove_hugetlb_folio(h, folio, false);
2299 h->max_huge_pages--;
2300 spin_unlock_irq(&hugetlb_lock);
2303 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2304 * before freeing the page. update_and_free_hugtlb_folio will fail to
2305 * free the page if it can not allocate required vmemmap. We
2306 * need to adjust max_huge_pages if the page is not freed.
2307 * Attempt to allocate vmemmmap here so that we can take
2308 * appropriate action on failure.
2310 rc = hugetlb_vmemmap_restore(h, &folio->page);
2312 update_and_free_hugetlb_folio(h, folio, false);
2314 spin_lock_irq(&hugetlb_lock);
2315 add_hugetlb_folio(h, folio, false);
2316 h->max_huge_pages++;
2317 spin_unlock_irq(&hugetlb_lock);
2323 spin_unlock_irq(&hugetlb_lock);
2328 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2329 * make specified memory blocks removable from the system.
2330 * Note that this will dissolve a free gigantic hugepage completely, if any
2331 * part of it lies within the given range.
2332 * Also note that if dissolve_free_huge_page() returns with an error, all
2333 * free hugepages that were dissolved before that error are lost.
2335 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2343 if (!hugepages_supported())
2346 order = huge_page_order(&default_hstate);
2348 order = min(order, huge_page_order(h));
2350 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2351 page = pfn_to_page(pfn);
2352 rc = dissolve_free_huge_page(page);
2361 * Allocates a fresh surplus page from the page allocator.
2363 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2364 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2366 struct folio *folio = NULL;
2368 if (hstate_is_gigantic(h))
2371 spin_lock_irq(&hugetlb_lock);
2372 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2374 spin_unlock_irq(&hugetlb_lock);
2376 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2380 spin_lock_irq(&hugetlb_lock);
2382 * We could have raced with the pool size change.
2383 * Double check that and simply deallocate the new page
2384 * if we would end up overcommiting the surpluses. Abuse
2385 * temporary page to workaround the nasty free_huge_folio
2388 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2389 folio_set_hugetlb_temporary(folio);
2390 spin_unlock_irq(&hugetlb_lock);
2391 free_huge_folio(folio);
2395 h->surplus_huge_pages++;
2396 h->surplus_huge_pages_node[folio_nid(folio)]++;
2399 spin_unlock_irq(&hugetlb_lock);
2404 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2405 int nid, nodemask_t *nmask)
2407 struct folio *folio;
2409 if (hstate_is_gigantic(h))
2412 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2416 /* fresh huge pages are frozen */
2417 folio_ref_unfreeze(folio, 1);
2419 * We do not account these pages as surplus because they are only
2420 * temporary and will be released properly on the last reference
2422 folio_set_hugetlb_temporary(folio);
2428 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2431 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2432 struct vm_area_struct *vma, unsigned long addr)
2434 struct folio *folio = NULL;
2435 struct mempolicy *mpol;
2436 gfp_t gfp_mask = htlb_alloc_mask(h);
2438 nodemask_t *nodemask;
2440 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2441 if (mpol_is_preferred_many(mpol)) {
2442 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2444 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2445 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2447 /* Fallback to all nodes if page==NULL */
2452 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2453 mpol_cond_put(mpol);
2457 /* folio migration callback function */
2458 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2459 nodemask_t *nmask, gfp_t gfp_mask)
2461 spin_lock_irq(&hugetlb_lock);
2462 if (available_huge_pages(h)) {
2463 struct folio *folio;
2465 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2466 preferred_nid, nmask);
2468 spin_unlock_irq(&hugetlb_lock);
2472 spin_unlock_irq(&hugetlb_lock);
2474 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2477 /* mempolicy aware migration callback */
2478 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2479 unsigned long address)
2481 struct mempolicy *mpol;
2482 nodemask_t *nodemask;
2483 struct folio *folio;
2487 gfp_mask = htlb_alloc_mask(h);
2488 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2489 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2490 mpol_cond_put(mpol);
2496 * Increase the hugetlb pool such that it can accommodate a reservation
2499 static int gather_surplus_pages(struct hstate *h, long delta)
2500 __must_hold(&hugetlb_lock)
2502 LIST_HEAD(surplus_list);
2503 struct folio *folio, *tmp;
2506 long needed, allocated;
2507 bool alloc_ok = true;
2509 lockdep_assert_held(&hugetlb_lock);
2510 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2512 h->resv_huge_pages += delta;
2520 spin_unlock_irq(&hugetlb_lock);
2521 for (i = 0; i < needed; i++) {
2522 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2523 NUMA_NO_NODE, NULL);
2528 list_add(&folio->lru, &surplus_list);
2534 * After retaking hugetlb_lock, we need to recalculate 'needed'
2535 * because either resv_huge_pages or free_huge_pages may have changed.
2537 spin_lock_irq(&hugetlb_lock);
2538 needed = (h->resv_huge_pages + delta) -
2539 (h->free_huge_pages + allocated);
2544 * We were not able to allocate enough pages to
2545 * satisfy the entire reservation so we free what
2546 * we've allocated so far.
2551 * The surplus_list now contains _at_least_ the number of extra pages
2552 * needed to accommodate the reservation. Add the appropriate number
2553 * of pages to the hugetlb pool and free the extras back to the buddy
2554 * allocator. Commit the entire reservation here to prevent another
2555 * process from stealing the pages as they are added to the pool but
2556 * before they are reserved.
2558 needed += allocated;
2559 h->resv_huge_pages += delta;
2562 /* Free the needed pages to the hugetlb pool */
2563 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2566 /* Add the page to the hugetlb allocator */
2567 enqueue_hugetlb_folio(h, folio);
2570 spin_unlock_irq(&hugetlb_lock);
2573 * Free unnecessary surplus pages to the buddy allocator.
2574 * Pages have no ref count, call free_huge_folio directly.
2576 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2577 free_huge_folio(folio);
2578 spin_lock_irq(&hugetlb_lock);
2584 * This routine has two main purposes:
2585 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2586 * in unused_resv_pages. This corresponds to the prior adjustments made
2587 * to the associated reservation map.
2588 * 2) Free any unused surplus pages that may have been allocated to satisfy
2589 * the reservation. As many as unused_resv_pages may be freed.
2591 static void return_unused_surplus_pages(struct hstate *h,
2592 unsigned long unused_resv_pages)
2594 unsigned long nr_pages;
2596 LIST_HEAD(page_list);
2598 lockdep_assert_held(&hugetlb_lock);
2599 /* Uncommit the reservation */
2600 h->resv_huge_pages -= unused_resv_pages;
2602 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2606 * Part (or even all) of the reservation could have been backed
2607 * by pre-allocated pages. Only free surplus pages.
2609 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2612 * We want to release as many surplus pages as possible, spread
2613 * evenly across all nodes with memory. Iterate across these nodes
2614 * until we can no longer free unreserved surplus pages. This occurs
2615 * when the nodes with surplus pages have no free pages.
2616 * remove_pool_huge_page() will balance the freed pages across the
2617 * on-line nodes with memory and will handle the hstate accounting.
2619 while (nr_pages--) {
2620 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2624 list_add(&page->lru, &page_list);
2628 spin_unlock_irq(&hugetlb_lock);
2629 update_and_free_pages_bulk(h, &page_list);
2630 spin_lock_irq(&hugetlb_lock);
2635 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2636 * are used by the huge page allocation routines to manage reservations.
2638 * vma_needs_reservation is called to determine if the huge page at addr
2639 * within the vma has an associated reservation. If a reservation is
2640 * needed, the value 1 is returned. The caller is then responsible for
2641 * managing the global reservation and subpool usage counts. After
2642 * the huge page has been allocated, vma_commit_reservation is called
2643 * to add the page to the reservation map. If the page allocation fails,
2644 * the reservation must be ended instead of committed. vma_end_reservation
2645 * is called in such cases.
2647 * In the normal case, vma_commit_reservation returns the same value
2648 * as the preceding vma_needs_reservation call. The only time this
2649 * is not the case is if a reserve map was changed between calls. It
2650 * is the responsibility of the caller to notice the difference and
2651 * take appropriate action.
2653 * vma_add_reservation is used in error paths where a reservation must
2654 * be restored when a newly allocated huge page must be freed. It is
2655 * to be called after calling vma_needs_reservation to determine if a
2656 * reservation exists.
2658 * vma_del_reservation is used in error paths where an entry in the reserve
2659 * map was created during huge page allocation and must be removed. It is to
2660 * be called after calling vma_needs_reservation to determine if a reservation
2663 enum vma_resv_mode {
2670 static long __vma_reservation_common(struct hstate *h,
2671 struct vm_area_struct *vma, unsigned long addr,
2672 enum vma_resv_mode mode)
2674 struct resv_map *resv;
2677 long dummy_out_regions_needed;
2679 resv = vma_resv_map(vma);
2683 idx = vma_hugecache_offset(h, vma, addr);
2685 case VMA_NEEDS_RESV:
2686 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2687 /* We assume that vma_reservation_* routines always operate on
2688 * 1 page, and that adding to resv map a 1 page entry can only
2689 * ever require 1 region.
2691 VM_BUG_ON(dummy_out_regions_needed != 1);
2693 case VMA_COMMIT_RESV:
2694 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2695 /* region_add calls of range 1 should never fail. */
2699 region_abort(resv, idx, idx + 1, 1);
2703 if (vma->vm_flags & VM_MAYSHARE) {
2704 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2705 /* region_add calls of range 1 should never fail. */
2708 region_abort(resv, idx, idx + 1, 1);
2709 ret = region_del(resv, idx, idx + 1);
2713 if (vma->vm_flags & VM_MAYSHARE) {
2714 region_abort(resv, idx, idx + 1, 1);
2715 ret = region_del(resv, idx, idx + 1);
2717 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2718 /* region_add calls of range 1 should never fail. */
2726 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2729 * We know private mapping must have HPAGE_RESV_OWNER set.
2731 * In most cases, reserves always exist for private mappings.
2732 * However, a file associated with mapping could have been
2733 * hole punched or truncated after reserves were consumed.
2734 * As subsequent fault on such a range will not use reserves.
2735 * Subtle - The reserve map for private mappings has the
2736 * opposite meaning than that of shared mappings. If NO
2737 * entry is in the reserve map, it means a reservation exists.
2738 * If an entry exists in the reserve map, it means the
2739 * reservation has already been consumed. As a result, the
2740 * return value of this routine is the opposite of the
2741 * value returned from reserve map manipulation routines above.
2750 static long vma_needs_reservation(struct hstate *h,
2751 struct vm_area_struct *vma, unsigned long addr)
2753 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2756 static long vma_commit_reservation(struct hstate *h,
2757 struct vm_area_struct *vma, unsigned long addr)
2759 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2762 static void vma_end_reservation(struct hstate *h,
2763 struct vm_area_struct *vma, unsigned long addr)
2765 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2768 static long vma_add_reservation(struct hstate *h,
2769 struct vm_area_struct *vma, unsigned long addr)
2771 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2774 static long vma_del_reservation(struct hstate *h,
2775 struct vm_area_struct *vma, unsigned long addr)
2777 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2781 * This routine is called to restore reservation information on error paths.
2782 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2783 * and the hugetlb mutex should remain held when calling this routine.
2785 * It handles two specific cases:
2786 * 1) A reservation was in place and the folio consumed the reservation.
2787 * hugetlb_restore_reserve is set in the folio.
2788 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2789 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2791 * In case 1, free_huge_folio later in the error path will increment the
2792 * global reserve count. But, free_huge_folio does not have enough context
2793 * to adjust the reservation map. This case deals primarily with private
2794 * mappings. Adjust the reserve map here to be consistent with global
2795 * reserve count adjustments to be made by free_huge_folio. Make sure the
2796 * reserve map indicates there is a reservation present.
2798 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2800 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2801 unsigned long address, struct folio *folio)
2803 long rc = vma_needs_reservation(h, vma, address);
2805 if (folio_test_hugetlb_restore_reserve(folio)) {
2806 if (unlikely(rc < 0))
2808 * Rare out of memory condition in reserve map
2809 * manipulation. Clear hugetlb_restore_reserve so
2810 * that global reserve count will not be incremented
2811 * by free_huge_folio. This will make it appear
2812 * as though the reservation for this folio was
2813 * consumed. This may prevent the task from
2814 * faulting in the folio at a later time. This
2815 * is better than inconsistent global huge page
2816 * accounting of reserve counts.
2818 folio_clear_hugetlb_restore_reserve(folio);
2820 (void)vma_add_reservation(h, vma, address);
2822 vma_end_reservation(h, vma, address);
2826 * This indicates there is an entry in the reserve map
2827 * not added by alloc_hugetlb_folio. We know it was added
2828 * before the alloc_hugetlb_folio call, otherwise
2829 * hugetlb_restore_reserve would be set on the folio.
2830 * Remove the entry so that a subsequent allocation
2831 * does not consume a reservation.
2833 rc = vma_del_reservation(h, vma, address);
2836 * VERY rare out of memory condition. Since
2837 * we can not delete the entry, set
2838 * hugetlb_restore_reserve so that the reserve
2839 * count will be incremented when the folio
2840 * is freed. This reserve will be consumed
2841 * on a subsequent allocation.
2843 folio_set_hugetlb_restore_reserve(folio);
2844 } else if (rc < 0) {
2846 * Rare out of memory condition from
2847 * vma_needs_reservation call. Memory allocation is
2848 * only attempted if a new entry is needed. Therefore,
2849 * this implies there is not an entry in the
2852 * For shared mappings, no entry in the map indicates
2853 * no reservation. We are done.
2855 if (!(vma->vm_flags & VM_MAYSHARE))
2857 * For private mappings, no entry indicates
2858 * a reservation is present. Since we can
2859 * not add an entry, set hugetlb_restore_reserve
2860 * on the folio so reserve count will be
2861 * incremented when freed. This reserve will
2862 * be consumed on a subsequent allocation.
2864 folio_set_hugetlb_restore_reserve(folio);
2867 * No reservation present, do nothing
2869 vma_end_reservation(h, vma, address);
2874 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2876 * @h: struct hstate old page belongs to
2877 * @old_folio: Old folio to dissolve
2878 * @list: List to isolate the page in case we need to
2879 * Returns 0 on success, otherwise negated error.
2881 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2882 struct folio *old_folio, struct list_head *list)
2884 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2885 int nid = folio_nid(old_folio);
2886 struct folio *new_folio;
2890 * Before dissolving the folio, we need to allocate a new one for the
2891 * pool to remain stable. Here, we allocate the folio and 'prep' it
2892 * by doing everything but actually updating counters and adding to
2893 * the pool. This simplifies and let us do most of the processing
2896 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2899 __prep_new_hugetlb_folio(h, new_folio);
2902 spin_lock_irq(&hugetlb_lock);
2903 if (!folio_test_hugetlb(old_folio)) {
2905 * Freed from under us. Drop new_folio too.
2908 } else if (folio_ref_count(old_folio)) {
2912 * Someone has grabbed the folio, try to isolate it here.
2913 * Fail with -EBUSY if not possible.
2915 spin_unlock_irq(&hugetlb_lock);
2916 isolated = isolate_hugetlb(old_folio, list);
2917 ret = isolated ? 0 : -EBUSY;
2918 spin_lock_irq(&hugetlb_lock);
2920 } else if (!folio_test_hugetlb_freed(old_folio)) {
2922 * Folio's refcount is 0 but it has not been enqueued in the
2923 * freelist yet. Race window is small, so we can succeed here if
2926 spin_unlock_irq(&hugetlb_lock);
2931 * Ok, old_folio is still a genuine free hugepage. Remove it from
2932 * the freelist and decrease the counters. These will be
2933 * incremented again when calling __prep_account_new_huge_page()
2934 * and enqueue_hugetlb_folio() for new_folio. The counters will
2935 * remain stable since this happens under the lock.
2937 remove_hugetlb_folio(h, old_folio, false);
2940 * Ref count on new_folio is already zero as it was dropped
2941 * earlier. It can be directly added to the pool free list.
2943 __prep_account_new_huge_page(h, nid);
2944 enqueue_hugetlb_folio(h, new_folio);
2947 * Folio has been replaced, we can safely free the old one.
2949 spin_unlock_irq(&hugetlb_lock);
2950 update_and_free_hugetlb_folio(h, old_folio, false);
2956 spin_unlock_irq(&hugetlb_lock);
2957 /* Folio has a zero ref count, but needs a ref to be freed */
2958 folio_ref_unfreeze(new_folio, 1);
2959 update_and_free_hugetlb_folio(h, new_folio, false);
2964 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2967 struct folio *folio = page_folio(page);
2971 * The page might have been dissolved from under our feet, so make sure
2972 * to carefully check the state under the lock.
2973 * Return success when racing as if we dissolved the page ourselves.
2975 spin_lock_irq(&hugetlb_lock);
2976 if (folio_test_hugetlb(folio)) {
2977 h = folio_hstate(folio);
2979 spin_unlock_irq(&hugetlb_lock);
2982 spin_unlock_irq(&hugetlb_lock);
2985 * Fence off gigantic pages as there is a cyclic dependency between
2986 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2987 * of bailing out right away without further retrying.
2989 if (hstate_is_gigantic(h))
2992 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
2994 else if (!folio_ref_count(folio))
2995 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3000 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3001 unsigned long addr, int avoid_reserve)
3003 struct hugepage_subpool *spool = subpool_vma(vma);
3004 struct hstate *h = hstate_vma(vma);
3005 struct folio *folio;
3006 long map_chg, map_commit;
3009 struct hugetlb_cgroup *h_cg = NULL;
3010 bool deferred_reserve;
3012 idx = hstate_index(h);
3014 * Examine the region/reserve map to determine if the process
3015 * has a reservation for the page to be allocated. A return
3016 * code of zero indicates a reservation exists (no change).
3018 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3020 return ERR_PTR(-ENOMEM);
3023 * Processes that did not create the mapping will have no
3024 * reserves as indicated by the region/reserve map. Check
3025 * that the allocation will not exceed the subpool limit.
3026 * Allocations for MAP_NORESERVE mappings also need to be
3027 * checked against any subpool limit.
3029 if (map_chg || avoid_reserve) {
3030 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3032 vma_end_reservation(h, vma, addr);
3033 return ERR_PTR(-ENOSPC);
3037 * Even though there was no reservation in the region/reserve
3038 * map, there could be reservations associated with the
3039 * subpool that can be used. This would be indicated if the
3040 * return value of hugepage_subpool_get_pages() is zero.
3041 * However, if avoid_reserve is specified we still avoid even
3042 * the subpool reservations.
3048 /* If this allocation is not consuming a reservation, charge it now.
3050 deferred_reserve = map_chg || avoid_reserve;
3051 if (deferred_reserve) {
3052 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3053 idx, pages_per_huge_page(h), &h_cg);
3055 goto out_subpool_put;
3058 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3060 goto out_uncharge_cgroup_reservation;
3062 spin_lock_irq(&hugetlb_lock);
3064 * glb_chg is passed to indicate whether or not a page must be taken
3065 * from the global free pool (global change). gbl_chg == 0 indicates
3066 * a reservation exists for the allocation.
3068 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3070 spin_unlock_irq(&hugetlb_lock);
3071 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3073 goto out_uncharge_cgroup;
3074 spin_lock_irq(&hugetlb_lock);
3075 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3076 folio_set_hugetlb_restore_reserve(folio);
3077 h->resv_huge_pages--;
3079 list_add(&folio->lru, &h->hugepage_activelist);
3080 folio_ref_unfreeze(folio, 1);
3084 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3085 /* If allocation is not consuming a reservation, also store the
3086 * hugetlb_cgroup pointer on the page.
3088 if (deferred_reserve) {
3089 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3093 spin_unlock_irq(&hugetlb_lock);
3095 hugetlb_set_folio_subpool(folio, spool);
3097 map_commit = vma_commit_reservation(h, vma, addr);
3098 if (unlikely(map_chg > map_commit)) {
3100 * The page was added to the reservation map between
3101 * vma_needs_reservation and vma_commit_reservation.
3102 * This indicates a race with hugetlb_reserve_pages.
3103 * Adjust for the subpool count incremented above AND
3104 * in hugetlb_reserve_pages for the same page. Also,
3105 * the reservation count added in hugetlb_reserve_pages
3106 * no longer applies.
3110 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3111 hugetlb_acct_memory(h, -rsv_adjust);
3112 if (deferred_reserve)
3113 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3114 pages_per_huge_page(h), folio);
3118 out_uncharge_cgroup:
3119 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3120 out_uncharge_cgroup_reservation:
3121 if (deferred_reserve)
3122 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3125 if (map_chg || avoid_reserve)
3126 hugepage_subpool_put_pages(spool, 1);
3127 vma_end_reservation(h, vma, addr);
3128 return ERR_PTR(-ENOSPC);
3131 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3132 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3133 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3135 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3138 /* do node specific alloc */
3139 if (nid != NUMA_NO_NODE) {
3140 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3141 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3146 /* allocate from next node when distributing huge pages */
3147 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3148 m = memblock_alloc_try_nid_raw(
3149 huge_page_size(h), huge_page_size(h),
3150 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3152 * Use the beginning of the huge page to store the
3153 * huge_bootmem_page struct (until gather_bootmem
3154 * puts them into the mem_map).
3162 /* Put them into a private list first because mem_map is not up yet */
3163 INIT_LIST_HEAD(&m->list);
3164 list_add(&m->list, &huge_boot_pages);
3170 * Put bootmem huge pages into the standard lists after mem_map is up.
3171 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3173 static void __init gather_bootmem_prealloc(void)
3175 struct huge_bootmem_page *m;
3177 list_for_each_entry(m, &huge_boot_pages, list) {
3178 struct page *page = virt_to_page(m);
3179 struct folio *folio = page_folio(page);
3180 struct hstate *h = m->hstate;
3182 VM_BUG_ON(!hstate_is_gigantic(h));
3183 WARN_ON(folio_ref_count(folio) != 1);
3184 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3185 WARN_ON(folio_test_reserved(folio));
3186 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3187 free_huge_folio(folio); /* add to the hugepage allocator */
3189 /* VERY unlikely inflated ref count on a tail page */
3190 free_gigantic_folio(folio, huge_page_order(h));
3194 * We need to restore the 'stolen' pages to totalram_pages
3195 * in order to fix confusing memory reports from free(1) and
3196 * other side-effects, like CommitLimit going negative.
3198 adjust_managed_page_count(page, pages_per_huge_page(h));
3202 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3207 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3208 if (hstate_is_gigantic(h)) {
3209 if (!alloc_bootmem_huge_page(h, nid))
3212 struct folio *folio;
3213 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3215 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3216 &node_states[N_MEMORY], NULL);
3219 free_huge_folio(folio); /* free it into the hugepage allocator */
3223 if (i == h->max_huge_pages_node[nid])
3226 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3227 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3228 h->max_huge_pages_node[nid], buf, nid, i);
3229 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3230 h->max_huge_pages_node[nid] = i;
3233 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3236 nodemask_t *node_alloc_noretry;
3237 bool node_specific_alloc = false;
3239 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3240 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3241 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3245 /* do node specific alloc */
3246 for_each_online_node(i) {
3247 if (h->max_huge_pages_node[i] > 0) {
3248 hugetlb_hstate_alloc_pages_onenode(h, i);
3249 node_specific_alloc = true;
3253 if (node_specific_alloc)
3256 /* below will do all node balanced alloc */
3257 if (!hstate_is_gigantic(h)) {
3259 * Bit mask controlling how hard we retry per-node allocations.
3260 * Ignore errors as lower level routines can deal with
3261 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3262 * time, we are likely in bigger trouble.
3264 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3267 /* allocations done at boot time */
3268 node_alloc_noretry = NULL;
3271 /* bit mask controlling how hard we retry per-node allocations */
3272 if (node_alloc_noretry)
3273 nodes_clear(*node_alloc_noretry);
3275 for (i = 0; i < h->max_huge_pages; ++i) {
3276 if (hstate_is_gigantic(h)) {
3277 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3279 } else if (!alloc_pool_huge_page(h,
3280 &node_states[N_MEMORY],
3281 node_alloc_noretry))
3285 if (i < h->max_huge_pages) {
3288 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3289 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3290 h->max_huge_pages, buf, i);
3291 h->max_huge_pages = i;
3293 kfree(node_alloc_noretry);
3296 static void __init hugetlb_init_hstates(void)
3298 struct hstate *h, *h2;
3300 for_each_hstate(h) {
3301 /* oversize hugepages were init'ed in early boot */
3302 if (!hstate_is_gigantic(h))
3303 hugetlb_hstate_alloc_pages(h);
3306 * Set demote order for each hstate. Note that
3307 * h->demote_order is initially 0.
3308 * - We can not demote gigantic pages if runtime freeing
3309 * is not supported, so skip this.
3310 * - If CMA allocation is possible, we can not demote
3311 * HUGETLB_PAGE_ORDER or smaller size pages.
3313 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3315 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3317 for_each_hstate(h2) {
3320 if (h2->order < h->order &&
3321 h2->order > h->demote_order)
3322 h->demote_order = h2->order;
3327 static void __init report_hugepages(void)
3331 for_each_hstate(h) {
3334 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3335 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3336 buf, h->free_huge_pages);
3337 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3338 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3342 #ifdef CONFIG_HIGHMEM
3343 static void try_to_free_low(struct hstate *h, unsigned long count,
3344 nodemask_t *nodes_allowed)
3347 LIST_HEAD(page_list);
3349 lockdep_assert_held(&hugetlb_lock);
3350 if (hstate_is_gigantic(h))
3354 * Collect pages to be freed on a list, and free after dropping lock
3356 for_each_node_mask(i, *nodes_allowed) {
3357 struct page *page, *next;
3358 struct list_head *freel = &h->hugepage_freelists[i];
3359 list_for_each_entry_safe(page, next, freel, lru) {
3360 if (count >= h->nr_huge_pages)
3362 if (PageHighMem(page))
3364 remove_hugetlb_folio(h, page_folio(page), false);
3365 list_add(&page->lru, &page_list);
3370 spin_unlock_irq(&hugetlb_lock);
3371 update_and_free_pages_bulk(h, &page_list);
3372 spin_lock_irq(&hugetlb_lock);
3375 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3376 nodemask_t *nodes_allowed)
3382 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3383 * balanced by operating on them in a round-robin fashion.
3384 * Returns 1 if an adjustment was made.
3386 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3391 lockdep_assert_held(&hugetlb_lock);
3392 VM_BUG_ON(delta != -1 && delta != 1);
3395 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3396 if (h->surplus_huge_pages_node[node])
3400 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3401 if (h->surplus_huge_pages_node[node] <
3402 h->nr_huge_pages_node[node])
3409 h->surplus_huge_pages += delta;
3410 h->surplus_huge_pages_node[node] += delta;
3414 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3415 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3416 nodemask_t *nodes_allowed)
3418 unsigned long min_count, ret;
3420 LIST_HEAD(page_list);
3421 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3424 * Bit mask controlling how hard we retry per-node allocations.
3425 * If we can not allocate the bit mask, do not attempt to allocate
3426 * the requested huge pages.
3428 if (node_alloc_noretry)
3429 nodes_clear(*node_alloc_noretry);
3434 * resize_lock mutex prevents concurrent adjustments to number of
3435 * pages in hstate via the proc/sysfs interfaces.
3437 mutex_lock(&h->resize_lock);
3438 flush_free_hpage_work(h);
3439 spin_lock_irq(&hugetlb_lock);
3442 * Check for a node specific request.
3443 * Changing node specific huge page count may require a corresponding
3444 * change to the global count. In any case, the passed node mask
3445 * (nodes_allowed) will restrict alloc/free to the specified node.
3447 if (nid != NUMA_NO_NODE) {
3448 unsigned long old_count = count;
3450 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3452 * User may have specified a large count value which caused the
3453 * above calculation to overflow. In this case, they wanted
3454 * to allocate as many huge pages as possible. Set count to
3455 * largest possible value to align with their intention.
3457 if (count < old_count)
3462 * Gigantic pages runtime allocation depend on the capability for large
3463 * page range allocation.
3464 * If the system does not provide this feature, return an error when
3465 * the user tries to allocate gigantic pages but let the user free the
3466 * boottime allocated gigantic pages.
3468 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3469 if (count > persistent_huge_pages(h)) {
3470 spin_unlock_irq(&hugetlb_lock);
3471 mutex_unlock(&h->resize_lock);
3472 NODEMASK_FREE(node_alloc_noretry);
3475 /* Fall through to decrease pool */
3479 * Increase the pool size
3480 * First take pages out of surplus state. Then make up the
3481 * remaining difference by allocating fresh huge pages.
3483 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3484 * to convert a surplus huge page to a normal huge page. That is
3485 * not critical, though, it just means the overall size of the
3486 * pool might be one hugepage larger than it needs to be, but
3487 * within all the constraints specified by the sysctls.
3489 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3490 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3494 while (count > persistent_huge_pages(h)) {
3496 * If this allocation races such that we no longer need the
3497 * page, free_huge_folio will handle it by freeing the page
3498 * and reducing the surplus.
3500 spin_unlock_irq(&hugetlb_lock);
3502 /* yield cpu to avoid soft lockup */
3505 ret = alloc_pool_huge_page(h, nodes_allowed,
3506 node_alloc_noretry);
3507 spin_lock_irq(&hugetlb_lock);
3511 /* Bail for signals. Probably ctrl-c from user */
3512 if (signal_pending(current))
3517 * Decrease the pool size
3518 * First return free pages to the buddy allocator (being careful
3519 * to keep enough around to satisfy reservations). Then place
3520 * pages into surplus state as needed so the pool will shrink
3521 * to the desired size as pages become free.
3523 * By placing pages into the surplus state independent of the
3524 * overcommit value, we are allowing the surplus pool size to
3525 * exceed overcommit. There are few sane options here. Since
3526 * alloc_surplus_hugetlb_folio() is checking the global counter,
3527 * though, we'll note that we're not allowed to exceed surplus
3528 * and won't grow the pool anywhere else. Not until one of the
3529 * sysctls are changed, or the surplus pages go out of use.
3531 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3532 min_count = max(count, min_count);
3533 try_to_free_low(h, min_count, nodes_allowed);
3536 * Collect pages to be removed on list without dropping lock
3538 while (min_count < persistent_huge_pages(h)) {
3539 page = remove_pool_huge_page(h, nodes_allowed, 0);
3543 list_add(&page->lru, &page_list);
3545 /* free the pages after dropping lock */
3546 spin_unlock_irq(&hugetlb_lock);
3547 update_and_free_pages_bulk(h, &page_list);
3548 flush_free_hpage_work(h);
3549 spin_lock_irq(&hugetlb_lock);
3551 while (count < persistent_huge_pages(h)) {
3552 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3556 h->max_huge_pages = persistent_huge_pages(h);
3557 spin_unlock_irq(&hugetlb_lock);
3558 mutex_unlock(&h->resize_lock);
3560 NODEMASK_FREE(node_alloc_noretry);
3565 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3567 int i, nid = folio_nid(folio);
3568 struct hstate *target_hstate;
3569 struct page *subpage;
3570 struct folio *inner_folio;
3573 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3575 remove_hugetlb_folio_for_demote(h, folio, false);
3576 spin_unlock_irq(&hugetlb_lock);
3578 rc = hugetlb_vmemmap_restore(h, &folio->page);
3580 /* Allocation of vmemmmap failed, we can not demote folio */
3581 spin_lock_irq(&hugetlb_lock);
3582 folio_ref_unfreeze(folio, 1);
3583 add_hugetlb_folio(h, folio, false);
3588 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3589 * sizes as it will not ref count folios.
3591 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3594 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3595 * Without the mutex, pages added to target hstate could be marked
3598 * Note that we already hold h->resize_lock. To prevent deadlock,
3599 * use the convention of always taking larger size hstate mutex first.
3601 mutex_lock(&target_hstate->resize_lock);
3602 for (i = 0; i < pages_per_huge_page(h);
3603 i += pages_per_huge_page(target_hstate)) {
3604 subpage = folio_page(folio, i);
3605 inner_folio = page_folio(subpage);
3606 if (hstate_is_gigantic(target_hstate))
3607 prep_compound_gigantic_folio_for_demote(inner_folio,
3608 target_hstate->order);
3610 prep_compound_page(subpage, target_hstate->order);
3611 folio_change_private(inner_folio, NULL);
3612 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3613 free_huge_folio(inner_folio);
3615 mutex_unlock(&target_hstate->resize_lock);
3617 spin_lock_irq(&hugetlb_lock);
3620 * Not absolutely necessary, but for consistency update max_huge_pages
3621 * based on pool changes for the demoted page.
3623 h->max_huge_pages--;
3624 target_hstate->max_huge_pages +=
3625 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3630 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3631 __must_hold(&hugetlb_lock)
3634 struct folio *folio;
3636 lockdep_assert_held(&hugetlb_lock);
3638 /* We should never get here if no demote order */
3639 if (!h->demote_order) {
3640 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3641 return -EINVAL; /* internal error */
3644 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3645 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3646 if (folio_test_hwpoison(folio))
3648 return demote_free_hugetlb_folio(h, folio);
3653 * Only way to get here is if all pages on free lists are poisoned.
3654 * Return -EBUSY so that caller will not retry.
3659 #define HSTATE_ATTR_RO(_name) \
3660 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3662 #define HSTATE_ATTR_WO(_name) \
3663 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3665 #define HSTATE_ATTR(_name) \
3666 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3668 static struct kobject *hugepages_kobj;
3669 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3671 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3673 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3677 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3678 if (hstate_kobjs[i] == kobj) {
3680 *nidp = NUMA_NO_NODE;
3684 return kobj_to_node_hstate(kobj, nidp);
3687 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3688 struct kobj_attribute *attr, char *buf)
3691 unsigned long nr_huge_pages;
3694 h = kobj_to_hstate(kobj, &nid);
3695 if (nid == NUMA_NO_NODE)
3696 nr_huge_pages = h->nr_huge_pages;
3698 nr_huge_pages = h->nr_huge_pages_node[nid];
3700 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3703 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3704 struct hstate *h, int nid,
3705 unsigned long count, size_t len)
3708 nodemask_t nodes_allowed, *n_mask;
3710 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3713 if (nid == NUMA_NO_NODE) {
3715 * global hstate attribute
3717 if (!(obey_mempolicy &&
3718 init_nodemask_of_mempolicy(&nodes_allowed)))
3719 n_mask = &node_states[N_MEMORY];
3721 n_mask = &nodes_allowed;
3724 * Node specific request. count adjustment happens in
3725 * set_max_huge_pages() after acquiring hugetlb_lock.
3727 init_nodemask_of_node(&nodes_allowed, nid);
3728 n_mask = &nodes_allowed;
3731 err = set_max_huge_pages(h, count, nid, n_mask);
3733 return err ? err : len;
3736 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3737 struct kobject *kobj, const char *buf,
3741 unsigned long count;
3745 err = kstrtoul(buf, 10, &count);
3749 h = kobj_to_hstate(kobj, &nid);
3750 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3753 static ssize_t nr_hugepages_show(struct kobject *kobj,
3754 struct kobj_attribute *attr, char *buf)
3756 return nr_hugepages_show_common(kobj, attr, buf);
3759 static ssize_t nr_hugepages_store(struct kobject *kobj,
3760 struct kobj_attribute *attr, const char *buf, size_t len)
3762 return nr_hugepages_store_common(false, kobj, buf, len);
3764 HSTATE_ATTR(nr_hugepages);
3769 * hstate attribute for optionally mempolicy-based constraint on persistent
3770 * huge page alloc/free.
3772 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3773 struct kobj_attribute *attr,
3776 return nr_hugepages_show_common(kobj, attr, buf);
3779 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3780 struct kobj_attribute *attr, const char *buf, size_t len)
3782 return nr_hugepages_store_common(true, kobj, buf, len);
3784 HSTATE_ATTR(nr_hugepages_mempolicy);
3788 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3789 struct kobj_attribute *attr, char *buf)
3791 struct hstate *h = kobj_to_hstate(kobj, NULL);
3792 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3795 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3796 struct kobj_attribute *attr, const char *buf, size_t count)
3799 unsigned long input;
3800 struct hstate *h = kobj_to_hstate(kobj, NULL);
3802 if (hstate_is_gigantic(h))
3805 err = kstrtoul(buf, 10, &input);
3809 spin_lock_irq(&hugetlb_lock);
3810 h->nr_overcommit_huge_pages = input;
3811 spin_unlock_irq(&hugetlb_lock);
3815 HSTATE_ATTR(nr_overcommit_hugepages);
3817 static ssize_t free_hugepages_show(struct kobject *kobj,
3818 struct kobj_attribute *attr, char *buf)
3821 unsigned long free_huge_pages;
3824 h = kobj_to_hstate(kobj, &nid);
3825 if (nid == NUMA_NO_NODE)
3826 free_huge_pages = h->free_huge_pages;
3828 free_huge_pages = h->free_huge_pages_node[nid];
3830 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3832 HSTATE_ATTR_RO(free_hugepages);
3834 static ssize_t resv_hugepages_show(struct kobject *kobj,
3835 struct kobj_attribute *attr, char *buf)
3837 struct hstate *h = kobj_to_hstate(kobj, NULL);
3838 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3840 HSTATE_ATTR_RO(resv_hugepages);
3842 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3843 struct kobj_attribute *attr, char *buf)
3846 unsigned long surplus_huge_pages;
3849 h = kobj_to_hstate(kobj, &nid);
3850 if (nid == NUMA_NO_NODE)
3851 surplus_huge_pages = h->surplus_huge_pages;
3853 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3855 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3857 HSTATE_ATTR_RO(surplus_hugepages);
3859 static ssize_t demote_store(struct kobject *kobj,
3860 struct kobj_attribute *attr, const char *buf, size_t len)
3862 unsigned long nr_demote;
3863 unsigned long nr_available;
3864 nodemask_t nodes_allowed, *n_mask;
3869 err = kstrtoul(buf, 10, &nr_demote);
3872 h = kobj_to_hstate(kobj, &nid);
3874 if (nid != NUMA_NO_NODE) {
3875 init_nodemask_of_node(&nodes_allowed, nid);
3876 n_mask = &nodes_allowed;
3878 n_mask = &node_states[N_MEMORY];
3881 /* Synchronize with other sysfs operations modifying huge pages */
3882 mutex_lock(&h->resize_lock);
3883 spin_lock_irq(&hugetlb_lock);
3887 * Check for available pages to demote each time thorough the
3888 * loop as demote_pool_huge_page will drop hugetlb_lock.
3890 if (nid != NUMA_NO_NODE)
3891 nr_available = h->free_huge_pages_node[nid];
3893 nr_available = h->free_huge_pages;
3894 nr_available -= h->resv_huge_pages;
3898 err = demote_pool_huge_page(h, n_mask);
3905 spin_unlock_irq(&hugetlb_lock);
3906 mutex_unlock(&h->resize_lock);
3912 HSTATE_ATTR_WO(demote);
3914 static ssize_t demote_size_show(struct kobject *kobj,
3915 struct kobj_attribute *attr, char *buf)
3917 struct hstate *h = kobj_to_hstate(kobj, NULL);
3918 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3920 return sysfs_emit(buf, "%lukB\n", demote_size);
3923 static ssize_t demote_size_store(struct kobject *kobj,
3924 struct kobj_attribute *attr,
3925 const char *buf, size_t count)
3927 struct hstate *h, *demote_hstate;
3928 unsigned long demote_size;
3929 unsigned int demote_order;
3931 demote_size = (unsigned long)memparse(buf, NULL);
3933 demote_hstate = size_to_hstate(demote_size);
3936 demote_order = demote_hstate->order;
3937 if (demote_order < HUGETLB_PAGE_ORDER)
3940 /* demote order must be smaller than hstate order */
3941 h = kobj_to_hstate(kobj, NULL);
3942 if (demote_order >= h->order)
3945 /* resize_lock synchronizes access to demote size and writes */
3946 mutex_lock(&h->resize_lock);
3947 h->demote_order = demote_order;
3948 mutex_unlock(&h->resize_lock);
3952 HSTATE_ATTR(demote_size);
3954 static struct attribute *hstate_attrs[] = {
3955 &nr_hugepages_attr.attr,
3956 &nr_overcommit_hugepages_attr.attr,
3957 &free_hugepages_attr.attr,
3958 &resv_hugepages_attr.attr,
3959 &surplus_hugepages_attr.attr,
3961 &nr_hugepages_mempolicy_attr.attr,
3966 static const struct attribute_group hstate_attr_group = {
3967 .attrs = hstate_attrs,
3970 static struct attribute *hstate_demote_attrs[] = {
3971 &demote_size_attr.attr,
3976 static const struct attribute_group hstate_demote_attr_group = {
3977 .attrs = hstate_demote_attrs,
3980 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3981 struct kobject **hstate_kobjs,
3982 const struct attribute_group *hstate_attr_group)
3985 int hi = hstate_index(h);
3987 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3988 if (!hstate_kobjs[hi])
3991 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3993 kobject_put(hstate_kobjs[hi]);
3994 hstate_kobjs[hi] = NULL;
3998 if (h->demote_order) {
3999 retval = sysfs_create_group(hstate_kobjs[hi],
4000 &hstate_demote_attr_group);
4002 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4003 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4004 kobject_put(hstate_kobjs[hi]);
4005 hstate_kobjs[hi] = NULL;
4014 static bool hugetlb_sysfs_initialized __ro_after_init;
4017 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4018 * with node devices in node_devices[] using a parallel array. The array
4019 * index of a node device or _hstate == node id.
4020 * This is here to avoid any static dependency of the node device driver, in
4021 * the base kernel, on the hugetlb module.
4023 struct node_hstate {
4024 struct kobject *hugepages_kobj;
4025 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4027 static struct node_hstate node_hstates[MAX_NUMNODES];
4030 * A subset of global hstate attributes for node devices
4032 static struct attribute *per_node_hstate_attrs[] = {
4033 &nr_hugepages_attr.attr,
4034 &free_hugepages_attr.attr,
4035 &surplus_hugepages_attr.attr,
4039 static const struct attribute_group per_node_hstate_attr_group = {
4040 .attrs = per_node_hstate_attrs,
4044 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4045 * Returns node id via non-NULL nidp.
4047 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4051 for (nid = 0; nid < nr_node_ids; nid++) {
4052 struct node_hstate *nhs = &node_hstates[nid];
4054 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4055 if (nhs->hstate_kobjs[i] == kobj) {
4067 * Unregister hstate attributes from a single node device.
4068 * No-op if no hstate attributes attached.
4070 void hugetlb_unregister_node(struct node *node)
4073 struct node_hstate *nhs = &node_hstates[node->dev.id];
4075 if (!nhs->hugepages_kobj)
4076 return; /* no hstate attributes */
4078 for_each_hstate(h) {
4079 int idx = hstate_index(h);
4080 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4084 if (h->demote_order)
4085 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4086 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4087 kobject_put(hstate_kobj);
4088 nhs->hstate_kobjs[idx] = NULL;
4091 kobject_put(nhs->hugepages_kobj);
4092 nhs->hugepages_kobj = NULL;
4097 * Register hstate attributes for a single node device.
4098 * No-op if attributes already registered.
4100 void hugetlb_register_node(struct node *node)
4103 struct node_hstate *nhs = &node_hstates[node->dev.id];
4106 if (!hugetlb_sysfs_initialized)
4109 if (nhs->hugepages_kobj)
4110 return; /* already allocated */
4112 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4114 if (!nhs->hugepages_kobj)
4117 for_each_hstate(h) {
4118 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4120 &per_node_hstate_attr_group);
4122 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4123 h->name, node->dev.id);
4124 hugetlb_unregister_node(node);
4131 * hugetlb init time: register hstate attributes for all registered node
4132 * devices of nodes that have memory. All on-line nodes should have
4133 * registered their associated device by this time.
4135 static void __init hugetlb_register_all_nodes(void)
4139 for_each_online_node(nid)
4140 hugetlb_register_node(node_devices[nid]);
4142 #else /* !CONFIG_NUMA */
4144 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4152 static void hugetlb_register_all_nodes(void) { }
4157 static void __init hugetlb_cma_check(void);
4159 static inline __init void hugetlb_cma_check(void)
4164 static void __init hugetlb_sysfs_init(void)
4169 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4170 if (!hugepages_kobj)
4173 for_each_hstate(h) {
4174 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4175 hstate_kobjs, &hstate_attr_group);
4177 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4181 hugetlb_sysfs_initialized = true;
4183 hugetlb_register_all_nodes();
4186 #ifdef CONFIG_SYSCTL
4187 static void hugetlb_sysctl_init(void);
4189 static inline void hugetlb_sysctl_init(void) { }
4192 static int __init hugetlb_init(void)
4196 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4199 if (!hugepages_supported()) {
4200 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4201 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4206 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4207 * architectures depend on setup being done here.
4209 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4210 if (!parsed_default_hugepagesz) {
4212 * If we did not parse a default huge page size, set
4213 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4214 * number of huge pages for this default size was implicitly
4215 * specified, set that here as well.
4216 * Note that the implicit setting will overwrite an explicit
4217 * setting. A warning will be printed in this case.
4219 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4220 if (default_hstate_max_huge_pages) {
4221 if (default_hstate.max_huge_pages) {
4224 string_get_size(huge_page_size(&default_hstate),
4225 1, STRING_UNITS_2, buf, 32);
4226 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4227 default_hstate.max_huge_pages, buf);
4228 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4229 default_hstate_max_huge_pages);
4231 default_hstate.max_huge_pages =
4232 default_hstate_max_huge_pages;
4234 for_each_online_node(i)
4235 default_hstate.max_huge_pages_node[i] =
4236 default_hugepages_in_node[i];
4240 hugetlb_cma_check();
4241 hugetlb_init_hstates();
4242 gather_bootmem_prealloc();
4245 hugetlb_sysfs_init();
4246 hugetlb_cgroup_file_init();
4247 hugetlb_sysctl_init();
4250 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4252 num_fault_mutexes = 1;
4254 hugetlb_fault_mutex_table =
4255 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4257 BUG_ON(!hugetlb_fault_mutex_table);
4259 for (i = 0; i < num_fault_mutexes; i++)
4260 mutex_init(&hugetlb_fault_mutex_table[i]);
4263 subsys_initcall(hugetlb_init);
4265 /* Overwritten by architectures with more huge page sizes */
4266 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4268 return size == HPAGE_SIZE;
4271 void __init hugetlb_add_hstate(unsigned int order)
4276 if (size_to_hstate(PAGE_SIZE << order)) {
4279 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4281 h = &hstates[hugetlb_max_hstate++];
4282 mutex_init(&h->resize_lock);
4284 h->mask = ~(huge_page_size(h) - 1);
4285 for (i = 0; i < MAX_NUMNODES; ++i)
4286 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4287 INIT_LIST_HEAD(&h->hugepage_activelist);
4288 h->next_nid_to_alloc = first_memory_node;
4289 h->next_nid_to_free = first_memory_node;
4290 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4291 huge_page_size(h)/SZ_1K);
4296 bool __init __weak hugetlb_node_alloc_supported(void)
4301 static void __init hugepages_clear_pages_in_node(void)
4303 if (!hugetlb_max_hstate) {
4304 default_hstate_max_huge_pages = 0;
4305 memset(default_hugepages_in_node, 0,
4306 sizeof(default_hugepages_in_node));
4308 parsed_hstate->max_huge_pages = 0;
4309 memset(parsed_hstate->max_huge_pages_node, 0,
4310 sizeof(parsed_hstate->max_huge_pages_node));
4315 * hugepages command line processing
4316 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4317 * specification. If not, ignore the hugepages value. hugepages can also
4318 * be the first huge page command line option in which case it implicitly
4319 * specifies the number of huge pages for the default size.
4321 static int __init hugepages_setup(char *s)
4324 static unsigned long *last_mhp;
4325 int node = NUMA_NO_NODE;
4330 if (!parsed_valid_hugepagesz) {
4331 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4332 parsed_valid_hugepagesz = true;
4337 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4338 * yet, so this hugepages= parameter goes to the "default hstate".
4339 * Otherwise, it goes with the previously parsed hugepagesz or
4340 * default_hugepagesz.
4342 else if (!hugetlb_max_hstate)
4343 mhp = &default_hstate_max_huge_pages;
4345 mhp = &parsed_hstate->max_huge_pages;
4347 if (mhp == last_mhp) {
4348 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4354 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4356 /* Parameter is node format */
4357 if (p[count] == ':') {
4358 if (!hugetlb_node_alloc_supported()) {
4359 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4362 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4364 node = array_index_nospec(tmp, MAX_NUMNODES);
4366 /* Parse hugepages */
4367 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4369 if (!hugetlb_max_hstate)
4370 default_hugepages_in_node[node] = tmp;
4372 parsed_hstate->max_huge_pages_node[node] = tmp;
4374 /* Go to parse next node*/
4375 if (p[count] == ',')
4388 * Global state is always initialized later in hugetlb_init.
4389 * But we need to allocate gigantic hstates here early to still
4390 * use the bootmem allocator.
4392 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4393 hugetlb_hstate_alloc_pages(parsed_hstate);
4400 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4401 hugepages_clear_pages_in_node();
4404 __setup("hugepages=", hugepages_setup);
4407 * hugepagesz command line processing
4408 * A specific huge page size can only be specified once with hugepagesz.
4409 * hugepagesz is followed by hugepages on the command line. The global
4410 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4411 * hugepagesz argument was valid.
4413 static int __init hugepagesz_setup(char *s)
4418 parsed_valid_hugepagesz = false;
4419 size = (unsigned long)memparse(s, NULL);
4421 if (!arch_hugetlb_valid_size(size)) {
4422 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4426 h = size_to_hstate(size);
4429 * hstate for this size already exists. This is normally
4430 * an error, but is allowed if the existing hstate is the
4431 * default hstate. More specifically, it is only allowed if
4432 * the number of huge pages for the default hstate was not
4433 * previously specified.
4435 if (!parsed_default_hugepagesz || h != &default_hstate ||
4436 default_hstate.max_huge_pages) {
4437 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4442 * No need to call hugetlb_add_hstate() as hstate already
4443 * exists. But, do set parsed_hstate so that a following
4444 * hugepages= parameter will be applied to this hstate.
4447 parsed_valid_hugepagesz = true;
4451 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4452 parsed_valid_hugepagesz = true;
4455 __setup("hugepagesz=", hugepagesz_setup);
4458 * default_hugepagesz command line input
4459 * Only one instance of default_hugepagesz allowed on command line.
4461 static int __init default_hugepagesz_setup(char *s)
4466 parsed_valid_hugepagesz = false;
4467 if (parsed_default_hugepagesz) {
4468 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4472 size = (unsigned long)memparse(s, NULL);
4474 if (!arch_hugetlb_valid_size(size)) {
4475 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4479 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4480 parsed_valid_hugepagesz = true;
4481 parsed_default_hugepagesz = true;
4482 default_hstate_idx = hstate_index(size_to_hstate(size));
4485 * The number of default huge pages (for this size) could have been
4486 * specified as the first hugetlb parameter: hugepages=X. If so,
4487 * then default_hstate_max_huge_pages is set. If the default huge
4488 * page size is gigantic (> MAX_ORDER), then the pages must be
4489 * allocated here from bootmem allocator.
4491 if (default_hstate_max_huge_pages) {
4492 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4493 for_each_online_node(i)
4494 default_hstate.max_huge_pages_node[i] =
4495 default_hugepages_in_node[i];
4496 if (hstate_is_gigantic(&default_hstate))
4497 hugetlb_hstate_alloc_pages(&default_hstate);
4498 default_hstate_max_huge_pages = 0;
4503 __setup("default_hugepagesz=", default_hugepagesz_setup);
4505 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4508 struct mempolicy *mpol = get_task_policy(current);
4511 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4512 * (from policy_nodemask) specifically for hugetlb case
4514 if (mpol->mode == MPOL_BIND &&
4515 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4516 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4517 return &mpol->nodes;
4522 static unsigned int allowed_mems_nr(struct hstate *h)
4525 unsigned int nr = 0;
4526 nodemask_t *mbind_nodemask;
4527 unsigned int *array = h->free_huge_pages_node;
4528 gfp_t gfp_mask = htlb_alloc_mask(h);
4530 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4531 for_each_node_mask(node, cpuset_current_mems_allowed) {
4532 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4539 #ifdef CONFIG_SYSCTL
4540 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4541 void *buffer, size_t *length,
4542 loff_t *ppos, unsigned long *out)
4544 struct ctl_table dup_table;
4547 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4548 * can duplicate the @table and alter the duplicate of it.
4551 dup_table.data = out;
4553 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4556 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4557 struct ctl_table *table, int write,
4558 void *buffer, size_t *length, loff_t *ppos)
4560 struct hstate *h = &default_hstate;
4561 unsigned long tmp = h->max_huge_pages;
4564 if (!hugepages_supported())
4567 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4573 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4574 NUMA_NO_NODE, tmp, *length);
4579 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4580 void *buffer, size_t *length, loff_t *ppos)
4583 return hugetlb_sysctl_handler_common(false, table, write,
4584 buffer, length, ppos);
4588 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4589 void *buffer, size_t *length, loff_t *ppos)
4591 return hugetlb_sysctl_handler_common(true, table, write,
4592 buffer, length, ppos);
4594 #endif /* CONFIG_NUMA */
4596 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4597 void *buffer, size_t *length, loff_t *ppos)
4599 struct hstate *h = &default_hstate;
4603 if (!hugepages_supported())
4606 tmp = h->nr_overcommit_huge_pages;
4608 if (write && hstate_is_gigantic(h))
4611 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4617 spin_lock_irq(&hugetlb_lock);
4618 h->nr_overcommit_huge_pages = tmp;
4619 spin_unlock_irq(&hugetlb_lock);
4625 static struct ctl_table hugetlb_table[] = {
4627 .procname = "nr_hugepages",
4629 .maxlen = sizeof(unsigned long),
4631 .proc_handler = hugetlb_sysctl_handler,
4635 .procname = "nr_hugepages_mempolicy",
4637 .maxlen = sizeof(unsigned long),
4639 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4643 .procname = "hugetlb_shm_group",
4644 .data = &sysctl_hugetlb_shm_group,
4645 .maxlen = sizeof(gid_t),
4647 .proc_handler = proc_dointvec,
4650 .procname = "nr_overcommit_hugepages",
4652 .maxlen = sizeof(unsigned long),
4654 .proc_handler = hugetlb_overcommit_handler,
4659 static void hugetlb_sysctl_init(void)
4661 register_sysctl_init("vm", hugetlb_table);
4663 #endif /* CONFIG_SYSCTL */
4665 void hugetlb_report_meminfo(struct seq_file *m)
4668 unsigned long total = 0;
4670 if (!hugepages_supported())
4673 for_each_hstate(h) {
4674 unsigned long count = h->nr_huge_pages;
4676 total += huge_page_size(h) * count;
4678 if (h == &default_hstate)
4680 "HugePages_Total: %5lu\n"
4681 "HugePages_Free: %5lu\n"
4682 "HugePages_Rsvd: %5lu\n"
4683 "HugePages_Surp: %5lu\n"
4684 "Hugepagesize: %8lu kB\n",
4688 h->surplus_huge_pages,
4689 huge_page_size(h) / SZ_1K);
4692 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4695 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4697 struct hstate *h = &default_hstate;
4699 if (!hugepages_supported())
4702 return sysfs_emit_at(buf, len,
4703 "Node %d HugePages_Total: %5u\n"
4704 "Node %d HugePages_Free: %5u\n"
4705 "Node %d HugePages_Surp: %5u\n",
4706 nid, h->nr_huge_pages_node[nid],
4707 nid, h->free_huge_pages_node[nid],
4708 nid, h->surplus_huge_pages_node[nid]);
4711 void hugetlb_show_meminfo_node(int nid)
4715 if (!hugepages_supported())
4719 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4721 h->nr_huge_pages_node[nid],
4722 h->free_huge_pages_node[nid],
4723 h->surplus_huge_pages_node[nid],
4724 huge_page_size(h) / SZ_1K);
4727 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4729 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4730 K(atomic_long_read(&mm->hugetlb_usage)));
4733 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4734 unsigned long hugetlb_total_pages(void)
4737 unsigned long nr_total_pages = 0;
4740 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4741 return nr_total_pages;
4744 static int hugetlb_acct_memory(struct hstate *h, long delta)
4751 spin_lock_irq(&hugetlb_lock);
4753 * When cpuset is configured, it breaks the strict hugetlb page
4754 * reservation as the accounting is done on a global variable. Such
4755 * reservation is completely rubbish in the presence of cpuset because
4756 * the reservation is not checked against page availability for the
4757 * current cpuset. Application can still potentially OOM'ed by kernel
4758 * with lack of free htlb page in cpuset that the task is in.
4759 * Attempt to enforce strict accounting with cpuset is almost
4760 * impossible (or too ugly) because cpuset is too fluid that
4761 * task or memory node can be dynamically moved between cpusets.
4763 * The change of semantics for shared hugetlb mapping with cpuset is
4764 * undesirable. However, in order to preserve some of the semantics,
4765 * we fall back to check against current free page availability as
4766 * a best attempt and hopefully to minimize the impact of changing
4767 * semantics that cpuset has.
4769 * Apart from cpuset, we also have memory policy mechanism that
4770 * also determines from which node the kernel will allocate memory
4771 * in a NUMA system. So similar to cpuset, we also should consider
4772 * the memory policy of the current task. Similar to the description
4776 if (gather_surplus_pages(h, delta) < 0)
4779 if (delta > allowed_mems_nr(h)) {
4780 return_unused_surplus_pages(h, delta);
4787 return_unused_surplus_pages(h, (unsigned long) -delta);
4790 spin_unlock_irq(&hugetlb_lock);
4794 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4796 struct resv_map *resv = vma_resv_map(vma);
4799 * HPAGE_RESV_OWNER indicates a private mapping.
4800 * This new VMA should share its siblings reservation map if present.
4801 * The VMA will only ever have a valid reservation map pointer where
4802 * it is being copied for another still existing VMA. As that VMA
4803 * has a reference to the reservation map it cannot disappear until
4804 * after this open call completes. It is therefore safe to take a
4805 * new reference here without additional locking.
4807 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4808 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4809 kref_get(&resv->refs);
4813 * vma_lock structure for sharable mappings is vma specific.
4814 * Clear old pointer (if copied via vm_area_dup) and allocate
4815 * new structure. Before clearing, make sure vma_lock is not
4818 if (vma->vm_flags & VM_MAYSHARE) {
4819 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4822 if (vma_lock->vma != vma) {
4823 vma->vm_private_data = NULL;
4824 hugetlb_vma_lock_alloc(vma);
4826 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4828 hugetlb_vma_lock_alloc(vma);
4832 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4834 struct hstate *h = hstate_vma(vma);
4835 struct resv_map *resv;
4836 struct hugepage_subpool *spool = subpool_vma(vma);
4837 unsigned long reserve, start, end;
4840 hugetlb_vma_lock_free(vma);
4842 resv = vma_resv_map(vma);
4843 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4846 start = vma_hugecache_offset(h, vma, vma->vm_start);
4847 end = vma_hugecache_offset(h, vma, vma->vm_end);
4849 reserve = (end - start) - region_count(resv, start, end);
4850 hugetlb_cgroup_uncharge_counter(resv, start, end);
4853 * Decrement reserve counts. The global reserve count may be
4854 * adjusted if the subpool has a minimum size.
4856 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4857 hugetlb_acct_memory(h, -gbl_reserve);
4860 kref_put(&resv->refs, resv_map_release);
4863 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4865 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4869 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4870 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4871 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4873 if (addr & ~PUD_MASK) {
4875 * hugetlb_vm_op_split is called right before we attempt to
4876 * split the VMA. We will need to unshare PMDs in the old and
4877 * new VMAs, so let's unshare before we split.
4879 unsigned long floor = addr & PUD_MASK;
4880 unsigned long ceil = floor + PUD_SIZE;
4882 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4883 hugetlb_unshare_pmds(vma, floor, ceil);
4889 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4891 return huge_page_size(hstate_vma(vma));
4895 * We cannot handle pagefaults against hugetlb pages at all. They cause
4896 * handle_mm_fault() to try to instantiate regular-sized pages in the
4897 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4900 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4907 * When a new function is introduced to vm_operations_struct and added
4908 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4909 * This is because under System V memory model, mappings created via
4910 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4911 * their original vm_ops are overwritten with shm_vm_ops.
4913 const struct vm_operations_struct hugetlb_vm_ops = {
4914 .fault = hugetlb_vm_op_fault,
4915 .open = hugetlb_vm_op_open,
4916 .close = hugetlb_vm_op_close,
4917 .may_split = hugetlb_vm_op_split,
4918 .pagesize = hugetlb_vm_op_pagesize,
4921 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4925 unsigned int shift = huge_page_shift(hstate_vma(vma));
4928 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4929 vma->vm_page_prot)));
4931 entry = huge_pte_wrprotect(mk_huge_pte(page,
4932 vma->vm_page_prot));
4934 entry = pte_mkyoung(entry);
4935 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4940 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4941 unsigned long address, pte_t *ptep)
4945 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4946 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4947 update_mmu_cache(vma, address, ptep);
4950 bool is_hugetlb_entry_migration(pte_t pte)
4954 if (huge_pte_none(pte) || pte_present(pte))
4956 swp = pte_to_swp_entry(pte);
4957 if (is_migration_entry(swp))
4963 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4967 if (huge_pte_none(pte) || pte_present(pte))
4969 swp = pte_to_swp_entry(pte);
4970 if (is_hwpoison_entry(swp))
4977 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4978 struct folio *new_folio, pte_t old)
4980 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
4982 __folio_mark_uptodate(new_folio);
4983 hugepage_add_new_anon_rmap(new_folio, vma, addr);
4984 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
4985 newpte = huge_pte_mkuffd_wp(newpte);
4986 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte);
4987 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4988 folio_set_hugetlb_migratable(new_folio);
4991 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4992 struct vm_area_struct *dst_vma,
4993 struct vm_area_struct *src_vma)
4995 pte_t *src_pte, *dst_pte, entry;
4996 struct folio *pte_folio;
4998 bool cow = is_cow_mapping(src_vma->vm_flags);
4999 struct hstate *h = hstate_vma(src_vma);
5000 unsigned long sz = huge_page_size(h);
5001 unsigned long npages = pages_per_huge_page(h);
5002 struct mmu_notifier_range range;
5003 unsigned long last_addr_mask;
5007 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5010 mmu_notifier_invalidate_range_start(&range);
5011 vma_assert_write_locked(src_vma);
5012 raw_write_seqcount_begin(&src->write_protect_seq);
5015 * For shared mappings the vma lock must be held before
5016 * calling hugetlb_walk() in the src vma. Otherwise, the
5017 * returned ptep could go away if part of a shared pmd and
5018 * another thread calls huge_pmd_unshare.
5020 hugetlb_vma_lock_read(src_vma);
5023 last_addr_mask = hugetlb_mask_last_page(h);
5024 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5025 spinlock_t *src_ptl, *dst_ptl;
5026 src_pte = hugetlb_walk(src_vma, addr, sz);
5028 addr |= last_addr_mask;
5031 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5038 * If the pagetables are shared don't copy or take references.
5040 * dst_pte == src_pte is the common case of src/dest sharing.
5041 * However, src could have 'unshared' and dst shares with
5042 * another vma. So page_count of ptep page is checked instead
5043 * to reliably determine whether pte is shared.
5045 if (page_count(virt_to_page(dst_pte)) > 1) {
5046 addr |= last_addr_mask;
5050 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5051 src_ptl = huge_pte_lockptr(h, src, src_pte);
5052 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5053 entry = huge_ptep_get(src_pte);
5055 if (huge_pte_none(entry)) {
5057 * Skip if src entry none.
5060 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5061 if (!userfaultfd_wp(dst_vma))
5062 entry = huge_pte_clear_uffd_wp(entry);
5063 set_huge_pte_at(dst, addr, dst_pte, entry);
5064 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5065 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5066 bool uffd_wp = pte_swp_uffd_wp(entry);
5068 if (!is_readable_migration_entry(swp_entry) && cow) {
5070 * COW mappings require pages in both
5071 * parent and child to be set to read.
5073 swp_entry = make_readable_migration_entry(
5074 swp_offset(swp_entry));
5075 entry = swp_entry_to_pte(swp_entry);
5076 if (userfaultfd_wp(src_vma) && uffd_wp)
5077 entry = pte_swp_mkuffd_wp(entry);
5078 set_huge_pte_at(src, addr, src_pte, entry);
5080 if (!userfaultfd_wp(dst_vma))
5081 entry = huge_pte_clear_uffd_wp(entry);
5082 set_huge_pte_at(dst, addr, dst_pte, entry);
5083 } else if (unlikely(is_pte_marker(entry))) {
5084 pte_marker marker = copy_pte_marker(
5085 pte_to_swp_entry(entry), dst_vma);
5088 set_huge_pte_at(dst, addr, dst_pte,
5089 make_pte_marker(marker));
5091 entry = huge_ptep_get(src_pte);
5092 pte_folio = page_folio(pte_page(entry));
5093 folio_get(pte_folio);
5096 * Failing to duplicate the anon rmap is a rare case
5097 * where we see pinned hugetlb pages while they're
5098 * prone to COW. We need to do the COW earlier during
5101 * When pre-allocating the page or copying data, we
5102 * need to be without the pgtable locks since we could
5103 * sleep during the process.
5105 if (!folio_test_anon(pte_folio)) {
5106 page_dup_file_rmap(&pte_folio->page, true);
5107 } else if (page_try_dup_anon_rmap(&pte_folio->page,
5109 pte_t src_pte_old = entry;
5110 struct folio *new_folio;
5112 spin_unlock(src_ptl);
5113 spin_unlock(dst_ptl);
5114 /* Do not use reserve as it's private owned */
5115 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5116 if (IS_ERR(new_folio)) {
5117 folio_put(pte_folio);
5118 ret = PTR_ERR(new_folio);
5121 ret = copy_user_large_folio(new_folio,
5124 folio_put(pte_folio);
5126 folio_put(new_folio);
5130 /* Install the new hugetlb folio if src pte stable */
5131 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5132 src_ptl = huge_pte_lockptr(h, src, src_pte);
5133 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5134 entry = huge_ptep_get(src_pte);
5135 if (!pte_same(src_pte_old, entry)) {
5136 restore_reserve_on_error(h, dst_vma, addr,
5138 folio_put(new_folio);
5139 /* huge_ptep of dst_pte won't change as in child */
5142 hugetlb_install_folio(dst_vma, dst_pte, addr,
5143 new_folio, src_pte_old);
5144 spin_unlock(src_ptl);
5145 spin_unlock(dst_ptl);
5151 * No need to notify as we are downgrading page
5152 * table protection not changing it to point
5155 * See Documentation/mm/mmu_notifier.rst
5157 huge_ptep_set_wrprotect(src, addr, src_pte);
5158 entry = huge_pte_wrprotect(entry);
5161 if (!userfaultfd_wp(dst_vma))
5162 entry = huge_pte_clear_uffd_wp(entry);
5164 set_huge_pte_at(dst, addr, dst_pte, entry);
5165 hugetlb_count_add(npages, dst);
5167 spin_unlock(src_ptl);
5168 spin_unlock(dst_ptl);
5172 raw_write_seqcount_end(&src->write_protect_seq);
5173 mmu_notifier_invalidate_range_end(&range);
5175 hugetlb_vma_unlock_read(src_vma);
5181 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5182 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5184 struct hstate *h = hstate_vma(vma);
5185 struct mm_struct *mm = vma->vm_mm;
5186 spinlock_t *src_ptl, *dst_ptl;
5189 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5190 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5193 * We don't have to worry about the ordering of src and dst ptlocks
5194 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5196 if (src_ptl != dst_ptl)
5197 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5199 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5200 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5202 if (src_ptl != dst_ptl)
5203 spin_unlock(src_ptl);
5204 spin_unlock(dst_ptl);
5207 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5208 struct vm_area_struct *new_vma,
5209 unsigned long old_addr, unsigned long new_addr,
5212 struct hstate *h = hstate_vma(vma);
5213 struct address_space *mapping = vma->vm_file->f_mapping;
5214 unsigned long sz = huge_page_size(h);
5215 struct mm_struct *mm = vma->vm_mm;
5216 unsigned long old_end = old_addr + len;
5217 unsigned long last_addr_mask;
5218 pte_t *src_pte, *dst_pte;
5219 struct mmu_notifier_range range;
5220 bool shared_pmd = false;
5222 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5224 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5226 * In case of shared PMDs, we should cover the maximum possible
5229 flush_cache_range(vma, range.start, range.end);
5231 mmu_notifier_invalidate_range_start(&range);
5232 last_addr_mask = hugetlb_mask_last_page(h);
5233 /* Prevent race with file truncation */
5234 hugetlb_vma_lock_write(vma);
5235 i_mmap_lock_write(mapping);
5236 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5237 src_pte = hugetlb_walk(vma, old_addr, sz);
5239 old_addr |= last_addr_mask;
5240 new_addr |= last_addr_mask;
5243 if (huge_pte_none(huge_ptep_get(src_pte)))
5246 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5248 old_addr |= last_addr_mask;
5249 new_addr |= last_addr_mask;
5253 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5257 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5261 flush_hugetlb_tlb_range(vma, range.start, range.end);
5263 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5264 mmu_notifier_invalidate_range_end(&range);
5265 i_mmap_unlock_write(mapping);
5266 hugetlb_vma_unlock_write(vma);
5268 return len + old_addr - old_end;
5271 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5272 unsigned long start, unsigned long end,
5273 struct page *ref_page, zap_flags_t zap_flags)
5275 struct mm_struct *mm = vma->vm_mm;
5276 unsigned long address;
5281 struct hstate *h = hstate_vma(vma);
5282 unsigned long sz = huge_page_size(h);
5283 unsigned long last_addr_mask;
5284 bool force_flush = false;
5286 WARN_ON(!is_vm_hugetlb_page(vma));
5287 BUG_ON(start & ~huge_page_mask(h));
5288 BUG_ON(end & ~huge_page_mask(h));
5291 * This is a hugetlb vma, all the pte entries should point
5294 tlb_change_page_size(tlb, sz);
5295 tlb_start_vma(tlb, vma);
5297 last_addr_mask = hugetlb_mask_last_page(h);
5299 for (; address < end; address += sz) {
5300 ptep = hugetlb_walk(vma, address, sz);
5302 address |= last_addr_mask;
5306 ptl = huge_pte_lock(h, mm, ptep);
5307 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5309 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5311 address |= last_addr_mask;
5315 pte = huge_ptep_get(ptep);
5316 if (huge_pte_none(pte)) {
5322 * Migrating hugepage or HWPoisoned hugepage is already
5323 * unmapped and its refcount is dropped, so just clear pte here.
5325 if (unlikely(!pte_present(pte))) {
5327 * If the pte was wr-protected by uffd-wp in any of the
5328 * swap forms, meanwhile the caller does not want to
5329 * drop the uffd-wp bit in this zap, then replace the
5330 * pte with a marker.
5332 if (pte_swp_uffd_wp_any(pte) &&
5333 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5334 set_huge_pte_at(mm, address, ptep,
5335 make_pte_marker(PTE_MARKER_UFFD_WP));
5337 huge_pte_clear(mm, address, ptep, sz);
5342 page = pte_page(pte);
5344 * If a reference page is supplied, it is because a specific
5345 * page is being unmapped, not a range. Ensure the page we
5346 * are about to unmap is the actual page of interest.
5349 if (page != ref_page) {
5354 * Mark the VMA as having unmapped its page so that
5355 * future faults in this VMA will fail rather than
5356 * looking like data was lost
5358 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5361 pte = huge_ptep_get_and_clear(mm, address, ptep);
5362 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5363 if (huge_pte_dirty(pte))
5364 set_page_dirty(page);
5365 /* Leave a uffd-wp pte marker if needed */
5366 if (huge_pte_uffd_wp(pte) &&
5367 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5368 set_huge_pte_at(mm, address, ptep,
5369 make_pte_marker(PTE_MARKER_UFFD_WP));
5370 hugetlb_count_sub(pages_per_huge_page(h), mm);
5371 page_remove_rmap(page, vma, true);
5374 tlb_remove_page_size(tlb, page, huge_page_size(h));
5376 * Bail out after unmapping reference page if supplied
5381 tlb_end_vma(tlb, vma);
5384 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5385 * could defer the flush until now, since by holding i_mmap_rwsem we
5386 * guaranteed that the last refernece would not be dropped. But we must
5387 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5388 * dropped and the last reference to the shared PMDs page might be
5391 * In theory we could defer the freeing of the PMD pages as well, but
5392 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5393 * detect sharing, so we cannot defer the release of the page either.
5394 * Instead, do flush now.
5397 tlb_flush_mmu_tlbonly(tlb);
5400 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5401 struct vm_area_struct *vma, unsigned long start,
5402 unsigned long end, struct page *ref_page,
5403 zap_flags_t zap_flags)
5405 hugetlb_vma_lock_write(vma);
5406 i_mmap_lock_write(vma->vm_file->f_mapping);
5408 /* mmu notification performed in caller */
5409 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5411 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5413 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5414 * When the vma_lock is freed, this makes the vma ineligible
5415 * for pmd sharing. And, i_mmap_rwsem is required to set up
5416 * pmd sharing. This is important as page tables for this
5417 * unmapped range will be asynchrously deleted. If the page
5418 * tables are shared, there will be issues when accessed by
5421 __hugetlb_vma_unlock_write_free(vma);
5422 i_mmap_unlock_write(vma->vm_file->f_mapping);
5424 i_mmap_unlock_write(vma->vm_file->f_mapping);
5425 hugetlb_vma_unlock_write(vma);
5429 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5430 unsigned long end, struct page *ref_page,
5431 zap_flags_t zap_flags)
5433 struct mmu_notifier_range range;
5434 struct mmu_gather tlb;
5436 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5438 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5439 mmu_notifier_invalidate_range_start(&range);
5440 tlb_gather_mmu(&tlb, vma->vm_mm);
5442 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5444 mmu_notifier_invalidate_range_end(&range);
5445 tlb_finish_mmu(&tlb);
5449 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5450 * mapping it owns the reserve page for. The intention is to unmap the page
5451 * from other VMAs and let the children be SIGKILLed if they are faulting the
5454 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5455 struct page *page, unsigned long address)
5457 struct hstate *h = hstate_vma(vma);
5458 struct vm_area_struct *iter_vma;
5459 struct address_space *mapping;
5463 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5464 * from page cache lookup which is in HPAGE_SIZE units.
5466 address = address & huge_page_mask(h);
5467 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5469 mapping = vma->vm_file->f_mapping;
5472 * Take the mapping lock for the duration of the table walk. As
5473 * this mapping should be shared between all the VMAs,
5474 * __unmap_hugepage_range() is called as the lock is already held
5476 i_mmap_lock_write(mapping);
5477 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5478 /* Do not unmap the current VMA */
5479 if (iter_vma == vma)
5483 * Shared VMAs have their own reserves and do not affect
5484 * MAP_PRIVATE accounting but it is possible that a shared
5485 * VMA is using the same page so check and skip such VMAs.
5487 if (iter_vma->vm_flags & VM_MAYSHARE)
5491 * Unmap the page from other VMAs without their own reserves.
5492 * They get marked to be SIGKILLed if they fault in these
5493 * areas. This is because a future no-page fault on this VMA
5494 * could insert a zeroed page instead of the data existing
5495 * from the time of fork. This would look like data corruption
5497 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5498 unmap_hugepage_range(iter_vma, address,
5499 address + huge_page_size(h), page, 0);
5501 i_mmap_unlock_write(mapping);
5505 * hugetlb_wp() should be called with page lock of the original hugepage held.
5506 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5507 * cannot race with other handlers or page migration.
5508 * Keep the pte_same checks anyway to make transition from the mutex easier.
5510 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5511 unsigned long address, pte_t *ptep, unsigned int flags,
5512 struct folio *pagecache_folio, spinlock_t *ptl)
5514 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5515 pte_t pte = huge_ptep_get(ptep);
5516 struct hstate *h = hstate_vma(vma);
5517 struct folio *old_folio;
5518 struct folio *new_folio;
5519 int outside_reserve = 0;
5521 unsigned long haddr = address & huge_page_mask(h);
5522 struct mmu_notifier_range range;
5525 * Never handle CoW for uffd-wp protected pages. It should be only
5526 * handled when the uffd-wp protection is removed.
5528 * Note that only the CoW optimization path (in hugetlb_no_page())
5529 * can trigger this, because hugetlb_fault() will always resolve
5530 * uffd-wp bit first.
5532 if (!unshare && huge_pte_uffd_wp(pte))
5536 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5537 * PTE mapped R/O such as maybe_mkwrite() would do.
5539 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5540 return VM_FAULT_SIGSEGV;
5542 /* Let's take out MAP_SHARED mappings first. */
5543 if (vma->vm_flags & VM_MAYSHARE) {
5544 set_huge_ptep_writable(vma, haddr, ptep);
5548 old_folio = page_folio(pte_page(pte));
5550 delayacct_wpcopy_start();
5554 * If no-one else is actually using this page, we're the exclusive
5555 * owner and can reuse this page.
5557 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5558 if (!PageAnonExclusive(&old_folio->page))
5559 page_move_anon_rmap(&old_folio->page, vma);
5560 if (likely(!unshare))
5561 set_huge_ptep_writable(vma, haddr, ptep);
5563 delayacct_wpcopy_end();
5566 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5567 PageAnonExclusive(&old_folio->page), &old_folio->page);
5570 * If the process that created a MAP_PRIVATE mapping is about to
5571 * perform a COW due to a shared page count, attempt to satisfy
5572 * the allocation without using the existing reserves. The pagecache
5573 * page is used to determine if the reserve at this address was
5574 * consumed or not. If reserves were used, a partial faulted mapping
5575 * at the time of fork() could consume its reserves on COW instead
5576 * of the full address range.
5578 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5579 old_folio != pagecache_folio)
5580 outside_reserve = 1;
5582 folio_get(old_folio);
5585 * Drop page table lock as buddy allocator may be called. It will
5586 * be acquired again before returning to the caller, as expected.
5589 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5591 if (IS_ERR(new_folio)) {
5593 * If a process owning a MAP_PRIVATE mapping fails to COW,
5594 * it is due to references held by a child and an insufficient
5595 * huge page pool. To guarantee the original mappers
5596 * reliability, unmap the page from child processes. The child
5597 * may get SIGKILLed if it later faults.
5599 if (outside_reserve) {
5600 struct address_space *mapping = vma->vm_file->f_mapping;
5604 folio_put(old_folio);
5606 * Drop hugetlb_fault_mutex and vma_lock before
5607 * unmapping. unmapping needs to hold vma_lock
5608 * in write mode. Dropping vma_lock in read mode
5609 * here is OK as COW mappings do not interact with
5612 * Reacquire both after unmap operation.
5614 idx = vma_hugecache_offset(h, vma, haddr);
5615 hash = hugetlb_fault_mutex_hash(mapping, idx);
5616 hugetlb_vma_unlock_read(vma);
5617 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5619 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5621 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5622 hugetlb_vma_lock_read(vma);
5624 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5626 pte_same(huge_ptep_get(ptep), pte)))
5627 goto retry_avoidcopy;
5629 * race occurs while re-acquiring page table
5630 * lock, and our job is done.
5632 delayacct_wpcopy_end();
5636 ret = vmf_error(PTR_ERR(new_folio));
5637 goto out_release_old;
5641 * When the original hugepage is shared one, it does not have
5642 * anon_vma prepared.
5644 if (unlikely(anon_vma_prepare(vma))) {
5646 goto out_release_all;
5649 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5650 ret = VM_FAULT_HWPOISON_LARGE;
5651 goto out_release_all;
5653 __folio_mark_uptodate(new_folio);
5655 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5656 haddr + huge_page_size(h));
5657 mmu_notifier_invalidate_range_start(&range);
5660 * Retake the page table lock to check for racing updates
5661 * before the page tables are altered
5664 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5665 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5666 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5668 /* Break COW or unshare */
5669 huge_ptep_clear_flush(vma, haddr, ptep);
5670 page_remove_rmap(&old_folio->page, vma, true);
5671 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5672 if (huge_pte_uffd_wp(pte))
5673 newpte = huge_pte_mkuffd_wp(newpte);
5674 set_huge_pte_at(mm, haddr, ptep, newpte);
5675 folio_set_hugetlb_migratable(new_folio);
5676 /* Make the old page be freed below */
5677 new_folio = old_folio;
5680 mmu_notifier_invalidate_range_end(&range);
5683 * No restore in case of successful pagetable update (Break COW or
5686 if (new_folio != old_folio)
5687 restore_reserve_on_error(h, vma, haddr, new_folio);
5688 folio_put(new_folio);
5690 folio_put(old_folio);
5692 spin_lock(ptl); /* Caller expects lock to be held */
5694 delayacct_wpcopy_end();
5699 * Return whether there is a pagecache page to back given address within VMA.
5701 static bool hugetlbfs_pagecache_present(struct hstate *h,
5702 struct vm_area_struct *vma, unsigned long address)
5704 struct address_space *mapping = vma->vm_file->f_mapping;
5705 pgoff_t idx = vma_hugecache_offset(h, vma, address);
5706 struct folio *folio;
5708 folio = filemap_get_folio(mapping, idx);
5715 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5718 struct inode *inode = mapping->host;
5719 struct hstate *h = hstate_inode(inode);
5722 __folio_set_locked(folio);
5723 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5725 if (unlikely(err)) {
5726 __folio_clear_locked(folio);
5729 folio_clear_hugetlb_restore_reserve(folio);
5732 * mark folio dirty so that it will not be removed from cache/file
5733 * by non-hugetlbfs specific code paths.
5735 folio_mark_dirty(folio);
5737 spin_lock(&inode->i_lock);
5738 inode->i_blocks += blocks_per_huge_page(h);
5739 spin_unlock(&inode->i_lock);
5743 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5744 struct address_space *mapping,
5747 unsigned long haddr,
5749 unsigned long reason)
5752 struct vm_fault vmf = {
5755 .real_address = addr,
5759 * Hard to debug if it ends up being
5760 * used by a callee that assumes
5761 * something about the other
5762 * uninitialized fields... same as in
5768 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5769 * userfault. Also mmap_lock could be dropped due to handling
5770 * userfault, any vma operation should be careful from here.
5772 hugetlb_vma_unlock_read(vma);
5773 hash = hugetlb_fault_mutex_hash(mapping, idx);
5774 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5775 return handle_userfault(&vmf, reason);
5779 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5780 * false if pte changed or is changing.
5782 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5783 pte_t *ptep, pte_t old_pte)
5788 ptl = huge_pte_lock(h, mm, ptep);
5789 same = pte_same(huge_ptep_get(ptep), old_pte);
5795 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5796 struct vm_area_struct *vma,
5797 struct address_space *mapping, pgoff_t idx,
5798 unsigned long address, pte_t *ptep,
5799 pte_t old_pte, unsigned int flags)
5801 struct hstate *h = hstate_vma(vma);
5802 vm_fault_t ret = VM_FAULT_SIGBUS;
5805 struct folio *folio;
5808 unsigned long haddr = address & huge_page_mask(h);
5809 bool new_folio, new_pagecache_folio = false;
5810 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5813 * Currently, we are forced to kill the process in the event the
5814 * original mapper has unmapped pages from the child due to a failed
5815 * COW/unsharing. Warn that such a situation has occurred as it may not
5818 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5819 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5825 * Use page lock to guard against racing truncation
5826 * before we get page_table_lock.
5829 folio = filemap_lock_folio(mapping, idx);
5830 if (IS_ERR(folio)) {
5831 size = i_size_read(mapping->host) >> huge_page_shift(h);
5834 /* Check for page in userfault range */
5835 if (userfaultfd_missing(vma)) {
5837 * Since hugetlb_no_page() was examining pte
5838 * without pgtable lock, we need to re-test under
5839 * lock because the pte may not be stable and could
5840 * have changed from under us. Try to detect
5841 * either changed or during-changing ptes and retry
5842 * properly when needed.
5844 * Note that userfaultfd is actually fine with
5845 * false positives (e.g. caused by pte changed),
5846 * but not wrong logical events (e.g. caused by
5847 * reading a pte during changing). The latter can
5848 * confuse the userspace, so the strictness is very
5849 * much preferred. E.g., MISSING event should
5850 * never happen on the page after UFFDIO_COPY has
5851 * correctly installed the page and returned.
5853 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5858 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5863 folio = alloc_hugetlb_folio(vma, haddr, 0);
5864 if (IS_ERR(folio)) {
5866 * Returning error will result in faulting task being
5867 * sent SIGBUS. The hugetlb fault mutex prevents two
5868 * tasks from racing to fault in the same page which
5869 * could result in false unable to allocate errors.
5870 * Page migration does not take the fault mutex, but
5871 * does a clear then write of pte's under page table
5872 * lock. Page fault code could race with migration,
5873 * notice the clear pte and try to allocate a page
5874 * here. Before returning error, get ptl and make
5875 * sure there really is no pte entry.
5877 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5878 ret = vmf_error(PTR_ERR(folio));
5883 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5884 __folio_mark_uptodate(folio);
5887 if (vma->vm_flags & VM_MAYSHARE) {
5888 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5891 * err can't be -EEXIST which implies someone
5892 * else consumed the reservation since hugetlb
5893 * fault mutex is held when add a hugetlb page
5894 * to the page cache. So it's safe to call
5895 * restore_reserve_on_error() here.
5897 restore_reserve_on_error(h, vma, haddr, folio);
5901 new_pagecache_folio = true;
5904 if (unlikely(anon_vma_prepare(vma))) {
5906 goto backout_unlocked;
5912 * If memory error occurs between mmap() and fault, some process
5913 * don't have hwpoisoned swap entry for errored virtual address.
5914 * So we need to block hugepage fault by PG_hwpoison bit check.
5916 if (unlikely(folio_test_hwpoison(folio))) {
5917 ret = VM_FAULT_HWPOISON_LARGE |
5918 VM_FAULT_SET_HINDEX(hstate_index(h));
5919 goto backout_unlocked;
5922 /* Check for page in userfault range. */
5923 if (userfaultfd_minor(vma)) {
5924 folio_unlock(folio);
5926 /* See comment in userfaultfd_missing() block above */
5927 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5931 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5938 * If we are going to COW a private mapping later, we examine the
5939 * pending reservations for this page now. This will ensure that
5940 * any allocations necessary to record that reservation occur outside
5943 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5944 if (vma_needs_reservation(h, vma, haddr) < 0) {
5946 goto backout_unlocked;
5948 /* Just decrements count, does not deallocate */
5949 vma_end_reservation(h, vma, haddr);
5952 ptl = huge_pte_lock(h, mm, ptep);
5954 /* If pte changed from under us, retry */
5955 if (!pte_same(huge_ptep_get(ptep), old_pte))
5959 hugepage_add_new_anon_rmap(folio, vma, haddr);
5961 page_dup_file_rmap(&folio->page, true);
5962 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
5963 && (vma->vm_flags & VM_SHARED)));
5965 * If this pte was previously wr-protected, keep it wr-protected even
5968 if (unlikely(pte_marker_uffd_wp(old_pte)))
5969 new_pte = huge_pte_mkuffd_wp(new_pte);
5970 set_huge_pte_at(mm, haddr, ptep, new_pte);
5972 hugetlb_count_add(pages_per_huge_page(h), mm);
5973 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5974 /* Optimization, do the COW without a second fault */
5975 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
5981 * Only set hugetlb_migratable in newly allocated pages. Existing pages
5982 * found in the pagecache may not have hugetlb_migratable if they have
5983 * been isolated for migration.
5986 folio_set_hugetlb_migratable(folio);
5988 folio_unlock(folio);
5990 hugetlb_vma_unlock_read(vma);
5991 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5997 if (new_folio && !new_pagecache_folio)
5998 restore_reserve_on_error(h, vma, haddr, folio);
6000 folio_unlock(folio);
6006 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6008 unsigned long key[2];
6011 key[0] = (unsigned long) mapping;
6014 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6016 return hash & (num_fault_mutexes - 1);
6020 * For uniprocessor systems we always use a single mutex, so just
6021 * return 0 and avoid the hashing overhead.
6023 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6029 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6030 unsigned long address, unsigned int flags)
6037 struct folio *folio = NULL;
6038 struct folio *pagecache_folio = NULL;
6039 struct hstate *h = hstate_vma(vma);
6040 struct address_space *mapping;
6041 int need_wait_lock = 0;
6042 unsigned long haddr = address & huge_page_mask(h);
6044 /* TODO: Handle faults under the VMA lock */
6045 if (flags & FAULT_FLAG_VMA_LOCK) {
6047 return VM_FAULT_RETRY;
6051 * Serialize hugepage allocation and instantiation, so that we don't
6052 * get spurious allocation failures if two CPUs race to instantiate
6053 * the same page in the page cache.
6055 mapping = vma->vm_file->f_mapping;
6056 idx = vma_hugecache_offset(h, vma, haddr);
6057 hash = hugetlb_fault_mutex_hash(mapping, idx);
6058 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6061 * Acquire vma lock before calling huge_pte_alloc and hold
6062 * until finished with ptep. This prevents huge_pmd_unshare from
6063 * being called elsewhere and making the ptep no longer valid.
6065 hugetlb_vma_lock_read(vma);
6066 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6068 hugetlb_vma_unlock_read(vma);
6069 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6070 return VM_FAULT_OOM;
6073 entry = huge_ptep_get(ptep);
6074 if (huge_pte_none_mostly(entry)) {
6075 if (is_pte_marker(entry)) {
6077 pte_marker_get(pte_to_swp_entry(entry));
6079 if (marker & PTE_MARKER_POISONED) {
6080 ret = VM_FAULT_HWPOISON_LARGE;
6086 * Other PTE markers should be handled the same way as none PTE.
6088 * hugetlb_no_page will drop vma lock and hugetlb fault
6089 * mutex internally, which make us return immediately.
6091 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6098 * entry could be a migration/hwpoison entry at this point, so this
6099 * check prevents the kernel from going below assuming that we have
6100 * an active hugepage in pagecache. This goto expects the 2nd page
6101 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6102 * properly handle it.
6104 if (!pte_present(entry)) {
6105 if (unlikely(is_hugetlb_entry_migration(entry))) {
6107 * Release the hugetlb fault lock now, but retain
6108 * the vma lock, because it is needed to guard the
6109 * huge_pte_lockptr() later in
6110 * migration_entry_wait_huge(). The vma lock will
6111 * be released there.
6113 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6114 migration_entry_wait_huge(vma, ptep);
6116 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6117 ret = VM_FAULT_HWPOISON_LARGE |
6118 VM_FAULT_SET_HINDEX(hstate_index(h));
6123 * If we are going to COW/unshare the mapping later, we examine the
6124 * pending reservations for this page now. This will ensure that any
6125 * allocations necessary to record that reservation occur outside the
6126 * spinlock. Also lookup the pagecache page now as it is used to
6127 * determine if a reservation has been consumed.
6129 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6130 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6131 if (vma_needs_reservation(h, vma, haddr) < 0) {
6135 /* Just decrements count, does not deallocate */
6136 vma_end_reservation(h, vma, haddr);
6138 pagecache_folio = filemap_lock_folio(mapping, idx);
6139 if (IS_ERR(pagecache_folio))
6140 pagecache_folio = NULL;
6143 ptl = huge_pte_lock(h, mm, ptep);
6145 /* Check for a racing update before calling hugetlb_wp() */
6146 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6149 /* Handle userfault-wp first, before trying to lock more pages */
6150 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6151 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6152 struct vm_fault vmf = {
6155 .real_address = address,
6160 if (pagecache_folio) {
6161 folio_unlock(pagecache_folio);
6162 folio_put(pagecache_folio);
6164 hugetlb_vma_unlock_read(vma);
6165 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6166 return handle_userfault(&vmf, VM_UFFD_WP);
6170 * hugetlb_wp() requires page locks of pte_page(entry) and
6171 * pagecache_folio, so here we need take the former one
6172 * when folio != pagecache_folio or !pagecache_folio.
6174 folio = page_folio(pte_page(entry));
6175 if (folio != pagecache_folio)
6176 if (!folio_trylock(folio)) {
6183 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6184 if (!huge_pte_write(entry)) {
6185 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6186 pagecache_folio, ptl);
6188 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6189 entry = huge_pte_mkdirty(entry);
6192 entry = pte_mkyoung(entry);
6193 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6194 flags & FAULT_FLAG_WRITE))
6195 update_mmu_cache(vma, haddr, ptep);
6197 if (folio != pagecache_folio)
6198 folio_unlock(folio);
6203 if (pagecache_folio) {
6204 folio_unlock(pagecache_folio);
6205 folio_put(pagecache_folio);
6208 hugetlb_vma_unlock_read(vma);
6209 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6211 * Generally it's safe to hold refcount during waiting page lock. But
6212 * here we just wait to defer the next page fault to avoid busy loop and
6213 * the page is not used after unlocked before returning from the current
6214 * page fault. So we are safe from accessing freed page, even if we wait
6215 * here without taking refcount.
6218 folio_wait_locked(folio);
6222 #ifdef CONFIG_USERFAULTFD
6224 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6225 * with modifications for hugetlb pages.
6227 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6228 struct vm_area_struct *dst_vma,
6229 unsigned long dst_addr,
6230 unsigned long src_addr,
6232 struct folio **foliop)
6234 struct mm_struct *dst_mm = dst_vma->vm_mm;
6235 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6236 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6237 struct hstate *h = hstate_vma(dst_vma);
6238 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6239 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6241 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6245 struct folio *folio;
6247 bool folio_in_pagecache = false;
6249 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6250 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6252 /* Don't overwrite any existing PTEs (even markers) */
6253 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6258 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6259 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6261 /* No need to invalidate - it was non-present before */
6262 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6270 folio = filemap_lock_folio(mapping, idx);
6273 folio_in_pagecache = true;
6274 } else if (!*foliop) {
6275 /* If a folio already exists, then it's UFFDIO_COPY for
6276 * a non-missing case. Return -EEXIST.
6279 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6284 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6285 if (IS_ERR(folio)) {
6290 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6293 /* fallback to copy_from_user outside mmap_lock */
6294 if (unlikely(ret)) {
6296 /* Free the allocated folio which may have
6297 * consumed a reservation.
6299 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6302 /* Allocate a temporary folio to hold the copied
6305 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6311 /* Set the outparam foliop and return to the caller to
6312 * copy the contents outside the lock. Don't free the
6319 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6326 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6327 if (IS_ERR(folio)) {
6333 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6343 * The memory barrier inside __folio_mark_uptodate makes sure that
6344 * preceding stores to the page contents become visible before
6345 * the set_pte_at() write.
6347 __folio_mark_uptodate(folio);
6349 /* Add shared, newly allocated pages to the page cache. */
6350 if (vm_shared && !is_continue) {
6351 size = i_size_read(mapping->host) >> huge_page_shift(h);
6354 goto out_release_nounlock;
6357 * Serialization between remove_inode_hugepages() and
6358 * hugetlb_add_to_page_cache() below happens through the
6359 * hugetlb_fault_mutex_table that here must be hold by
6362 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6364 goto out_release_nounlock;
6365 folio_in_pagecache = true;
6368 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6371 if (folio_test_hwpoison(folio))
6372 goto out_release_unlock;
6375 * We allow to overwrite a pte marker: consider when both MISSING|WP
6376 * registered, we firstly wr-protect a none pte which has no page cache
6377 * page backing it, then access the page.
6380 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6381 goto out_release_unlock;
6383 if (folio_in_pagecache)
6384 page_dup_file_rmap(&folio->page, true);
6386 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6389 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6390 * with wp flag set, don't set pte write bit.
6392 if (wp_enabled || (is_continue && !vm_shared))
6395 writable = dst_vma->vm_flags & VM_WRITE;
6397 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6399 * Always mark UFFDIO_COPY page dirty; note that this may not be
6400 * extremely important for hugetlbfs for now since swapping is not
6401 * supported, but we should still be clear in that this page cannot be
6402 * thrown away at will, even if write bit not set.
6404 _dst_pte = huge_pte_mkdirty(_dst_pte);
6405 _dst_pte = pte_mkyoung(_dst_pte);
6408 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6410 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6412 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6414 /* No need to invalidate - it was non-present before */
6415 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6419 folio_set_hugetlb_migratable(folio);
6420 if (vm_shared || is_continue)
6421 folio_unlock(folio);
6427 if (vm_shared || is_continue)
6428 folio_unlock(folio);
6429 out_release_nounlock:
6430 if (!folio_in_pagecache)
6431 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6435 #endif /* CONFIG_USERFAULTFD */
6437 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6438 unsigned long address, unsigned int flags,
6439 unsigned int *page_mask)
6441 struct hstate *h = hstate_vma(vma);
6442 struct mm_struct *mm = vma->vm_mm;
6443 unsigned long haddr = address & huge_page_mask(h);
6444 struct page *page = NULL;
6449 hugetlb_vma_lock_read(vma);
6450 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6454 ptl = huge_pte_lock(h, mm, pte);
6455 entry = huge_ptep_get(pte);
6456 if (pte_present(entry)) {
6457 page = pte_page(entry);
6459 if (!huge_pte_write(entry)) {
6460 if (flags & FOLL_WRITE) {
6465 if (gup_must_unshare(vma, flags, page)) {
6466 /* Tell the caller to do unsharing */
6467 page = ERR_PTR(-EMLINK);
6472 page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6475 * Note that page may be a sub-page, and with vmemmap
6476 * optimizations the page struct may be read only.
6477 * try_grab_page() will increase the ref count on the
6478 * head page, so this will be OK.
6480 * try_grab_page() should always be able to get the page here,
6481 * because we hold the ptl lock and have verified pte_present().
6483 ret = try_grab_page(page, flags);
6485 if (WARN_ON_ONCE(ret)) {
6486 page = ERR_PTR(ret);
6490 *page_mask = (1U << huge_page_order(h)) - 1;
6495 hugetlb_vma_unlock_read(vma);
6498 * Fixup retval for dump requests: if pagecache doesn't exist,
6499 * don't try to allocate a new page but just skip it.
6501 if (!page && (flags & FOLL_DUMP) &&
6502 !hugetlbfs_pagecache_present(h, vma, address))
6503 page = ERR_PTR(-EFAULT);
6508 long hugetlb_change_protection(struct vm_area_struct *vma,
6509 unsigned long address, unsigned long end,
6510 pgprot_t newprot, unsigned long cp_flags)
6512 struct mm_struct *mm = vma->vm_mm;
6513 unsigned long start = address;
6516 struct hstate *h = hstate_vma(vma);
6517 long pages = 0, psize = huge_page_size(h);
6518 bool shared_pmd = false;
6519 struct mmu_notifier_range range;
6520 unsigned long last_addr_mask;
6521 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6522 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6525 * In the case of shared PMDs, the area to flush could be beyond
6526 * start/end. Set range.start/range.end to cover the maximum possible
6527 * range if PMD sharing is possible.
6529 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6531 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6533 BUG_ON(address >= end);
6534 flush_cache_range(vma, range.start, range.end);
6536 mmu_notifier_invalidate_range_start(&range);
6537 hugetlb_vma_lock_write(vma);
6538 i_mmap_lock_write(vma->vm_file->f_mapping);
6539 last_addr_mask = hugetlb_mask_last_page(h);
6540 for (; address < end; address += psize) {
6542 ptep = hugetlb_walk(vma, address, psize);
6545 address |= last_addr_mask;
6549 * Userfaultfd wr-protect requires pgtable
6550 * pre-allocations to install pte markers.
6552 ptep = huge_pte_alloc(mm, vma, address, psize);
6558 ptl = huge_pte_lock(h, mm, ptep);
6559 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6561 * When uffd-wp is enabled on the vma, unshare
6562 * shouldn't happen at all. Warn about it if it
6563 * happened due to some reason.
6565 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6569 address |= last_addr_mask;
6572 pte = huge_ptep_get(ptep);
6573 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6574 /* Nothing to do. */
6575 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6576 swp_entry_t entry = pte_to_swp_entry(pte);
6577 struct page *page = pfn_swap_entry_to_page(entry);
6580 if (is_writable_migration_entry(entry)) {
6582 entry = make_readable_exclusive_migration_entry(
6585 entry = make_readable_migration_entry(
6587 newpte = swp_entry_to_pte(entry);
6592 newpte = pte_swp_mkuffd_wp(newpte);
6593 else if (uffd_wp_resolve)
6594 newpte = pte_swp_clear_uffd_wp(newpte);
6595 if (!pte_same(pte, newpte))
6596 set_huge_pte_at(mm, address, ptep, newpte);
6597 } else if (unlikely(is_pte_marker(pte))) {
6598 /* No other markers apply for now. */
6599 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6600 if (uffd_wp_resolve)
6601 /* Safe to modify directly (non-present->none). */
6602 huge_pte_clear(mm, address, ptep, psize);
6603 } else if (!huge_pte_none(pte)) {
6605 unsigned int shift = huge_page_shift(hstate_vma(vma));
6607 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6608 pte = huge_pte_modify(old_pte, newprot);
6609 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6611 pte = huge_pte_mkuffd_wp(pte);
6612 else if (uffd_wp_resolve)
6613 pte = huge_pte_clear_uffd_wp(pte);
6614 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6618 if (unlikely(uffd_wp))
6619 /* Safe to modify directly (none->non-present). */
6620 set_huge_pte_at(mm, address, ptep,
6621 make_pte_marker(PTE_MARKER_UFFD_WP));
6626 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6627 * may have cleared our pud entry and done put_page on the page table:
6628 * once we release i_mmap_rwsem, another task can do the final put_page
6629 * and that page table be reused and filled with junk. If we actually
6630 * did unshare a page of pmds, flush the range corresponding to the pud.
6633 flush_hugetlb_tlb_range(vma, range.start, range.end);
6635 flush_hugetlb_tlb_range(vma, start, end);
6637 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6638 * downgrading page table protection not changing it to point to a new
6641 * See Documentation/mm/mmu_notifier.rst
6643 i_mmap_unlock_write(vma->vm_file->f_mapping);
6644 hugetlb_vma_unlock_write(vma);
6645 mmu_notifier_invalidate_range_end(&range);
6647 return pages > 0 ? (pages << h->order) : pages;
6650 /* Return true if reservation was successful, false otherwise. */
6651 bool hugetlb_reserve_pages(struct inode *inode,
6653 struct vm_area_struct *vma,
6654 vm_flags_t vm_flags)
6656 long chg = -1, add = -1;
6657 struct hstate *h = hstate_inode(inode);
6658 struct hugepage_subpool *spool = subpool_inode(inode);
6659 struct resv_map *resv_map;
6660 struct hugetlb_cgroup *h_cg = NULL;
6661 long gbl_reserve, regions_needed = 0;
6663 /* This should never happen */
6665 VM_WARN(1, "%s called with a negative range\n", __func__);
6670 * vma specific semaphore used for pmd sharing and fault/truncation
6673 hugetlb_vma_lock_alloc(vma);
6676 * Only apply hugepage reservation if asked. At fault time, an
6677 * attempt will be made for VM_NORESERVE to allocate a page
6678 * without using reserves
6680 if (vm_flags & VM_NORESERVE)
6684 * Shared mappings base their reservation on the number of pages that
6685 * are already allocated on behalf of the file. Private mappings need
6686 * to reserve the full area even if read-only as mprotect() may be
6687 * called to make the mapping read-write. Assume !vma is a shm mapping
6689 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6691 * resv_map can not be NULL as hugetlb_reserve_pages is only
6692 * called for inodes for which resv_maps were created (see
6693 * hugetlbfs_get_inode).
6695 resv_map = inode_resv_map(inode);
6697 chg = region_chg(resv_map, from, to, ®ions_needed);
6699 /* Private mapping. */
6700 resv_map = resv_map_alloc();
6706 set_vma_resv_map(vma, resv_map);
6707 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6713 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6714 chg * pages_per_huge_page(h), &h_cg) < 0)
6717 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6718 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6721 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6725 * There must be enough pages in the subpool for the mapping. If
6726 * the subpool has a minimum size, there may be some global
6727 * reservations already in place (gbl_reserve).
6729 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6730 if (gbl_reserve < 0)
6731 goto out_uncharge_cgroup;
6734 * Check enough hugepages are available for the reservation.
6735 * Hand the pages back to the subpool if there are not
6737 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6741 * Account for the reservations made. Shared mappings record regions
6742 * that have reservations as they are shared by multiple VMAs.
6743 * When the last VMA disappears, the region map says how much
6744 * the reservation was and the page cache tells how much of
6745 * the reservation was consumed. Private mappings are per-VMA and
6746 * only the consumed reservations are tracked. When the VMA
6747 * disappears, the original reservation is the VMA size and the
6748 * consumed reservations are stored in the map. Hence, nothing
6749 * else has to be done for private mappings here
6751 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6752 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6754 if (unlikely(add < 0)) {
6755 hugetlb_acct_memory(h, -gbl_reserve);
6757 } else if (unlikely(chg > add)) {
6759 * pages in this range were added to the reserve
6760 * map between region_chg and region_add. This
6761 * indicates a race with alloc_hugetlb_folio. Adjust
6762 * the subpool and reserve counts modified above
6763 * based on the difference.
6768 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6769 * reference to h_cg->css. See comment below for detail.
6771 hugetlb_cgroup_uncharge_cgroup_rsvd(
6773 (chg - add) * pages_per_huge_page(h), h_cg);
6775 rsv_adjust = hugepage_subpool_put_pages(spool,
6777 hugetlb_acct_memory(h, -rsv_adjust);
6780 * The file_regions will hold their own reference to
6781 * h_cg->css. So we should release the reference held
6782 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6785 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6791 /* put back original number of pages, chg */
6792 (void)hugepage_subpool_put_pages(spool, chg);
6793 out_uncharge_cgroup:
6794 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6795 chg * pages_per_huge_page(h), h_cg);
6797 hugetlb_vma_lock_free(vma);
6798 if (!vma || vma->vm_flags & VM_MAYSHARE)
6799 /* Only call region_abort if the region_chg succeeded but the
6800 * region_add failed or didn't run.
6802 if (chg >= 0 && add < 0)
6803 region_abort(resv_map, from, to, regions_needed);
6804 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6805 kref_put(&resv_map->refs, resv_map_release);
6809 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6812 struct hstate *h = hstate_inode(inode);
6813 struct resv_map *resv_map = inode_resv_map(inode);
6815 struct hugepage_subpool *spool = subpool_inode(inode);
6819 * Since this routine can be called in the evict inode path for all
6820 * hugetlbfs inodes, resv_map could be NULL.
6823 chg = region_del(resv_map, start, end);
6825 * region_del() can fail in the rare case where a region
6826 * must be split and another region descriptor can not be
6827 * allocated. If end == LONG_MAX, it will not fail.
6833 spin_lock(&inode->i_lock);
6834 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6835 spin_unlock(&inode->i_lock);
6838 * If the subpool has a minimum size, the number of global
6839 * reservations to be released may be adjusted.
6841 * Note that !resv_map implies freed == 0. So (chg - freed)
6842 * won't go negative.
6844 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6845 hugetlb_acct_memory(h, -gbl_reserve);
6850 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6851 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6852 struct vm_area_struct *vma,
6853 unsigned long addr, pgoff_t idx)
6855 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6857 unsigned long sbase = saddr & PUD_MASK;
6858 unsigned long s_end = sbase + PUD_SIZE;
6860 /* Allow segments to share if only one is marked locked */
6861 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6862 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6865 * match the virtual addresses, permission and the alignment of the
6868 * Also, vma_lock (vm_private_data) is required for sharing.
6870 if (pmd_index(addr) != pmd_index(saddr) ||
6871 vm_flags != svm_flags ||
6872 !range_in_vma(svma, sbase, s_end) ||
6873 !svma->vm_private_data)
6879 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6881 unsigned long start = addr & PUD_MASK;
6882 unsigned long end = start + PUD_SIZE;
6884 #ifdef CONFIG_USERFAULTFD
6885 if (uffd_disable_huge_pmd_share(vma))
6889 * check on proper vm_flags and page table alignment
6891 if (!(vma->vm_flags & VM_MAYSHARE))
6893 if (!vma->vm_private_data) /* vma lock required for sharing */
6895 if (!range_in_vma(vma, start, end))
6901 * Determine if start,end range within vma could be mapped by shared pmd.
6902 * If yes, adjust start and end to cover range associated with possible
6903 * shared pmd mappings.
6905 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6906 unsigned long *start, unsigned long *end)
6908 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6909 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6912 * vma needs to span at least one aligned PUD size, and the range
6913 * must be at least partially within in.
6915 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6916 (*end <= v_start) || (*start >= v_end))
6919 /* Extend the range to be PUD aligned for a worst case scenario */
6920 if (*start > v_start)
6921 *start = ALIGN_DOWN(*start, PUD_SIZE);
6924 *end = ALIGN(*end, PUD_SIZE);
6928 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6929 * and returns the corresponding pte. While this is not necessary for the
6930 * !shared pmd case because we can allocate the pmd later as well, it makes the
6931 * code much cleaner. pmd allocation is essential for the shared case because
6932 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6933 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6934 * bad pmd for sharing.
6936 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6937 unsigned long addr, pud_t *pud)
6939 struct address_space *mapping = vma->vm_file->f_mapping;
6940 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6942 struct vm_area_struct *svma;
6943 unsigned long saddr;
6947 i_mmap_lock_read(mapping);
6948 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6952 saddr = page_table_shareable(svma, vma, addr, idx);
6954 spte = hugetlb_walk(svma, saddr,
6955 vma_mmu_pagesize(svma));
6957 get_page(virt_to_page(spte));
6966 spin_lock(&mm->page_table_lock);
6967 if (pud_none(*pud)) {
6968 pud_populate(mm, pud,
6969 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6972 put_page(virt_to_page(spte));
6974 spin_unlock(&mm->page_table_lock);
6976 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6977 i_mmap_unlock_read(mapping);
6982 * unmap huge page backed by shared pte.
6984 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6985 * indicated by page_count > 1, unmap is achieved by clearing pud and
6986 * decrementing the ref count. If count == 1, the pte page is not shared.
6988 * Called with page table lock held.
6990 * returns: 1 successfully unmapped a shared pte page
6991 * 0 the underlying pte page is not shared, or it is the last user
6993 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6994 unsigned long addr, pte_t *ptep)
6996 pgd_t *pgd = pgd_offset(mm, addr);
6997 p4d_t *p4d = p4d_offset(pgd, addr);
6998 pud_t *pud = pud_offset(p4d, addr);
7000 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7001 hugetlb_vma_assert_locked(vma);
7002 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7003 if (page_count(virt_to_page(ptep)) == 1)
7007 put_page(virt_to_page(ptep));
7012 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7014 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7015 unsigned long addr, pud_t *pud)
7020 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7021 unsigned long addr, pte_t *ptep)
7026 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7027 unsigned long *start, unsigned long *end)
7031 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7035 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7037 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7038 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7039 unsigned long addr, unsigned long sz)
7046 pgd = pgd_offset(mm, addr);
7047 p4d = p4d_alloc(mm, pgd, addr);
7050 pud = pud_alloc(mm, p4d, addr);
7052 if (sz == PUD_SIZE) {
7055 BUG_ON(sz != PMD_SIZE);
7056 if (want_pmd_share(vma, addr) && pud_none(*pud))
7057 pte = huge_pmd_share(mm, vma, addr, pud);
7059 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7064 pte_t pteval = ptep_get_lockless(pte);
7066 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7073 * huge_pte_offset() - Walk the page table to resolve the hugepage
7074 * entry at address @addr
7076 * Return: Pointer to page table entry (PUD or PMD) for
7077 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7078 * size @sz doesn't match the hugepage size at this level of the page
7081 pte_t *huge_pte_offset(struct mm_struct *mm,
7082 unsigned long addr, unsigned long sz)
7089 pgd = pgd_offset(mm, addr);
7090 if (!pgd_present(*pgd))
7092 p4d = p4d_offset(pgd, addr);
7093 if (!p4d_present(*p4d))
7096 pud = pud_offset(p4d, addr);
7098 /* must be pud huge, non-present or none */
7099 return (pte_t *)pud;
7100 if (!pud_present(*pud))
7102 /* must have a valid entry and size to go further */
7104 pmd = pmd_offset(pud, addr);
7105 /* must be pmd huge, non-present or none */
7106 return (pte_t *)pmd;
7110 * Return a mask that can be used to update an address to the last huge
7111 * page in a page table page mapping size. Used to skip non-present
7112 * page table entries when linearly scanning address ranges. Architectures
7113 * with unique huge page to page table relationships can define their own
7114 * version of this routine.
7116 unsigned long hugetlb_mask_last_page(struct hstate *h)
7118 unsigned long hp_size = huge_page_size(h);
7120 if (hp_size == PUD_SIZE)
7121 return P4D_SIZE - PUD_SIZE;
7122 else if (hp_size == PMD_SIZE)
7123 return PUD_SIZE - PMD_SIZE;
7130 /* See description above. Architectures can provide their own version. */
7131 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7133 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7134 if (huge_page_size(h) == PMD_SIZE)
7135 return PUD_SIZE - PMD_SIZE;
7140 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7143 * These functions are overwritable if your architecture needs its own
7146 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7150 spin_lock_irq(&hugetlb_lock);
7151 if (!folio_test_hugetlb(folio) ||
7152 !folio_test_hugetlb_migratable(folio) ||
7153 !folio_try_get(folio)) {
7157 folio_clear_hugetlb_migratable(folio);
7158 list_move_tail(&folio->lru, list);
7160 spin_unlock_irq(&hugetlb_lock);
7164 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7169 spin_lock_irq(&hugetlb_lock);
7170 if (folio_test_hugetlb(folio)) {
7172 if (folio_test_hugetlb_freed(folio))
7174 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7175 ret = folio_try_get(folio);
7179 spin_unlock_irq(&hugetlb_lock);
7183 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7184 bool *migratable_cleared)
7188 spin_lock_irq(&hugetlb_lock);
7189 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7190 spin_unlock_irq(&hugetlb_lock);
7194 void folio_putback_active_hugetlb(struct folio *folio)
7196 spin_lock_irq(&hugetlb_lock);
7197 folio_set_hugetlb_migratable(folio);
7198 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7199 spin_unlock_irq(&hugetlb_lock);
7203 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7205 struct hstate *h = folio_hstate(old_folio);
7207 hugetlb_cgroup_migrate(old_folio, new_folio);
7208 set_page_owner_migrate_reason(&new_folio->page, reason);
7211 * transfer temporary state of the new hugetlb folio. This is
7212 * reverse to other transitions because the newpage is going to
7213 * be final while the old one will be freed so it takes over
7214 * the temporary status.
7216 * Also note that we have to transfer the per-node surplus state
7217 * here as well otherwise the global surplus count will not match
7220 if (folio_test_hugetlb_temporary(new_folio)) {
7221 int old_nid = folio_nid(old_folio);
7222 int new_nid = folio_nid(new_folio);
7224 folio_set_hugetlb_temporary(old_folio);
7225 folio_clear_hugetlb_temporary(new_folio);
7229 * There is no need to transfer the per-node surplus state
7230 * when we do not cross the node.
7232 if (new_nid == old_nid)
7234 spin_lock_irq(&hugetlb_lock);
7235 if (h->surplus_huge_pages_node[old_nid]) {
7236 h->surplus_huge_pages_node[old_nid]--;
7237 h->surplus_huge_pages_node[new_nid]++;
7239 spin_unlock_irq(&hugetlb_lock);
7243 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7244 unsigned long start,
7247 struct hstate *h = hstate_vma(vma);
7248 unsigned long sz = huge_page_size(h);
7249 struct mm_struct *mm = vma->vm_mm;
7250 struct mmu_notifier_range range;
7251 unsigned long address;
7255 if (!(vma->vm_flags & VM_MAYSHARE))
7261 flush_cache_range(vma, start, end);
7263 * No need to call adjust_range_if_pmd_sharing_possible(), because
7264 * we have already done the PUD_SIZE alignment.
7266 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7268 mmu_notifier_invalidate_range_start(&range);
7269 hugetlb_vma_lock_write(vma);
7270 i_mmap_lock_write(vma->vm_file->f_mapping);
7271 for (address = start; address < end; address += PUD_SIZE) {
7272 ptep = hugetlb_walk(vma, address, sz);
7275 ptl = huge_pte_lock(h, mm, ptep);
7276 huge_pmd_unshare(mm, vma, address, ptep);
7279 flush_hugetlb_tlb_range(vma, start, end);
7280 i_mmap_unlock_write(vma->vm_file->f_mapping);
7281 hugetlb_vma_unlock_write(vma);
7283 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7284 * Documentation/mm/mmu_notifier.rst.
7286 mmu_notifier_invalidate_range_end(&range);
7290 * This function will unconditionally remove all the shared pmd pgtable entries
7291 * within the specific vma for a hugetlbfs memory range.
7293 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7295 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7296 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7300 static bool cma_reserve_called __initdata;
7302 static int __init cmdline_parse_hugetlb_cma(char *p)
7309 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7312 if (s[count] == ':') {
7313 if (tmp >= MAX_NUMNODES)
7315 nid = array_index_nospec(tmp, MAX_NUMNODES);
7318 tmp = memparse(s, &s);
7319 hugetlb_cma_size_in_node[nid] = tmp;
7320 hugetlb_cma_size += tmp;
7323 * Skip the separator if have one, otherwise
7324 * break the parsing.
7331 hugetlb_cma_size = memparse(p, &p);
7339 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7341 void __init hugetlb_cma_reserve(int order)
7343 unsigned long size, reserved, per_node;
7344 bool node_specific_cma_alloc = false;
7347 cma_reserve_called = true;
7349 if (!hugetlb_cma_size)
7352 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7353 if (hugetlb_cma_size_in_node[nid] == 0)
7356 if (!node_online(nid)) {
7357 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7358 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7359 hugetlb_cma_size_in_node[nid] = 0;
7363 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7364 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7365 nid, (PAGE_SIZE << order) / SZ_1M);
7366 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7367 hugetlb_cma_size_in_node[nid] = 0;
7369 node_specific_cma_alloc = true;
7373 /* Validate the CMA size again in case some invalid nodes specified. */
7374 if (!hugetlb_cma_size)
7377 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7378 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7379 (PAGE_SIZE << order) / SZ_1M);
7380 hugetlb_cma_size = 0;
7384 if (!node_specific_cma_alloc) {
7386 * If 3 GB area is requested on a machine with 4 numa nodes,
7387 * let's allocate 1 GB on first three nodes and ignore the last one.
7389 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7390 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7391 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7395 for_each_online_node(nid) {
7397 char name[CMA_MAX_NAME];
7399 if (node_specific_cma_alloc) {
7400 if (hugetlb_cma_size_in_node[nid] == 0)
7403 size = hugetlb_cma_size_in_node[nid];
7405 size = min(per_node, hugetlb_cma_size - reserved);
7408 size = round_up(size, PAGE_SIZE << order);
7410 snprintf(name, sizeof(name), "hugetlb%d", nid);
7412 * Note that 'order per bit' is based on smallest size that
7413 * may be returned to CMA allocator in the case of
7414 * huge page demotion.
7416 res = cma_declare_contiguous_nid(0, size, 0,
7417 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7419 &hugetlb_cma[nid], nid);
7421 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7427 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7430 if (reserved >= hugetlb_cma_size)
7436 * hugetlb_cma_size is used to determine if allocations from
7437 * cma are possible. Set to zero if no cma regions are set up.
7439 hugetlb_cma_size = 0;
7442 static void __init hugetlb_cma_check(void)
7444 if (!hugetlb_cma_size || cma_reserve_called)
7447 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7450 #endif /* CONFIG_CMA */