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 * vma_kernel_pagesize - Page size granularity for this VMA.
973 * @vma: The user mapping.
975 * Folios in this VMA will be aligned to, and at least the size of the
976 * number of bytes returned by this function.
978 * Return: The default size of the folios allocated when backing a VMA.
980 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
982 if (vma->vm_ops && vma->vm_ops->pagesize)
983 return vma->vm_ops->pagesize(vma);
986 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
989 * Return the page size being used by the MMU to back a VMA. In the majority
990 * of cases, the page size used by the kernel matches the MMU size. On
991 * architectures where it differs, an architecture-specific 'strong'
992 * version of this symbol is required.
994 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
996 return vma_kernel_pagesize(vma);
1000 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1001 * bits of the reservation map pointer, which are always clear due to
1004 #define HPAGE_RESV_OWNER (1UL << 0)
1005 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1006 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1009 * These helpers are used to track how many pages are reserved for
1010 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1011 * is guaranteed to have their future faults succeed.
1013 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1014 * the reserve counters are updated with the hugetlb_lock held. It is safe
1015 * to reset the VMA at fork() time as it is not in use yet and there is no
1016 * chance of the global counters getting corrupted as a result of the values.
1018 * The private mapping reservation is represented in a subtly different
1019 * manner to a shared mapping. A shared mapping has a region map associated
1020 * with the underlying file, this region map represents the backing file
1021 * pages which have ever had a reservation assigned which this persists even
1022 * after the page is instantiated. A private mapping has a region map
1023 * associated with the original mmap which is attached to all VMAs which
1024 * reference it, this region map represents those offsets which have consumed
1025 * reservation ie. where pages have been instantiated.
1027 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1029 return (unsigned long)vma->vm_private_data;
1032 static void set_vma_private_data(struct vm_area_struct *vma,
1033 unsigned long value)
1035 vma->vm_private_data = (void *)value;
1039 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1040 struct hugetlb_cgroup *h_cg,
1043 #ifdef CONFIG_CGROUP_HUGETLB
1045 resv_map->reservation_counter = NULL;
1046 resv_map->pages_per_hpage = 0;
1047 resv_map->css = NULL;
1049 resv_map->reservation_counter =
1050 &h_cg->rsvd_hugepage[hstate_index(h)];
1051 resv_map->pages_per_hpage = pages_per_huge_page(h);
1052 resv_map->css = &h_cg->css;
1057 struct resv_map *resv_map_alloc(void)
1059 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1060 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1062 if (!resv_map || !rg) {
1068 kref_init(&resv_map->refs);
1069 spin_lock_init(&resv_map->lock);
1070 INIT_LIST_HEAD(&resv_map->regions);
1072 resv_map->adds_in_progress = 0;
1074 * Initialize these to 0. On shared mappings, 0's here indicate these
1075 * fields don't do cgroup accounting. On private mappings, these will be
1076 * re-initialized to the proper values, to indicate that hugetlb cgroup
1077 * reservations are to be un-charged from here.
1079 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1081 INIT_LIST_HEAD(&resv_map->region_cache);
1082 list_add(&rg->link, &resv_map->region_cache);
1083 resv_map->region_cache_count = 1;
1088 void resv_map_release(struct kref *ref)
1090 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1091 struct list_head *head = &resv_map->region_cache;
1092 struct file_region *rg, *trg;
1094 /* Clear out any active regions before we release the map. */
1095 region_del(resv_map, 0, LONG_MAX);
1097 /* ... and any entries left in the cache */
1098 list_for_each_entry_safe(rg, trg, head, link) {
1099 list_del(&rg->link);
1103 VM_BUG_ON(resv_map->adds_in_progress);
1108 static inline struct resv_map *inode_resv_map(struct inode *inode)
1111 * At inode evict time, i_mapping may not point to the original
1112 * address space within the inode. This original address space
1113 * contains the pointer to the resv_map. So, always use the
1114 * address space embedded within the inode.
1115 * The VERY common case is inode->mapping == &inode->i_data but,
1116 * this may not be true for device special inodes.
1118 return (struct resv_map *)(&inode->i_data)->private_data;
1121 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1123 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1124 if (vma->vm_flags & VM_MAYSHARE) {
1125 struct address_space *mapping = vma->vm_file->f_mapping;
1126 struct inode *inode = mapping->host;
1128 return inode_resv_map(inode);
1131 return (struct resv_map *)(get_vma_private_data(vma) &
1136 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1138 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1139 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1141 set_vma_private_data(vma, (unsigned long)map);
1144 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1146 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1147 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1149 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1152 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1154 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1156 return (get_vma_private_data(vma) & flag) != 0;
1159 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1161 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1163 * Clear vm_private_data
1164 * - For shared mappings this is a per-vma semaphore that may be
1165 * allocated in a subsequent call to hugetlb_vm_op_open.
1166 * Before clearing, make sure pointer is not associated with vma
1167 * as this will leak the structure. This is the case when called
1168 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1169 * been called to allocate a new structure.
1170 * - For MAP_PRIVATE mappings, this is the reserve map which does
1171 * not apply to children. Faults generated by the children are
1172 * not guaranteed to succeed, even if read-only.
1174 if (vma->vm_flags & VM_MAYSHARE) {
1175 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1177 if (vma_lock && vma_lock->vma != vma)
1178 vma->vm_private_data = NULL;
1180 vma->vm_private_data = NULL;
1184 * Reset and decrement one ref on hugepage private reservation.
1185 * Called with mm->mmap_lock writer semaphore held.
1186 * This function should be only used by move_vma() and operate on
1187 * same sized vma. It should never come here with last ref on the
1190 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1193 * Clear the old hugetlb private page reservation.
1194 * It has already been transferred to new_vma.
1196 * During a mremap() operation of a hugetlb vma we call move_vma()
1197 * which copies vma into new_vma and unmaps vma. After the copy
1198 * operation both new_vma and vma share a reference to the resv_map
1199 * struct, and at that point vma is about to be unmapped. We don't
1200 * want to return the reservation to the pool at unmap of vma because
1201 * the reservation still lives on in new_vma, so simply decrement the
1202 * ref here and remove the resv_map reference from this vma.
1204 struct resv_map *reservations = vma_resv_map(vma);
1206 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1207 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1208 kref_put(&reservations->refs, resv_map_release);
1211 hugetlb_dup_vma_private(vma);
1214 /* Returns true if the VMA has associated reserve pages */
1215 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1217 if (vma->vm_flags & VM_NORESERVE) {
1219 * This address is already reserved by other process(chg == 0),
1220 * so, we should decrement reserved count. Without decrementing,
1221 * reserve count remains after releasing inode, because this
1222 * allocated page will go into page cache and is regarded as
1223 * coming from reserved pool in releasing step. Currently, we
1224 * don't have any other solution to deal with this situation
1225 * properly, so add work-around here.
1227 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1233 /* Shared mappings always use reserves */
1234 if (vma->vm_flags & VM_MAYSHARE) {
1236 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1237 * be a region map for all pages. The only situation where
1238 * there is no region map is if a hole was punched via
1239 * fallocate. In this case, there really are no reserves to
1240 * use. This situation is indicated if chg != 0.
1249 * Only the process that called mmap() has reserves for
1252 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1254 * Like the shared case above, a hole punch or truncate
1255 * could have been performed on the private mapping.
1256 * Examine the value of chg to determine if reserves
1257 * actually exist or were previously consumed.
1258 * Very Subtle - The value of chg comes from a previous
1259 * call to vma_needs_reserves(). The reserve map for
1260 * private mappings has different (opposite) semantics
1261 * than that of shared mappings. vma_needs_reserves()
1262 * has already taken this difference in semantics into
1263 * account. Therefore, the meaning of chg is the same
1264 * as in the shared case above. Code could easily be
1265 * combined, but keeping it separate draws attention to
1266 * subtle differences.
1277 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1279 int nid = folio_nid(folio);
1281 lockdep_assert_held(&hugetlb_lock);
1282 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1284 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1285 h->free_huge_pages++;
1286 h->free_huge_pages_node[nid]++;
1287 folio_set_hugetlb_freed(folio);
1290 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1293 struct folio *folio;
1294 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1296 lockdep_assert_held(&hugetlb_lock);
1297 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1298 if (pin && !folio_is_longterm_pinnable(folio))
1301 if (folio_test_hwpoison(folio))
1304 list_move(&folio->lru, &h->hugepage_activelist);
1305 folio_ref_unfreeze(folio, 1);
1306 folio_clear_hugetlb_freed(folio);
1307 h->free_huge_pages--;
1308 h->free_huge_pages_node[nid]--;
1315 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1316 int nid, nodemask_t *nmask)
1318 unsigned int cpuset_mems_cookie;
1319 struct zonelist *zonelist;
1322 int node = NUMA_NO_NODE;
1324 zonelist = node_zonelist(nid, gfp_mask);
1327 cpuset_mems_cookie = read_mems_allowed_begin();
1328 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1329 struct folio *folio;
1331 if (!cpuset_zone_allowed(zone, gfp_mask))
1334 * no need to ask again on the same node. Pool is node rather than
1337 if (zone_to_nid(zone) == node)
1339 node = zone_to_nid(zone);
1341 folio = dequeue_hugetlb_folio_node_exact(h, node);
1345 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1351 static unsigned long available_huge_pages(struct hstate *h)
1353 return h->free_huge_pages - h->resv_huge_pages;
1356 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1357 struct vm_area_struct *vma,
1358 unsigned long address, int avoid_reserve,
1361 struct folio *folio = NULL;
1362 struct mempolicy *mpol;
1364 nodemask_t *nodemask;
1368 * A child process with MAP_PRIVATE mappings created by their parent
1369 * have no page reserves. This check ensures that reservations are
1370 * not "stolen". The child may still get SIGKILLed
1372 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1375 /* If reserves cannot be used, ensure enough pages are in the pool */
1376 if (avoid_reserve && !available_huge_pages(h))
1379 gfp_mask = htlb_alloc_mask(h);
1380 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1382 if (mpol_is_preferred_many(mpol)) {
1383 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1386 /* Fallback to all nodes if page==NULL */
1391 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1394 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1395 folio_set_hugetlb_restore_reserve(folio);
1396 h->resv_huge_pages--;
1399 mpol_cond_put(mpol);
1407 * common helper functions for hstate_next_node_to_{alloc|free}.
1408 * We may have allocated or freed a huge page based on a different
1409 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1410 * be outside of *nodes_allowed. Ensure that we use an allowed
1411 * node for alloc or free.
1413 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1415 nid = next_node_in(nid, *nodes_allowed);
1416 VM_BUG_ON(nid >= MAX_NUMNODES);
1421 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1423 if (!node_isset(nid, *nodes_allowed))
1424 nid = next_node_allowed(nid, nodes_allowed);
1429 * returns the previously saved node ["this node"] from which to
1430 * allocate a persistent huge page for the pool and advance the
1431 * next node from which to allocate, handling wrap at end of node
1434 static int hstate_next_node_to_alloc(struct hstate *h,
1435 nodemask_t *nodes_allowed)
1439 VM_BUG_ON(!nodes_allowed);
1441 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1442 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1448 * helper for remove_pool_huge_page() - return the previously saved
1449 * node ["this node"] from which to free a huge page. Advance the
1450 * next node id whether or not we find a free huge page to free so
1451 * that the next attempt to free addresses the next node.
1453 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1457 VM_BUG_ON(!nodes_allowed);
1459 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1460 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1465 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1466 for (nr_nodes = nodes_weight(*mask); \
1468 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1471 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1472 for (nr_nodes = nodes_weight(*mask); \
1474 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1477 /* used to demote non-gigantic_huge pages as well */
1478 static void __destroy_compound_gigantic_folio(struct folio *folio,
1479 unsigned int order, bool demote)
1482 int nr_pages = 1 << order;
1485 atomic_set(&folio->_entire_mapcount, 0);
1486 atomic_set(&folio->_nr_pages_mapped, 0);
1487 atomic_set(&folio->_pincount, 0);
1489 for (i = 1; i < nr_pages; i++) {
1490 p = folio_page(folio, i);
1491 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1493 clear_compound_head(p);
1495 set_page_refcounted(p);
1498 __folio_clear_head(folio);
1501 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1504 __destroy_compound_gigantic_folio(folio, order, true);
1507 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1508 static void destroy_compound_gigantic_folio(struct folio *folio,
1511 __destroy_compound_gigantic_folio(folio, order, false);
1514 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1517 * If the page isn't allocated using the cma allocator,
1518 * cma_release() returns false.
1521 int nid = folio_nid(folio);
1523 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1527 free_contig_range(folio_pfn(folio), 1 << order);
1530 #ifdef CONFIG_CONTIG_ALLOC
1531 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1532 int nid, nodemask_t *nodemask)
1535 unsigned long nr_pages = pages_per_huge_page(h);
1536 if (nid == NUMA_NO_NODE)
1537 nid = numa_mem_id();
1543 if (hugetlb_cma[nid]) {
1544 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1545 huge_page_order(h), true);
1547 return page_folio(page);
1550 if (!(gfp_mask & __GFP_THISNODE)) {
1551 for_each_node_mask(node, *nodemask) {
1552 if (node == nid || !hugetlb_cma[node])
1555 page = cma_alloc(hugetlb_cma[node], nr_pages,
1556 huge_page_order(h), true);
1558 return page_folio(page);
1564 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1565 return page ? page_folio(page) : NULL;
1568 #else /* !CONFIG_CONTIG_ALLOC */
1569 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1570 int nid, nodemask_t *nodemask)
1574 #endif /* CONFIG_CONTIG_ALLOC */
1576 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1577 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1578 int nid, nodemask_t *nodemask)
1582 static inline void free_gigantic_folio(struct folio *folio,
1583 unsigned int order) { }
1584 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1585 unsigned int order) { }
1588 static inline void __clear_hugetlb_destructor(struct hstate *h,
1589 struct folio *folio)
1591 lockdep_assert_held(&hugetlb_lock);
1593 folio_clear_hugetlb(folio);
1597 * Remove hugetlb folio from lists.
1598 * If vmemmap exists for the folio, update dtor so that the folio appears
1599 * as just a compound page. Otherwise, wait until after allocating vmemmap
1602 * A reference is held on the folio, except in the case of demote.
1604 * Must be called with hugetlb lock held.
1606 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1607 bool adjust_surplus,
1610 int nid = folio_nid(folio);
1612 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1613 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1615 lockdep_assert_held(&hugetlb_lock);
1616 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1619 list_del(&folio->lru);
1621 if (folio_test_hugetlb_freed(folio)) {
1622 h->free_huge_pages--;
1623 h->free_huge_pages_node[nid]--;
1625 if (adjust_surplus) {
1626 h->surplus_huge_pages--;
1627 h->surplus_huge_pages_node[nid]--;
1631 * We can only clear the hugetlb destructor after allocating vmemmap
1632 * pages. Otherwise, someone (memory error handling) may try to write
1633 * to tail struct pages.
1635 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1636 __clear_hugetlb_destructor(h, folio);
1639 * In the case of demote we do not ref count the page as it will soon
1640 * be turned into a page of smaller size.
1643 folio_ref_unfreeze(folio, 1);
1646 h->nr_huge_pages_node[nid]--;
1649 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1650 bool adjust_surplus)
1652 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1655 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1656 bool adjust_surplus)
1658 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1661 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1662 bool adjust_surplus)
1665 int nid = folio_nid(folio);
1667 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1669 lockdep_assert_held(&hugetlb_lock);
1671 INIT_LIST_HEAD(&folio->lru);
1673 h->nr_huge_pages_node[nid]++;
1675 if (adjust_surplus) {
1676 h->surplus_huge_pages++;
1677 h->surplus_huge_pages_node[nid]++;
1680 folio_set_hugetlb(folio);
1681 folio_change_private(folio, NULL);
1683 * We have to set hugetlb_vmemmap_optimized again as above
1684 * folio_change_private(folio, NULL) cleared it.
1686 folio_set_hugetlb_vmemmap_optimized(folio);
1689 * This folio is about to be managed by the hugetlb allocator and
1690 * should have no users. Drop our reference, and check for others
1693 zeroed = folio_put_testzero(folio);
1694 if (unlikely(!zeroed))
1696 * It is VERY unlikely soneone else has taken a ref
1697 * on the folio. In this case, we simply return as
1698 * free_huge_folio() will be called when this other ref
1703 arch_clear_hugepage_flags(&folio->page);
1704 enqueue_hugetlb_folio(h, folio);
1707 static void __update_and_free_hugetlb_folio(struct hstate *h,
1708 struct folio *folio)
1710 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1712 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1716 * If we don't know which subpages are hwpoisoned, we can't free
1717 * the hugepage, so it's leaked intentionally.
1719 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1722 if (hugetlb_vmemmap_restore(h, &folio->page)) {
1723 spin_lock_irq(&hugetlb_lock);
1725 * If we cannot allocate vmemmap pages, just refuse to free the
1726 * page and put the page back on the hugetlb free list and treat
1727 * as a surplus page.
1729 add_hugetlb_folio(h, folio, true);
1730 spin_unlock_irq(&hugetlb_lock);
1735 * Move PageHWPoison flag from head page to the raw error pages,
1736 * which makes any healthy subpages reusable.
1738 if (unlikely(folio_test_hwpoison(folio)))
1739 folio_clear_hugetlb_hwpoison(folio);
1742 * If vmemmap pages were allocated above, then we need to clear the
1743 * hugetlb destructor under the hugetlb lock.
1746 spin_lock_irq(&hugetlb_lock);
1747 __clear_hugetlb_destructor(h, folio);
1748 spin_unlock_irq(&hugetlb_lock);
1752 * Non-gigantic pages demoted from CMA allocated gigantic pages
1753 * need to be given back to CMA in free_gigantic_folio.
1755 if (hstate_is_gigantic(h) ||
1756 hugetlb_cma_folio(folio, huge_page_order(h))) {
1757 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1758 free_gigantic_folio(folio, huge_page_order(h));
1760 __free_pages(&folio->page, huge_page_order(h));
1765 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1766 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1767 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1768 * the vmemmap pages.
1770 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1771 * freed and frees them one-by-one. As the page->mapping pointer is going
1772 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1773 * structure of a lockless linked list of huge pages to be freed.
1775 static LLIST_HEAD(hpage_freelist);
1777 static void free_hpage_workfn(struct work_struct *work)
1779 struct llist_node *node;
1781 node = llist_del_all(&hpage_freelist);
1787 page = container_of((struct address_space **)node,
1788 struct page, mapping);
1790 page->mapping = NULL;
1792 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1793 * folio_hstate() is going to trigger because a previous call to
1794 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1795 * not use folio_hstate() directly.
1797 h = size_to_hstate(page_size(page));
1799 __update_and_free_hugetlb_folio(h, page_folio(page));
1804 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1806 static inline void flush_free_hpage_work(struct hstate *h)
1808 if (hugetlb_vmemmap_optimizable(h))
1809 flush_work(&free_hpage_work);
1812 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1815 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1816 __update_and_free_hugetlb_folio(h, folio);
1821 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1823 * Only call schedule_work() if hpage_freelist is previously
1824 * empty. Otherwise, schedule_work() had been called but the workfn
1825 * hasn't retrieved the list yet.
1827 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1828 schedule_work(&free_hpage_work);
1831 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1833 struct page *page, *t_page;
1834 struct folio *folio;
1836 list_for_each_entry_safe(page, t_page, list, lru) {
1837 folio = page_folio(page);
1838 update_and_free_hugetlb_folio(h, folio, false);
1843 struct hstate *size_to_hstate(unsigned long size)
1847 for_each_hstate(h) {
1848 if (huge_page_size(h) == size)
1854 void free_huge_folio(struct folio *folio)
1857 * Can't pass hstate in here because it is called from the
1858 * compound page destructor.
1860 struct hstate *h = folio_hstate(folio);
1861 int nid = folio_nid(folio);
1862 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1863 bool restore_reserve;
1864 unsigned long flags;
1866 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1867 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1869 hugetlb_set_folio_subpool(folio, NULL);
1870 if (folio_test_anon(folio))
1871 __ClearPageAnonExclusive(&folio->page);
1872 folio->mapping = NULL;
1873 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1874 folio_clear_hugetlb_restore_reserve(folio);
1877 * If HPageRestoreReserve was set on page, page allocation consumed a
1878 * reservation. If the page was associated with a subpool, there
1879 * would have been a page reserved in the subpool before allocation
1880 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1881 * reservation, do not call hugepage_subpool_put_pages() as this will
1882 * remove the reserved page from the subpool.
1884 if (!restore_reserve) {
1886 * A return code of zero implies that the subpool will be
1887 * under its minimum size if the reservation is not restored
1888 * after page is free. Therefore, force restore_reserve
1891 if (hugepage_subpool_put_pages(spool, 1) == 0)
1892 restore_reserve = true;
1895 spin_lock_irqsave(&hugetlb_lock, flags);
1896 folio_clear_hugetlb_migratable(folio);
1897 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1898 pages_per_huge_page(h), folio);
1899 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1900 pages_per_huge_page(h), folio);
1901 if (restore_reserve)
1902 h->resv_huge_pages++;
1904 if (folio_test_hugetlb_temporary(folio)) {
1905 remove_hugetlb_folio(h, folio, false);
1906 spin_unlock_irqrestore(&hugetlb_lock, flags);
1907 update_and_free_hugetlb_folio(h, folio, true);
1908 } else if (h->surplus_huge_pages_node[nid]) {
1909 /* remove the page from active list */
1910 remove_hugetlb_folio(h, folio, true);
1911 spin_unlock_irqrestore(&hugetlb_lock, flags);
1912 update_and_free_hugetlb_folio(h, folio, true);
1914 arch_clear_hugepage_flags(&folio->page);
1915 enqueue_hugetlb_folio(h, folio);
1916 spin_unlock_irqrestore(&hugetlb_lock, flags);
1921 * Must be called with the hugetlb lock held
1923 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1925 lockdep_assert_held(&hugetlb_lock);
1927 h->nr_huge_pages_node[nid]++;
1930 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1932 hugetlb_vmemmap_optimize(h, &folio->page);
1933 INIT_LIST_HEAD(&folio->lru);
1934 folio_set_hugetlb(folio);
1935 hugetlb_set_folio_subpool(folio, NULL);
1936 set_hugetlb_cgroup(folio, NULL);
1937 set_hugetlb_cgroup_rsvd(folio, NULL);
1940 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1942 __prep_new_hugetlb_folio(h, folio);
1943 spin_lock_irq(&hugetlb_lock);
1944 __prep_account_new_huge_page(h, nid);
1945 spin_unlock_irq(&hugetlb_lock);
1948 static bool __prep_compound_gigantic_folio(struct folio *folio,
1949 unsigned int order, bool demote)
1952 int nr_pages = 1 << order;
1955 __folio_clear_reserved(folio);
1956 for (i = 0; i < nr_pages; i++) {
1957 p = folio_page(folio, i);
1960 * For gigantic hugepages allocated through bootmem at
1961 * boot, it's safer to be consistent with the not-gigantic
1962 * hugepages and clear the PG_reserved bit from all tail pages
1963 * too. Otherwise drivers using get_user_pages() to access tail
1964 * pages may get the reference counting wrong if they see
1965 * PG_reserved set on a tail page (despite the head page not
1966 * having PG_reserved set). Enforcing this consistency between
1967 * head and tail pages allows drivers to optimize away a check
1968 * on the head page when they need know if put_page() is needed
1969 * after get_user_pages().
1971 if (i != 0) /* head page cleared above */
1972 __ClearPageReserved(p);
1974 * Subtle and very unlikely
1976 * Gigantic 'page allocators' such as memblock or cma will
1977 * return a set of pages with each page ref counted. We need
1978 * to turn this set of pages into a compound page with tail
1979 * page ref counts set to zero. Code such as speculative page
1980 * cache adding could take a ref on a 'to be' tail page.
1981 * We need to respect any increased ref count, and only set
1982 * the ref count to zero if count is currently 1. If count
1983 * is not 1, we return an error. An error return indicates
1984 * the set of pages can not be converted to a gigantic page.
1985 * The caller who allocated the pages should then discard the
1986 * pages using the appropriate free interface.
1988 * In the case of demote, the ref count will be zero.
1991 if (!page_ref_freeze(p, 1)) {
1992 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1996 VM_BUG_ON_PAGE(page_count(p), p);
1999 set_compound_head(p, &folio->page);
2001 __folio_set_head(folio);
2002 /* we rely on prep_new_hugetlb_folio to set the destructor */
2003 folio_set_order(folio, order);
2004 atomic_set(&folio->_entire_mapcount, -1);
2005 atomic_set(&folio->_nr_pages_mapped, 0);
2006 atomic_set(&folio->_pincount, 0);
2010 /* undo page modifications made above */
2011 for (j = 0; j < i; j++) {
2012 p = folio_page(folio, j);
2014 clear_compound_head(p);
2015 set_page_refcounted(p);
2017 /* need to clear PG_reserved on remaining tail pages */
2018 for (; j < nr_pages; j++) {
2019 p = folio_page(folio, j);
2020 __ClearPageReserved(p);
2025 static bool prep_compound_gigantic_folio(struct folio *folio,
2028 return __prep_compound_gigantic_folio(folio, order, false);
2031 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2034 return __prep_compound_gigantic_folio(folio, order, true);
2038 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2039 * transparent huge pages. See the PageTransHuge() documentation for more
2042 int PageHuge(struct page *page)
2044 struct folio *folio;
2046 if (!PageCompound(page))
2048 folio = page_folio(page);
2049 return folio_test_hugetlb(folio);
2051 EXPORT_SYMBOL_GPL(PageHuge);
2054 * Find and lock address space (mapping) in write mode.
2056 * Upon entry, the page is locked which means that page_mapping() is
2057 * stable. Due to locking order, we can only trylock_write. If we can
2058 * not get the lock, simply return NULL to caller.
2060 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2062 struct address_space *mapping = page_mapping(hpage);
2067 if (i_mmap_trylock_write(mapping))
2073 pgoff_t hugetlb_basepage_index(struct page *page)
2075 struct page *page_head = compound_head(page);
2076 pgoff_t index = page_index(page_head);
2077 unsigned long compound_idx;
2079 if (compound_order(page_head) > MAX_ORDER)
2080 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2082 compound_idx = page - page_head;
2084 return (index << compound_order(page_head)) + compound_idx;
2087 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2088 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2089 nodemask_t *node_alloc_noretry)
2091 int order = huge_page_order(h);
2093 bool alloc_try_hard = true;
2097 * By default we always try hard to allocate the page with
2098 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2099 * a loop (to adjust global huge page counts) and previous allocation
2100 * failed, do not continue to try hard on the same node. Use the
2101 * node_alloc_noretry bitmap to manage this state information.
2103 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2104 alloc_try_hard = false;
2105 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2107 gfp_mask |= __GFP_RETRY_MAYFAIL;
2108 if (nid == NUMA_NO_NODE)
2109 nid = numa_mem_id();
2111 page = __alloc_pages(gfp_mask, order, nid, nmask);
2113 /* Freeze head page */
2114 if (page && !page_ref_freeze(page, 1)) {
2115 __free_pages(page, order);
2116 if (retry) { /* retry once */
2120 /* WOW! twice in a row. */
2121 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2126 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2127 * indicates an overall state change. Clear bit so that we resume
2128 * normal 'try hard' allocations.
2130 if (node_alloc_noretry && page && !alloc_try_hard)
2131 node_clear(nid, *node_alloc_noretry);
2134 * If we tried hard to get a page but failed, set bit so that
2135 * subsequent attempts will not try as hard until there is an
2136 * overall state change.
2138 if (node_alloc_noretry && !page && alloc_try_hard)
2139 node_set(nid, *node_alloc_noretry);
2142 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2146 __count_vm_event(HTLB_BUDDY_PGALLOC);
2147 return page_folio(page);
2151 * Common helper to allocate a fresh hugetlb page. All specific allocators
2152 * should use this function to get new hugetlb pages
2154 * Note that returned page is 'frozen': ref count of head page and all tail
2157 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2158 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2159 nodemask_t *node_alloc_noretry)
2161 struct folio *folio;
2165 if (hstate_is_gigantic(h))
2166 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2168 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2169 nid, nmask, node_alloc_noretry);
2172 if (hstate_is_gigantic(h)) {
2173 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2175 * Rare failure to convert pages to compound page.
2176 * Free pages and try again - ONCE!
2178 free_gigantic_folio(folio, huge_page_order(h));
2186 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2192 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2195 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2196 nodemask_t *node_alloc_noretry)
2198 struct folio *folio;
2200 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2202 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2203 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2204 nodes_allowed, node_alloc_noretry);
2206 free_huge_folio(folio); /* free it into the hugepage allocator */
2215 * Remove huge page from pool from next node to free. Attempt to keep
2216 * persistent huge pages more or less balanced over allowed nodes.
2217 * This routine only 'removes' the hugetlb page. The caller must make
2218 * an additional call to free the page to low level allocators.
2219 * Called with hugetlb_lock locked.
2221 static struct page *remove_pool_huge_page(struct hstate *h,
2222 nodemask_t *nodes_allowed,
2226 struct page *page = NULL;
2227 struct folio *folio;
2229 lockdep_assert_held(&hugetlb_lock);
2230 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2232 * If we're returning unused surplus pages, only examine
2233 * nodes with surplus pages.
2235 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2236 !list_empty(&h->hugepage_freelists[node])) {
2237 page = list_entry(h->hugepage_freelists[node].next,
2239 folio = page_folio(page);
2240 remove_hugetlb_folio(h, folio, acct_surplus);
2249 * Dissolve a given free hugepage into free buddy pages. This function does
2250 * nothing for in-use hugepages and non-hugepages.
2251 * This function returns values like below:
2253 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2254 * when the system is under memory pressure and the feature of
2255 * freeing unused vmemmap pages associated with each hugetlb page
2257 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2258 * (allocated or reserved.)
2259 * 0: successfully dissolved free hugepages or the page is not a
2260 * hugepage (considered as already dissolved)
2262 int dissolve_free_huge_page(struct page *page)
2265 struct folio *folio = page_folio(page);
2268 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2269 if (!folio_test_hugetlb(folio))
2272 spin_lock_irq(&hugetlb_lock);
2273 if (!folio_test_hugetlb(folio)) {
2278 if (!folio_ref_count(folio)) {
2279 struct hstate *h = folio_hstate(folio);
2280 if (!available_huge_pages(h))
2284 * We should make sure that the page is already on the free list
2285 * when it is dissolved.
2287 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2288 spin_unlock_irq(&hugetlb_lock);
2292 * Theoretically, we should return -EBUSY when we
2293 * encounter this race. In fact, we have a chance
2294 * to successfully dissolve the page if we do a
2295 * retry. Because the race window is quite small.
2296 * If we seize this opportunity, it is an optimization
2297 * for increasing the success rate of dissolving page.
2302 remove_hugetlb_folio(h, folio, false);
2303 h->max_huge_pages--;
2304 spin_unlock_irq(&hugetlb_lock);
2307 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2308 * before freeing the page. update_and_free_hugtlb_folio will fail to
2309 * free the page if it can not allocate required vmemmap. We
2310 * need to adjust max_huge_pages if the page is not freed.
2311 * Attempt to allocate vmemmmap here so that we can take
2312 * appropriate action on failure.
2314 rc = hugetlb_vmemmap_restore(h, &folio->page);
2316 update_and_free_hugetlb_folio(h, folio, false);
2318 spin_lock_irq(&hugetlb_lock);
2319 add_hugetlb_folio(h, folio, false);
2320 h->max_huge_pages++;
2321 spin_unlock_irq(&hugetlb_lock);
2327 spin_unlock_irq(&hugetlb_lock);
2332 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2333 * make specified memory blocks removable from the system.
2334 * Note that this will dissolve a free gigantic hugepage completely, if any
2335 * part of it lies within the given range.
2336 * Also note that if dissolve_free_huge_page() returns with an error, all
2337 * free hugepages that were dissolved before that error are lost.
2339 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2347 if (!hugepages_supported())
2350 order = huge_page_order(&default_hstate);
2352 order = min(order, huge_page_order(h));
2354 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2355 page = pfn_to_page(pfn);
2356 rc = dissolve_free_huge_page(page);
2365 * Allocates a fresh surplus page from the page allocator.
2367 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2368 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2370 struct folio *folio = NULL;
2372 if (hstate_is_gigantic(h))
2375 spin_lock_irq(&hugetlb_lock);
2376 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2378 spin_unlock_irq(&hugetlb_lock);
2380 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2384 spin_lock_irq(&hugetlb_lock);
2386 * We could have raced with the pool size change.
2387 * Double check that and simply deallocate the new page
2388 * if we would end up overcommiting the surpluses. Abuse
2389 * temporary page to workaround the nasty free_huge_folio
2392 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2393 folio_set_hugetlb_temporary(folio);
2394 spin_unlock_irq(&hugetlb_lock);
2395 free_huge_folio(folio);
2399 h->surplus_huge_pages++;
2400 h->surplus_huge_pages_node[folio_nid(folio)]++;
2403 spin_unlock_irq(&hugetlb_lock);
2408 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2409 int nid, nodemask_t *nmask)
2411 struct folio *folio;
2413 if (hstate_is_gigantic(h))
2416 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2420 /* fresh huge pages are frozen */
2421 folio_ref_unfreeze(folio, 1);
2423 * We do not account these pages as surplus because they are only
2424 * temporary and will be released properly on the last reference
2426 folio_set_hugetlb_temporary(folio);
2432 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2435 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2436 struct vm_area_struct *vma, unsigned long addr)
2438 struct folio *folio = NULL;
2439 struct mempolicy *mpol;
2440 gfp_t gfp_mask = htlb_alloc_mask(h);
2442 nodemask_t *nodemask;
2444 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2445 if (mpol_is_preferred_many(mpol)) {
2446 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2448 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2449 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2451 /* Fallback to all nodes if page==NULL */
2456 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2457 mpol_cond_put(mpol);
2461 /* folio migration callback function */
2462 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2463 nodemask_t *nmask, gfp_t gfp_mask)
2465 spin_lock_irq(&hugetlb_lock);
2466 if (available_huge_pages(h)) {
2467 struct folio *folio;
2469 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2470 preferred_nid, nmask);
2472 spin_unlock_irq(&hugetlb_lock);
2476 spin_unlock_irq(&hugetlb_lock);
2478 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2481 /* mempolicy aware migration callback */
2482 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2483 unsigned long address)
2485 struct mempolicy *mpol;
2486 nodemask_t *nodemask;
2487 struct folio *folio;
2491 gfp_mask = htlb_alloc_mask(h);
2492 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2493 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2494 mpol_cond_put(mpol);
2500 * Increase the hugetlb pool such that it can accommodate a reservation
2503 static int gather_surplus_pages(struct hstate *h, long delta)
2504 __must_hold(&hugetlb_lock)
2506 LIST_HEAD(surplus_list);
2507 struct folio *folio, *tmp;
2510 long needed, allocated;
2511 bool alloc_ok = true;
2513 lockdep_assert_held(&hugetlb_lock);
2514 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2516 h->resv_huge_pages += delta;
2524 spin_unlock_irq(&hugetlb_lock);
2525 for (i = 0; i < needed; i++) {
2526 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2527 NUMA_NO_NODE, NULL);
2532 list_add(&folio->lru, &surplus_list);
2538 * After retaking hugetlb_lock, we need to recalculate 'needed'
2539 * because either resv_huge_pages or free_huge_pages may have changed.
2541 spin_lock_irq(&hugetlb_lock);
2542 needed = (h->resv_huge_pages + delta) -
2543 (h->free_huge_pages + allocated);
2548 * We were not able to allocate enough pages to
2549 * satisfy the entire reservation so we free what
2550 * we've allocated so far.
2555 * The surplus_list now contains _at_least_ the number of extra pages
2556 * needed to accommodate the reservation. Add the appropriate number
2557 * of pages to the hugetlb pool and free the extras back to the buddy
2558 * allocator. Commit the entire reservation here to prevent another
2559 * process from stealing the pages as they are added to the pool but
2560 * before they are reserved.
2562 needed += allocated;
2563 h->resv_huge_pages += delta;
2566 /* Free the needed pages to the hugetlb pool */
2567 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2570 /* Add the page to the hugetlb allocator */
2571 enqueue_hugetlb_folio(h, folio);
2574 spin_unlock_irq(&hugetlb_lock);
2577 * Free unnecessary surplus pages to the buddy allocator.
2578 * Pages have no ref count, call free_huge_folio directly.
2580 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2581 free_huge_folio(folio);
2582 spin_lock_irq(&hugetlb_lock);
2588 * This routine has two main purposes:
2589 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2590 * in unused_resv_pages. This corresponds to the prior adjustments made
2591 * to the associated reservation map.
2592 * 2) Free any unused surplus pages that may have been allocated to satisfy
2593 * the reservation. As many as unused_resv_pages may be freed.
2595 static void return_unused_surplus_pages(struct hstate *h,
2596 unsigned long unused_resv_pages)
2598 unsigned long nr_pages;
2600 LIST_HEAD(page_list);
2602 lockdep_assert_held(&hugetlb_lock);
2603 /* Uncommit the reservation */
2604 h->resv_huge_pages -= unused_resv_pages;
2606 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2610 * Part (or even all) of the reservation could have been backed
2611 * by pre-allocated pages. Only free surplus pages.
2613 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2616 * We want to release as many surplus pages as possible, spread
2617 * evenly across all nodes with memory. Iterate across these nodes
2618 * until we can no longer free unreserved surplus pages. This occurs
2619 * when the nodes with surplus pages have no free pages.
2620 * remove_pool_huge_page() will balance the freed pages across the
2621 * on-line nodes with memory and will handle the hstate accounting.
2623 while (nr_pages--) {
2624 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2628 list_add(&page->lru, &page_list);
2632 spin_unlock_irq(&hugetlb_lock);
2633 update_and_free_pages_bulk(h, &page_list);
2634 spin_lock_irq(&hugetlb_lock);
2639 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2640 * are used by the huge page allocation routines to manage reservations.
2642 * vma_needs_reservation is called to determine if the huge page at addr
2643 * within the vma has an associated reservation. If a reservation is
2644 * needed, the value 1 is returned. The caller is then responsible for
2645 * managing the global reservation and subpool usage counts. After
2646 * the huge page has been allocated, vma_commit_reservation is called
2647 * to add the page to the reservation map. If the page allocation fails,
2648 * the reservation must be ended instead of committed. vma_end_reservation
2649 * is called in such cases.
2651 * In the normal case, vma_commit_reservation returns the same value
2652 * as the preceding vma_needs_reservation call. The only time this
2653 * is not the case is if a reserve map was changed between calls. It
2654 * is the responsibility of the caller to notice the difference and
2655 * take appropriate action.
2657 * vma_add_reservation is used in error paths where a reservation must
2658 * be restored when a newly allocated huge page must be freed. It is
2659 * to be called after calling vma_needs_reservation to determine if a
2660 * reservation exists.
2662 * vma_del_reservation is used in error paths where an entry in the reserve
2663 * map was created during huge page allocation and must be removed. It is to
2664 * be called after calling vma_needs_reservation to determine if a reservation
2667 enum vma_resv_mode {
2674 static long __vma_reservation_common(struct hstate *h,
2675 struct vm_area_struct *vma, unsigned long addr,
2676 enum vma_resv_mode mode)
2678 struct resv_map *resv;
2681 long dummy_out_regions_needed;
2683 resv = vma_resv_map(vma);
2687 idx = vma_hugecache_offset(h, vma, addr);
2689 case VMA_NEEDS_RESV:
2690 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2691 /* We assume that vma_reservation_* routines always operate on
2692 * 1 page, and that adding to resv map a 1 page entry can only
2693 * ever require 1 region.
2695 VM_BUG_ON(dummy_out_regions_needed != 1);
2697 case VMA_COMMIT_RESV:
2698 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2699 /* region_add calls of range 1 should never fail. */
2703 region_abort(resv, idx, idx + 1, 1);
2707 if (vma->vm_flags & VM_MAYSHARE) {
2708 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2709 /* region_add calls of range 1 should never fail. */
2712 region_abort(resv, idx, idx + 1, 1);
2713 ret = region_del(resv, idx, idx + 1);
2717 if (vma->vm_flags & VM_MAYSHARE) {
2718 region_abort(resv, idx, idx + 1, 1);
2719 ret = region_del(resv, idx, idx + 1);
2721 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2722 /* region_add calls of range 1 should never fail. */
2730 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2733 * We know private mapping must have HPAGE_RESV_OWNER set.
2735 * In most cases, reserves always exist for private mappings.
2736 * However, a file associated with mapping could have been
2737 * hole punched or truncated after reserves were consumed.
2738 * As subsequent fault on such a range will not use reserves.
2739 * Subtle - The reserve map for private mappings has the
2740 * opposite meaning than that of shared mappings. If NO
2741 * entry is in the reserve map, it means a reservation exists.
2742 * If an entry exists in the reserve map, it means the
2743 * reservation has already been consumed. As a result, the
2744 * return value of this routine is the opposite of the
2745 * value returned from reserve map manipulation routines above.
2754 static long vma_needs_reservation(struct hstate *h,
2755 struct vm_area_struct *vma, unsigned long addr)
2757 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2760 static long vma_commit_reservation(struct hstate *h,
2761 struct vm_area_struct *vma, unsigned long addr)
2763 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2766 static void vma_end_reservation(struct hstate *h,
2767 struct vm_area_struct *vma, unsigned long addr)
2769 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2772 static long vma_add_reservation(struct hstate *h,
2773 struct vm_area_struct *vma, unsigned long addr)
2775 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2778 static long vma_del_reservation(struct hstate *h,
2779 struct vm_area_struct *vma, unsigned long addr)
2781 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2785 * This routine is called to restore reservation information on error paths.
2786 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2787 * and the hugetlb mutex should remain held when calling this routine.
2789 * It handles two specific cases:
2790 * 1) A reservation was in place and the folio consumed the reservation.
2791 * hugetlb_restore_reserve is set in the folio.
2792 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2793 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2795 * In case 1, free_huge_folio later in the error path will increment the
2796 * global reserve count. But, free_huge_folio does not have enough context
2797 * to adjust the reservation map. This case deals primarily with private
2798 * mappings. Adjust the reserve map here to be consistent with global
2799 * reserve count adjustments to be made by free_huge_folio. Make sure the
2800 * reserve map indicates there is a reservation present.
2802 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2804 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2805 unsigned long address, struct folio *folio)
2807 long rc = vma_needs_reservation(h, vma, address);
2809 if (folio_test_hugetlb_restore_reserve(folio)) {
2810 if (unlikely(rc < 0))
2812 * Rare out of memory condition in reserve map
2813 * manipulation. Clear hugetlb_restore_reserve so
2814 * that global reserve count will not be incremented
2815 * by free_huge_folio. This will make it appear
2816 * as though the reservation for this folio was
2817 * consumed. This may prevent the task from
2818 * faulting in the folio at a later time. This
2819 * is better than inconsistent global huge page
2820 * accounting of reserve counts.
2822 folio_clear_hugetlb_restore_reserve(folio);
2824 (void)vma_add_reservation(h, vma, address);
2826 vma_end_reservation(h, vma, address);
2830 * This indicates there is an entry in the reserve map
2831 * not added by alloc_hugetlb_folio. We know it was added
2832 * before the alloc_hugetlb_folio call, otherwise
2833 * hugetlb_restore_reserve would be set on the folio.
2834 * Remove the entry so that a subsequent allocation
2835 * does not consume a reservation.
2837 rc = vma_del_reservation(h, vma, address);
2840 * VERY rare out of memory condition. Since
2841 * we can not delete the entry, set
2842 * hugetlb_restore_reserve so that the reserve
2843 * count will be incremented when the folio
2844 * is freed. This reserve will be consumed
2845 * on a subsequent allocation.
2847 folio_set_hugetlb_restore_reserve(folio);
2848 } else if (rc < 0) {
2850 * Rare out of memory condition from
2851 * vma_needs_reservation call. Memory allocation is
2852 * only attempted if a new entry is needed. Therefore,
2853 * this implies there is not an entry in the
2856 * For shared mappings, no entry in the map indicates
2857 * no reservation. We are done.
2859 if (!(vma->vm_flags & VM_MAYSHARE))
2861 * For private mappings, no entry indicates
2862 * a reservation is present. Since we can
2863 * not add an entry, set hugetlb_restore_reserve
2864 * on the folio so reserve count will be
2865 * incremented when freed. This reserve will
2866 * be consumed on a subsequent allocation.
2868 folio_set_hugetlb_restore_reserve(folio);
2871 * No reservation present, do nothing
2873 vma_end_reservation(h, vma, address);
2878 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2880 * @h: struct hstate old page belongs to
2881 * @old_folio: Old folio to dissolve
2882 * @list: List to isolate the page in case we need to
2883 * Returns 0 on success, otherwise negated error.
2885 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2886 struct folio *old_folio, struct list_head *list)
2888 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2889 int nid = folio_nid(old_folio);
2890 struct folio *new_folio;
2894 * Before dissolving the folio, we need to allocate a new one for the
2895 * pool to remain stable. Here, we allocate the folio and 'prep' it
2896 * by doing everything but actually updating counters and adding to
2897 * the pool. This simplifies and let us do most of the processing
2900 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2903 __prep_new_hugetlb_folio(h, new_folio);
2906 spin_lock_irq(&hugetlb_lock);
2907 if (!folio_test_hugetlb(old_folio)) {
2909 * Freed from under us. Drop new_folio too.
2912 } else if (folio_ref_count(old_folio)) {
2916 * Someone has grabbed the folio, try to isolate it here.
2917 * Fail with -EBUSY if not possible.
2919 spin_unlock_irq(&hugetlb_lock);
2920 isolated = isolate_hugetlb(old_folio, list);
2921 ret = isolated ? 0 : -EBUSY;
2922 spin_lock_irq(&hugetlb_lock);
2924 } else if (!folio_test_hugetlb_freed(old_folio)) {
2926 * Folio's refcount is 0 but it has not been enqueued in the
2927 * freelist yet. Race window is small, so we can succeed here if
2930 spin_unlock_irq(&hugetlb_lock);
2935 * Ok, old_folio is still a genuine free hugepage. Remove it from
2936 * the freelist and decrease the counters. These will be
2937 * incremented again when calling __prep_account_new_huge_page()
2938 * and enqueue_hugetlb_folio() for new_folio. The counters will
2939 * remain stable since this happens under the lock.
2941 remove_hugetlb_folio(h, old_folio, false);
2944 * Ref count on new_folio is already zero as it was dropped
2945 * earlier. It can be directly added to the pool free list.
2947 __prep_account_new_huge_page(h, nid);
2948 enqueue_hugetlb_folio(h, new_folio);
2951 * Folio has been replaced, we can safely free the old one.
2953 spin_unlock_irq(&hugetlb_lock);
2954 update_and_free_hugetlb_folio(h, old_folio, false);
2960 spin_unlock_irq(&hugetlb_lock);
2961 /* Folio has a zero ref count, but needs a ref to be freed */
2962 folio_ref_unfreeze(new_folio, 1);
2963 update_and_free_hugetlb_folio(h, new_folio, false);
2968 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2971 struct folio *folio = page_folio(page);
2975 * The page might have been dissolved from under our feet, so make sure
2976 * to carefully check the state under the lock.
2977 * Return success when racing as if we dissolved the page ourselves.
2979 spin_lock_irq(&hugetlb_lock);
2980 if (folio_test_hugetlb(folio)) {
2981 h = folio_hstate(folio);
2983 spin_unlock_irq(&hugetlb_lock);
2986 spin_unlock_irq(&hugetlb_lock);
2989 * Fence off gigantic pages as there is a cyclic dependency between
2990 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2991 * of bailing out right away without further retrying.
2993 if (hstate_is_gigantic(h))
2996 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
2998 else if (!folio_ref_count(folio))
2999 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3004 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3005 unsigned long addr, int avoid_reserve)
3007 struct hugepage_subpool *spool = subpool_vma(vma);
3008 struct hstate *h = hstate_vma(vma);
3009 struct folio *folio;
3010 long map_chg, map_commit;
3013 struct hugetlb_cgroup *h_cg = NULL;
3014 bool deferred_reserve;
3016 idx = hstate_index(h);
3018 * Examine the region/reserve map to determine if the process
3019 * has a reservation for the page to be allocated. A return
3020 * code of zero indicates a reservation exists (no change).
3022 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3024 return ERR_PTR(-ENOMEM);
3027 * Processes that did not create the mapping will have no
3028 * reserves as indicated by the region/reserve map. Check
3029 * that the allocation will not exceed the subpool limit.
3030 * Allocations for MAP_NORESERVE mappings also need to be
3031 * checked against any subpool limit.
3033 if (map_chg || avoid_reserve) {
3034 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3036 vma_end_reservation(h, vma, addr);
3037 return ERR_PTR(-ENOSPC);
3041 * Even though there was no reservation in the region/reserve
3042 * map, there could be reservations associated with the
3043 * subpool that can be used. This would be indicated if the
3044 * return value of hugepage_subpool_get_pages() is zero.
3045 * However, if avoid_reserve is specified we still avoid even
3046 * the subpool reservations.
3052 /* If this allocation is not consuming a reservation, charge it now.
3054 deferred_reserve = map_chg || avoid_reserve;
3055 if (deferred_reserve) {
3056 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3057 idx, pages_per_huge_page(h), &h_cg);
3059 goto out_subpool_put;
3062 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3064 goto out_uncharge_cgroup_reservation;
3066 spin_lock_irq(&hugetlb_lock);
3068 * glb_chg is passed to indicate whether or not a page must be taken
3069 * from the global free pool (global change). gbl_chg == 0 indicates
3070 * a reservation exists for the allocation.
3072 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3074 spin_unlock_irq(&hugetlb_lock);
3075 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3077 goto out_uncharge_cgroup;
3078 spin_lock_irq(&hugetlb_lock);
3079 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3080 folio_set_hugetlb_restore_reserve(folio);
3081 h->resv_huge_pages--;
3083 list_add(&folio->lru, &h->hugepage_activelist);
3084 folio_ref_unfreeze(folio, 1);
3088 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3089 /* If allocation is not consuming a reservation, also store the
3090 * hugetlb_cgroup pointer on the page.
3092 if (deferred_reserve) {
3093 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3097 spin_unlock_irq(&hugetlb_lock);
3099 hugetlb_set_folio_subpool(folio, spool);
3101 map_commit = vma_commit_reservation(h, vma, addr);
3102 if (unlikely(map_chg > map_commit)) {
3104 * The page was added to the reservation map between
3105 * vma_needs_reservation and vma_commit_reservation.
3106 * This indicates a race with hugetlb_reserve_pages.
3107 * Adjust for the subpool count incremented above AND
3108 * in hugetlb_reserve_pages for the same page. Also,
3109 * the reservation count added in hugetlb_reserve_pages
3110 * no longer applies.
3114 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3115 hugetlb_acct_memory(h, -rsv_adjust);
3116 if (deferred_reserve)
3117 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3118 pages_per_huge_page(h), folio);
3122 out_uncharge_cgroup:
3123 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3124 out_uncharge_cgroup_reservation:
3125 if (deferred_reserve)
3126 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3129 if (map_chg || avoid_reserve)
3130 hugepage_subpool_put_pages(spool, 1);
3131 vma_end_reservation(h, vma, addr);
3132 return ERR_PTR(-ENOSPC);
3135 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3136 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3137 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3139 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3142 /* do node specific alloc */
3143 if (nid != NUMA_NO_NODE) {
3144 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3145 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3150 /* allocate from next node when distributing huge pages */
3151 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3152 m = memblock_alloc_try_nid_raw(
3153 huge_page_size(h), huge_page_size(h),
3154 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3156 * Use the beginning of the huge page to store the
3157 * huge_bootmem_page struct (until gather_bootmem
3158 * puts them into the mem_map).
3166 /* Put them into a private list first because mem_map is not up yet */
3167 INIT_LIST_HEAD(&m->list);
3168 list_add(&m->list, &huge_boot_pages);
3174 * Put bootmem huge pages into the standard lists after mem_map is up.
3175 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3177 static void __init gather_bootmem_prealloc(void)
3179 struct huge_bootmem_page *m;
3181 list_for_each_entry(m, &huge_boot_pages, list) {
3182 struct page *page = virt_to_page(m);
3183 struct folio *folio = page_folio(page);
3184 struct hstate *h = m->hstate;
3186 VM_BUG_ON(!hstate_is_gigantic(h));
3187 WARN_ON(folio_ref_count(folio) != 1);
3188 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3189 WARN_ON(folio_test_reserved(folio));
3190 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3191 free_huge_folio(folio); /* add to the hugepage allocator */
3193 /* VERY unlikely inflated ref count on a tail page */
3194 free_gigantic_folio(folio, huge_page_order(h));
3198 * We need to restore the 'stolen' pages to totalram_pages
3199 * in order to fix confusing memory reports from free(1) and
3200 * other side-effects, like CommitLimit going negative.
3202 adjust_managed_page_count(page, pages_per_huge_page(h));
3206 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3211 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3212 if (hstate_is_gigantic(h)) {
3213 if (!alloc_bootmem_huge_page(h, nid))
3216 struct folio *folio;
3217 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3219 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3220 &node_states[N_MEMORY], NULL);
3223 free_huge_folio(folio); /* free it into the hugepage allocator */
3227 if (i == h->max_huge_pages_node[nid])
3230 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3231 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3232 h->max_huge_pages_node[nid], buf, nid, i);
3233 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3234 h->max_huge_pages_node[nid] = i;
3237 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3240 nodemask_t *node_alloc_noretry;
3241 bool node_specific_alloc = false;
3243 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3244 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3245 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3249 /* do node specific alloc */
3250 for_each_online_node(i) {
3251 if (h->max_huge_pages_node[i] > 0) {
3252 hugetlb_hstate_alloc_pages_onenode(h, i);
3253 node_specific_alloc = true;
3257 if (node_specific_alloc)
3260 /* below will do all node balanced alloc */
3261 if (!hstate_is_gigantic(h)) {
3263 * Bit mask controlling how hard we retry per-node allocations.
3264 * Ignore errors as lower level routines can deal with
3265 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3266 * time, we are likely in bigger trouble.
3268 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3271 /* allocations done at boot time */
3272 node_alloc_noretry = NULL;
3275 /* bit mask controlling how hard we retry per-node allocations */
3276 if (node_alloc_noretry)
3277 nodes_clear(*node_alloc_noretry);
3279 for (i = 0; i < h->max_huge_pages; ++i) {
3280 if (hstate_is_gigantic(h)) {
3281 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3283 } else if (!alloc_pool_huge_page(h,
3284 &node_states[N_MEMORY],
3285 node_alloc_noretry))
3289 if (i < h->max_huge_pages) {
3292 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3293 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3294 h->max_huge_pages, buf, i);
3295 h->max_huge_pages = i;
3297 kfree(node_alloc_noretry);
3300 static void __init hugetlb_init_hstates(void)
3302 struct hstate *h, *h2;
3304 for_each_hstate(h) {
3305 /* oversize hugepages were init'ed in early boot */
3306 if (!hstate_is_gigantic(h))
3307 hugetlb_hstate_alloc_pages(h);
3310 * Set demote order for each hstate. Note that
3311 * h->demote_order is initially 0.
3312 * - We can not demote gigantic pages if runtime freeing
3313 * is not supported, so skip this.
3314 * - If CMA allocation is possible, we can not demote
3315 * HUGETLB_PAGE_ORDER or smaller size pages.
3317 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3319 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3321 for_each_hstate(h2) {
3324 if (h2->order < h->order &&
3325 h2->order > h->demote_order)
3326 h->demote_order = h2->order;
3331 static void __init report_hugepages(void)
3335 for_each_hstate(h) {
3338 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3339 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3340 buf, h->free_huge_pages);
3341 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3342 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3346 #ifdef CONFIG_HIGHMEM
3347 static void try_to_free_low(struct hstate *h, unsigned long count,
3348 nodemask_t *nodes_allowed)
3351 LIST_HEAD(page_list);
3353 lockdep_assert_held(&hugetlb_lock);
3354 if (hstate_is_gigantic(h))
3358 * Collect pages to be freed on a list, and free after dropping lock
3360 for_each_node_mask(i, *nodes_allowed) {
3361 struct page *page, *next;
3362 struct list_head *freel = &h->hugepage_freelists[i];
3363 list_for_each_entry_safe(page, next, freel, lru) {
3364 if (count >= h->nr_huge_pages)
3366 if (PageHighMem(page))
3368 remove_hugetlb_folio(h, page_folio(page), false);
3369 list_add(&page->lru, &page_list);
3374 spin_unlock_irq(&hugetlb_lock);
3375 update_and_free_pages_bulk(h, &page_list);
3376 spin_lock_irq(&hugetlb_lock);
3379 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3380 nodemask_t *nodes_allowed)
3386 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3387 * balanced by operating on them in a round-robin fashion.
3388 * Returns 1 if an adjustment was made.
3390 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3395 lockdep_assert_held(&hugetlb_lock);
3396 VM_BUG_ON(delta != -1 && delta != 1);
3399 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3400 if (h->surplus_huge_pages_node[node])
3404 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3405 if (h->surplus_huge_pages_node[node] <
3406 h->nr_huge_pages_node[node])
3413 h->surplus_huge_pages += delta;
3414 h->surplus_huge_pages_node[node] += delta;
3418 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3419 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3420 nodemask_t *nodes_allowed)
3422 unsigned long min_count, ret;
3424 LIST_HEAD(page_list);
3425 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3428 * Bit mask controlling how hard we retry per-node allocations.
3429 * If we can not allocate the bit mask, do not attempt to allocate
3430 * the requested huge pages.
3432 if (node_alloc_noretry)
3433 nodes_clear(*node_alloc_noretry);
3438 * resize_lock mutex prevents concurrent adjustments to number of
3439 * pages in hstate via the proc/sysfs interfaces.
3441 mutex_lock(&h->resize_lock);
3442 flush_free_hpage_work(h);
3443 spin_lock_irq(&hugetlb_lock);
3446 * Check for a node specific request.
3447 * Changing node specific huge page count may require a corresponding
3448 * change to the global count. In any case, the passed node mask
3449 * (nodes_allowed) will restrict alloc/free to the specified node.
3451 if (nid != NUMA_NO_NODE) {
3452 unsigned long old_count = count;
3454 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3456 * User may have specified a large count value which caused the
3457 * above calculation to overflow. In this case, they wanted
3458 * to allocate as many huge pages as possible. Set count to
3459 * largest possible value to align with their intention.
3461 if (count < old_count)
3466 * Gigantic pages runtime allocation depend on the capability for large
3467 * page range allocation.
3468 * If the system does not provide this feature, return an error when
3469 * the user tries to allocate gigantic pages but let the user free the
3470 * boottime allocated gigantic pages.
3472 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3473 if (count > persistent_huge_pages(h)) {
3474 spin_unlock_irq(&hugetlb_lock);
3475 mutex_unlock(&h->resize_lock);
3476 NODEMASK_FREE(node_alloc_noretry);
3479 /* Fall through to decrease pool */
3483 * Increase the pool size
3484 * First take pages out of surplus state. Then make up the
3485 * remaining difference by allocating fresh huge pages.
3487 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3488 * to convert a surplus huge page to a normal huge page. That is
3489 * not critical, though, it just means the overall size of the
3490 * pool might be one hugepage larger than it needs to be, but
3491 * within all the constraints specified by the sysctls.
3493 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3494 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3498 while (count > persistent_huge_pages(h)) {
3500 * If this allocation races such that we no longer need the
3501 * page, free_huge_folio will handle it by freeing the page
3502 * and reducing the surplus.
3504 spin_unlock_irq(&hugetlb_lock);
3506 /* yield cpu to avoid soft lockup */
3509 ret = alloc_pool_huge_page(h, nodes_allowed,
3510 node_alloc_noretry);
3511 spin_lock_irq(&hugetlb_lock);
3515 /* Bail for signals. Probably ctrl-c from user */
3516 if (signal_pending(current))
3521 * Decrease the pool size
3522 * First return free pages to the buddy allocator (being careful
3523 * to keep enough around to satisfy reservations). Then place
3524 * pages into surplus state as needed so the pool will shrink
3525 * to the desired size as pages become free.
3527 * By placing pages into the surplus state independent of the
3528 * overcommit value, we are allowing the surplus pool size to
3529 * exceed overcommit. There are few sane options here. Since
3530 * alloc_surplus_hugetlb_folio() is checking the global counter,
3531 * though, we'll note that we're not allowed to exceed surplus
3532 * and won't grow the pool anywhere else. Not until one of the
3533 * sysctls are changed, or the surplus pages go out of use.
3535 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3536 min_count = max(count, min_count);
3537 try_to_free_low(h, min_count, nodes_allowed);
3540 * Collect pages to be removed on list without dropping lock
3542 while (min_count < persistent_huge_pages(h)) {
3543 page = remove_pool_huge_page(h, nodes_allowed, 0);
3547 list_add(&page->lru, &page_list);
3549 /* free the pages after dropping lock */
3550 spin_unlock_irq(&hugetlb_lock);
3551 update_and_free_pages_bulk(h, &page_list);
3552 flush_free_hpage_work(h);
3553 spin_lock_irq(&hugetlb_lock);
3555 while (count < persistent_huge_pages(h)) {
3556 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3560 h->max_huge_pages = persistent_huge_pages(h);
3561 spin_unlock_irq(&hugetlb_lock);
3562 mutex_unlock(&h->resize_lock);
3564 NODEMASK_FREE(node_alloc_noretry);
3569 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3571 int i, nid = folio_nid(folio);
3572 struct hstate *target_hstate;
3573 struct page *subpage;
3574 struct folio *inner_folio;
3577 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3579 remove_hugetlb_folio_for_demote(h, folio, false);
3580 spin_unlock_irq(&hugetlb_lock);
3582 rc = hugetlb_vmemmap_restore(h, &folio->page);
3584 /* Allocation of vmemmmap failed, we can not demote folio */
3585 spin_lock_irq(&hugetlb_lock);
3586 folio_ref_unfreeze(folio, 1);
3587 add_hugetlb_folio(h, folio, false);
3592 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3593 * sizes as it will not ref count folios.
3595 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3598 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3599 * Without the mutex, pages added to target hstate could be marked
3602 * Note that we already hold h->resize_lock. To prevent deadlock,
3603 * use the convention of always taking larger size hstate mutex first.
3605 mutex_lock(&target_hstate->resize_lock);
3606 for (i = 0; i < pages_per_huge_page(h);
3607 i += pages_per_huge_page(target_hstate)) {
3608 subpage = folio_page(folio, i);
3609 inner_folio = page_folio(subpage);
3610 if (hstate_is_gigantic(target_hstate))
3611 prep_compound_gigantic_folio_for_demote(inner_folio,
3612 target_hstate->order);
3614 prep_compound_page(subpage, target_hstate->order);
3615 folio_change_private(inner_folio, NULL);
3616 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3617 free_huge_folio(inner_folio);
3619 mutex_unlock(&target_hstate->resize_lock);
3621 spin_lock_irq(&hugetlb_lock);
3624 * Not absolutely necessary, but for consistency update max_huge_pages
3625 * based on pool changes for the demoted page.
3627 h->max_huge_pages--;
3628 target_hstate->max_huge_pages +=
3629 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3634 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3635 __must_hold(&hugetlb_lock)
3638 struct folio *folio;
3640 lockdep_assert_held(&hugetlb_lock);
3642 /* We should never get here if no demote order */
3643 if (!h->demote_order) {
3644 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3645 return -EINVAL; /* internal error */
3648 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3649 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3650 if (folio_test_hwpoison(folio))
3652 return demote_free_hugetlb_folio(h, folio);
3657 * Only way to get here is if all pages on free lists are poisoned.
3658 * Return -EBUSY so that caller will not retry.
3663 #define HSTATE_ATTR_RO(_name) \
3664 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3666 #define HSTATE_ATTR_WO(_name) \
3667 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3669 #define HSTATE_ATTR(_name) \
3670 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3672 static struct kobject *hugepages_kobj;
3673 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3675 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3677 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3681 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3682 if (hstate_kobjs[i] == kobj) {
3684 *nidp = NUMA_NO_NODE;
3688 return kobj_to_node_hstate(kobj, nidp);
3691 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3692 struct kobj_attribute *attr, char *buf)
3695 unsigned long nr_huge_pages;
3698 h = kobj_to_hstate(kobj, &nid);
3699 if (nid == NUMA_NO_NODE)
3700 nr_huge_pages = h->nr_huge_pages;
3702 nr_huge_pages = h->nr_huge_pages_node[nid];
3704 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3707 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3708 struct hstate *h, int nid,
3709 unsigned long count, size_t len)
3712 nodemask_t nodes_allowed, *n_mask;
3714 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3717 if (nid == NUMA_NO_NODE) {
3719 * global hstate attribute
3721 if (!(obey_mempolicy &&
3722 init_nodemask_of_mempolicy(&nodes_allowed)))
3723 n_mask = &node_states[N_MEMORY];
3725 n_mask = &nodes_allowed;
3728 * Node specific request. count adjustment happens in
3729 * set_max_huge_pages() after acquiring hugetlb_lock.
3731 init_nodemask_of_node(&nodes_allowed, nid);
3732 n_mask = &nodes_allowed;
3735 err = set_max_huge_pages(h, count, nid, n_mask);
3737 return err ? err : len;
3740 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3741 struct kobject *kobj, const char *buf,
3745 unsigned long count;
3749 err = kstrtoul(buf, 10, &count);
3753 h = kobj_to_hstate(kobj, &nid);
3754 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3757 static ssize_t nr_hugepages_show(struct kobject *kobj,
3758 struct kobj_attribute *attr, char *buf)
3760 return nr_hugepages_show_common(kobj, attr, buf);
3763 static ssize_t nr_hugepages_store(struct kobject *kobj,
3764 struct kobj_attribute *attr, const char *buf, size_t len)
3766 return nr_hugepages_store_common(false, kobj, buf, len);
3768 HSTATE_ATTR(nr_hugepages);
3773 * hstate attribute for optionally mempolicy-based constraint on persistent
3774 * huge page alloc/free.
3776 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3777 struct kobj_attribute *attr,
3780 return nr_hugepages_show_common(kobj, attr, buf);
3783 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3784 struct kobj_attribute *attr, const char *buf, size_t len)
3786 return nr_hugepages_store_common(true, kobj, buf, len);
3788 HSTATE_ATTR(nr_hugepages_mempolicy);
3792 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3793 struct kobj_attribute *attr, char *buf)
3795 struct hstate *h = kobj_to_hstate(kobj, NULL);
3796 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3799 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3800 struct kobj_attribute *attr, const char *buf, size_t count)
3803 unsigned long input;
3804 struct hstate *h = kobj_to_hstate(kobj, NULL);
3806 if (hstate_is_gigantic(h))
3809 err = kstrtoul(buf, 10, &input);
3813 spin_lock_irq(&hugetlb_lock);
3814 h->nr_overcommit_huge_pages = input;
3815 spin_unlock_irq(&hugetlb_lock);
3819 HSTATE_ATTR(nr_overcommit_hugepages);
3821 static ssize_t free_hugepages_show(struct kobject *kobj,
3822 struct kobj_attribute *attr, char *buf)
3825 unsigned long free_huge_pages;
3828 h = kobj_to_hstate(kobj, &nid);
3829 if (nid == NUMA_NO_NODE)
3830 free_huge_pages = h->free_huge_pages;
3832 free_huge_pages = h->free_huge_pages_node[nid];
3834 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3836 HSTATE_ATTR_RO(free_hugepages);
3838 static ssize_t resv_hugepages_show(struct kobject *kobj,
3839 struct kobj_attribute *attr, char *buf)
3841 struct hstate *h = kobj_to_hstate(kobj, NULL);
3842 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3844 HSTATE_ATTR_RO(resv_hugepages);
3846 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3847 struct kobj_attribute *attr, char *buf)
3850 unsigned long surplus_huge_pages;
3853 h = kobj_to_hstate(kobj, &nid);
3854 if (nid == NUMA_NO_NODE)
3855 surplus_huge_pages = h->surplus_huge_pages;
3857 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3859 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3861 HSTATE_ATTR_RO(surplus_hugepages);
3863 static ssize_t demote_store(struct kobject *kobj,
3864 struct kobj_attribute *attr, const char *buf, size_t len)
3866 unsigned long nr_demote;
3867 unsigned long nr_available;
3868 nodemask_t nodes_allowed, *n_mask;
3873 err = kstrtoul(buf, 10, &nr_demote);
3876 h = kobj_to_hstate(kobj, &nid);
3878 if (nid != NUMA_NO_NODE) {
3879 init_nodemask_of_node(&nodes_allowed, nid);
3880 n_mask = &nodes_allowed;
3882 n_mask = &node_states[N_MEMORY];
3885 /* Synchronize with other sysfs operations modifying huge pages */
3886 mutex_lock(&h->resize_lock);
3887 spin_lock_irq(&hugetlb_lock);
3891 * Check for available pages to demote each time thorough the
3892 * loop as demote_pool_huge_page will drop hugetlb_lock.
3894 if (nid != NUMA_NO_NODE)
3895 nr_available = h->free_huge_pages_node[nid];
3897 nr_available = h->free_huge_pages;
3898 nr_available -= h->resv_huge_pages;
3902 err = demote_pool_huge_page(h, n_mask);
3909 spin_unlock_irq(&hugetlb_lock);
3910 mutex_unlock(&h->resize_lock);
3916 HSTATE_ATTR_WO(demote);
3918 static ssize_t demote_size_show(struct kobject *kobj,
3919 struct kobj_attribute *attr, char *buf)
3921 struct hstate *h = kobj_to_hstate(kobj, NULL);
3922 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3924 return sysfs_emit(buf, "%lukB\n", demote_size);
3927 static ssize_t demote_size_store(struct kobject *kobj,
3928 struct kobj_attribute *attr,
3929 const char *buf, size_t count)
3931 struct hstate *h, *demote_hstate;
3932 unsigned long demote_size;
3933 unsigned int demote_order;
3935 demote_size = (unsigned long)memparse(buf, NULL);
3937 demote_hstate = size_to_hstate(demote_size);
3940 demote_order = demote_hstate->order;
3941 if (demote_order < HUGETLB_PAGE_ORDER)
3944 /* demote order must be smaller than hstate order */
3945 h = kobj_to_hstate(kobj, NULL);
3946 if (demote_order >= h->order)
3949 /* resize_lock synchronizes access to demote size and writes */
3950 mutex_lock(&h->resize_lock);
3951 h->demote_order = demote_order;
3952 mutex_unlock(&h->resize_lock);
3956 HSTATE_ATTR(demote_size);
3958 static struct attribute *hstate_attrs[] = {
3959 &nr_hugepages_attr.attr,
3960 &nr_overcommit_hugepages_attr.attr,
3961 &free_hugepages_attr.attr,
3962 &resv_hugepages_attr.attr,
3963 &surplus_hugepages_attr.attr,
3965 &nr_hugepages_mempolicy_attr.attr,
3970 static const struct attribute_group hstate_attr_group = {
3971 .attrs = hstate_attrs,
3974 static struct attribute *hstate_demote_attrs[] = {
3975 &demote_size_attr.attr,
3980 static const struct attribute_group hstate_demote_attr_group = {
3981 .attrs = hstate_demote_attrs,
3984 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3985 struct kobject **hstate_kobjs,
3986 const struct attribute_group *hstate_attr_group)
3989 int hi = hstate_index(h);
3991 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3992 if (!hstate_kobjs[hi])
3995 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3997 kobject_put(hstate_kobjs[hi]);
3998 hstate_kobjs[hi] = NULL;
4002 if (h->demote_order) {
4003 retval = sysfs_create_group(hstate_kobjs[hi],
4004 &hstate_demote_attr_group);
4006 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4007 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4008 kobject_put(hstate_kobjs[hi]);
4009 hstate_kobjs[hi] = NULL;
4018 static bool hugetlb_sysfs_initialized __ro_after_init;
4021 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4022 * with node devices in node_devices[] using a parallel array. The array
4023 * index of a node device or _hstate == node id.
4024 * This is here to avoid any static dependency of the node device driver, in
4025 * the base kernel, on the hugetlb module.
4027 struct node_hstate {
4028 struct kobject *hugepages_kobj;
4029 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4031 static struct node_hstate node_hstates[MAX_NUMNODES];
4034 * A subset of global hstate attributes for node devices
4036 static struct attribute *per_node_hstate_attrs[] = {
4037 &nr_hugepages_attr.attr,
4038 &free_hugepages_attr.attr,
4039 &surplus_hugepages_attr.attr,
4043 static const struct attribute_group per_node_hstate_attr_group = {
4044 .attrs = per_node_hstate_attrs,
4048 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4049 * Returns node id via non-NULL nidp.
4051 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4055 for (nid = 0; nid < nr_node_ids; nid++) {
4056 struct node_hstate *nhs = &node_hstates[nid];
4058 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4059 if (nhs->hstate_kobjs[i] == kobj) {
4071 * Unregister hstate attributes from a single node device.
4072 * No-op if no hstate attributes attached.
4074 void hugetlb_unregister_node(struct node *node)
4077 struct node_hstate *nhs = &node_hstates[node->dev.id];
4079 if (!nhs->hugepages_kobj)
4080 return; /* no hstate attributes */
4082 for_each_hstate(h) {
4083 int idx = hstate_index(h);
4084 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4088 if (h->demote_order)
4089 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4090 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4091 kobject_put(hstate_kobj);
4092 nhs->hstate_kobjs[idx] = NULL;
4095 kobject_put(nhs->hugepages_kobj);
4096 nhs->hugepages_kobj = NULL;
4101 * Register hstate attributes for a single node device.
4102 * No-op if attributes already registered.
4104 void hugetlb_register_node(struct node *node)
4107 struct node_hstate *nhs = &node_hstates[node->dev.id];
4110 if (!hugetlb_sysfs_initialized)
4113 if (nhs->hugepages_kobj)
4114 return; /* already allocated */
4116 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4118 if (!nhs->hugepages_kobj)
4121 for_each_hstate(h) {
4122 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4124 &per_node_hstate_attr_group);
4126 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4127 h->name, node->dev.id);
4128 hugetlb_unregister_node(node);
4135 * hugetlb init time: register hstate attributes for all registered node
4136 * devices of nodes that have memory. All on-line nodes should have
4137 * registered their associated device by this time.
4139 static void __init hugetlb_register_all_nodes(void)
4143 for_each_online_node(nid)
4144 hugetlb_register_node(node_devices[nid]);
4146 #else /* !CONFIG_NUMA */
4148 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4156 static void hugetlb_register_all_nodes(void) { }
4161 static void __init hugetlb_cma_check(void);
4163 static inline __init void hugetlb_cma_check(void)
4168 static void __init hugetlb_sysfs_init(void)
4173 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4174 if (!hugepages_kobj)
4177 for_each_hstate(h) {
4178 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4179 hstate_kobjs, &hstate_attr_group);
4181 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4185 hugetlb_sysfs_initialized = true;
4187 hugetlb_register_all_nodes();
4190 #ifdef CONFIG_SYSCTL
4191 static void hugetlb_sysctl_init(void);
4193 static inline void hugetlb_sysctl_init(void) { }
4196 static int __init hugetlb_init(void)
4200 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4203 if (!hugepages_supported()) {
4204 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4205 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4210 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4211 * architectures depend on setup being done here.
4213 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4214 if (!parsed_default_hugepagesz) {
4216 * If we did not parse a default huge page size, set
4217 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4218 * number of huge pages for this default size was implicitly
4219 * specified, set that here as well.
4220 * Note that the implicit setting will overwrite an explicit
4221 * setting. A warning will be printed in this case.
4223 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4224 if (default_hstate_max_huge_pages) {
4225 if (default_hstate.max_huge_pages) {
4228 string_get_size(huge_page_size(&default_hstate),
4229 1, STRING_UNITS_2, buf, 32);
4230 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4231 default_hstate.max_huge_pages, buf);
4232 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4233 default_hstate_max_huge_pages);
4235 default_hstate.max_huge_pages =
4236 default_hstate_max_huge_pages;
4238 for_each_online_node(i)
4239 default_hstate.max_huge_pages_node[i] =
4240 default_hugepages_in_node[i];
4244 hugetlb_cma_check();
4245 hugetlb_init_hstates();
4246 gather_bootmem_prealloc();
4249 hugetlb_sysfs_init();
4250 hugetlb_cgroup_file_init();
4251 hugetlb_sysctl_init();
4254 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4256 num_fault_mutexes = 1;
4258 hugetlb_fault_mutex_table =
4259 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4261 BUG_ON(!hugetlb_fault_mutex_table);
4263 for (i = 0; i < num_fault_mutexes; i++)
4264 mutex_init(&hugetlb_fault_mutex_table[i]);
4267 subsys_initcall(hugetlb_init);
4269 /* Overwritten by architectures with more huge page sizes */
4270 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4272 return size == HPAGE_SIZE;
4275 void __init hugetlb_add_hstate(unsigned int order)
4280 if (size_to_hstate(PAGE_SIZE << order)) {
4283 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4285 h = &hstates[hugetlb_max_hstate++];
4286 mutex_init(&h->resize_lock);
4288 h->mask = ~(huge_page_size(h) - 1);
4289 for (i = 0; i < MAX_NUMNODES; ++i)
4290 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4291 INIT_LIST_HEAD(&h->hugepage_activelist);
4292 h->next_nid_to_alloc = first_memory_node;
4293 h->next_nid_to_free = first_memory_node;
4294 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4295 huge_page_size(h)/SZ_1K);
4300 bool __init __weak hugetlb_node_alloc_supported(void)
4305 static void __init hugepages_clear_pages_in_node(void)
4307 if (!hugetlb_max_hstate) {
4308 default_hstate_max_huge_pages = 0;
4309 memset(default_hugepages_in_node, 0,
4310 sizeof(default_hugepages_in_node));
4312 parsed_hstate->max_huge_pages = 0;
4313 memset(parsed_hstate->max_huge_pages_node, 0,
4314 sizeof(parsed_hstate->max_huge_pages_node));
4319 * hugepages command line processing
4320 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4321 * specification. If not, ignore the hugepages value. hugepages can also
4322 * be the first huge page command line option in which case it implicitly
4323 * specifies the number of huge pages for the default size.
4325 static int __init hugepages_setup(char *s)
4328 static unsigned long *last_mhp;
4329 int node = NUMA_NO_NODE;
4334 if (!parsed_valid_hugepagesz) {
4335 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4336 parsed_valid_hugepagesz = true;
4341 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4342 * yet, so this hugepages= parameter goes to the "default hstate".
4343 * Otherwise, it goes with the previously parsed hugepagesz or
4344 * default_hugepagesz.
4346 else if (!hugetlb_max_hstate)
4347 mhp = &default_hstate_max_huge_pages;
4349 mhp = &parsed_hstate->max_huge_pages;
4351 if (mhp == last_mhp) {
4352 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4358 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4360 /* Parameter is node format */
4361 if (p[count] == ':') {
4362 if (!hugetlb_node_alloc_supported()) {
4363 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4366 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4368 node = array_index_nospec(tmp, MAX_NUMNODES);
4370 /* Parse hugepages */
4371 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4373 if (!hugetlb_max_hstate)
4374 default_hugepages_in_node[node] = tmp;
4376 parsed_hstate->max_huge_pages_node[node] = tmp;
4378 /* Go to parse next node*/
4379 if (p[count] == ',')
4392 * Global state is always initialized later in hugetlb_init.
4393 * But we need to allocate gigantic hstates here early to still
4394 * use the bootmem allocator.
4396 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4397 hugetlb_hstate_alloc_pages(parsed_hstate);
4404 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4405 hugepages_clear_pages_in_node();
4408 __setup("hugepages=", hugepages_setup);
4411 * hugepagesz command line processing
4412 * A specific huge page size can only be specified once with hugepagesz.
4413 * hugepagesz is followed by hugepages on the command line. The global
4414 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4415 * hugepagesz argument was valid.
4417 static int __init hugepagesz_setup(char *s)
4422 parsed_valid_hugepagesz = false;
4423 size = (unsigned long)memparse(s, NULL);
4425 if (!arch_hugetlb_valid_size(size)) {
4426 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4430 h = size_to_hstate(size);
4433 * hstate for this size already exists. This is normally
4434 * an error, but is allowed if the existing hstate is the
4435 * default hstate. More specifically, it is only allowed if
4436 * the number of huge pages for the default hstate was not
4437 * previously specified.
4439 if (!parsed_default_hugepagesz || h != &default_hstate ||
4440 default_hstate.max_huge_pages) {
4441 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4446 * No need to call hugetlb_add_hstate() as hstate already
4447 * exists. But, do set parsed_hstate so that a following
4448 * hugepages= parameter will be applied to this hstate.
4451 parsed_valid_hugepagesz = true;
4455 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4456 parsed_valid_hugepagesz = true;
4459 __setup("hugepagesz=", hugepagesz_setup);
4462 * default_hugepagesz command line input
4463 * Only one instance of default_hugepagesz allowed on command line.
4465 static int __init default_hugepagesz_setup(char *s)
4470 parsed_valid_hugepagesz = false;
4471 if (parsed_default_hugepagesz) {
4472 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4476 size = (unsigned long)memparse(s, NULL);
4478 if (!arch_hugetlb_valid_size(size)) {
4479 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4483 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4484 parsed_valid_hugepagesz = true;
4485 parsed_default_hugepagesz = true;
4486 default_hstate_idx = hstate_index(size_to_hstate(size));
4489 * The number of default huge pages (for this size) could have been
4490 * specified as the first hugetlb parameter: hugepages=X. If so,
4491 * then default_hstate_max_huge_pages is set. If the default huge
4492 * page size is gigantic (> MAX_ORDER), then the pages must be
4493 * allocated here from bootmem allocator.
4495 if (default_hstate_max_huge_pages) {
4496 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4497 for_each_online_node(i)
4498 default_hstate.max_huge_pages_node[i] =
4499 default_hugepages_in_node[i];
4500 if (hstate_is_gigantic(&default_hstate))
4501 hugetlb_hstate_alloc_pages(&default_hstate);
4502 default_hstate_max_huge_pages = 0;
4507 __setup("default_hugepagesz=", default_hugepagesz_setup);
4509 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4512 struct mempolicy *mpol = get_task_policy(current);
4515 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4516 * (from policy_nodemask) specifically for hugetlb case
4518 if (mpol->mode == MPOL_BIND &&
4519 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4520 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4521 return &mpol->nodes;
4526 static unsigned int allowed_mems_nr(struct hstate *h)
4529 unsigned int nr = 0;
4530 nodemask_t *mbind_nodemask;
4531 unsigned int *array = h->free_huge_pages_node;
4532 gfp_t gfp_mask = htlb_alloc_mask(h);
4534 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4535 for_each_node_mask(node, cpuset_current_mems_allowed) {
4536 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4543 #ifdef CONFIG_SYSCTL
4544 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4545 void *buffer, size_t *length,
4546 loff_t *ppos, unsigned long *out)
4548 struct ctl_table dup_table;
4551 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4552 * can duplicate the @table and alter the duplicate of it.
4555 dup_table.data = out;
4557 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4560 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4561 struct ctl_table *table, int write,
4562 void *buffer, size_t *length, loff_t *ppos)
4564 struct hstate *h = &default_hstate;
4565 unsigned long tmp = h->max_huge_pages;
4568 if (!hugepages_supported())
4571 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4577 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4578 NUMA_NO_NODE, tmp, *length);
4583 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4584 void *buffer, size_t *length, loff_t *ppos)
4587 return hugetlb_sysctl_handler_common(false, table, write,
4588 buffer, length, ppos);
4592 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4593 void *buffer, size_t *length, loff_t *ppos)
4595 return hugetlb_sysctl_handler_common(true, table, write,
4596 buffer, length, ppos);
4598 #endif /* CONFIG_NUMA */
4600 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4601 void *buffer, size_t *length, loff_t *ppos)
4603 struct hstate *h = &default_hstate;
4607 if (!hugepages_supported())
4610 tmp = h->nr_overcommit_huge_pages;
4612 if (write && hstate_is_gigantic(h))
4615 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4621 spin_lock_irq(&hugetlb_lock);
4622 h->nr_overcommit_huge_pages = tmp;
4623 spin_unlock_irq(&hugetlb_lock);
4629 static struct ctl_table hugetlb_table[] = {
4631 .procname = "nr_hugepages",
4633 .maxlen = sizeof(unsigned long),
4635 .proc_handler = hugetlb_sysctl_handler,
4639 .procname = "nr_hugepages_mempolicy",
4641 .maxlen = sizeof(unsigned long),
4643 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4647 .procname = "hugetlb_shm_group",
4648 .data = &sysctl_hugetlb_shm_group,
4649 .maxlen = sizeof(gid_t),
4651 .proc_handler = proc_dointvec,
4654 .procname = "nr_overcommit_hugepages",
4656 .maxlen = sizeof(unsigned long),
4658 .proc_handler = hugetlb_overcommit_handler,
4663 static void hugetlb_sysctl_init(void)
4665 register_sysctl_init("vm", hugetlb_table);
4667 #endif /* CONFIG_SYSCTL */
4669 void hugetlb_report_meminfo(struct seq_file *m)
4672 unsigned long total = 0;
4674 if (!hugepages_supported())
4677 for_each_hstate(h) {
4678 unsigned long count = h->nr_huge_pages;
4680 total += huge_page_size(h) * count;
4682 if (h == &default_hstate)
4684 "HugePages_Total: %5lu\n"
4685 "HugePages_Free: %5lu\n"
4686 "HugePages_Rsvd: %5lu\n"
4687 "HugePages_Surp: %5lu\n"
4688 "Hugepagesize: %8lu kB\n",
4692 h->surplus_huge_pages,
4693 huge_page_size(h) / SZ_1K);
4696 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4699 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4701 struct hstate *h = &default_hstate;
4703 if (!hugepages_supported())
4706 return sysfs_emit_at(buf, len,
4707 "Node %d HugePages_Total: %5u\n"
4708 "Node %d HugePages_Free: %5u\n"
4709 "Node %d HugePages_Surp: %5u\n",
4710 nid, h->nr_huge_pages_node[nid],
4711 nid, h->free_huge_pages_node[nid],
4712 nid, h->surplus_huge_pages_node[nid]);
4715 void hugetlb_show_meminfo_node(int nid)
4719 if (!hugepages_supported())
4723 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4725 h->nr_huge_pages_node[nid],
4726 h->free_huge_pages_node[nid],
4727 h->surplus_huge_pages_node[nid],
4728 huge_page_size(h) / SZ_1K);
4731 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4733 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4734 K(atomic_long_read(&mm->hugetlb_usage)));
4737 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4738 unsigned long hugetlb_total_pages(void)
4741 unsigned long nr_total_pages = 0;
4744 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4745 return nr_total_pages;
4748 static int hugetlb_acct_memory(struct hstate *h, long delta)
4755 spin_lock_irq(&hugetlb_lock);
4757 * When cpuset is configured, it breaks the strict hugetlb page
4758 * reservation as the accounting is done on a global variable. Such
4759 * reservation is completely rubbish in the presence of cpuset because
4760 * the reservation is not checked against page availability for the
4761 * current cpuset. Application can still potentially OOM'ed by kernel
4762 * with lack of free htlb page in cpuset that the task is in.
4763 * Attempt to enforce strict accounting with cpuset is almost
4764 * impossible (or too ugly) because cpuset is too fluid that
4765 * task or memory node can be dynamically moved between cpusets.
4767 * The change of semantics for shared hugetlb mapping with cpuset is
4768 * undesirable. However, in order to preserve some of the semantics,
4769 * we fall back to check against current free page availability as
4770 * a best attempt and hopefully to minimize the impact of changing
4771 * semantics that cpuset has.
4773 * Apart from cpuset, we also have memory policy mechanism that
4774 * also determines from which node the kernel will allocate memory
4775 * in a NUMA system. So similar to cpuset, we also should consider
4776 * the memory policy of the current task. Similar to the description
4780 if (gather_surplus_pages(h, delta) < 0)
4783 if (delta > allowed_mems_nr(h)) {
4784 return_unused_surplus_pages(h, delta);
4791 return_unused_surplus_pages(h, (unsigned long) -delta);
4794 spin_unlock_irq(&hugetlb_lock);
4798 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4800 struct resv_map *resv = vma_resv_map(vma);
4803 * HPAGE_RESV_OWNER indicates a private mapping.
4804 * This new VMA should share its siblings reservation map if present.
4805 * The VMA will only ever have a valid reservation map pointer where
4806 * it is being copied for another still existing VMA. As that VMA
4807 * has a reference to the reservation map it cannot disappear until
4808 * after this open call completes. It is therefore safe to take a
4809 * new reference here without additional locking.
4811 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4812 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4813 kref_get(&resv->refs);
4817 * vma_lock structure for sharable mappings is vma specific.
4818 * Clear old pointer (if copied via vm_area_dup) and allocate
4819 * new structure. Before clearing, make sure vma_lock is not
4822 if (vma->vm_flags & VM_MAYSHARE) {
4823 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4826 if (vma_lock->vma != vma) {
4827 vma->vm_private_data = NULL;
4828 hugetlb_vma_lock_alloc(vma);
4830 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4832 hugetlb_vma_lock_alloc(vma);
4836 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4838 struct hstate *h = hstate_vma(vma);
4839 struct resv_map *resv;
4840 struct hugepage_subpool *spool = subpool_vma(vma);
4841 unsigned long reserve, start, end;
4844 hugetlb_vma_lock_free(vma);
4846 resv = vma_resv_map(vma);
4847 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4850 start = vma_hugecache_offset(h, vma, vma->vm_start);
4851 end = vma_hugecache_offset(h, vma, vma->vm_end);
4853 reserve = (end - start) - region_count(resv, start, end);
4854 hugetlb_cgroup_uncharge_counter(resv, start, end);
4857 * Decrement reserve counts. The global reserve count may be
4858 * adjusted if the subpool has a minimum size.
4860 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4861 hugetlb_acct_memory(h, -gbl_reserve);
4864 kref_put(&resv->refs, resv_map_release);
4867 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4869 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4873 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4874 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4875 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4877 if (addr & ~PUD_MASK) {
4879 * hugetlb_vm_op_split is called right before we attempt to
4880 * split the VMA. We will need to unshare PMDs in the old and
4881 * new VMAs, so let's unshare before we split.
4883 unsigned long floor = addr & PUD_MASK;
4884 unsigned long ceil = floor + PUD_SIZE;
4886 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4887 hugetlb_unshare_pmds(vma, floor, ceil);
4893 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4895 return huge_page_size(hstate_vma(vma));
4899 * We cannot handle pagefaults against hugetlb pages at all. They cause
4900 * handle_mm_fault() to try to instantiate regular-sized pages in the
4901 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4904 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4911 * When a new function is introduced to vm_operations_struct and added
4912 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4913 * This is because under System V memory model, mappings created via
4914 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4915 * their original vm_ops are overwritten with shm_vm_ops.
4917 const struct vm_operations_struct hugetlb_vm_ops = {
4918 .fault = hugetlb_vm_op_fault,
4919 .open = hugetlb_vm_op_open,
4920 .close = hugetlb_vm_op_close,
4921 .may_split = hugetlb_vm_op_split,
4922 .pagesize = hugetlb_vm_op_pagesize,
4925 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4929 unsigned int shift = huge_page_shift(hstate_vma(vma));
4932 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4933 vma->vm_page_prot)));
4935 entry = huge_pte_wrprotect(mk_huge_pte(page,
4936 vma->vm_page_prot));
4938 entry = pte_mkyoung(entry);
4939 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4944 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4945 unsigned long address, pte_t *ptep)
4949 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4950 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4951 update_mmu_cache(vma, address, ptep);
4954 bool is_hugetlb_entry_migration(pte_t pte)
4958 if (huge_pte_none(pte) || pte_present(pte))
4960 swp = pte_to_swp_entry(pte);
4961 if (is_migration_entry(swp))
4967 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4971 if (huge_pte_none(pte) || pte_present(pte))
4973 swp = pte_to_swp_entry(pte);
4974 if (is_hwpoison_entry(swp))
4981 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4982 struct folio *new_folio, pte_t old, unsigned long sz)
4984 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
4986 __folio_mark_uptodate(new_folio);
4987 hugepage_add_new_anon_rmap(new_folio, vma, addr);
4988 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
4989 newpte = huge_pte_mkuffd_wp(newpte);
4990 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
4991 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4992 folio_set_hugetlb_migratable(new_folio);
4995 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4996 struct vm_area_struct *dst_vma,
4997 struct vm_area_struct *src_vma)
4999 pte_t *src_pte, *dst_pte, entry;
5000 struct folio *pte_folio;
5002 bool cow = is_cow_mapping(src_vma->vm_flags);
5003 struct hstate *h = hstate_vma(src_vma);
5004 unsigned long sz = huge_page_size(h);
5005 unsigned long npages = pages_per_huge_page(h);
5006 struct mmu_notifier_range range;
5007 unsigned long last_addr_mask;
5011 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5014 mmu_notifier_invalidate_range_start(&range);
5015 vma_assert_write_locked(src_vma);
5016 raw_write_seqcount_begin(&src->write_protect_seq);
5019 * For shared mappings the vma lock must be held before
5020 * calling hugetlb_walk() in the src vma. Otherwise, the
5021 * returned ptep could go away if part of a shared pmd and
5022 * another thread calls huge_pmd_unshare.
5024 hugetlb_vma_lock_read(src_vma);
5027 last_addr_mask = hugetlb_mask_last_page(h);
5028 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5029 spinlock_t *src_ptl, *dst_ptl;
5030 src_pte = hugetlb_walk(src_vma, addr, sz);
5032 addr |= last_addr_mask;
5035 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5042 * If the pagetables are shared don't copy or take references.
5044 * dst_pte == src_pte is the common case of src/dest sharing.
5045 * However, src could have 'unshared' and dst shares with
5046 * another vma. So page_count of ptep page is checked instead
5047 * to reliably determine whether pte is shared.
5049 if (page_count(virt_to_page(dst_pte)) > 1) {
5050 addr |= last_addr_mask;
5054 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5055 src_ptl = huge_pte_lockptr(h, src, src_pte);
5056 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5057 entry = huge_ptep_get(src_pte);
5059 if (huge_pte_none(entry)) {
5061 * Skip if src entry none.
5064 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5065 if (!userfaultfd_wp(dst_vma))
5066 entry = huge_pte_clear_uffd_wp(entry);
5067 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5068 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5069 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5070 bool uffd_wp = pte_swp_uffd_wp(entry);
5072 if (!is_readable_migration_entry(swp_entry) && cow) {
5074 * COW mappings require pages in both
5075 * parent and child to be set to read.
5077 swp_entry = make_readable_migration_entry(
5078 swp_offset(swp_entry));
5079 entry = swp_entry_to_pte(swp_entry);
5080 if (userfaultfd_wp(src_vma) && uffd_wp)
5081 entry = pte_swp_mkuffd_wp(entry);
5082 set_huge_pte_at(src, addr, src_pte, entry, sz);
5084 if (!userfaultfd_wp(dst_vma))
5085 entry = huge_pte_clear_uffd_wp(entry);
5086 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5087 } else if (unlikely(is_pte_marker(entry))) {
5088 pte_marker marker = copy_pte_marker(
5089 pte_to_swp_entry(entry), dst_vma);
5092 set_huge_pte_at(dst, addr, dst_pte,
5093 make_pte_marker(marker), sz);
5095 entry = huge_ptep_get(src_pte);
5096 pte_folio = page_folio(pte_page(entry));
5097 folio_get(pte_folio);
5100 * Failing to duplicate the anon rmap is a rare case
5101 * where we see pinned hugetlb pages while they're
5102 * prone to COW. We need to do the COW earlier during
5105 * When pre-allocating the page or copying data, we
5106 * need to be without the pgtable locks since we could
5107 * sleep during the process.
5109 if (!folio_test_anon(pte_folio)) {
5110 page_dup_file_rmap(&pte_folio->page, true);
5111 } else if (page_try_dup_anon_rmap(&pte_folio->page,
5113 pte_t src_pte_old = entry;
5114 struct folio *new_folio;
5116 spin_unlock(src_ptl);
5117 spin_unlock(dst_ptl);
5118 /* Do not use reserve as it's private owned */
5119 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5120 if (IS_ERR(new_folio)) {
5121 folio_put(pte_folio);
5122 ret = PTR_ERR(new_folio);
5125 ret = copy_user_large_folio(new_folio,
5128 folio_put(pte_folio);
5130 folio_put(new_folio);
5134 /* Install the new hugetlb folio if src pte stable */
5135 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5136 src_ptl = huge_pte_lockptr(h, src, src_pte);
5137 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5138 entry = huge_ptep_get(src_pte);
5139 if (!pte_same(src_pte_old, entry)) {
5140 restore_reserve_on_error(h, dst_vma, addr,
5142 folio_put(new_folio);
5143 /* huge_ptep of dst_pte won't change as in child */
5146 hugetlb_install_folio(dst_vma, dst_pte, addr,
5147 new_folio, src_pte_old, sz);
5148 spin_unlock(src_ptl);
5149 spin_unlock(dst_ptl);
5155 * No need to notify as we are downgrading page
5156 * table protection not changing it to point
5159 * See Documentation/mm/mmu_notifier.rst
5161 huge_ptep_set_wrprotect(src, addr, src_pte);
5162 entry = huge_pte_wrprotect(entry);
5165 if (!userfaultfd_wp(dst_vma))
5166 entry = huge_pte_clear_uffd_wp(entry);
5168 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5169 hugetlb_count_add(npages, dst);
5171 spin_unlock(src_ptl);
5172 spin_unlock(dst_ptl);
5176 raw_write_seqcount_end(&src->write_protect_seq);
5177 mmu_notifier_invalidate_range_end(&range);
5179 hugetlb_vma_unlock_read(src_vma);
5185 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5186 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5189 struct hstate *h = hstate_vma(vma);
5190 struct mm_struct *mm = vma->vm_mm;
5191 spinlock_t *src_ptl, *dst_ptl;
5194 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5195 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5198 * We don't have to worry about the ordering of src and dst ptlocks
5199 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5201 if (src_ptl != dst_ptl)
5202 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5204 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5205 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5207 if (src_ptl != dst_ptl)
5208 spin_unlock(src_ptl);
5209 spin_unlock(dst_ptl);
5212 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5213 struct vm_area_struct *new_vma,
5214 unsigned long old_addr, unsigned long new_addr,
5217 struct hstate *h = hstate_vma(vma);
5218 struct address_space *mapping = vma->vm_file->f_mapping;
5219 unsigned long sz = huge_page_size(h);
5220 struct mm_struct *mm = vma->vm_mm;
5221 unsigned long old_end = old_addr + len;
5222 unsigned long last_addr_mask;
5223 pte_t *src_pte, *dst_pte;
5224 struct mmu_notifier_range range;
5225 bool shared_pmd = false;
5227 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5229 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5231 * In case of shared PMDs, we should cover the maximum possible
5234 flush_cache_range(vma, range.start, range.end);
5236 mmu_notifier_invalidate_range_start(&range);
5237 last_addr_mask = hugetlb_mask_last_page(h);
5238 /* Prevent race with file truncation */
5239 hugetlb_vma_lock_write(vma);
5240 i_mmap_lock_write(mapping);
5241 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5242 src_pte = hugetlb_walk(vma, old_addr, sz);
5244 old_addr |= last_addr_mask;
5245 new_addr |= last_addr_mask;
5248 if (huge_pte_none(huge_ptep_get(src_pte)))
5251 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5253 old_addr |= last_addr_mask;
5254 new_addr |= last_addr_mask;
5258 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5262 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5266 flush_hugetlb_tlb_range(vma, range.start, range.end);
5268 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5269 mmu_notifier_invalidate_range_end(&range);
5270 i_mmap_unlock_write(mapping);
5271 hugetlb_vma_unlock_write(vma);
5273 return len + old_addr - old_end;
5276 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5277 unsigned long start, unsigned long end,
5278 struct page *ref_page, zap_flags_t zap_flags)
5280 struct mm_struct *mm = vma->vm_mm;
5281 unsigned long address;
5286 struct hstate *h = hstate_vma(vma);
5287 unsigned long sz = huge_page_size(h);
5288 unsigned long last_addr_mask;
5289 bool force_flush = false;
5291 WARN_ON(!is_vm_hugetlb_page(vma));
5292 BUG_ON(start & ~huge_page_mask(h));
5293 BUG_ON(end & ~huge_page_mask(h));
5296 * This is a hugetlb vma, all the pte entries should point
5299 tlb_change_page_size(tlb, sz);
5300 tlb_start_vma(tlb, vma);
5302 last_addr_mask = hugetlb_mask_last_page(h);
5304 for (; address < end; address += sz) {
5305 ptep = hugetlb_walk(vma, address, sz);
5307 address |= last_addr_mask;
5311 ptl = huge_pte_lock(h, mm, ptep);
5312 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5314 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5316 address |= last_addr_mask;
5320 pte = huge_ptep_get(ptep);
5321 if (huge_pte_none(pte)) {
5327 * Migrating hugepage or HWPoisoned hugepage is already
5328 * unmapped and its refcount is dropped, so just clear pte here.
5330 if (unlikely(!pte_present(pte))) {
5332 * If the pte was wr-protected by uffd-wp in any of the
5333 * swap forms, meanwhile the caller does not want to
5334 * drop the uffd-wp bit in this zap, then replace the
5335 * pte with a marker.
5337 if (pte_swp_uffd_wp_any(pte) &&
5338 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5339 set_huge_pte_at(mm, address, ptep,
5340 make_pte_marker(PTE_MARKER_UFFD_WP),
5343 huge_pte_clear(mm, address, ptep, sz);
5348 page = pte_page(pte);
5350 * If a reference page is supplied, it is because a specific
5351 * page is being unmapped, not a range. Ensure the page we
5352 * are about to unmap is the actual page of interest.
5355 if (page != ref_page) {
5360 * Mark the VMA as having unmapped its page so that
5361 * future faults in this VMA will fail rather than
5362 * looking like data was lost
5364 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5367 pte = huge_ptep_get_and_clear(mm, address, ptep);
5368 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5369 if (huge_pte_dirty(pte))
5370 set_page_dirty(page);
5371 /* Leave a uffd-wp pte marker if needed */
5372 if (huge_pte_uffd_wp(pte) &&
5373 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5374 set_huge_pte_at(mm, address, ptep,
5375 make_pte_marker(PTE_MARKER_UFFD_WP),
5377 hugetlb_count_sub(pages_per_huge_page(h), mm);
5378 page_remove_rmap(page, vma, true);
5381 tlb_remove_page_size(tlb, page, huge_page_size(h));
5383 * Bail out after unmapping reference page if supplied
5388 tlb_end_vma(tlb, vma);
5391 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5392 * could defer the flush until now, since by holding i_mmap_rwsem we
5393 * guaranteed that the last refernece would not be dropped. But we must
5394 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5395 * dropped and the last reference to the shared PMDs page might be
5398 * In theory we could defer the freeing of the PMD pages as well, but
5399 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5400 * detect sharing, so we cannot defer the release of the page either.
5401 * Instead, do flush now.
5404 tlb_flush_mmu_tlbonly(tlb);
5407 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5408 struct vm_area_struct *vma, unsigned long start,
5409 unsigned long end, struct page *ref_page,
5410 zap_flags_t zap_flags)
5412 hugetlb_vma_lock_write(vma);
5413 i_mmap_lock_write(vma->vm_file->f_mapping);
5415 /* mmu notification performed in caller */
5416 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5418 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5420 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5421 * When the vma_lock is freed, this makes the vma ineligible
5422 * for pmd sharing. And, i_mmap_rwsem is required to set up
5423 * pmd sharing. This is important as page tables for this
5424 * unmapped range will be asynchrously deleted. If the page
5425 * tables are shared, there will be issues when accessed by
5428 __hugetlb_vma_unlock_write_free(vma);
5429 i_mmap_unlock_write(vma->vm_file->f_mapping);
5431 i_mmap_unlock_write(vma->vm_file->f_mapping);
5432 hugetlb_vma_unlock_write(vma);
5436 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5437 unsigned long end, struct page *ref_page,
5438 zap_flags_t zap_flags)
5440 struct mmu_notifier_range range;
5441 struct mmu_gather tlb;
5443 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5445 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5446 mmu_notifier_invalidate_range_start(&range);
5447 tlb_gather_mmu(&tlb, vma->vm_mm);
5449 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5451 mmu_notifier_invalidate_range_end(&range);
5452 tlb_finish_mmu(&tlb);
5456 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5457 * mapping it owns the reserve page for. The intention is to unmap the page
5458 * from other VMAs and let the children be SIGKILLed if they are faulting the
5461 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5462 struct page *page, unsigned long address)
5464 struct hstate *h = hstate_vma(vma);
5465 struct vm_area_struct *iter_vma;
5466 struct address_space *mapping;
5470 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5471 * from page cache lookup which is in HPAGE_SIZE units.
5473 address = address & huge_page_mask(h);
5474 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5476 mapping = vma->vm_file->f_mapping;
5479 * Take the mapping lock for the duration of the table walk. As
5480 * this mapping should be shared between all the VMAs,
5481 * __unmap_hugepage_range() is called as the lock is already held
5483 i_mmap_lock_write(mapping);
5484 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5485 /* Do not unmap the current VMA */
5486 if (iter_vma == vma)
5490 * Shared VMAs have their own reserves and do not affect
5491 * MAP_PRIVATE accounting but it is possible that a shared
5492 * VMA is using the same page so check and skip such VMAs.
5494 if (iter_vma->vm_flags & VM_MAYSHARE)
5498 * Unmap the page from other VMAs without their own reserves.
5499 * They get marked to be SIGKILLed if they fault in these
5500 * areas. This is because a future no-page fault on this VMA
5501 * could insert a zeroed page instead of the data existing
5502 * from the time of fork. This would look like data corruption
5504 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5505 unmap_hugepage_range(iter_vma, address,
5506 address + huge_page_size(h), page, 0);
5508 i_mmap_unlock_write(mapping);
5512 * hugetlb_wp() should be called with page lock of the original hugepage held.
5513 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5514 * cannot race with other handlers or page migration.
5515 * Keep the pte_same checks anyway to make transition from the mutex easier.
5517 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5518 unsigned long address, pte_t *ptep, unsigned int flags,
5519 struct folio *pagecache_folio, spinlock_t *ptl)
5521 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5522 pte_t pte = huge_ptep_get(ptep);
5523 struct hstate *h = hstate_vma(vma);
5524 struct folio *old_folio;
5525 struct folio *new_folio;
5526 int outside_reserve = 0;
5528 unsigned long haddr = address & huge_page_mask(h);
5529 struct mmu_notifier_range range;
5532 * Never handle CoW for uffd-wp protected pages. It should be only
5533 * handled when the uffd-wp protection is removed.
5535 * Note that only the CoW optimization path (in hugetlb_no_page())
5536 * can trigger this, because hugetlb_fault() will always resolve
5537 * uffd-wp bit first.
5539 if (!unshare && huge_pte_uffd_wp(pte))
5543 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5544 * PTE mapped R/O such as maybe_mkwrite() would do.
5546 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5547 return VM_FAULT_SIGSEGV;
5549 /* Let's take out MAP_SHARED mappings first. */
5550 if (vma->vm_flags & VM_MAYSHARE) {
5551 set_huge_ptep_writable(vma, haddr, ptep);
5555 old_folio = page_folio(pte_page(pte));
5557 delayacct_wpcopy_start();
5561 * If no-one else is actually using this page, we're the exclusive
5562 * owner and can reuse this page.
5564 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5565 if (!PageAnonExclusive(&old_folio->page))
5566 page_move_anon_rmap(&old_folio->page, vma);
5567 if (likely(!unshare))
5568 set_huge_ptep_writable(vma, haddr, ptep);
5570 delayacct_wpcopy_end();
5573 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5574 PageAnonExclusive(&old_folio->page), &old_folio->page);
5577 * If the process that created a MAP_PRIVATE mapping is about to
5578 * perform a COW due to a shared page count, attempt to satisfy
5579 * the allocation without using the existing reserves. The pagecache
5580 * page is used to determine if the reserve at this address was
5581 * consumed or not. If reserves were used, a partial faulted mapping
5582 * at the time of fork() could consume its reserves on COW instead
5583 * of the full address range.
5585 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5586 old_folio != pagecache_folio)
5587 outside_reserve = 1;
5589 folio_get(old_folio);
5592 * Drop page table lock as buddy allocator may be called. It will
5593 * be acquired again before returning to the caller, as expected.
5596 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5598 if (IS_ERR(new_folio)) {
5600 * If a process owning a MAP_PRIVATE mapping fails to COW,
5601 * it is due to references held by a child and an insufficient
5602 * huge page pool. To guarantee the original mappers
5603 * reliability, unmap the page from child processes. The child
5604 * may get SIGKILLed if it later faults.
5606 if (outside_reserve) {
5607 struct address_space *mapping = vma->vm_file->f_mapping;
5611 folio_put(old_folio);
5613 * Drop hugetlb_fault_mutex and vma_lock before
5614 * unmapping. unmapping needs to hold vma_lock
5615 * in write mode. Dropping vma_lock in read mode
5616 * here is OK as COW mappings do not interact with
5619 * Reacquire both after unmap operation.
5621 idx = vma_hugecache_offset(h, vma, haddr);
5622 hash = hugetlb_fault_mutex_hash(mapping, idx);
5623 hugetlb_vma_unlock_read(vma);
5624 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5626 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5628 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5629 hugetlb_vma_lock_read(vma);
5631 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5633 pte_same(huge_ptep_get(ptep), pte)))
5634 goto retry_avoidcopy;
5636 * race occurs while re-acquiring page table
5637 * lock, and our job is done.
5639 delayacct_wpcopy_end();
5643 ret = vmf_error(PTR_ERR(new_folio));
5644 goto out_release_old;
5648 * When the original hugepage is shared one, it does not have
5649 * anon_vma prepared.
5651 if (unlikely(anon_vma_prepare(vma))) {
5653 goto out_release_all;
5656 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5657 ret = VM_FAULT_HWPOISON_LARGE;
5658 goto out_release_all;
5660 __folio_mark_uptodate(new_folio);
5662 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5663 haddr + huge_page_size(h));
5664 mmu_notifier_invalidate_range_start(&range);
5667 * Retake the page table lock to check for racing updates
5668 * before the page tables are altered
5671 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5672 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5673 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5675 /* Break COW or unshare */
5676 huge_ptep_clear_flush(vma, haddr, ptep);
5677 page_remove_rmap(&old_folio->page, vma, true);
5678 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5679 if (huge_pte_uffd_wp(pte))
5680 newpte = huge_pte_mkuffd_wp(newpte);
5681 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5682 folio_set_hugetlb_migratable(new_folio);
5683 /* Make the old page be freed below */
5684 new_folio = old_folio;
5687 mmu_notifier_invalidate_range_end(&range);
5690 * No restore in case of successful pagetable update (Break COW or
5693 if (new_folio != old_folio)
5694 restore_reserve_on_error(h, vma, haddr, new_folio);
5695 folio_put(new_folio);
5697 folio_put(old_folio);
5699 spin_lock(ptl); /* Caller expects lock to be held */
5701 delayacct_wpcopy_end();
5706 * Return whether there is a pagecache page to back given address within VMA.
5708 static bool hugetlbfs_pagecache_present(struct hstate *h,
5709 struct vm_area_struct *vma, unsigned long address)
5711 struct address_space *mapping = vma->vm_file->f_mapping;
5712 pgoff_t idx = vma_hugecache_offset(h, vma, address);
5713 struct folio *folio;
5715 folio = filemap_get_folio(mapping, idx);
5722 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5725 struct inode *inode = mapping->host;
5726 struct hstate *h = hstate_inode(inode);
5729 __folio_set_locked(folio);
5730 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5732 if (unlikely(err)) {
5733 __folio_clear_locked(folio);
5736 folio_clear_hugetlb_restore_reserve(folio);
5739 * mark folio dirty so that it will not be removed from cache/file
5740 * by non-hugetlbfs specific code paths.
5742 folio_mark_dirty(folio);
5744 spin_lock(&inode->i_lock);
5745 inode->i_blocks += blocks_per_huge_page(h);
5746 spin_unlock(&inode->i_lock);
5750 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5751 struct address_space *mapping,
5754 unsigned long haddr,
5756 unsigned long reason)
5759 struct vm_fault vmf = {
5762 .real_address = addr,
5766 * Hard to debug if it ends up being
5767 * used by a callee that assumes
5768 * something about the other
5769 * uninitialized fields... same as in
5775 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5776 * userfault. Also mmap_lock could be dropped due to handling
5777 * userfault, any vma operation should be careful from here.
5779 hugetlb_vma_unlock_read(vma);
5780 hash = hugetlb_fault_mutex_hash(mapping, idx);
5781 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5782 return handle_userfault(&vmf, reason);
5786 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5787 * false if pte changed or is changing.
5789 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5790 pte_t *ptep, pte_t old_pte)
5795 ptl = huge_pte_lock(h, mm, ptep);
5796 same = pte_same(huge_ptep_get(ptep), old_pte);
5802 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5803 struct vm_area_struct *vma,
5804 struct address_space *mapping, pgoff_t idx,
5805 unsigned long address, pte_t *ptep,
5806 pte_t old_pte, unsigned int flags)
5808 struct hstate *h = hstate_vma(vma);
5809 vm_fault_t ret = VM_FAULT_SIGBUS;
5812 struct folio *folio;
5815 unsigned long haddr = address & huge_page_mask(h);
5816 bool new_folio, new_pagecache_folio = false;
5817 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5820 * Currently, we are forced to kill the process in the event the
5821 * original mapper has unmapped pages from the child due to a failed
5822 * COW/unsharing. Warn that such a situation has occurred as it may not
5825 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5826 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5832 * Use page lock to guard against racing truncation
5833 * before we get page_table_lock.
5836 folio = filemap_lock_folio(mapping, idx);
5837 if (IS_ERR(folio)) {
5838 size = i_size_read(mapping->host) >> huge_page_shift(h);
5841 /* Check for page in userfault range */
5842 if (userfaultfd_missing(vma)) {
5844 * Since hugetlb_no_page() was examining pte
5845 * without pgtable lock, we need to re-test under
5846 * lock because the pte may not be stable and could
5847 * have changed from under us. Try to detect
5848 * either changed or during-changing ptes and retry
5849 * properly when needed.
5851 * Note that userfaultfd is actually fine with
5852 * false positives (e.g. caused by pte changed),
5853 * but not wrong logical events (e.g. caused by
5854 * reading a pte during changing). The latter can
5855 * confuse the userspace, so the strictness is very
5856 * much preferred. E.g., MISSING event should
5857 * never happen on the page after UFFDIO_COPY has
5858 * correctly installed the page and returned.
5860 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5865 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5870 folio = alloc_hugetlb_folio(vma, haddr, 0);
5871 if (IS_ERR(folio)) {
5873 * Returning error will result in faulting task being
5874 * sent SIGBUS. The hugetlb fault mutex prevents two
5875 * tasks from racing to fault in the same page which
5876 * could result in false unable to allocate errors.
5877 * Page migration does not take the fault mutex, but
5878 * does a clear then write of pte's under page table
5879 * lock. Page fault code could race with migration,
5880 * notice the clear pte and try to allocate a page
5881 * here. Before returning error, get ptl and make
5882 * sure there really is no pte entry.
5884 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5885 ret = vmf_error(PTR_ERR(folio));
5890 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5891 __folio_mark_uptodate(folio);
5894 if (vma->vm_flags & VM_MAYSHARE) {
5895 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5898 * err can't be -EEXIST which implies someone
5899 * else consumed the reservation since hugetlb
5900 * fault mutex is held when add a hugetlb page
5901 * to the page cache. So it's safe to call
5902 * restore_reserve_on_error() here.
5904 restore_reserve_on_error(h, vma, haddr, folio);
5908 new_pagecache_folio = true;
5911 if (unlikely(anon_vma_prepare(vma))) {
5913 goto backout_unlocked;
5919 * If memory error occurs between mmap() and fault, some process
5920 * don't have hwpoisoned swap entry for errored virtual address.
5921 * So we need to block hugepage fault by PG_hwpoison bit check.
5923 if (unlikely(folio_test_hwpoison(folio))) {
5924 ret = VM_FAULT_HWPOISON_LARGE |
5925 VM_FAULT_SET_HINDEX(hstate_index(h));
5926 goto backout_unlocked;
5929 /* Check for page in userfault range. */
5930 if (userfaultfd_minor(vma)) {
5931 folio_unlock(folio);
5933 /* See comment in userfaultfd_missing() block above */
5934 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5938 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5945 * If we are going to COW a private mapping later, we examine the
5946 * pending reservations for this page now. This will ensure that
5947 * any allocations necessary to record that reservation occur outside
5950 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5951 if (vma_needs_reservation(h, vma, haddr) < 0) {
5953 goto backout_unlocked;
5955 /* Just decrements count, does not deallocate */
5956 vma_end_reservation(h, vma, haddr);
5959 ptl = huge_pte_lock(h, mm, ptep);
5961 /* If pte changed from under us, retry */
5962 if (!pte_same(huge_ptep_get(ptep), old_pte))
5966 hugepage_add_new_anon_rmap(folio, vma, haddr);
5968 page_dup_file_rmap(&folio->page, true);
5969 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
5970 && (vma->vm_flags & VM_SHARED)));
5972 * If this pte was previously wr-protected, keep it wr-protected even
5975 if (unlikely(pte_marker_uffd_wp(old_pte)))
5976 new_pte = huge_pte_mkuffd_wp(new_pte);
5977 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
5979 hugetlb_count_add(pages_per_huge_page(h), mm);
5980 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5981 /* Optimization, do the COW without a second fault */
5982 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
5988 * Only set hugetlb_migratable in newly allocated pages. Existing pages
5989 * found in the pagecache may not have hugetlb_migratable if they have
5990 * been isolated for migration.
5993 folio_set_hugetlb_migratable(folio);
5995 folio_unlock(folio);
5997 hugetlb_vma_unlock_read(vma);
5998 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6004 if (new_folio && !new_pagecache_folio)
6005 restore_reserve_on_error(h, vma, haddr, folio);
6007 folio_unlock(folio);
6013 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6015 unsigned long key[2];
6018 key[0] = (unsigned long) mapping;
6021 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6023 return hash & (num_fault_mutexes - 1);
6027 * For uniprocessor systems we always use a single mutex, so just
6028 * return 0 and avoid the hashing overhead.
6030 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6036 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6037 unsigned long address, unsigned int flags)
6044 struct folio *folio = NULL;
6045 struct folio *pagecache_folio = NULL;
6046 struct hstate *h = hstate_vma(vma);
6047 struct address_space *mapping;
6048 int need_wait_lock = 0;
6049 unsigned long haddr = address & huge_page_mask(h);
6051 /* TODO: Handle faults under the VMA lock */
6052 if (flags & FAULT_FLAG_VMA_LOCK) {
6054 return VM_FAULT_RETRY;
6058 * Serialize hugepage allocation and instantiation, so that we don't
6059 * get spurious allocation failures if two CPUs race to instantiate
6060 * the same page in the page cache.
6062 mapping = vma->vm_file->f_mapping;
6063 idx = vma_hugecache_offset(h, vma, haddr);
6064 hash = hugetlb_fault_mutex_hash(mapping, idx);
6065 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6068 * Acquire vma lock before calling huge_pte_alloc and hold
6069 * until finished with ptep. This prevents huge_pmd_unshare from
6070 * being called elsewhere and making the ptep no longer valid.
6072 hugetlb_vma_lock_read(vma);
6073 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6075 hugetlb_vma_unlock_read(vma);
6076 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6077 return VM_FAULT_OOM;
6080 entry = huge_ptep_get(ptep);
6081 if (huge_pte_none_mostly(entry)) {
6082 if (is_pte_marker(entry)) {
6084 pte_marker_get(pte_to_swp_entry(entry));
6086 if (marker & PTE_MARKER_POISONED) {
6087 ret = VM_FAULT_HWPOISON_LARGE;
6093 * Other PTE markers should be handled the same way as none PTE.
6095 * hugetlb_no_page will drop vma lock and hugetlb fault
6096 * mutex internally, which make us return immediately.
6098 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6105 * entry could be a migration/hwpoison entry at this point, so this
6106 * check prevents the kernel from going below assuming that we have
6107 * an active hugepage in pagecache. This goto expects the 2nd page
6108 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6109 * properly handle it.
6111 if (!pte_present(entry)) {
6112 if (unlikely(is_hugetlb_entry_migration(entry))) {
6114 * Release the hugetlb fault lock now, but retain
6115 * the vma lock, because it is needed to guard the
6116 * huge_pte_lockptr() later in
6117 * migration_entry_wait_huge(). The vma lock will
6118 * be released there.
6120 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6121 migration_entry_wait_huge(vma, ptep);
6123 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6124 ret = VM_FAULT_HWPOISON_LARGE |
6125 VM_FAULT_SET_HINDEX(hstate_index(h));
6130 * If we are going to COW/unshare the mapping later, we examine the
6131 * pending reservations for this page now. This will ensure that any
6132 * allocations necessary to record that reservation occur outside the
6133 * spinlock. Also lookup the pagecache page now as it is used to
6134 * determine if a reservation has been consumed.
6136 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6137 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6138 if (vma_needs_reservation(h, vma, haddr) < 0) {
6142 /* Just decrements count, does not deallocate */
6143 vma_end_reservation(h, vma, haddr);
6145 pagecache_folio = filemap_lock_folio(mapping, idx);
6146 if (IS_ERR(pagecache_folio))
6147 pagecache_folio = NULL;
6150 ptl = huge_pte_lock(h, mm, ptep);
6152 /* Check for a racing update before calling hugetlb_wp() */
6153 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6156 /* Handle userfault-wp first, before trying to lock more pages */
6157 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6158 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6159 struct vm_fault vmf = {
6162 .real_address = address,
6167 if (pagecache_folio) {
6168 folio_unlock(pagecache_folio);
6169 folio_put(pagecache_folio);
6171 hugetlb_vma_unlock_read(vma);
6172 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6173 return handle_userfault(&vmf, VM_UFFD_WP);
6177 * hugetlb_wp() requires page locks of pte_page(entry) and
6178 * pagecache_folio, so here we need take the former one
6179 * when folio != pagecache_folio or !pagecache_folio.
6181 folio = page_folio(pte_page(entry));
6182 if (folio != pagecache_folio)
6183 if (!folio_trylock(folio)) {
6190 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6191 if (!huge_pte_write(entry)) {
6192 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6193 pagecache_folio, ptl);
6195 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6196 entry = huge_pte_mkdirty(entry);
6199 entry = pte_mkyoung(entry);
6200 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6201 flags & FAULT_FLAG_WRITE))
6202 update_mmu_cache(vma, haddr, ptep);
6204 if (folio != pagecache_folio)
6205 folio_unlock(folio);
6210 if (pagecache_folio) {
6211 folio_unlock(pagecache_folio);
6212 folio_put(pagecache_folio);
6215 hugetlb_vma_unlock_read(vma);
6216 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6218 * Generally it's safe to hold refcount during waiting page lock. But
6219 * here we just wait to defer the next page fault to avoid busy loop and
6220 * the page is not used after unlocked before returning from the current
6221 * page fault. So we are safe from accessing freed page, even if we wait
6222 * here without taking refcount.
6225 folio_wait_locked(folio);
6229 #ifdef CONFIG_USERFAULTFD
6231 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6232 * with modifications for hugetlb pages.
6234 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6235 struct vm_area_struct *dst_vma,
6236 unsigned long dst_addr,
6237 unsigned long src_addr,
6239 struct folio **foliop)
6241 struct mm_struct *dst_mm = dst_vma->vm_mm;
6242 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6243 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6244 struct hstate *h = hstate_vma(dst_vma);
6245 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6246 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6248 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6252 struct folio *folio;
6254 bool folio_in_pagecache = false;
6256 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6257 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6259 /* Don't overwrite any existing PTEs (even markers) */
6260 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6265 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6266 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6269 /* No need to invalidate - it was non-present before */
6270 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6278 folio = filemap_lock_folio(mapping, idx);
6281 folio_in_pagecache = true;
6282 } else if (!*foliop) {
6283 /* If a folio already exists, then it's UFFDIO_COPY for
6284 * a non-missing case. Return -EEXIST.
6287 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6292 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6293 if (IS_ERR(folio)) {
6298 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6301 /* fallback to copy_from_user outside mmap_lock */
6302 if (unlikely(ret)) {
6304 /* Free the allocated folio which may have
6305 * consumed a reservation.
6307 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6310 /* Allocate a temporary folio to hold the copied
6313 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6319 /* Set the outparam foliop and return to the caller to
6320 * copy the contents outside the lock. Don't free the
6327 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6334 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6335 if (IS_ERR(folio)) {
6341 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6351 * The memory barrier inside __folio_mark_uptodate makes sure that
6352 * preceding stores to the page contents become visible before
6353 * the set_pte_at() write.
6355 __folio_mark_uptodate(folio);
6357 /* Add shared, newly allocated pages to the page cache. */
6358 if (vm_shared && !is_continue) {
6359 size = i_size_read(mapping->host) >> huge_page_shift(h);
6362 goto out_release_nounlock;
6365 * Serialization between remove_inode_hugepages() and
6366 * hugetlb_add_to_page_cache() below happens through the
6367 * hugetlb_fault_mutex_table that here must be hold by
6370 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6372 goto out_release_nounlock;
6373 folio_in_pagecache = true;
6376 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6379 if (folio_test_hwpoison(folio))
6380 goto out_release_unlock;
6383 * We allow to overwrite a pte marker: consider when both MISSING|WP
6384 * registered, we firstly wr-protect a none pte which has no page cache
6385 * page backing it, then access the page.
6388 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6389 goto out_release_unlock;
6391 if (folio_in_pagecache)
6392 page_dup_file_rmap(&folio->page, true);
6394 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6397 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6398 * with wp flag set, don't set pte write bit.
6400 if (wp_enabled || (is_continue && !vm_shared))
6403 writable = dst_vma->vm_flags & VM_WRITE;
6405 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6407 * Always mark UFFDIO_COPY page dirty; note that this may not be
6408 * extremely important for hugetlbfs for now since swapping is not
6409 * supported, but we should still be clear in that this page cannot be
6410 * thrown away at will, even if write bit not set.
6412 _dst_pte = huge_pte_mkdirty(_dst_pte);
6413 _dst_pte = pte_mkyoung(_dst_pte);
6416 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6418 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6420 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6422 /* No need to invalidate - it was non-present before */
6423 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6427 folio_set_hugetlb_migratable(folio);
6428 if (vm_shared || is_continue)
6429 folio_unlock(folio);
6435 if (vm_shared || is_continue)
6436 folio_unlock(folio);
6437 out_release_nounlock:
6438 if (!folio_in_pagecache)
6439 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6443 #endif /* CONFIG_USERFAULTFD */
6445 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6446 unsigned long address, unsigned int flags,
6447 unsigned int *page_mask)
6449 struct hstate *h = hstate_vma(vma);
6450 struct mm_struct *mm = vma->vm_mm;
6451 unsigned long haddr = address & huge_page_mask(h);
6452 struct page *page = NULL;
6457 hugetlb_vma_lock_read(vma);
6458 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6462 ptl = huge_pte_lock(h, mm, pte);
6463 entry = huge_ptep_get(pte);
6464 if (pte_present(entry)) {
6465 page = pte_page(entry);
6467 if (!huge_pte_write(entry)) {
6468 if (flags & FOLL_WRITE) {
6473 if (gup_must_unshare(vma, flags, page)) {
6474 /* Tell the caller to do unsharing */
6475 page = ERR_PTR(-EMLINK);
6480 page += ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6483 * Note that page may be a sub-page, and with vmemmap
6484 * optimizations the page struct may be read only.
6485 * try_grab_page() will increase the ref count on the
6486 * head page, so this will be OK.
6488 * try_grab_page() should always be able to get the page here,
6489 * because we hold the ptl lock and have verified pte_present().
6491 ret = try_grab_page(page, flags);
6493 if (WARN_ON_ONCE(ret)) {
6494 page = ERR_PTR(ret);
6498 *page_mask = (1U << huge_page_order(h)) - 1;
6503 hugetlb_vma_unlock_read(vma);
6506 * Fixup retval for dump requests: if pagecache doesn't exist,
6507 * don't try to allocate a new page but just skip it.
6509 if (!page && (flags & FOLL_DUMP) &&
6510 !hugetlbfs_pagecache_present(h, vma, address))
6511 page = ERR_PTR(-EFAULT);
6516 long hugetlb_change_protection(struct vm_area_struct *vma,
6517 unsigned long address, unsigned long end,
6518 pgprot_t newprot, unsigned long cp_flags)
6520 struct mm_struct *mm = vma->vm_mm;
6521 unsigned long start = address;
6524 struct hstate *h = hstate_vma(vma);
6525 long pages = 0, psize = huge_page_size(h);
6526 bool shared_pmd = false;
6527 struct mmu_notifier_range range;
6528 unsigned long last_addr_mask;
6529 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6530 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6533 * In the case of shared PMDs, the area to flush could be beyond
6534 * start/end. Set range.start/range.end to cover the maximum possible
6535 * range if PMD sharing is possible.
6537 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6539 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6541 BUG_ON(address >= end);
6542 flush_cache_range(vma, range.start, range.end);
6544 mmu_notifier_invalidate_range_start(&range);
6545 hugetlb_vma_lock_write(vma);
6546 i_mmap_lock_write(vma->vm_file->f_mapping);
6547 last_addr_mask = hugetlb_mask_last_page(h);
6548 for (; address < end; address += psize) {
6550 ptep = hugetlb_walk(vma, address, psize);
6553 address |= last_addr_mask;
6557 * Userfaultfd wr-protect requires pgtable
6558 * pre-allocations to install pte markers.
6560 ptep = huge_pte_alloc(mm, vma, address, psize);
6566 ptl = huge_pte_lock(h, mm, ptep);
6567 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6569 * When uffd-wp is enabled on the vma, unshare
6570 * shouldn't happen at all. Warn about it if it
6571 * happened due to some reason.
6573 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6577 address |= last_addr_mask;
6580 pte = huge_ptep_get(ptep);
6581 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6582 /* Nothing to do. */
6583 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6584 swp_entry_t entry = pte_to_swp_entry(pte);
6585 struct page *page = pfn_swap_entry_to_page(entry);
6588 if (is_writable_migration_entry(entry)) {
6590 entry = make_readable_exclusive_migration_entry(
6593 entry = make_readable_migration_entry(
6595 newpte = swp_entry_to_pte(entry);
6600 newpte = pte_swp_mkuffd_wp(newpte);
6601 else if (uffd_wp_resolve)
6602 newpte = pte_swp_clear_uffd_wp(newpte);
6603 if (!pte_same(pte, newpte))
6604 set_huge_pte_at(mm, address, ptep, newpte, psize);
6605 } else if (unlikely(is_pte_marker(pte))) {
6606 /* No other markers apply for now. */
6607 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6608 if (uffd_wp_resolve)
6609 /* Safe to modify directly (non-present->none). */
6610 huge_pte_clear(mm, address, ptep, psize);
6611 } else if (!huge_pte_none(pte)) {
6613 unsigned int shift = huge_page_shift(hstate_vma(vma));
6615 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6616 pte = huge_pte_modify(old_pte, newprot);
6617 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6619 pte = huge_pte_mkuffd_wp(pte);
6620 else if (uffd_wp_resolve)
6621 pte = huge_pte_clear_uffd_wp(pte);
6622 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6626 if (unlikely(uffd_wp))
6627 /* Safe to modify directly (none->non-present). */
6628 set_huge_pte_at(mm, address, ptep,
6629 make_pte_marker(PTE_MARKER_UFFD_WP),
6635 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6636 * may have cleared our pud entry and done put_page on the page table:
6637 * once we release i_mmap_rwsem, another task can do the final put_page
6638 * and that page table be reused and filled with junk. If we actually
6639 * did unshare a page of pmds, flush the range corresponding to the pud.
6642 flush_hugetlb_tlb_range(vma, range.start, range.end);
6644 flush_hugetlb_tlb_range(vma, start, end);
6646 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6647 * downgrading page table protection not changing it to point to a new
6650 * See Documentation/mm/mmu_notifier.rst
6652 i_mmap_unlock_write(vma->vm_file->f_mapping);
6653 hugetlb_vma_unlock_write(vma);
6654 mmu_notifier_invalidate_range_end(&range);
6656 return pages > 0 ? (pages << h->order) : pages;
6659 /* Return true if reservation was successful, false otherwise. */
6660 bool hugetlb_reserve_pages(struct inode *inode,
6662 struct vm_area_struct *vma,
6663 vm_flags_t vm_flags)
6665 long chg = -1, add = -1;
6666 struct hstate *h = hstate_inode(inode);
6667 struct hugepage_subpool *spool = subpool_inode(inode);
6668 struct resv_map *resv_map;
6669 struct hugetlb_cgroup *h_cg = NULL;
6670 long gbl_reserve, regions_needed = 0;
6672 /* This should never happen */
6674 VM_WARN(1, "%s called with a negative range\n", __func__);
6679 * vma specific semaphore used for pmd sharing and fault/truncation
6682 hugetlb_vma_lock_alloc(vma);
6685 * Only apply hugepage reservation if asked. At fault time, an
6686 * attempt will be made for VM_NORESERVE to allocate a page
6687 * without using reserves
6689 if (vm_flags & VM_NORESERVE)
6693 * Shared mappings base their reservation on the number of pages that
6694 * are already allocated on behalf of the file. Private mappings need
6695 * to reserve the full area even if read-only as mprotect() may be
6696 * called to make the mapping read-write. Assume !vma is a shm mapping
6698 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6700 * resv_map can not be NULL as hugetlb_reserve_pages is only
6701 * called for inodes for which resv_maps were created (see
6702 * hugetlbfs_get_inode).
6704 resv_map = inode_resv_map(inode);
6706 chg = region_chg(resv_map, from, to, ®ions_needed);
6708 /* Private mapping. */
6709 resv_map = resv_map_alloc();
6715 set_vma_resv_map(vma, resv_map);
6716 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6722 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6723 chg * pages_per_huge_page(h), &h_cg) < 0)
6726 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6727 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6730 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6734 * There must be enough pages in the subpool for the mapping. If
6735 * the subpool has a minimum size, there may be some global
6736 * reservations already in place (gbl_reserve).
6738 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6739 if (gbl_reserve < 0)
6740 goto out_uncharge_cgroup;
6743 * Check enough hugepages are available for the reservation.
6744 * Hand the pages back to the subpool if there are not
6746 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6750 * Account for the reservations made. Shared mappings record regions
6751 * that have reservations as they are shared by multiple VMAs.
6752 * When the last VMA disappears, the region map says how much
6753 * the reservation was and the page cache tells how much of
6754 * the reservation was consumed. Private mappings are per-VMA and
6755 * only the consumed reservations are tracked. When the VMA
6756 * disappears, the original reservation is the VMA size and the
6757 * consumed reservations are stored in the map. Hence, nothing
6758 * else has to be done for private mappings here
6760 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6761 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6763 if (unlikely(add < 0)) {
6764 hugetlb_acct_memory(h, -gbl_reserve);
6766 } else if (unlikely(chg > add)) {
6768 * pages in this range were added to the reserve
6769 * map between region_chg and region_add. This
6770 * indicates a race with alloc_hugetlb_folio. Adjust
6771 * the subpool and reserve counts modified above
6772 * based on the difference.
6777 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6778 * reference to h_cg->css. See comment below for detail.
6780 hugetlb_cgroup_uncharge_cgroup_rsvd(
6782 (chg - add) * pages_per_huge_page(h), h_cg);
6784 rsv_adjust = hugepage_subpool_put_pages(spool,
6786 hugetlb_acct_memory(h, -rsv_adjust);
6789 * The file_regions will hold their own reference to
6790 * h_cg->css. So we should release the reference held
6791 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6794 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6800 /* put back original number of pages, chg */
6801 (void)hugepage_subpool_put_pages(spool, chg);
6802 out_uncharge_cgroup:
6803 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6804 chg * pages_per_huge_page(h), h_cg);
6806 hugetlb_vma_lock_free(vma);
6807 if (!vma || vma->vm_flags & VM_MAYSHARE)
6808 /* Only call region_abort if the region_chg succeeded but the
6809 * region_add failed or didn't run.
6811 if (chg >= 0 && add < 0)
6812 region_abort(resv_map, from, to, regions_needed);
6813 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
6814 kref_put(&resv_map->refs, resv_map_release);
6815 set_vma_resv_map(vma, NULL);
6820 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6823 struct hstate *h = hstate_inode(inode);
6824 struct resv_map *resv_map = inode_resv_map(inode);
6826 struct hugepage_subpool *spool = subpool_inode(inode);
6830 * Since this routine can be called in the evict inode path for all
6831 * hugetlbfs inodes, resv_map could be NULL.
6834 chg = region_del(resv_map, start, end);
6836 * region_del() can fail in the rare case where a region
6837 * must be split and another region descriptor can not be
6838 * allocated. If end == LONG_MAX, it will not fail.
6844 spin_lock(&inode->i_lock);
6845 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6846 spin_unlock(&inode->i_lock);
6849 * If the subpool has a minimum size, the number of global
6850 * reservations to be released may be adjusted.
6852 * Note that !resv_map implies freed == 0. So (chg - freed)
6853 * won't go negative.
6855 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6856 hugetlb_acct_memory(h, -gbl_reserve);
6861 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6862 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6863 struct vm_area_struct *vma,
6864 unsigned long addr, pgoff_t idx)
6866 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6868 unsigned long sbase = saddr & PUD_MASK;
6869 unsigned long s_end = sbase + PUD_SIZE;
6871 /* Allow segments to share if only one is marked locked */
6872 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6873 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6876 * match the virtual addresses, permission and the alignment of the
6879 * Also, vma_lock (vm_private_data) is required for sharing.
6881 if (pmd_index(addr) != pmd_index(saddr) ||
6882 vm_flags != svm_flags ||
6883 !range_in_vma(svma, sbase, s_end) ||
6884 !svma->vm_private_data)
6890 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6892 unsigned long start = addr & PUD_MASK;
6893 unsigned long end = start + PUD_SIZE;
6895 #ifdef CONFIG_USERFAULTFD
6896 if (uffd_disable_huge_pmd_share(vma))
6900 * check on proper vm_flags and page table alignment
6902 if (!(vma->vm_flags & VM_MAYSHARE))
6904 if (!vma->vm_private_data) /* vma lock required for sharing */
6906 if (!range_in_vma(vma, start, end))
6912 * Determine if start,end range within vma could be mapped by shared pmd.
6913 * If yes, adjust start and end to cover range associated with possible
6914 * shared pmd mappings.
6916 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6917 unsigned long *start, unsigned long *end)
6919 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6920 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6923 * vma needs to span at least one aligned PUD size, and the range
6924 * must be at least partially within in.
6926 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6927 (*end <= v_start) || (*start >= v_end))
6930 /* Extend the range to be PUD aligned for a worst case scenario */
6931 if (*start > v_start)
6932 *start = ALIGN_DOWN(*start, PUD_SIZE);
6935 *end = ALIGN(*end, PUD_SIZE);
6939 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6940 * and returns the corresponding pte. While this is not necessary for the
6941 * !shared pmd case because we can allocate the pmd later as well, it makes the
6942 * code much cleaner. pmd allocation is essential for the shared case because
6943 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6944 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6945 * bad pmd for sharing.
6947 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6948 unsigned long addr, pud_t *pud)
6950 struct address_space *mapping = vma->vm_file->f_mapping;
6951 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6953 struct vm_area_struct *svma;
6954 unsigned long saddr;
6958 i_mmap_lock_read(mapping);
6959 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6963 saddr = page_table_shareable(svma, vma, addr, idx);
6965 spte = hugetlb_walk(svma, saddr,
6966 vma_mmu_pagesize(svma));
6968 get_page(virt_to_page(spte));
6977 spin_lock(&mm->page_table_lock);
6978 if (pud_none(*pud)) {
6979 pud_populate(mm, pud,
6980 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6983 put_page(virt_to_page(spte));
6985 spin_unlock(&mm->page_table_lock);
6987 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6988 i_mmap_unlock_read(mapping);
6993 * unmap huge page backed by shared pte.
6995 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6996 * indicated by page_count > 1, unmap is achieved by clearing pud and
6997 * decrementing the ref count. If count == 1, the pte page is not shared.
6999 * Called with page table lock held.
7001 * returns: 1 successfully unmapped a shared pte page
7002 * 0 the underlying pte page is not shared, or it is the last user
7004 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7005 unsigned long addr, pte_t *ptep)
7007 pgd_t *pgd = pgd_offset(mm, addr);
7008 p4d_t *p4d = p4d_offset(pgd, addr);
7009 pud_t *pud = pud_offset(p4d, addr);
7011 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7012 hugetlb_vma_assert_locked(vma);
7013 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7014 if (page_count(virt_to_page(ptep)) == 1)
7018 put_page(virt_to_page(ptep));
7023 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7025 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7026 unsigned long addr, pud_t *pud)
7031 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7032 unsigned long addr, pte_t *ptep)
7037 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7038 unsigned long *start, unsigned long *end)
7042 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7046 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7048 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7049 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7050 unsigned long addr, unsigned long sz)
7057 pgd = pgd_offset(mm, addr);
7058 p4d = p4d_alloc(mm, pgd, addr);
7061 pud = pud_alloc(mm, p4d, addr);
7063 if (sz == PUD_SIZE) {
7066 BUG_ON(sz != PMD_SIZE);
7067 if (want_pmd_share(vma, addr) && pud_none(*pud))
7068 pte = huge_pmd_share(mm, vma, addr, pud);
7070 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7075 pte_t pteval = ptep_get_lockless(pte);
7077 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7084 * huge_pte_offset() - Walk the page table to resolve the hugepage
7085 * entry at address @addr
7087 * Return: Pointer to page table entry (PUD or PMD) for
7088 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7089 * size @sz doesn't match the hugepage size at this level of the page
7092 pte_t *huge_pte_offset(struct mm_struct *mm,
7093 unsigned long addr, unsigned long sz)
7100 pgd = pgd_offset(mm, addr);
7101 if (!pgd_present(*pgd))
7103 p4d = p4d_offset(pgd, addr);
7104 if (!p4d_present(*p4d))
7107 pud = pud_offset(p4d, addr);
7109 /* must be pud huge, non-present or none */
7110 return (pte_t *)pud;
7111 if (!pud_present(*pud))
7113 /* must have a valid entry and size to go further */
7115 pmd = pmd_offset(pud, addr);
7116 /* must be pmd huge, non-present or none */
7117 return (pte_t *)pmd;
7121 * Return a mask that can be used to update an address to the last huge
7122 * page in a page table page mapping size. Used to skip non-present
7123 * page table entries when linearly scanning address ranges. Architectures
7124 * with unique huge page to page table relationships can define their own
7125 * version of this routine.
7127 unsigned long hugetlb_mask_last_page(struct hstate *h)
7129 unsigned long hp_size = huge_page_size(h);
7131 if (hp_size == PUD_SIZE)
7132 return P4D_SIZE - PUD_SIZE;
7133 else if (hp_size == PMD_SIZE)
7134 return PUD_SIZE - PMD_SIZE;
7141 /* See description above. Architectures can provide their own version. */
7142 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7144 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7145 if (huge_page_size(h) == PMD_SIZE)
7146 return PUD_SIZE - PMD_SIZE;
7151 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7154 * These functions are overwritable if your architecture needs its own
7157 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7161 spin_lock_irq(&hugetlb_lock);
7162 if (!folio_test_hugetlb(folio) ||
7163 !folio_test_hugetlb_migratable(folio) ||
7164 !folio_try_get(folio)) {
7168 folio_clear_hugetlb_migratable(folio);
7169 list_move_tail(&folio->lru, list);
7171 spin_unlock_irq(&hugetlb_lock);
7175 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7180 spin_lock_irq(&hugetlb_lock);
7181 if (folio_test_hugetlb(folio)) {
7183 if (folio_test_hugetlb_freed(folio))
7185 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7186 ret = folio_try_get(folio);
7190 spin_unlock_irq(&hugetlb_lock);
7194 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7195 bool *migratable_cleared)
7199 spin_lock_irq(&hugetlb_lock);
7200 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7201 spin_unlock_irq(&hugetlb_lock);
7205 void folio_putback_active_hugetlb(struct folio *folio)
7207 spin_lock_irq(&hugetlb_lock);
7208 folio_set_hugetlb_migratable(folio);
7209 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7210 spin_unlock_irq(&hugetlb_lock);
7214 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7216 struct hstate *h = folio_hstate(old_folio);
7218 hugetlb_cgroup_migrate(old_folio, new_folio);
7219 set_page_owner_migrate_reason(&new_folio->page, reason);
7222 * transfer temporary state of the new hugetlb folio. This is
7223 * reverse to other transitions because the newpage is going to
7224 * be final while the old one will be freed so it takes over
7225 * the temporary status.
7227 * Also note that we have to transfer the per-node surplus state
7228 * here as well otherwise the global surplus count will not match
7231 if (folio_test_hugetlb_temporary(new_folio)) {
7232 int old_nid = folio_nid(old_folio);
7233 int new_nid = folio_nid(new_folio);
7235 folio_set_hugetlb_temporary(old_folio);
7236 folio_clear_hugetlb_temporary(new_folio);
7240 * There is no need to transfer the per-node surplus state
7241 * when we do not cross the node.
7243 if (new_nid == old_nid)
7245 spin_lock_irq(&hugetlb_lock);
7246 if (h->surplus_huge_pages_node[old_nid]) {
7247 h->surplus_huge_pages_node[old_nid]--;
7248 h->surplus_huge_pages_node[new_nid]++;
7250 spin_unlock_irq(&hugetlb_lock);
7254 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7255 unsigned long start,
7258 struct hstate *h = hstate_vma(vma);
7259 unsigned long sz = huge_page_size(h);
7260 struct mm_struct *mm = vma->vm_mm;
7261 struct mmu_notifier_range range;
7262 unsigned long address;
7266 if (!(vma->vm_flags & VM_MAYSHARE))
7272 flush_cache_range(vma, start, end);
7274 * No need to call adjust_range_if_pmd_sharing_possible(), because
7275 * we have already done the PUD_SIZE alignment.
7277 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7279 mmu_notifier_invalidate_range_start(&range);
7280 hugetlb_vma_lock_write(vma);
7281 i_mmap_lock_write(vma->vm_file->f_mapping);
7282 for (address = start; address < end; address += PUD_SIZE) {
7283 ptep = hugetlb_walk(vma, address, sz);
7286 ptl = huge_pte_lock(h, mm, ptep);
7287 huge_pmd_unshare(mm, vma, address, ptep);
7290 flush_hugetlb_tlb_range(vma, start, end);
7291 i_mmap_unlock_write(vma->vm_file->f_mapping);
7292 hugetlb_vma_unlock_write(vma);
7294 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7295 * Documentation/mm/mmu_notifier.rst.
7297 mmu_notifier_invalidate_range_end(&range);
7301 * This function will unconditionally remove all the shared pmd pgtable entries
7302 * within the specific vma for a hugetlbfs memory range.
7304 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7306 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7307 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7311 static bool cma_reserve_called __initdata;
7313 static int __init cmdline_parse_hugetlb_cma(char *p)
7320 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7323 if (s[count] == ':') {
7324 if (tmp >= MAX_NUMNODES)
7326 nid = array_index_nospec(tmp, MAX_NUMNODES);
7329 tmp = memparse(s, &s);
7330 hugetlb_cma_size_in_node[nid] = tmp;
7331 hugetlb_cma_size += tmp;
7334 * Skip the separator if have one, otherwise
7335 * break the parsing.
7342 hugetlb_cma_size = memparse(p, &p);
7350 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7352 void __init hugetlb_cma_reserve(int order)
7354 unsigned long size, reserved, per_node;
7355 bool node_specific_cma_alloc = false;
7358 cma_reserve_called = true;
7360 if (!hugetlb_cma_size)
7363 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7364 if (hugetlb_cma_size_in_node[nid] == 0)
7367 if (!node_online(nid)) {
7368 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7369 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7370 hugetlb_cma_size_in_node[nid] = 0;
7374 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7375 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7376 nid, (PAGE_SIZE << order) / SZ_1M);
7377 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7378 hugetlb_cma_size_in_node[nid] = 0;
7380 node_specific_cma_alloc = true;
7384 /* Validate the CMA size again in case some invalid nodes specified. */
7385 if (!hugetlb_cma_size)
7388 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7389 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7390 (PAGE_SIZE << order) / SZ_1M);
7391 hugetlb_cma_size = 0;
7395 if (!node_specific_cma_alloc) {
7397 * If 3 GB area is requested on a machine with 4 numa nodes,
7398 * let's allocate 1 GB on first three nodes and ignore the last one.
7400 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7401 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7402 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7406 for_each_online_node(nid) {
7408 char name[CMA_MAX_NAME];
7410 if (node_specific_cma_alloc) {
7411 if (hugetlb_cma_size_in_node[nid] == 0)
7414 size = hugetlb_cma_size_in_node[nid];
7416 size = min(per_node, hugetlb_cma_size - reserved);
7419 size = round_up(size, PAGE_SIZE << order);
7421 snprintf(name, sizeof(name), "hugetlb%d", nid);
7423 * Note that 'order per bit' is based on smallest size that
7424 * may be returned to CMA allocator in the case of
7425 * huge page demotion.
7427 res = cma_declare_contiguous_nid(0, size, 0,
7428 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7430 &hugetlb_cma[nid], nid);
7432 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7438 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7441 if (reserved >= hugetlb_cma_size)
7447 * hugetlb_cma_size is used to determine if allocations from
7448 * cma are possible. Set to zero if no cma regions are set up.
7450 hugetlb_cma_size = 0;
7453 static void __init hugetlb_cma_check(void)
7455 if (!hugetlb_cma_size || cma_reserve_called)
7458 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7461 #endif /* CONFIG_CMA */