1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
39 #include <asm/pgalloc.h>
43 #include <linux/hugetlb.h>
44 #include <linux/hugetlb_cgroup.h>
45 #include <linux/node.h>
46 #include <linux/page_owner.h>
48 #include "hugetlb_vmemmap.h"
50 int hugetlb_max_hstate __read_mostly;
51 unsigned int default_hstate_idx;
52 struct hstate hstates[HUGE_MAX_HSTATE];
55 static struct cma *hugetlb_cma[MAX_NUMNODES];
56 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
57 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
63 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
68 static unsigned long hugetlb_cma_size __initdata;
70 __initdata LIST_HEAD(huge_boot_pages);
72 /* for command line parsing */
73 static struct hstate * __initdata parsed_hstate;
74 static unsigned long __initdata default_hstate_max_huge_pages;
75 static bool __initdata parsed_valid_hugepagesz = true;
76 static bool __initdata parsed_default_hugepagesz;
77 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
80 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
81 * free_huge_pages, and surplus_huge_pages.
83 DEFINE_SPINLOCK(hugetlb_lock);
86 * Serializes faults on the same logical page. This is used to
87 * prevent spurious OOMs when the hugepage pool is fully utilized.
89 static int num_fault_mutexes;
90 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92 /* Forward declaration */
93 static int hugetlb_acct_memory(struct hstate *h, long delta);
94 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
95 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
96 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
97 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
98 unsigned long start, unsigned long end);
100 static inline bool subpool_is_free(struct hugepage_subpool *spool)
104 if (spool->max_hpages != -1)
105 return spool->used_hpages == 0;
106 if (spool->min_hpages != -1)
107 return spool->rsv_hpages == spool->min_hpages;
112 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
113 unsigned long irq_flags)
115 spin_unlock_irqrestore(&spool->lock, irq_flags);
117 /* If no pages are used, and no other handles to the subpool
118 * remain, give up any reservations based on minimum size and
119 * free the subpool */
120 if (subpool_is_free(spool)) {
121 if (spool->min_hpages != -1)
122 hugetlb_acct_memory(spool->hstate,
128 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 struct hugepage_subpool *spool;
133 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
137 spin_lock_init(&spool->lock);
139 spool->max_hpages = max_hpages;
141 spool->min_hpages = min_hpages;
143 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
147 spool->rsv_hpages = min_hpages;
152 void hugepage_put_subpool(struct hugepage_subpool *spool)
156 spin_lock_irqsave(&spool->lock, flags);
157 BUG_ON(!spool->count);
159 unlock_or_release_subpool(spool, flags);
163 * Subpool accounting for allocating and reserving pages.
164 * Return -ENOMEM if there are not enough resources to satisfy the
165 * request. Otherwise, return the number of pages by which the
166 * global pools must be adjusted (upward). The returned value may
167 * only be different than the passed value (delta) in the case where
168 * a subpool minimum size must be maintained.
170 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
178 spin_lock_irq(&spool->lock);
180 if (spool->max_hpages != -1) { /* maximum size accounting */
181 if ((spool->used_hpages + delta) <= spool->max_hpages)
182 spool->used_hpages += delta;
189 /* minimum size accounting */
190 if (spool->min_hpages != -1 && spool->rsv_hpages) {
191 if (delta > spool->rsv_hpages) {
193 * Asking for more reserves than those already taken on
194 * behalf of subpool. Return difference.
196 ret = delta - spool->rsv_hpages;
197 spool->rsv_hpages = 0;
199 ret = 0; /* reserves already accounted for */
200 spool->rsv_hpages -= delta;
205 spin_unlock_irq(&spool->lock);
210 * Subpool accounting for freeing and unreserving pages.
211 * Return the number of global page reservations that must be dropped.
212 * The return value may only be different than the passed value (delta)
213 * in the case where a subpool minimum size must be maintained.
215 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
224 spin_lock_irqsave(&spool->lock, flags);
226 if (spool->max_hpages != -1) /* maximum size accounting */
227 spool->used_hpages -= delta;
229 /* minimum size accounting */
230 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
231 if (spool->rsv_hpages + delta <= spool->min_hpages)
234 ret = spool->rsv_hpages + delta - spool->min_hpages;
236 spool->rsv_hpages += delta;
237 if (spool->rsv_hpages > spool->min_hpages)
238 spool->rsv_hpages = spool->min_hpages;
242 * If hugetlbfs_put_super couldn't free spool due to an outstanding
243 * quota reference, free it now.
245 unlock_or_release_subpool(spool, flags);
250 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
252 return HUGETLBFS_SB(inode->i_sb)->spool;
255 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
257 return subpool_inode(file_inode(vma->vm_file));
261 * hugetlb vma_lock helper routines
263 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
265 if (__vma_shareable_lock(vma)) {
266 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
268 down_read(&vma_lock->rw_sema);
272 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
274 if (__vma_shareable_lock(vma)) {
275 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
277 up_read(&vma_lock->rw_sema);
281 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
283 if (__vma_shareable_lock(vma)) {
284 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
286 down_write(&vma_lock->rw_sema);
290 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
292 if (__vma_shareable_lock(vma)) {
293 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
295 up_write(&vma_lock->rw_sema);
299 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
301 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
303 if (!__vma_shareable_lock(vma))
306 return down_write_trylock(&vma_lock->rw_sema);
309 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
311 if (__vma_shareable_lock(vma)) {
312 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
314 lockdep_assert_held(&vma_lock->rw_sema);
318 void hugetlb_vma_lock_release(struct kref *kref)
320 struct hugetlb_vma_lock *vma_lock = container_of(kref,
321 struct hugetlb_vma_lock, refs);
326 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
328 struct vm_area_struct *vma = vma_lock->vma;
331 * vma_lock structure may or not be released as a result of put,
332 * it certainly will no longer be attached to vma so clear pointer.
333 * Semaphore synchronizes access to vma_lock->vma field.
335 vma_lock->vma = NULL;
336 vma->vm_private_data = NULL;
337 up_write(&vma_lock->rw_sema);
338 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
341 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
343 if (__vma_shareable_lock(vma)) {
344 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
346 __hugetlb_vma_unlock_write_put(vma_lock);
350 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
353 * Only present in sharable vmas.
355 if (!vma || !__vma_shareable_lock(vma))
358 if (vma->vm_private_data) {
359 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
361 down_write(&vma_lock->rw_sema);
362 __hugetlb_vma_unlock_write_put(vma_lock);
366 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
368 struct hugetlb_vma_lock *vma_lock;
370 /* Only establish in (flags) sharable vmas */
371 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
374 /* Should never get here with non-NULL vm_private_data */
375 if (vma->vm_private_data)
378 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
381 * If we can not allocate structure, then vma can not
382 * participate in pmd sharing. This is only a possible
383 * performance enhancement and memory saving issue.
384 * However, the lock is also used to synchronize page
385 * faults with truncation. If the lock is not present,
386 * unlikely races could leave pages in a file past i_size
387 * until the file is removed. Warn in the unlikely case of
388 * allocation failure.
390 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
394 kref_init(&vma_lock->refs);
395 init_rwsem(&vma_lock->rw_sema);
397 vma->vm_private_data = vma_lock;
400 /* Helper that removes a struct file_region from the resv_map cache and returns
403 static struct file_region *
404 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
406 struct file_region *nrg;
408 VM_BUG_ON(resv->region_cache_count <= 0);
410 resv->region_cache_count--;
411 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
412 list_del(&nrg->link);
420 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
421 struct file_region *rg)
423 #ifdef CONFIG_CGROUP_HUGETLB
424 nrg->reservation_counter = rg->reservation_counter;
431 /* Helper that records hugetlb_cgroup uncharge info. */
432 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
434 struct resv_map *resv,
435 struct file_region *nrg)
437 #ifdef CONFIG_CGROUP_HUGETLB
439 nrg->reservation_counter =
440 &h_cg->rsvd_hugepage[hstate_index(h)];
441 nrg->css = &h_cg->css;
443 * The caller will hold exactly one h_cg->css reference for the
444 * whole contiguous reservation region. But this area might be
445 * scattered when there are already some file_regions reside in
446 * it. As a result, many file_regions may share only one css
447 * reference. In order to ensure that one file_region must hold
448 * exactly one h_cg->css reference, we should do css_get for
449 * each file_region and leave the reference held by caller
453 if (!resv->pages_per_hpage)
454 resv->pages_per_hpage = pages_per_huge_page(h);
455 /* pages_per_hpage should be the same for all entries in
458 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
460 nrg->reservation_counter = NULL;
466 static void put_uncharge_info(struct file_region *rg)
468 #ifdef CONFIG_CGROUP_HUGETLB
474 static bool has_same_uncharge_info(struct file_region *rg,
475 struct file_region *org)
477 #ifdef CONFIG_CGROUP_HUGETLB
478 return rg->reservation_counter == org->reservation_counter &&
486 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
488 struct file_region *nrg, *prg;
490 prg = list_prev_entry(rg, link);
491 if (&prg->link != &resv->regions && prg->to == rg->from &&
492 has_same_uncharge_info(prg, rg)) {
496 put_uncharge_info(rg);
502 nrg = list_next_entry(rg, link);
503 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
504 has_same_uncharge_info(nrg, rg)) {
505 nrg->from = rg->from;
508 put_uncharge_info(rg);
514 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
515 long to, struct hstate *h, struct hugetlb_cgroup *cg,
516 long *regions_needed)
518 struct file_region *nrg;
520 if (!regions_needed) {
521 nrg = get_file_region_entry_from_cache(map, from, to);
522 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
523 list_add(&nrg->link, rg);
524 coalesce_file_region(map, nrg);
526 *regions_needed += 1;
532 * Must be called with resv->lock held.
534 * Calling this with regions_needed != NULL will count the number of pages
535 * to be added but will not modify the linked list. And regions_needed will
536 * indicate the number of file_regions needed in the cache to carry out to add
537 * the regions for this range.
539 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
540 struct hugetlb_cgroup *h_cg,
541 struct hstate *h, long *regions_needed)
544 struct list_head *head = &resv->regions;
545 long last_accounted_offset = f;
546 struct file_region *iter, *trg = NULL;
547 struct list_head *rg = NULL;
552 /* In this loop, we essentially handle an entry for the range
553 * [last_accounted_offset, iter->from), at every iteration, with some
556 list_for_each_entry_safe(iter, trg, head, link) {
557 /* Skip irrelevant regions that start before our range. */
558 if (iter->from < f) {
559 /* If this region ends after the last accounted offset,
560 * then we need to update last_accounted_offset.
562 if (iter->to > last_accounted_offset)
563 last_accounted_offset = iter->to;
567 /* When we find a region that starts beyond our range, we've
570 if (iter->from >= t) {
571 rg = iter->link.prev;
575 /* Add an entry for last_accounted_offset -> iter->from, and
576 * update last_accounted_offset.
578 if (iter->from > last_accounted_offset)
579 add += hugetlb_resv_map_add(resv, iter->link.prev,
580 last_accounted_offset,
584 last_accounted_offset = iter->to;
587 /* Handle the case where our range extends beyond
588 * last_accounted_offset.
592 if (last_accounted_offset < t)
593 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
594 t, h, h_cg, regions_needed);
599 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
601 static int allocate_file_region_entries(struct resv_map *resv,
603 __must_hold(&resv->lock)
605 LIST_HEAD(allocated_regions);
606 int to_allocate = 0, i = 0;
607 struct file_region *trg = NULL, *rg = NULL;
609 VM_BUG_ON(regions_needed < 0);
612 * Check for sufficient descriptors in the cache to accommodate
613 * the number of in progress add operations plus regions_needed.
615 * This is a while loop because when we drop the lock, some other call
616 * to region_add or region_del may have consumed some region_entries,
617 * so we keep looping here until we finally have enough entries for
618 * (adds_in_progress + regions_needed).
620 while (resv->region_cache_count <
621 (resv->adds_in_progress + regions_needed)) {
622 to_allocate = resv->adds_in_progress + regions_needed -
623 resv->region_cache_count;
625 /* At this point, we should have enough entries in the cache
626 * for all the existing adds_in_progress. We should only be
627 * needing to allocate for regions_needed.
629 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
631 spin_unlock(&resv->lock);
632 for (i = 0; i < to_allocate; i++) {
633 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
636 list_add(&trg->link, &allocated_regions);
639 spin_lock(&resv->lock);
641 list_splice(&allocated_regions, &resv->region_cache);
642 resv->region_cache_count += to_allocate;
648 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
656 * Add the huge page range represented by [f, t) to the reserve
657 * map. Regions will be taken from the cache to fill in this range.
658 * Sufficient regions should exist in the cache due to the previous
659 * call to region_chg with the same range, but in some cases the cache will not
660 * have sufficient entries due to races with other code doing region_add or
661 * region_del. The extra needed entries will be allocated.
663 * regions_needed is the out value provided by a previous call to region_chg.
665 * Return the number of new huge pages added to the map. This number is greater
666 * than or equal to zero. If file_region entries needed to be allocated for
667 * this operation and we were not able to allocate, it returns -ENOMEM.
668 * region_add of regions of length 1 never allocate file_regions and cannot
669 * fail; region_chg will always allocate at least 1 entry and a region_add for
670 * 1 page will only require at most 1 entry.
672 static long region_add(struct resv_map *resv, long f, long t,
673 long in_regions_needed, struct hstate *h,
674 struct hugetlb_cgroup *h_cg)
676 long add = 0, actual_regions_needed = 0;
678 spin_lock(&resv->lock);
681 /* Count how many regions are actually needed to execute this add. */
682 add_reservation_in_range(resv, f, t, NULL, NULL,
683 &actual_regions_needed);
686 * Check for sufficient descriptors in the cache to accommodate
687 * this add operation. Note that actual_regions_needed may be greater
688 * than in_regions_needed, as the resv_map may have been modified since
689 * the region_chg call. In this case, we need to make sure that we
690 * allocate extra entries, such that we have enough for all the
691 * existing adds_in_progress, plus the excess needed for this
694 if (actual_regions_needed > in_regions_needed &&
695 resv->region_cache_count <
696 resv->adds_in_progress +
697 (actual_regions_needed - in_regions_needed)) {
698 /* region_add operation of range 1 should never need to
699 * allocate file_region entries.
701 VM_BUG_ON(t - f <= 1);
703 if (allocate_file_region_entries(
704 resv, actual_regions_needed - in_regions_needed)) {
711 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
713 resv->adds_in_progress -= in_regions_needed;
715 spin_unlock(&resv->lock);
720 * Examine the existing reserve map and determine how many
721 * huge pages in the specified range [f, t) are NOT currently
722 * represented. This routine is called before a subsequent
723 * call to region_add that will actually modify the reserve
724 * map to add the specified range [f, t). region_chg does
725 * not change the number of huge pages represented by the
726 * map. A number of new file_region structures is added to the cache as a
727 * placeholder, for the subsequent region_add call to use. At least 1
728 * file_region structure is added.
730 * out_regions_needed is the number of regions added to the
731 * resv->adds_in_progress. This value needs to be provided to a follow up call
732 * to region_add or region_abort for proper accounting.
734 * Returns the number of huge pages that need to be added to the existing
735 * reservation map for the range [f, t). This number is greater or equal to
736 * zero. -ENOMEM is returned if a new file_region structure or cache entry
737 * is needed and can not be allocated.
739 static long region_chg(struct resv_map *resv, long f, long t,
740 long *out_regions_needed)
744 spin_lock(&resv->lock);
746 /* Count how many hugepages in this range are NOT represented. */
747 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
750 if (*out_regions_needed == 0)
751 *out_regions_needed = 1;
753 if (allocate_file_region_entries(resv, *out_regions_needed))
756 resv->adds_in_progress += *out_regions_needed;
758 spin_unlock(&resv->lock);
763 * Abort the in progress add operation. The adds_in_progress field
764 * of the resv_map keeps track of the operations in progress between
765 * calls to region_chg and region_add. Operations are sometimes
766 * aborted after the call to region_chg. In such cases, region_abort
767 * is called to decrement the adds_in_progress counter. regions_needed
768 * is the value returned by the region_chg call, it is used to decrement
769 * the adds_in_progress counter.
771 * NOTE: The range arguments [f, t) are not needed or used in this
772 * routine. They are kept to make reading the calling code easier as
773 * arguments will match the associated region_chg call.
775 static void region_abort(struct resv_map *resv, long f, long t,
778 spin_lock(&resv->lock);
779 VM_BUG_ON(!resv->region_cache_count);
780 resv->adds_in_progress -= regions_needed;
781 spin_unlock(&resv->lock);
785 * Delete the specified range [f, t) from the reserve map. If the
786 * t parameter is LONG_MAX, this indicates that ALL regions after f
787 * should be deleted. Locate the regions which intersect [f, t)
788 * and either trim, delete or split the existing regions.
790 * Returns the number of huge pages deleted from the reserve map.
791 * In the normal case, the return value is zero or more. In the
792 * case where a region must be split, a new region descriptor must
793 * be allocated. If the allocation fails, -ENOMEM will be returned.
794 * NOTE: If the parameter t == LONG_MAX, then we will never split
795 * a region and possibly return -ENOMEM. Callers specifying
796 * t == LONG_MAX do not need to check for -ENOMEM error.
798 static long region_del(struct resv_map *resv, long f, long t)
800 struct list_head *head = &resv->regions;
801 struct file_region *rg, *trg;
802 struct file_region *nrg = NULL;
806 spin_lock(&resv->lock);
807 list_for_each_entry_safe(rg, trg, head, link) {
809 * Skip regions before the range to be deleted. file_region
810 * ranges are normally of the form [from, to). However, there
811 * may be a "placeholder" entry in the map which is of the form
812 * (from, to) with from == to. Check for placeholder entries
813 * at the beginning of the range to be deleted.
815 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
821 if (f > rg->from && t < rg->to) { /* Must split region */
823 * Check for an entry in the cache before dropping
824 * lock and attempting allocation.
827 resv->region_cache_count > resv->adds_in_progress) {
828 nrg = list_first_entry(&resv->region_cache,
831 list_del(&nrg->link);
832 resv->region_cache_count--;
836 spin_unlock(&resv->lock);
837 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
844 hugetlb_cgroup_uncharge_file_region(
845 resv, rg, t - f, false);
847 /* New entry for end of split region */
851 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
853 INIT_LIST_HEAD(&nrg->link);
855 /* Original entry is trimmed */
858 list_add(&nrg->link, &rg->link);
863 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
864 del += rg->to - rg->from;
865 hugetlb_cgroup_uncharge_file_region(resv, rg,
866 rg->to - rg->from, true);
872 if (f <= rg->from) { /* Trim beginning of region */
873 hugetlb_cgroup_uncharge_file_region(resv, rg,
874 t - rg->from, false);
878 } else { /* Trim end of region */
879 hugetlb_cgroup_uncharge_file_region(resv, rg,
887 spin_unlock(&resv->lock);
893 * A rare out of memory error was encountered which prevented removal of
894 * the reserve map region for a page. The huge page itself was free'ed
895 * and removed from the page cache. This routine will adjust the subpool
896 * usage count, and the global reserve count if needed. By incrementing
897 * these counts, the reserve map entry which could not be deleted will
898 * appear as a "reserved" entry instead of simply dangling with incorrect
901 void hugetlb_fix_reserve_counts(struct inode *inode)
903 struct hugepage_subpool *spool = subpool_inode(inode);
905 bool reserved = false;
907 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
908 if (rsv_adjust > 0) {
909 struct hstate *h = hstate_inode(inode);
911 if (!hugetlb_acct_memory(h, 1))
913 } else if (!rsv_adjust) {
918 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
922 * Count and return the number of huge pages in the reserve map
923 * that intersect with the range [f, t).
925 static long region_count(struct resv_map *resv, long f, long t)
927 struct list_head *head = &resv->regions;
928 struct file_region *rg;
931 spin_lock(&resv->lock);
932 /* Locate each segment we overlap with, and count that overlap. */
933 list_for_each_entry(rg, head, link) {
942 seg_from = max(rg->from, f);
943 seg_to = min(rg->to, t);
945 chg += seg_to - seg_from;
947 spin_unlock(&resv->lock);
953 * Convert the address within this vma to the page offset within
954 * the mapping, in pagecache page units; huge pages here.
956 static pgoff_t vma_hugecache_offset(struct hstate *h,
957 struct vm_area_struct *vma, unsigned long address)
959 return ((address - vma->vm_start) >> huge_page_shift(h)) +
960 (vma->vm_pgoff >> huge_page_order(h));
963 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
964 unsigned long address)
966 return vma_hugecache_offset(hstate_vma(vma), vma, address);
968 EXPORT_SYMBOL_GPL(linear_hugepage_index);
971 * Return the size of the pages allocated when backing a VMA. In the majority
972 * cases this will be same size as used by the page table entries.
974 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
976 if (vma->vm_ops && vma->vm_ops->pagesize)
977 return vma->vm_ops->pagesize(vma);
980 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
983 * Return the page size being used by the MMU to back a VMA. In the majority
984 * of cases, the page size used by the kernel matches the MMU size. On
985 * architectures where it differs, an architecture-specific 'strong'
986 * version of this symbol is required.
988 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
990 return vma_kernel_pagesize(vma);
994 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
995 * bits of the reservation map pointer, which are always clear due to
998 #define HPAGE_RESV_OWNER (1UL << 0)
999 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1000 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1003 * These helpers are used to track how many pages are reserved for
1004 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1005 * is guaranteed to have their future faults succeed.
1007 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1008 * the reserve counters are updated with the hugetlb_lock held. It is safe
1009 * to reset the VMA at fork() time as it is not in use yet and there is no
1010 * chance of the global counters getting corrupted as a result of the values.
1012 * The private mapping reservation is represented in a subtly different
1013 * manner to a shared mapping. A shared mapping has a region map associated
1014 * with the underlying file, this region map represents the backing file
1015 * pages which have ever had a reservation assigned which this persists even
1016 * after the page is instantiated. A private mapping has a region map
1017 * associated with the original mmap which is attached to all VMAs which
1018 * reference it, this region map represents those offsets which have consumed
1019 * reservation ie. where pages have been instantiated.
1021 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1023 return (unsigned long)vma->vm_private_data;
1026 static void set_vma_private_data(struct vm_area_struct *vma,
1027 unsigned long value)
1029 vma->vm_private_data = (void *)value;
1033 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1034 struct hugetlb_cgroup *h_cg,
1037 #ifdef CONFIG_CGROUP_HUGETLB
1039 resv_map->reservation_counter = NULL;
1040 resv_map->pages_per_hpage = 0;
1041 resv_map->css = NULL;
1043 resv_map->reservation_counter =
1044 &h_cg->rsvd_hugepage[hstate_index(h)];
1045 resv_map->pages_per_hpage = pages_per_huge_page(h);
1046 resv_map->css = &h_cg->css;
1051 struct resv_map *resv_map_alloc(void)
1053 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1054 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1056 if (!resv_map || !rg) {
1062 kref_init(&resv_map->refs);
1063 spin_lock_init(&resv_map->lock);
1064 INIT_LIST_HEAD(&resv_map->regions);
1066 resv_map->adds_in_progress = 0;
1068 * Initialize these to 0. On shared mappings, 0's here indicate these
1069 * fields don't do cgroup accounting. On private mappings, these will be
1070 * re-initialized to the proper values, to indicate that hugetlb cgroup
1071 * reservations are to be un-charged from here.
1073 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1075 INIT_LIST_HEAD(&resv_map->region_cache);
1076 list_add(&rg->link, &resv_map->region_cache);
1077 resv_map->region_cache_count = 1;
1082 void resv_map_release(struct kref *ref)
1084 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1085 struct list_head *head = &resv_map->region_cache;
1086 struct file_region *rg, *trg;
1088 /* Clear out any active regions before we release the map. */
1089 region_del(resv_map, 0, LONG_MAX);
1091 /* ... and any entries left in the cache */
1092 list_for_each_entry_safe(rg, trg, head, link) {
1093 list_del(&rg->link);
1097 VM_BUG_ON(resv_map->adds_in_progress);
1102 static inline struct resv_map *inode_resv_map(struct inode *inode)
1105 * At inode evict time, i_mapping may not point to the original
1106 * address space within the inode. This original address space
1107 * contains the pointer to the resv_map. So, always use the
1108 * address space embedded within the inode.
1109 * The VERY common case is inode->mapping == &inode->i_data but,
1110 * this may not be true for device special inodes.
1112 return (struct resv_map *)(&inode->i_data)->private_data;
1115 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1117 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1118 if (vma->vm_flags & VM_MAYSHARE) {
1119 struct address_space *mapping = vma->vm_file->f_mapping;
1120 struct inode *inode = mapping->host;
1122 return inode_resv_map(inode);
1125 return (struct resv_map *)(get_vma_private_data(vma) &
1130 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1132 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1133 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1135 set_vma_private_data(vma, (get_vma_private_data(vma) &
1136 HPAGE_RESV_MASK) | (unsigned long)map);
1139 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1141 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1142 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1144 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1147 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1149 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1151 return (get_vma_private_data(vma) & flag) != 0;
1154 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1156 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1158 * Clear vm_private_data
1159 * - For shared mappings this is a per-vma semaphore that may be
1160 * allocated in a subsequent call to hugetlb_vm_op_open.
1161 * Before clearing, make sure pointer is not associated with vma
1162 * as this will leak the structure. This is the case when called
1163 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1164 * been called to allocate a new structure.
1165 * - For MAP_PRIVATE mappings, this is the reserve map which does
1166 * not apply to children. Faults generated by the children are
1167 * not guaranteed to succeed, even if read-only.
1169 if (vma->vm_flags & VM_MAYSHARE) {
1170 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1172 if (vma_lock && vma_lock->vma != vma)
1173 vma->vm_private_data = NULL;
1175 vma->vm_private_data = NULL;
1179 * Reset and decrement one ref on hugepage private reservation.
1180 * Called with mm->mmap_lock writer semaphore held.
1181 * This function should be only used by move_vma() and operate on
1182 * same sized vma. It should never come here with last ref on the
1185 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1188 * Clear the old hugetlb private page reservation.
1189 * It has already been transferred to new_vma.
1191 * During a mremap() operation of a hugetlb vma we call move_vma()
1192 * which copies vma into new_vma and unmaps vma. After the copy
1193 * operation both new_vma and vma share a reference to the resv_map
1194 * struct, and at that point vma is about to be unmapped. We don't
1195 * want to return the reservation to the pool at unmap of vma because
1196 * the reservation still lives on in new_vma, so simply decrement the
1197 * ref here and remove the resv_map reference from this vma.
1199 struct resv_map *reservations = vma_resv_map(vma);
1201 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1202 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1203 kref_put(&reservations->refs, resv_map_release);
1206 hugetlb_dup_vma_private(vma);
1209 /* Returns true if the VMA has associated reserve pages */
1210 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1212 if (vma->vm_flags & VM_NORESERVE) {
1214 * This address is already reserved by other process(chg == 0),
1215 * so, we should decrement reserved count. Without decrementing,
1216 * reserve count remains after releasing inode, because this
1217 * allocated page will go into page cache and is regarded as
1218 * coming from reserved pool in releasing step. Currently, we
1219 * don't have any other solution to deal with this situation
1220 * properly, so add work-around here.
1222 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1228 /* Shared mappings always use reserves */
1229 if (vma->vm_flags & VM_MAYSHARE) {
1231 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1232 * be a region map for all pages. The only situation where
1233 * there is no region map is if a hole was punched via
1234 * fallocate. In this case, there really are no reserves to
1235 * use. This situation is indicated if chg != 0.
1244 * Only the process that called mmap() has reserves for
1247 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1249 * Like the shared case above, a hole punch or truncate
1250 * could have been performed on the private mapping.
1251 * Examine the value of chg to determine if reserves
1252 * actually exist or were previously consumed.
1253 * Very Subtle - The value of chg comes from a previous
1254 * call to vma_needs_reserves(). The reserve map for
1255 * private mappings has different (opposite) semantics
1256 * than that of shared mappings. vma_needs_reserves()
1257 * has already taken this difference in semantics into
1258 * account. Therefore, the meaning of chg is the same
1259 * as in the shared case above. Code could easily be
1260 * combined, but keeping it separate draws attention to
1261 * subtle differences.
1272 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1274 int nid = folio_nid(folio);
1276 lockdep_assert_held(&hugetlb_lock);
1277 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1279 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1280 h->free_huge_pages++;
1281 h->free_huge_pages_node[nid]++;
1282 folio_set_hugetlb_freed(folio);
1285 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1288 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1290 lockdep_assert_held(&hugetlb_lock);
1291 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1292 if (pin && !is_longterm_pinnable_page(page))
1295 if (PageHWPoison(page))
1298 list_move(&page->lru, &h->hugepage_activelist);
1299 set_page_refcounted(page);
1300 ClearHPageFreed(page);
1301 h->free_huge_pages--;
1302 h->free_huge_pages_node[nid]--;
1309 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1312 unsigned int cpuset_mems_cookie;
1313 struct zonelist *zonelist;
1316 int node = NUMA_NO_NODE;
1318 zonelist = node_zonelist(nid, gfp_mask);
1321 cpuset_mems_cookie = read_mems_allowed_begin();
1322 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1325 if (!cpuset_zone_allowed(zone, gfp_mask))
1328 * no need to ask again on the same node. Pool is node rather than
1331 if (zone_to_nid(zone) == node)
1333 node = zone_to_nid(zone);
1335 page = dequeue_huge_page_node_exact(h, node);
1339 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1345 static unsigned long available_huge_pages(struct hstate *h)
1347 return h->free_huge_pages - h->resv_huge_pages;
1350 static struct page *dequeue_huge_page_vma(struct hstate *h,
1351 struct vm_area_struct *vma,
1352 unsigned long address, int avoid_reserve,
1355 struct page *page = NULL;
1356 struct mempolicy *mpol;
1358 nodemask_t *nodemask;
1362 * A child process with MAP_PRIVATE mappings created by their parent
1363 * have no page reserves. This check ensures that reservations are
1364 * not "stolen". The child may still get SIGKILLed
1366 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1369 /* If reserves cannot be used, ensure enough pages are in the pool */
1370 if (avoid_reserve && !available_huge_pages(h))
1373 gfp_mask = htlb_alloc_mask(h);
1374 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1376 if (mpol_is_preferred_many(mpol)) {
1377 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1379 /* Fallback to all nodes if page==NULL */
1384 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1386 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1387 SetHPageRestoreReserve(page);
1388 h->resv_huge_pages--;
1391 mpol_cond_put(mpol);
1399 * common helper functions for hstate_next_node_to_{alloc|free}.
1400 * We may have allocated or freed a huge page based on a different
1401 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1402 * be outside of *nodes_allowed. Ensure that we use an allowed
1403 * node for alloc or free.
1405 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1407 nid = next_node_in(nid, *nodes_allowed);
1408 VM_BUG_ON(nid >= MAX_NUMNODES);
1413 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1415 if (!node_isset(nid, *nodes_allowed))
1416 nid = next_node_allowed(nid, nodes_allowed);
1421 * returns the previously saved node ["this node"] from which to
1422 * allocate a persistent huge page for the pool and advance the
1423 * next node from which to allocate, handling wrap at end of node
1426 static int hstate_next_node_to_alloc(struct hstate *h,
1427 nodemask_t *nodes_allowed)
1431 VM_BUG_ON(!nodes_allowed);
1433 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1434 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1440 * helper for remove_pool_huge_page() - return the previously saved
1441 * node ["this node"] from which to free a huge page. Advance the
1442 * next node id whether or not we find a free huge page to free so
1443 * that the next attempt to free addresses the next node.
1445 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1449 VM_BUG_ON(!nodes_allowed);
1451 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1452 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1457 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1458 for (nr_nodes = nodes_weight(*mask); \
1460 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1463 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1464 for (nr_nodes = nodes_weight(*mask); \
1466 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1469 /* used to demote non-gigantic_huge pages as well */
1470 static void __destroy_compound_gigantic_folio(struct folio *folio,
1471 unsigned int order, bool demote)
1474 int nr_pages = 1 << order;
1477 atomic_set(folio_mapcount_ptr(folio), 0);
1478 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
1479 atomic_set(folio_pincount_ptr(folio), 0);
1481 for (i = 1; i < nr_pages; i++) {
1482 p = folio_page(folio, i);
1484 clear_compound_head(p);
1486 set_page_refcounted(p);
1489 folio_set_order(folio, 0);
1490 __folio_clear_head(folio);
1493 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1496 __destroy_compound_gigantic_folio(folio, order, true);
1499 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1500 static void destroy_compound_gigantic_folio(struct folio *folio,
1503 __destroy_compound_gigantic_folio(folio, order, false);
1506 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1509 * If the page isn't allocated using the cma allocator,
1510 * cma_release() returns false.
1513 int nid = folio_nid(folio);
1515 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1519 free_contig_range(folio_pfn(folio), 1 << order);
1522 #ifdef CONFIG_CONTIG_ALLOC
1523 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1524 int nid, nodemask_t *nodemask)
1527 unsigned long nr_pages = pages_per_huge_page(h);
1528 if (nid == NUMA_NO_NODE)
1529 nid = numa_mem_id();
1535 if (hugetlb_cma[nid]) {
1536 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1537 huge_page_order(h), true);
1539 return page_folio(page);
1542 if (!(gfp_mask & __GFP_THISNODE)) {
1543 for_each_node_mask(node, *nodemask) {
1544 if (node == nid || !hugetlb_cma[node])
1547 page = cma_alloc(hugetlb_cma[node], nr_pages,
1548 huge_page_order(h), true);
1550 return page_folio(page);
1556 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1557 return page ? page_folio(page) : NULL;
1560 #else /* !CONFIG_CONTIG_ALLOC */
1561 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1562 int nid, nodemask_t *nodemask)
1566 #endif /* CONFIG_CONTIG_ALLOC */
1568 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1569 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1570 int nid, nodemask_t *nodemask)
1574 static inline void free_gigantic_folio(struct folio *folio,
1575 unsigned int order) { }
1576 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1577 unsigned int order) { }
1581 * Remove hugetlb folio from lists, and update dtor so that the folio appears
1582 * as just a compound page.
1584 * A reference is held on the folio, except in the case of demote.
1586 * Must be called with hugetlb lock held.
1588 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1589 bool adjust_surplus,
1592 int nid = folio_nid(folio);
1594 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1595 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1597 lockdep_assert_held(&hugetlb_lock);
1598 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1601 list_del(&folio->lru);
1603 if (folio_test_hugetlb_freed(folio)) {
1604 h->free_huge_pages--;
1605 h->free_huge_pages_node[nid]--;
1607 if (adjust_surplus) {
1608 h->surplus_huge_pages--;
1609 h->surplus_huge_pages_node[nid]--;
1615 * For non-gigantic pages set the destructor to the normal compound
1616 * page dtor. This is needed in case someone takes an additional
1617 * temporary ref to the page, and freeing is delayed until they drop
1620 * For gigantic pages set the destructor to the null dtor. This
1621 * destructor will never be called. Before freeing the gigantic
1622 * page destroy_compound_gigantic_folio will turn the folio into a
1623 * simple group of pages. After this the destructor does not
1626 * This handles the case where more than one ref is held when and
1627 * after update_and_free_hugetlb_folio is called.
1629 * In the case of demote we do not ref count the page as it will soon
1630 * be turned into a page of smaller size.
1633 folio_ref_unfreeze(folio, 1);
1634 if (hstate_is_gigantic(h))
1635 folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR);
1637 folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR);
1640 h->nr_huge_pages_node[nid]--;
1643 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1644 bool adjust_surplus)
1646 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1649 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1650 bool adjust_surplus)
1652 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1655 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1656 bool adjust_surplus)
1659 int nid = folio_nid(folio);
1661 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1663 lockdep_assert_held(&hugetlb_lock);
1665 INIT_LIST_HEAD(&folio->lru);
1667 h->nr_huge_pages_node[nid]++;
1669 if (adjust_surplus) {
1670 h->surplus_huge_pages++;
1671 h->surplus_huge_pages_node[nid]++;
1674 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1675 folio_change_private(folio, NULL);
1677 * We have to set hugetlb_vmemmap_optimized again as above
1678 * folio_change_private(folio, NULL) cleared it.
1680 folio_set_hugetlb_vmemmap_optimized(folio);
1683 * This folio is about to be managed by the hugetlb allocator and
1684 * should have no users. Drop our reference, and check for others
1687 zeroed = folio_put_testzero(folio);
1688 if (unlikely(!zeroed))
1690 * It is VERY unlikely soneone else has taken a ref on
1691 * the page. In this case, we simply return as the
1692 * hugetlb destructor (free_huge_page) will be called
1693 * when this other ref is dropped.
1697 arch_clear_hugepage_flags(&folio->page);
1698 enqueue_hugetlb_folio(h, folio);
1701 static void __update_and_free_page(struct hstate *h, struct page *page)
1704 struct folio *folio = page_folio(page);
1705 struct page *subpage;
1707 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1711 * If we don't know which subpages are hwpoisoned, we can't free
1712 * the hugepage, so it's leaked intentionally.
1714 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1717 if (hugetlb_vmemmap_restore(h, page)) {
1718 spin_lock_irq(&hugetlb_lock);
1720 * If we cannot allocate vmemmap pages, just refuse to free the
1721 * page and put the page back on the hugetlb free list and treat
1722 * as a surplus page.
1724 add_hugetlb_folio(h, folio, true);
1725 spin_unlock_irq(&hugetlb_lock);
1730 * Move PageHWPoison flag from head page to the raw error pages,
1731 * which makes any healthy subpages reusable.
1733 if (unlikely(folio_test_hwpoison(folio)))
1734 hugetlb_clear_page_hwpoison(&folio->page);
1736 for (i = 0; i < pages_per_huge_page(h); i++) {
1737 subpage = folio_page(folio, i);
1738 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1739 1 << PG_referenced | 1 << PG_dirty |
1740 1 << PG_active | 1 << PG_private |
1745 * Non-gigantic pages demoted from CMA allocated gigantic pages
1746 * need to be given back to CMA in free_gigantic_folio.
1748 if (hstate_is_gigantic(h) ||
1749 hugetlb_cma_folio(folio, huge_page_order(h))) {
1750 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1751 free_gigantic_folio(folio, huge_page_order(h));
1753 __free_pages(page, huge_page_order(h));
1758 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1759 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1760 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1761 * the vmemmap pages.
1763 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1764 * freed and frees them one-by-one. As the page->mapping pointer is going
1765 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1766 * structure of a lockless linked list of huge pages to be freed.
1768 static LLIST_HEAD(hpage_freelist);
1770 static void free_hpage_workfn(struct work_struct *work)
1772 struct llist_node *node;
1774 node = llist_del_all(&hpage_freelist);
1780 page = container_of((struct address_space **)node,
1781 struct page, mapping);
1783 page->mapping = NULL;
1785 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1786 * is going to trigger because a previous call to
1787 * remove_hugetlb_folio() will call folio_set_compound_dtor
1788 * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate()
1791 h = size_to_hstate(page_size(page));
1793 __update_and_free_page(h, page);
1798 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1800 static inline void flush_free_hpage_work(struct hstate *h)
1802 if (hugetlb_vmemmap_optimizable(h))
1803 flush_work(&free_hpage_work);
1806 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1809 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1810 __update_and_free_page(h, &folio->page);
1815 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1817 * Only call schedule_work() if hpage_freelist is previously
1818 * empty. Otherwise, schedule_work() had been called but the workfn
1819 * hasn't retrieved the list yet.
1821 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1822 schedule_work(&free_hpage_work);
1825 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1827 struct page *page, *t_page;
1828 struct folio *folio;
1830 list_for_each_entry_safe(page, t_page, list, lru) {
1831 folio = page_folio(page);
1832 update_and_free_hugetlb_folio(h, folio, false);
1837 struct hstate *size_to_hstate(unsigned long size)
1841 for_each_hstate(h) {
1842 if (huge_page_size(h) == size)
1848 void free_huge_page(struct page *page)
1851 * Can't pass hstate in here because it is called from the
1852 * compound page destructor.
1854 struct folio *folio = page_folio(page);
1855 struct hstate *h = folio_hstate(folio);
1856 int nid = folio_nid(folio);
1857 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1858 bool restore_reserve;
1859 unsigned long flags;
1861 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1862 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1864 hugetlb_set_folio_subpool(folio, NULL);
1865 if (folio_test_anon(folio))
1866 __ClearPageAnonExclusive(&folio->page);
1867 folio->mapping = NULL;
1868 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1869 folio_clear_hugetlb_restore_reserve(folio);
1872 * If HPageRestoreReserve was set on page, page allocation consumed a
1873 * reservation. If the page was associated with a subpool, there
1874 * would have been a page reserved in the subpool before allocation
1875 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1876 * reservation, do not call hugepage_subpool_put_pages() as this will
1877 * remove the reserved page from the subpool.
1879 if (!restore_reserve) {
1881 * A return code of zero implies that the subpool will be
1882 * under its minimum size if the reservation is not restored
1883 * after page is free. Therefore, force restore_reserve
1886 if (hugepage_subpool_put_pages(spool, 1) == 0)
1887 restore_reserve = true;
1890 spin_lock_irqsave(&hugetlb_lock, flags);
1891 folio_clear_hugetlb_migratable(folio);
1892 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1893 pages_per_huge_page(h), folio);
1894 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1895 pages_per_huge_page(h), folio);
1896 if (restore_reserve)
1897 h->resv_huge_pages++;
1899 if (folio_test_hugetlb_temporary(folio)) {
1900 remove_hugetlb_folio(h, folio, false);
1901 spin_unlock_irqrestore(&hugetlb_lock, flags);
1902 update_and_free_hugetlb_folio(h, folio, true);
1903 } else if (h->surplus_huge_pages_node[nid]) {
1904 /* remove the page from active list */
1905 remove_hugetlb_folio(h, folio, true);
1906 spin_unlock_irqrestore(&hugetlb_lock, flags);
1907 update_and_free_hugetlb_folio(h, folio, true);
1909 arch_clear_hugepage_flags(page);
1910 enqueue_hugetlb_folio(h, folio);
1911 spin_unlock_irqrestore(&hugetlb_lock, flags);
1916 * Must be called with the hugetlb lock held
1918 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1920 lockdep_assert_held(&hugetlb_lock);
1922 h->nr_huge_pages_node[nid]++;
1925 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1927 hugetlb_vmemmap_optimize(h, &folio->page);
1928 INIT_LIST_HEAD(&folio->lru);
1929 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1930 hugetlb_set_folio_subpool(folio, NULL);
1931 set_hugetlb_cgroup(folio, NULL);
1932 set_hugetlb_cgroup_rsvd(folio, NULL);
1935 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1937 __prep_new_hugetlb_folio(h, folio);
1938 spin_lock_irq(&hugetlb_lock);
1939 __prep_account_new_huge_page(h, nid);
1940 spin_unlock_irq(&hugetlb_lock);
1943 static bool __prep_compound_gigantic_folio(struct folio *folio,
1944 unsigned int order, bool demote)
1947 int nr_pages = 1 << order;
1950 __folio_clear_reserved(folio);
1951 __folio_set_head(folio);
1952 /* we rely on prep_new_hugetlb_folio to set the destructor */
1953 folio_set_order(folio, order);
1954 for (i = 0; i < nr_pages; i++) {
1955 p = folio_page(folio, i);
1958 * For gigantic hugepages allocated through bootmem at
1959 * boot, it's safer to be consistent with the not-gigantic
1960 * hugepages and clear the PG_reserved bit from all tail pages
1961 * too. Otherwise drivers using get_user_pages() to access tail
1962 * pages may get the reference counting wrong if they see
1963 * PG_reserved set on a tail page (despite the head page not
1964 * having PG_reserved set). Enforcing this consistency between
1965 * head and tail pages allows drivers to optimize away a check
1966 * on the head page when they need know if put_page() is needed
1967 * after get_user_pages().
1969 if (i != 0) /* head page cleared above */
1970 __ClearPageReserved(p);
1972 * Subtle and very unlikely
1974 * Gigantic 'page allocators' such as memblock or cma will
1975 * return a set of pages with each page ref counted. We need
1976 * to turn this set of pages into a compound page with tail
1977 * page ref counts set to zero. Code such as speculative page
1978 * cache adding could take a ref on a 'to be' tail page.
1979 * We need to respect any increased ref count, and only set
1980 * the ref count to zero if count is currently 1. If count
1981 * is not 1, we return an error. An error return indicates
1982 * the set of pages can not be converted to a gigantic page.
1983 * The caller who allocated the pages should then discard the
1984 * pages using the appropriate free interface.
1986 * In the case of demote, the ref count will be zero.
1989 if (!page_ref_freeze(p, 1)) {
1990 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1994 VM_BUG_ON_PAGE(page_count(p), p);
1997 set_compound_head(p, &folio->page);
1999 atomic_set(folio_mapcount_ptr(folio), -1);
2000 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
2001 atomic_set(folio_pincount_ptr(folio), 0);
2005 /* undo page modifications made above */
2006 for (j = 0; j < i; j++) {
2007 p = folio_page(folio, j);
2009 clear_compound_head(p);
2010 set_page_refcounted(p);
2012 /* need to clear PG_reserved on remaining tail pages */
2013 for (; j < nr_pages; j++) {
2014 p = folio_page(folio, j);
2015 __ClearPageReserved(p);
2017 folio_set_order(folio, 0);
2018 __folio_clear_head(folio);
2022 static bool prep_compound_gigantic_folio(struct folio *folio,
2025 return __prep_compound_gigantic_folio(folio, order, false);
2028 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2031 return __prep_compound_gigantic_folio(folio, order, true);
2035 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2036 * transparent huge pages. See the PageTransHuge() documentation for more
2039 int PageHuge(struct page *page)
2041 if (!PageCompound(page))
2044 page = compound_head(page);
2045 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
2047 EXPORT_SYMBOL_GPL(PageHuge);
2050 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
2051 * normal or transparent huge pages.
2053 int PageHeadHuge(struct page *page_head)
2055 if (!PageHead(page_head))
2058 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
2060 EXPORT_SYMBOL_GPL(PageHeadHuge);
2063 * Find and lock address space (mapping) in write mode.
2065 * Upon entry, the page is locked which means that page_mapping() is
2066 * stable. Due to locking order, we can only trylock_write. If we can
2067 * not get the lock, simply return NULL to caller.
2069 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2071 struct address_space *mapping = page_mapping(hpage);
2076 if (i_mmap_trylock_write(mapping))
2082 pgoff_t hugetlb_basepage_index(struct page *page)
2084 struct page *page_head = compound_head(page);
2085 pgoff_t index = page_index(page_head);
2086 unsigned long compound_idx;
2088 if (compound_order(page_head) >= MAX_ORDER)
2089 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2091 compound_idx = page - page_head;
2093 return (index << compound_order(page_head)) + compound_idx;
2096 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2097 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2098 nodemask_t *node_alloc_noretry)
2100 int order = huge_page_order(h);
2102 bool alloc_try_hard = true;
2106 * By default we always try hard to allocate the page with
2107 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2108 * a loop (to adjust global huge page counts) and previous allocation
2109 * failed, do not continue to try hard on the same node. Use the
2110 * node_alloc_noretry bitmap to manage this state information.
2112 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2113 alloc_try_hard = false;
2114 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2116 gfp_mask |= __GFP_RETRY_MAYFAIL;
2117 if (nid == NUMA_NO_NODE)
2118 nid = numa_mem_id();
2120 page = __alloc_pages(gfp_mask, order, nid, nmask);
2122 /* Freeze head page */
2123 if (page && !page_ref_freeze(page, 1)) {
2124 __free_pages(page, order);
2125 if (retry) { /* retry once */
2129 /* WOW! twice in a row. */
2130 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2135 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2136 * indicates an overall state change. Clear bit so that we resume
2137 * normal 'try hard' allocations.
2139 if (node_alloc_noretry && page && !alloc_try_hard)
2140 node_clear(nid, *node_alloc_noretry);
2143 * If we tried hard to get a page but failed, set bit so that
2144 * subsequent attempts will not try as hard until there is an
2145 * overall state change.
2147 if (node_alloc_noretry && !page && alloc_try_hard)
2148 node_set(nid, *node_alloc_noretry);
2151 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2155 __count_vm_event(HTLB_BUDDY_PGALLOC);
2156 return page_folio(page);
2160 * Common helper to allocate a fresh hugetlb page. All specific allocators
2161 * should use this function to get new hugetlb pages
2163 * Note that returned page is 'frozen': ref count of head page and all tail
2166 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2167 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2168 nodemask_t *node_alloc_noretry)
2170 struct folio *folio;
2174 if (hstate_is_gigantic(h))
2175 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2177 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2178 nid, nmask, node_alloc_noretry);
2181 if (hstate_is_gigantic(h)) {
2182 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2184 * Rare failure to convert pages to compound page.
2185 * Free pages and try again - ONCE!
2187 free_gigantic_folio(folio, huge_page_order(h));
2195 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2201 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2204 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2205 nodemask_t *node_alloc_noretry)
2207 struct folio *folio;
2209 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2211 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2212 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2213 nodes_allowed, node_alloc_noretry);
2215 free_huge_page(&folio->page); /* free it into the hugepage allocator */
2224 * Remove huge page from pool from next node to free. Attempt to keep
2225 * persistent huge pages more or less balanced over allowed nodes.
2226 * This routine only 'removes' the hugetlb page. The caller must make
2227 * an additional call to free the page to low level allocators.
2228 * Called with hugetlb_lock locked.
2230 static struct page *remove_pool_huge_page(struct hstate *h,
2231 nodemask_t *nodes_allowed,
2235 struct page *page = NULL;
2236 struct folio *folio;
2238 lockdep_assert_held(&hugetlb_lock);
2239 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2241 * If we're returning unused surplus pages, only examine
2242 * nodes with surplus pages.
2244 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2245 !list_empty(&h->hugepage_freelists[node])) {
2246 page = list_entry(h->hugepage_freelists[node].next,
2248 folio = page_folio(page);
2249 remove_hugetlb_folio(h, folio, acct_surplus);
2258 * Dissolve a given free hugepage into free buddy pages. This function does
2259 * nothing for in-use hugepages and non-hugepages.
2260 * This function returns values like below:
2262 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2263 * when the system is under memory pressure and the feature of
2264 * freeing unused vmemmap pages associated with each hugetlb page
2266 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2267 * (allocated or reserved.)
2268 * 0: successfully dissolved free hugepages or the page is not a
2269 * hugepage (considered as already dissolved)
2271 int dissolve_free_huge_page(struct page *page)
2274 struct folio *folio = page_folio(page);
2277 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2278 if (!folio_test_hugetlb(folio))
2281 spin_lock_irq(&hugetlb_lock);
2282 if (!folio_test_hugetlb(folio)) {
2287 if (!folio_ref_count(folio)) {
2288 struct hstate *h = folio_hstate(folio);
2289 if (!available_huge_pages(h))
2293 * We should make sure that the page is already on the free list
2294 * when it is dissolved.
2296 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2297 spin_unlock_irq(&hugetlb_lock);
2301 * Theoretically, we should return -EBUSY when we
2302 * encounter this race. In fact, we have a chance
2303 * to successfully dissolve the page if we do a
2304 * retry. Because the race window is quite small.
2305 * If we seize this opportunity, it is an optimization
2306 * for increasing the success rate of dissolving page.
2311 remove_hugetlb_folio(h, folio, false);
2312 h->max_huge_pages--;
2313 spin_unlock_irq(&hugetlb_lock);
2316 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2317 * before freeing the page. update_and_free_hugtlb_folio will fail to
2318 * free the page if it can not allocate required vmemmap. We
2319 * need to adjust max_huge_pages if the page is not freed.
2320 * Attempt to allocate vmemmmap here so that we can take
2321 * appropriate action on failure.
2323 rc = hugetlb_vmemmap_restore(h, &folio->page);
2325 update_and_free_hugetlb_folio(h, folio, false);
2327 spin_lock_irq(&hugetlb_lock);
2328 add_hugetlb_folio(h, folio, false);
2329 h->max_huge_pages++;
2330 spin_unlock_irq(&hugetlb_lock);
2336 spin_unlock_irq(&hugetlb_lock);
2341 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2342 * make specified memory blocks removable from the system.
2343 * Note that this will dissolve a free gigantic hugepage completely, if any
2344 * part of it lies within the given range.
2345 * Also note that if dissolve_free_huge_page() returns with an error, all
2346 * free hugepages that were dissolved before that error are lost.
2348 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2356 if (!hugepages_supported())
2359 order = huge_page_order(&default_hstate);
2361 order = min(order, huge_page_order(h));
2363 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2364 page = pfn_to_page(pfn);
2365 rc = dissolve_free_huge_page(page);
2374 * Allocates a fresh surplus page from the page allocator.
2376 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2377 int nid, nodemask_t *nmask)
2379 struct folio *folio = NULL;
2381 if (hstate_is_gigantic(h))
2384 spin_lock_irq(&hugetlb_lock);
2385 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2387 spin_unlock_irq(&hugetlb_lock);
2389 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2393 spin_lock_irq(&hugetlb_lock);
2395 * We could have raced with the pool size change.
2396 * Double check that and simply deallocate the new page
2397 * if we would end up overcommiting the surpluses. Abuse
2398 * temporary page to workaround the nasty free_huge_page
2401 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2402 folio_set_hugetlb_temporary(folio);
2403 spin_unlock_irq(&hugetlb_lock);
2404 free_huge_page(&folio->page);
2408 h->surplus_huge_pages++;
2409 h->surplus_huge_pages_node[folio_nid(folio)]++;
2412 spin_unlock_irq(&hugetlb_lock);
2414 return &folio->page;
2417 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2418 int nid, nodemask_t *nmask)
2420 struct folio *folio;
2422 if (hstate_is_gigantic(h))
2425 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2429 /* fresh huge pages are frozen */
2430 folio_ref_unfreeze(folio, 1);
2432 * We do not account these pages as surplus because they are only
2433 * temporary and will be released properly on the last reference
2435 folio_set_hugetlb_temporary(folio);
2437 return &folio->page;
2441 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2444 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2445 struct vm_area_struct *vma, unsigned long addr)
2447 struct page *page = NULL;
2448 struct mempolicy *mpol;
2449 gfp_t gfp_mask = htlb_alloc_mask(h);
2451 nodemask_t *nodemask;
2453 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2454 if (mpol_is_preferred_many(mpol)) {
2455 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2457 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2458 page = alloc_surplus_huge_page(h, gfp, nid, nodemask);
2460 /* Fallback to all nodes if page==NULL */
2465 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2466 mpol_cond_put(mpol);
2470 /* page migration callback function */
2471 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2472 nodemask_t *nmask, gfp_t gfp_mask)
2474 spin_lock_irq(&hugetlb_lock);
2475 if (available_huge_pages(h)) {
2478 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2480 spin_unlock_irq(&hugetlb_lock);
2484 spin_unlock_irq(&hugetlb_lock);
2486 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2489 /* mempolicy aware migration callback */
2490 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2491 unsigned long address)
2493 struct mempolicy *mpol;
2494 nodemask_t *nodemask;
2499 gfp_mask = htlb_alloc_mask(h);
2500 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2501 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2502 mpol_cond_put(mpol);
2508 * Increase the hugetlb pool such that it can accommodate a reservation
2511 static int gather_surplus_pages(struct hstate *h, long delta)
2512 __must_hold(&hugetlb_lock)
2514 LIST_HEAD(surplus_list);
2515 struct page *page, *tmp;
2518 long needed, allocated;
2519 bool alloc_ok = true;
2521 lockdep_assert_held(&hugetlb_lock);
2522 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2524 h->resv_huge_pages += delta;
2532 spin_unlock_irq(&hugetlb_lock);
2533 for (i = 0; i < needed; i++) {
2534 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2535 NUMA_NO_NODE, NULL);
2540 list_add(&page->lru, &surplus_list);
2546 * After retaking hugetlb_lock, we need to recalculate 'needed'
2547 * because either resv_huge_pages or free_huge_pages may have changed.
2549 spin_lock_irq(&hugetlb_lock);
2550 needed = (h->resv_huge_pages + delta) -
2551 (h->free_huge_pages + allocated);
2556 * We were not able to allocate enough pages to
2557 * satisfy the entire reservation so we free what
2558 * we've allocated so far.
2563 * The surplus_list now contains _at_least_ the number of extra pages
2564 * needed to accommodate the reservation. Add the appropriate number
2565 * of pages to the hugetlb pool and free the extras back to the buddy
2566 * allocator. Commit the entire reservation here to prevent another
2567 * process from stealing the pages as they are added to the pool but
2568 * before they are reserved.
2570 needed += allocated;
2571 h->resv_huge_pages += delta;
2574 /* Free the needed pages to the hugetlb pool */
2575 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2578 /* Add the page to the hugetlb allocator */
2579 enqueue_hugetlb_folio(h, page_folio(page));
2582 spin_unlock_irq(&hugetlb_lock);
2585 * Free unnecessary surplus pages to the buddy allocator.
2586 * Pages have no ref count, call free_huge_page directly.
2588 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2589 free_huge_page(page);
2590 spin_lock_irq(&hugetlb_lock);
2596 * This routine has two main purposes:
2597 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2598 * in unused_resv_pages. This corresponds to the prior adjustments made
2599 * to the associated reservation map.
2600 * 2) Free any unused surplus pages that may have been allocated to satisfy
2601 * the reservation. As many as unused_resv_pages may be freed.
2603 static void return_unused_surplus_pages(struct hstate *h,
2604 unsigned long unused_resv_pages)
2606 unsigned long nr_pages;
2608 LIST_HEAD(page_list);
2610 lockdep_assert_held(&hugetlb_lock);
2611 /* Uncommit the reservation */
2612 h->resv_huge_pages -= unused_resv_pages;
2614 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2618 * Part (or even all) of the reservation could have been backed
2619 * by pre-allocated pages. Only free surplus pages.
2621 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2624 * We want to release as many surplus pages as possible, spread
2625 * evenly across all nodes with memory. Iterate across these nodes
2626 * until we can no longer free unreserved surplus pages. This occurs
2627 * when the nodes with surplus pages have no free pages.
2628 * remove_pool_huge_page() will balance the freed pages across the
2629 * on-line nodes with memory and will handle the hstate accounting.
2631 while (nr_pages--) {
2632 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2636 list_add(&page->lru, &page_list);
2640 spin_unlock_irq(&hugetlb_lock);
2641 update_and_free_pages_bulk(h, &page_list);
2642 spin_lock_irq(&hugetlb_lock);
2647 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2648 * are used by the huge page allocation routines to manage reservations.
2650 * vma_needs_reservation is called to determine if the huge page at addr
2651 * within the vma has an associated reservation. If a reservation is
2652 * needed, the value 1 is returned. The caller is then responsible for
2653 * managing the global reservation and subpool usage counts. After
2654 * the huge page has been allocated, vma_commit_reservation is called
2655 * to add the page to the reservation map. If the page allocation fails,
2656 * the reservation must be ended instead of committed. vma_end_reservation
2657 * is called in such cases.
2659 * In the normal case, vma_commit_reservation returns the same value
2660 * as the preceding vma_needs_reservation call. The only time this
2661 * is not the case is if a reserve map was changed between calls. It
2662 * is the responsibility of the caller to notice the difference and
2663 * take appropriate action.
2665 * vma_add_reservation is used in error paths where a reservation must
2666 * be restored when a newly allocated huge page must be freed. It is
2667 * to be called after calling vma_needs_reservation to determine if a
2668 * reservation exists.
2670 * vma_del_reservation is used in error paths where an entry in the reserve
2671 * map was created during huge page allocation and must be removed. It is to
2672 * be called after calling vma_needs_reservation to determine if a reservation
2675 enum vma_resv_mode {
2682 static long __vma_reservation_common(struct hstate *h,
2683 struct vm_area_struct *vma, unsigned long addr,
2684 enum vma_resv_mode mode)
2686 struct resv_map *resv;
2689 long dummy_out_regions_needed;
2691 resv = vma_resv_map(vma);
2695 idx = vma_hugecache_offset(h, vma, addr);
2697 case VMA_NEEDS_RESV:
2698 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2699 /* We assume that vma_reservation_* routines always operate on
2700 * 1 page, and that adding to resv map a 1 page entry can only
2701 * ever require 1 region.
2703 VM_BUG_ON(dummy_out_regions_needed != 1);
2705 case VMA_COMMIT_RESV:
2706 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2707 /* region_add calls of range 1 should never fail. */
2711 region_abort(resv, idx, idx + 1, 1);
2715 if (vma->vm_flags & VM_MAYSHARE) {
2716 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2717 /* region_add calls of range 1 should never fail. */
2720 region_abort(resv, idx, idx + 1, 1);
2721 ret = region_del(resv, idx, idx + 1);
2725 if (vma->vm_flags & VM_MAYSHARE) {
2726 region_abort(resv, idx, idx + 1, 1);
2727 ret = region_del(resv, idx, idx + 1);
2729 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2730 /* region_add calls of range 1 should never fail. */
2738 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2741 * We know private mapping must have HPAGE_RESV_OWNER set.
2743 * In most cases, reserves always exist for private mappings.
2744 * However, a file associated with mapping could have been
2745 * hole punched or truncated after reserves were consumed.
2746 * As subsequent fault on such a range will not use reserves.
2747 * Subtle - The reserve map for private mappings has the
2748 * opposite meaning than that of shared mappings. If NO
2749 * entry is in the reserve map, it means a reservation exists.
2750 * If an entry exists in the reserve map, it means the
2751 * reservation has already been consumed. As a result, the
2752 * return value of this routine is the opposite of the
2753 * value returned from reserve map manipulation routines above.
2762 static long vma_needs_reservation(struct hstate *h,
2763 struct vm_area_struct *vma, unsigned long addr)
2765 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2768 static long vma_commit_reservation(struct hstate *h,
2769 struct vm_area_struct *vma, unsigned long addr)
2771 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2774 static void vma_end_reservation(struct hstate *h,
2775 struct vm_area_struct *vma, unsigned long addr)
2777 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2780 static long vma_add_reservation(struct hstate *h,
2781 struct vm_area_struct *vma, unsigned long addr)
2783 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2786 static long vma_del_reservation(struct hstate *h,
2787 struct vm_area_struct *vma, unsigned long addr)
2789 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2793 * This routine is called to restore reservation information on error paths.
2794 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2795 * the hugetlb mutex should remain held when calling this routine.
2797 * It handles two specific cases:
2798 * 1) A reservation was in place and the page consumed the reservation.
2799 * HPageRestoreReserve is set in the page.
2800 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2801 * not set. However, alloc_huge_page always updates the reserve map.
2803 * In case 1, free_huge_page later in the error path will increment the
2804 * global reserve count. But, free_huge_page does not have enough context
2805 * to adjust the reservation map. This case deals primarily with private
2806 * mappings. Adjust the reserve map here to be consistent with global
2807 * reserve count adjustments to be made by free_huge_page. Make sure the
2808 * reserve map indicates there is a reservation present.
2810 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2812 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2813 unsigned long address, struct page *page)
2815 long rc = vma_needs_reservation(h, vma, address);
2817 if (HPageRestoreReserve(page)) {
2818 if (unlikely(rc < 0))
2820 * Rare out of memory condition in reserve map
2821 * manipulation. Clear HPageRestoreReserve so that
2822 * global reserve count will not be incremented
2823 * by free_huge_page. This will make it appear
2824 * as though the reservation for this page was
2825 * consumed. This may prevent the task from
2826 * faulting in the page at a later time. This
2827 * is better than inconsistent global huge page
2828 * accounting of reserve counts.
2830 ClearHPageRestoreReserve(page);
2832 (void)vma_add_reservation(h, vma, address);
2834 vma_end_reservation(h, vma, address);
2838 * This indicates there is an entry in the reserve map
2839 * not added by alloc_huge_page. We know it was added
2840 * before the alloc_huge_page call, otherwise
2841 * HPageRestoreReserve would be set on the page.
2842 * Remove the entry so that a subsequent allocation
2843 * does not consume a reservation.
2845 rc = vma_del_reservation(h, vma, address);
2848 * VERY rare out of memory condition. Since
2849 * we can not delete the entry, set
2850 * HPageRestoreReserve so that the reserve
2851 * count will be incremented when the page
2852 * is freed. This reserve will be consumed
2853 * on a subsequent allocation.
2855 SetHPageRestoreReserve(page);
2856 } else if (rc < 0) {
2858 * Rare out of memory condition from
2859 * vma_needs_reservation call. Memory allocation is
2860 * only attempted if a new entry is needed. Therefore,
2861 * this implies there is not an entry in the
2864 * For shared mappings, no entry in the map indicates
2865 * no reservation. We are done.
2867 if (!(vma->vm_flags & VM_MAYSHARE))
2869 * For private mappings, no entry indicates
2870 * a reservation is present. Since we can
2871 * not add an entry, set SetHPageRestoreReserve
2872 * on the page so reserve count will be
2873 * incremented when freed. This reserve will
2874 * be consumed on a subsequent allocation.
2876 SetHPageRestoreReserve(page);
2879 * No reservation present, do nothing
2881 vma_end_reservation(h, vma, address);
2886 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2888 * @h: struct hstate old page belongs to
2889 * @old_folio: Old folio to dissolve
2890 * @list: List to isolate the page in case we need to
2891 * Returns 0 on success, otherwise negated error.
2893 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2894 struct folio *old_folio, struct list_head *list)
2896 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2897 int nid = folio_nid(old_folio);
2898 struct folio *new_folio;
2902 * Before dissolving the folio, we need to allocate a new one for the
2903 * pool to remain stable. Here, we allocate the folio and 'prep' it
2904 * by doing everything but actually updating counters and adding to
2905 * the pool. This simplifies and let us do most of the processing
2908 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2911 __prep_new_hugetlb_folio(h, new_folio);
2914 spin_lock_irq(&hugetlb_lock);
2915 if (!folio_test_hugetlb(old_folio)) {
2917 * Freed from under us. Drop new_folio too.
2920 } else if (folio_ref_count(old_folio)) {
2922 * Someone has grabbed the folio, try to isolate it here.
2923 * Fail with -EBUSY if not possible.
2925 spin_unlock_irq(&hugetlb_lock);
2926 ret = isolate_hugetlb(&old_folio->page, list);
2927 spin_lock_irq(&hugetlb_lock);
2929 } else if (!folio_test_hugetlb_freed(old_folio)) {
2931 * Folio's refcount is 0 but it has not been enqueued in the
2932 * freelist yet. Race window is small, so we can succeed here if
2935 spin_unlock_irq(&hugetlb_lock);
2940 * Ok, old_folio is still a genuine free hugepage. Remove it from
2941 * the freelist and decrease the counters. These will be
2942 * incremented again when calling __prep_account_new_huge_page()
2943 * and enqueue_hugetlb_folio() for new_folio. The counters will
2944 * remain stable since this happens under the lock.
2946 remove_hugetlb_folio(h, old_folio, false);
2949 * Ref count on new_folio is already zero as it was dropped
2950 * earlier. It can be directly added to the pool free list.
2952 __prep_account_new_huge_page(h, nid);
2953 enqueue_hugetlb_folio(h, new_folio);
2956 * Folio has been replaced, we can safely free the old one.
2958 spin_unlock_irq(&hugetlb_lock);
2959 update_and_free_hugetlb_folio(h, old_folio, false);
2965 spin_unlock_irq(&hugetlb_lock);
2966 /* Folio has a zero ref count, but needs a ref to be freed */
2967 folio_ref_unfreeze(new_folio, 1);
2968 update_and_free_hugetlb_folio(h, new_folio, false);
2973 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2976 struct folio *folio = page_folio(page);
2980 * The page might have been dissolved from under our feet, so make sure
2981 * to carefully check the state under the lock.
2982 * Return success when racing as if we dissolved the page ourselves.
2984 spin_lock_irq(&hugetlb_lock);
2985 if (folio_test_hugetlb(folio)) {
2986 h = folio_hstate(folio);
2988 spin_unlock_irq(&hugetlb_lock);
2991 spin_unlock_irq(&hugetlb_lock);
2994 * Fence off gigantic pages as there is a cyclic dependency between
2995 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2996 * of bailing out right away without further retrying.
2998 if (hstate_is_gigantic(h))
3001 if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list))
3003 else if (!folio_ref_count(folio))
3004 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3009 struct page *alloc_huge_page(struct vm_area_struct *vma,
3010 unsigned long addr, int avoid_reserve)
3012 struct hugepage_subpool *spool = subpool_vma(vma);
3013 struct hstate *h = hstate_vma(vma);
3015 struct folio *folio;
3016 long map_chg, map_commit;
3019 struct hugetlb_cgroup *h_cg;
3020 bool deferred_reserve;
3022 idx = hstate_index(h);
3024 * Examine the region/reserve map to determine if the process
3025 * has a reservation for the page to be allocated. A return
3026 * code of zero indicates a reservation exists (no change).
3028 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3030 return ERR_PTR(-ENOMEM);
3033 * Processes that did not create the mapping will have no
3034 * reserves as indicated by the region/reserve map. Check
3035 * that the allocation will not exceed the subpool limit.
3036 * Allocations for MAP_NORESERVE mappings also need to be
3037 * checked against any subpool limit.
3039 if (map_chg || avoid_reserve) {
3040 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3042 vma_end_reservation(h, vma, addr);
3043 return ERR_PTR(-ENOSPC);
3047 * Even though there was no reservation in the region/reserve
3048 * map, there could be reservations associated with the
3049 * subpool that can be used. This would be indicated if the
3050 * return value of hugepage_subpool_get_pages() is zero.
3051 * However, if avoid_reserve is specified we still avoid even
3052 * the subpool reservations.
3058 /* If this allocation is not consuming a reservation, charge it now.
3060 deferred_reserve = map_chg || avoid_reserve;
3061 if (deferred_reserve) {
3062 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3063 idx, pages_per_huge_page(h), &h_cg);
3065 goto out_subpool_put;
3068 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3070 goto out_uncharge_cgroup_reservation;
3072 spin_lock_irq(&hugetlb_lock);
3074 * glb_chg is passed to indicate whether or not a page must be taken
3075 * from the global free pool (global change). gbl_chg == 0 indicates
3076 * a reservation exists for the allocation.
3078 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
3080 spin_unlock_irq(&hugetlb_lock);
3081 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
3083 goto out_uncharge_cgroup;
3084 spin_lock_irq(&hugetlb_lock);
3085 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3086 SetHPageRestoreReserve(page);
3087 h->resv_huge_pages--;
3089 list_add(&page->lru, &h->hugepage_activelist);
3090 set_page_refcounted(page);
3093 folio = page_folio(page);
3094 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
3095 /* If allocation is not consuming a reservation, also store the
3096 * hugetlb_cgroup pointer on the page.
3098 if (deferred_reserve) {
3099 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3103 spin_unlock_irq(&hugetlb_lock);
3105 hugetlb_set_page_subpool(page, spool);
3107 map_commit = vma_commit_reservation(h, vma, addr);
3108 if (unlikely(map_chg > map_commit)) {
3110 * The page was added to the reservation map between
3111 * vma_needs_reservation and vma_commit_reservation.
3112 * This indicates a race with hugetlb_reserve_pages.
3113 * Adjust for the subpool count incremented above AND
3114 * in hugetlb_reserve_pages for the same page. Also,
3115 * the reservation count added in hugetlb_reserve_pages
3116 * no longer applies.
3120 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3121 hugetlb_acct_memory(h, -rsv_adjust);
3122 if (deferred_reserve)
3123 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3124 pages_per_huge_page(h), folio);
3128 out_uncharge_cgroup:
3129 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3130 out_uncharge_cgroup_reservation:
3131 if (deferred_reserve)
3132 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3135 if (map_chg || avoid_reserve)
3136 hugepage_subpool_put_pages(spool, 1);
3137 vma_end_reservation(h, vma, addr);
3138 return ERR_PTR(-ENOSPC);
3141 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3142 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3143 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3145 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3148 /* do node specific alloc */
3149 if (nid != NUMA_NO_NODE) {
3150 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3151 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3156 /* allocate from next node when distributing huge pages */
3157 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3158 m = memblock_alloc_try_nid_raw(
3159 huge_page_size(h), huge_page_size(h),
3160 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3162 * Use the beginning of the huge page to store the
3163 * huge_bootmem_page struct (until gather_bootmem
3164 * puts them into the mem_map).
3172 /* Put them into a private list first because mem_map is not up yet */
3173 INIT_LIST_HEAD(&m->list);
3174 list_add(&m->list, &huge_boot_pages);
3180 * Put bootmem huge pages into the standard lists after mem_map is up.
3181 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3183 static void __init gather_bootmem_prealloc(void)
3185 struct huge_bootmem_page *m;
3187 list_for_each_entry(m, &huge_boot_pages, list) {
3188 struct page *page = virt_to_page(m);
3189 struct folio *folio = page_folio(page);
3190 struct hstate *h = m->hstate;
3192 VM_BUG_ON(!hstate_is_gigantic(h));
3193 WARN_ON(folio_ref_count(folio) != 1);
3194 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3195 WARN_ON(folio_test_reserved(folio));
3196 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3197 free_huge_page(page); /* add to the hugepage allocator */
3199 /* VERY unlikely inflated ref count on a tail page */
3200 free_gigantic_folio(folio, huge_page_order(h));
3204 * We need to restore the 'stolen' pages to totalram_pages
3205 * in order to fix confusing memory reports from free(1) and
3206 * other side-effects, like CommitLimit going negative.
3208 adjust_managed_page_count(page, pages_per_huge_page(h));
3212 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3217 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3218 if (hstate_is_gigantic(h)) {
3219 if (!alloc_bootmem_huge_page(h, nid))
3222 struct folio *folio;
3223 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3225 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3226 &node_states[N_MEMORY], NULL);
3229 free_huge_page(&folio->page); /* free it into the hugepage allocator */
3233 if (i == h->max_huge_pages_node[nid])
3236 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3237 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3238 h->max_huge_pages_node[nid], buf, nid, i);
3239 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3240 h->max_huge_pages_node[nid] = i;
3243 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3246 nodemask_t *node_alloc_noretry;
3247 bool node_specific_alloc = false;
3249 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3250 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3251 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3255 /* do node specific alloc */
3256 for_each_online_node(i) {
3257 if (h->max_huge_pages_node[i] > 0) {
3258 hugetlb_hstate_alloc_pages_onenode(h, i);
3259 node_specific_alloc = true;
3263 if (node_specific_alloc)
3266 /* below will do all node balanced alloc */
3267 if (!hstate_is_gigantic(h)) {
3269 * Bit mask controlling how hard we retry per-node allocations.
3270 * Ignore errors as lower level routines can deal with
3271 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3272 * time, we are likely in bigger trouble.
3274 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3277 /* allocations done at boot time */
3278 node_alloc_noretry = NULL;
3281 /* bit mask controlling how hard we retry per-node allocations */
3282 if (node_alloc_noretry)
3283 nodes_clear(*node_alloc_noretry);
3285 for (i = 0; i < h->max_huge_pages; ++i) {
3286 if (hstate_is_gigantic(h)) {
3287 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3289 } else if (!alloc_pool_huge_page(h,
3290 &node_states[N_MEMORY],
3291 node_alloc_noretry))
3295 if (i < h->max_huge_pages) {
3298 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3299 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3300 h->max_huge_pages, buf, i);
3301 h->max_huge_pages = i;
3303 kfree(node_alloc_noretry);
3306 static void __init hugetlb_init_hstates(void)
3308 struct hstate *h, *h2;
3310 for_each_hstate(h) {
3311 /* oversize hugepages were init'ed in early boot */
3312 if (!hstate_is_gigantic(h))
3313 hugetlb_hstate_alloc_pages(h);
3316 * Set demote order for each hstate. Note that
3317 * h->demote_order is initially 0.
3318 * - We can not demote gigantic pages if runtime freeing
3319 * is not supported, so skip this.
3320 * - If CMA allocation is possible, we can not demote
3321 * HUGETLB_PAGE_ORDER or smaller size pages.
3323 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3325 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3327 for_each_hstate(h2) {
3330 if (h2->order < h->order &&
3331 h2->order > h->demote_order)
3332 h->demote_order = h2->order;
3337 static void __init report_hugepages(void)
3341 for_each_hstate(h) {
3344 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3345 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3346 buf, h->free_huge_pages);
3347 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3348 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3352 #ifdef CONFIG_HIGHMEM
3353 static void try_to_free_low(struct hstate *h, unsigned long count,
3354 nodemask_t *nodes_allowed)
3357 LIST_HEAD(page_list);
3359 lockdep_assert_held(&hugetlb_lock);
3360 if (hstate_is_gigantic(h))
3364 * Collect pages to be freed on a list, and free after dropping lock
3366 for_each_node_mask(i, *nodes_allowed) {
3367 struct page *page, *next;
3368 struct list_head *freel = &h->hugepage_freelists[i];
3369 list_for_each_entry_safe(page, next, freel, lru) {
3370 if (count >= h->nr_huge_pages)
3372 if (PageHighMem(page))
3374 remove_hugetlb_folio(h, page_folio(page), false);
3375 list_add(&page->lru, &page_list);
3380 spin_unlock_irq(&hugetlb_lock);
3381 update_and_free_pages_bulk(h, &page_list);
3382 spin_lock_irq(&hugetlb_lock);
3385 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3386 nodemask_t *nodes_allowed)
3392 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3393 * balanced by operating on them in a round-robin fashion.
3394 * Returns 1 if an adjustment was made.
3396 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3401 lockdep_assert_held(&hugetlb_lock);
3402 VM_BUG_ON(delta != -1 && delta != 1);
3405 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3406 if (h->surplus_huge_pages_node[node])
3410 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3411 if (h->surplus_huge_pages_node[node] <
3412 h->nr_huge_pages_node[node])
3419 h->surplus_huge_pages += delta;
3420 h->surplus_huge_pages_node[node] += delta;
3424 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3425 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3426 nodemask_t *nodes_allowed)
3428 unsigned long min_count, ret;
3430 LIST_HEAD(page_list);
3431 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3434 * Bit mask controlling how hard we retry per-node allocations.
3435 * If we can not allocate the bit mask, do not attempt to allocate
3436 * the requested huge pages.
3438 if (node_alloc_noretry)
3439 nodes_clear(*node_alloc_noretry);
3444 * resize_lock mutex prevents concurrent adjustments to number of
3445 * pages in hstate via the proc/sysfs interfaces.
3447 mutex_lock(&h->resize_lock);
3448 flush_free_hpage_work(h);
3449 spin_lock_irq(&hugetlb_lock);
3452 * Check for a node specific request.
3453 * Changing node specific huge page count may require a corresponding
3454 * change to the global count. In any case, the passed node mask
3455 * (nodes_allowed) will restrict alloc/free to the specified node.
3457 if (nid != NUMA_NO_NODE) {
3458 unsigned long old_count = count;
3460 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3462 * User may have specified a large count value which caused the
3463 * above calculation to overflow. In this case, they wanted
3464 * to allocate as many huge pages as possible. Set count to
3465 * largest possible value to align with their intention.
3467 if (count < old_count)
3472 * Gigantic pages runtime allocation depend on the capability for large
3473 * page range allocation.
3474 * If the system does not provide this feature, return an error when
3475 * the user tries to allocate gigantic pages but let the user free the
3476 * boottime allocated gigantic pages.
3478 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3479 if (count > persistent_huge_pages(h)) {
3480 spin_unlock_irq(&hugetlb_lock);
3481 mutex_unlock(&h->resize_lock);
3482 NODEMASK_FREE(node_alloc_noretry);
3485 /* Fall through to decrease pool */
3489 * Increase the pool size
3490 * First take pages out of surplus state. Then make up the
3491 * remaining difference by allocating fresh huge pages.
3493 * We might race with alloc_surplus_huge_page() here and be unable
3494 * to convert a surplus huge page to a normal huge page. That is
3495 * not critical, though, it just means the overall size of the
3496 * pool might be one hugepage larger than it needs to be, but
3497 * within all the constraints specified by the sysctls.
3499 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3500 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3504 while (count > persistent_huge_pages(h)) {
3506 * If this allocation races such that we no longer need the
3507 * page, free_huge_page will handle it by freeing the page
3508 * and reducing the surplus.
3510 spin_unlock_irq(&hugetlb_lock);
3512 /* yield cpu to avoid soft lockup */
3515 ret = alloc_pool_huge_page(h, nodes_allowed,
3516 node_alloc_noretry);
3517 spin_lock_irq(&hugetlb_lock);
3521 /* Bail for signals. Probably ctrl-c from user */
3522 if (signal_pending(current))
3527 * Decrease the pool size
3528 * First return free pages to the buddy allocator (being careful
3529 * to keep enough around to satisfy reservations). Then place
3530 * pages into surplus state as needed so the pool will shrink
3531 * to the desired size as pages become free.
3533 * By placing pages into the surplus state independent of the
3534 * overcommit value, we are allowing the surplus pool size to
3535 * exceed overcommit. There are few sane options here. Since
3536 * alloc_surplus_huge_page() is checking the global counter,
3537 * though, we'll note that we're not allowed to exceed surplus
3538 * and won't grow the pool anywhere else. Not until one of the
3539 * sysctls are changed, or the surplus pages go out of use.
3541 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3542 min_count = max(count, min_count);
3543 try_to_free_low(h, min_count, nodes_allowed);
3546 * Collect pages to be removed on list without dropping lock
3548 while (min_count < persistent_huge_pages(h)) {
3549 page = remove_pool_huge_page(h, nodes_allowed, 0);
3553 list_add(&page->lru, &page_list);
3555 /* free the pages after dropping lock */
3556 spin_unlock_irq(&hugetlb_lock);
3557 update_and_free_pages_bulk(h, &page_list);
3558 flush_free_hpage_work(h);
3559 spin_lock_irq(&hugetlb_lock);
3561 while (count < persistent_huge_pages(h)) {
3562 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3566 h->max_huge_pages = persistent_huge_pages(h);
3567 spin_unlock_irq(&hugetlb_lock);
3568 mutex_unlock(&h->resize_lock);
3570 NODEMASK_FREE(node_alloc_noretry);
3575 static int demote_free_huge_page(struct hstate *h, struct page *page)
3577 int i, nid = page_to_nid(page);
3578 struct hstate *target_hstate;
3579 struct folio *folio = page_folio(page);
3580 struct page *subpage;
3583 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3585 remove_hugetlb_folio_for_demote(h, folio, false);
3586 spin_unlock_irq(&hugetlb_lock);
3588 rc = hugetlb_vmemmap_restore(h, page);
3590 /* Allocation of vmemmmap failed, we can not demote page */
3591 spin_lock_irq(&hugetlb_lock);
3592 set_page_refcounted(page);
3593 add_hugetlb_folio(h, page_folio(page), false);
3598 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3599 * sizes as it will not ref count pages.
3601 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3604 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3605 * Without the mutex, pages added to target hstate could be marked
3608 * Note that we already hold h->resize_lock. To prevent deadlock,
3609 * use the convention of always taking larger size hstate mutex first.
3611 mutex_lock(&target_hstate->resize_lock);
3612 for (i = 0; i < pages_per_huge_page(h);
3613 i += pages_per_huge_page(target_hstate)) {
3614 subpage = nth_page(page, i);
3615 folio = page_folio(subpage);
3616 if (hstate_is_gigantic(target_hstate))
3617 prep_compound_gigantic_folio_for_demote(folio,
3618 target_hstate->order);
3620 prep_compound_page(subpage, target_hstate->order);
3621 set_page_private(subpage, 0);
3622 prep_new_hugetlb_folio(target_hstate, folio, nid);
3623 free_huge_page(subpage);
3625 mutex_unlock(&target_hstate->resize_lock);
3627 spin_lock_irq(&hugetlb_lock);
3630 * Not absolutely necessary, but for consistency update max_huge_pages
3631 * based on pool changes for the demoted page.
3633 h->max_huge_pages--;
3634 target_hstate->max_huge_pages +=
3635 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3640 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3641 __must_hold(&hugetlb_lock)
3646 lockdep_assert_held(&hugetlb_lock);
3648 /* We should never get here if no demote order */
3649 if (!h->demote_order) {
3650 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3651 return -EINVAL; /* internal error */
3654 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3655 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3656 if (PageHWPoison(page))
3659 return demote_free_huge_page(h, page);
3664 * Only way to get here is if all pages on free lists are poisoned.
3665 * Return -EBUSY so that caller will not retry.
3670 #define HSTATE_ATTR_RO(_name) \
3671 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3673 #define HSTATE_ATTR_WO(_name) \
3674 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3676 #define HSTATE_ATTR(_name) \
3677 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3679 static struct kobject *hugepages_kobj;
3680 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3682 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3684 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3688 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3689 if (hstate_kobjs[i] == kobj) {
3691 *nidp = NUMA_NO_NODE;
3695 return kobj_to_node_hstate(kobj, nidp);
3698 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3699 struct kobj_attribute *attr, char *buf)
3702 unsigned long nr_huge_pages;
3705 h = kobj_to_hstate(kobj, &nid);
3706 if (nid == NUMA_NO_NODE)
3707 nr_huge_pages = h->nr_huge_pages;
3709 nr_huge_pages = h->nr_huge_pages_node[nid];
3711 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3714 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3715 struct hstate *h, int nid,
3716 unsigned long count, size_t len)
3719 nodemask_t nodes_allowed, *n_mask;
3721 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3724 if (nid == NUMA_NO_NODE) {
3726 * global hstate attribute
3728 if (!(obey_mempolicy &&
3729 init_nodemask_of_mempolicy(&nodes_allowed)))
3730 n_mask = &node_states[N_MEMORY];
3732 n_mask = &nodes_allowed;
3735 * Node specific request. count adjustment happens in
3736 * set_max_huge_pages() after acquiring hugetlb_lock.
3738 init_nodemask_of_node(&nodes_allowed, nid);
3739 n_mask = &nodes_allowed;
3742 err = set_max_huge_pages(h, count, nid, n_mask);
3744 return err ? err : len;
3747 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3748 struct kobject *kobj, const char *buf,
3752 unsigned long count;
3756 err = kstrtoul(buf, 10, &count);
3760 h = kobj_to_hstate(kobj, &nid);
3761 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3764 static ssize_t nr_hugepages_show(struct kobject *kobj,
3765 struct kobj_attribute *attr, char *buf)
3767 return nr_hugepages_show_common(kobj, attr, buf);
3770 static ssize_t nr_hugepages_store(struct kobject *kobj,
3771 struct kobj_attribute *attr, const char *buf, size_t len)
3773 return nr_hugepages_store_common(false, kobj, buf, len);
3775 HSTATE_ATTR(nr_hugepages);
3780 * hstate attribute for optionally mempolicy-based constraint on persistent
3781 * huge page alloc/free.
3783 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3784 struct kobj_attribute *attr,
3787 return nr_hugepages_show_common(kobj, attr, buf);
3790 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3791 struct kobj_attribute *attr, const char *buf, size_t len)
3793 return nr_hugepages_store_common(true, kobj, buf, len);
3795 HSTATE_ATTR(nr_hugepages_mempolicy);
3799 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3800 struct kobj_attribute *attr, char *buf)
3802 struct hstate *h = kobj_to_hstate(kobj, NULL);
3803 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3806 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3807 struct kobj_attribute *attr, const char *buf, size_t count)
3810 unsigned long input;
3811 struct hstate *h = kobj_to_hstate(kobj, NULL);
3813 if (hstate_is_gigantic(h))
3816 err = kstrtoul(buf, 10, &input);
3820 spin_lock_irq(&hugetlb_lock);
3821 h->nr_overcommit_huge_pages = input;
3822 spin_unlock_irq(&hugetlb_lock);
3826 HSTATE_ATTR(nr_overcommit_hugepages);
3828 static ssize_t free_hugepages_show(struct kobject *kobj,
3829 struct kobj_attribute *attr, char *buf)
3832 unsigned long free_huge_pages;
3835 h = kobj_to_hstate(kobj, &nid);
3836 if (nid == NUMA_NO_NODE)
3837 free_huge_pages = h->free_huge_pages;
3839 free_huge_pages = h->free_huge_pages_node[nid];
3841 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3843 HSTATE_ATTR_RO(free_hugepages);
3845 static ssize_t resv_hugepages_show(struct kobject *kobj,
3846 struct kobj_attribute *attr, char *buf)
3848 struct hstate *h = kobj_to_hstate(kobj, NULL);
3849 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3851 HSTATE_ATTR_RO(resv_hugepages);
3853 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3854 struct kobj_attribute *attr, char *buf)
3857 unsigned long surplus_huge_pages;
3860 h = kobj_to_hstate(kobj, &nid);
3861 if (nid == NUMA_NO_NODE)
3862 surplus_huge_pages = h->surplus_huge_pages;
3864 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3866 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3868 HSTATE_ATTR_RO(surplus_hugepages);
3870 static ssize_t demote_store(struct kobject *kobj,
3871 struct kobj_attribute *attr, const char *buf, size_t len)
3873 unsigned long nr_demote;
3874 unsigned long nr_available;
3875 nodemask_t nodes_allowed, *n_mask;
3880 err = kstrtoul(buf, 10, &nr_demote);
3883 h = kobj_to_hstate(kobj, &nid);
3885 if (nid != NUMA_NO_NODE) {
3886 init_nodemask_of_node(&nodes_allowed, nid);
3887 n_mask = &nodes_allowed;
3889 n_mask = &node_states[N_MEMORY];
3892 /* Synchronize with other sysfs operations modifying huge pages */
3893 mutex_lock(&h->resize_lock);
3894 spin_lock_irq(&hugetlb_lock);
3898 * Check for available pages to demote each time thorough the
3899 * loop as demote_pool_huge_page will drop hugetlb_lock.
3901 if (nid != NUMA_NO_NODE)
3902 nr_available = h->free_huge_pages_node[nid];
3904 nr_available = h->free_huge_pages;
3905 nr_available -= h->resv_huge_pages;
3909 err = demote_pool_huge_page(h, n_mask);
3916 spin_unlock_irq(&hugetlb_lock);
3917 mutex_unlock(&h->resize_lock);
3923 HSTATE_ATTR_WO(demote);
3925 static ssize_t demote_size_show(struct kobject *kobj,
3926 struct kobj_attribute *attr, char *buf)
3928 struct hstate *h = kobj_to_hstate(kobj, NULL);
3929 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3931 return sysfs_emit(buf, "%lukB\n", demote_size);
3934 static ssize_t demote_size_store(struct kobject *kobj,
3935 struct kobj_attribute *attr,
3936 const char *buf, size_t count)
3938 struct hstate *h, *demote_hstate;
3939 unsigned long demote_size;
3940 unsigned int demote_order;
3942 demote_size = (unsigned long)memparse(buf, NULL);
3944 demote_hstate = size_to_hstate(demote_size);
3947 demote_order = demote_hstate->order;
3948 if (demote_order < HUGETLB_PAGE_ORDER)
3951 /* demote order must be smaller than hstate order */
3952 h = kobj_to_hstate(kobj, NULL);
3953 if (demote_order >= h->order)
3956 /* resize_lock synchronizes access to demote size and writes */
3957 mutex_lock(&h->resize_lock);
3958 h->demote_order = demote_order;
3959 mutex_unlock(&h->resize_lock);
3963 HSTATE_ATTR(demote_size);
3965 static struct attribute *hstate_attrs[] = {
3966 &nr_hugepages_attr.attr,
3967 &nr_overcommit_hugepages_attr.attr,
3968 &free_hugepages_attr.attr,
3969 &resv_hugepages_attr.attr,
3970 &surplus_hugepages_attr.attr,
3972 &nr_hugepages_mempolicy_attr.attr,
3977 static const struct attribute_group hstate_attr_group = {
3978 .attrs = hstate_attrs,
3981 static struct attribute *hstate_demote_attrs[] = {
3982 &demote_size_attr.attr,
3987 static const struct attribute_group hstate_demote_attr_group = {
3988 .attrs = hstate_demote_attrs,
3991 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3992 struct kobject **hstate_kobjs,
3993 const struct attribute_group *hstate_attr_group)
3996 int hi = hstate_index(h);
3998 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3999 if (!hstate_kobjs[hi])
4002 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4004 kobject_put(hstate_kobjs[hi]);
4005 hstate_kobjs[hi] = NULL;
4009 if (h->demote_order) {
4010 retval = sysfs_create_group(hstate_kobjs[hi],
4011 &hstate_demote_attr_group);
4013 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4014 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4015 kobject_put(hstate_kobjs[hi]);
4016 hstate_kobjs[hi] = NULL;
4025 static bool hugetlb_sysfs_initialized __ro_after_init;
4028 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4029 * with node devices in node_devices[] using a parallel array. The array
4030 * index of a node device or _hstate == node id.
4031 * This is here to avoid any static dependency of the node device driver, in
4032 * the base kernel, on the hugetlb module.
4034 struct node_hstate {
4035 struct kobject *hugepages_kobj;
4036 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4038 static struct node_hstate node_hstates[MAX_NUMNODES];
4041 * A subset of global hstate attributes for node devices
4043 static struct attribute *per_node_hstate_attrs[] = {
4044 &nr_hugepages_attr.attr,
4045 &free_hugepages_attr.attr,
4046 &surplus_hugepages_attr.attr,
4050 static const struct attribute_group per_node_hstate_attr_group = {
4051 .attrs = per_node_hstate_attrs,
4055 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4056 * Returns node id via non-NULL nidp.
4058 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4062 for (nid = 0; nid < nr_node_ids; nid++) {
4063 struct node_hstate *nhs = &node_hstates[nid];
4065 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4066 if (nhs->hstate_kobjs[i] == kobj) {
4078 * Unregister hstate attributes from a single node device.
4079 * No-op if no hstate attributes attached.
4081 void hugetlb_unregister_node(struct node *node)
4084 struct node_hstate *nhs = &node_hstates[node->dev.id];
4086 if (!nhs->hugepages_kobj)
4087 return; /* no hstate attributes */
4089 for_each_hstate(h) {
4090 int idx = hstate_index(h);
4091 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4095 if (h->demote_order)
4096 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4097 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4098 kobject_put(hstate_kobj);
4099 nhs->hstate_kobjs[idx] = NULL;
4102 kobject_put(nhs->hugepages_kobj);
4103 nhs->hugepages_kobj = NULL;
4108 * Register hstate attributes for a single node device.
4109 * No-op if attributes already registered.
4111 void hugetlb_register_node(struct node *node)
4114 struct node_hstate *nhs = &node_hstates[node->dev.id];
4117 if (!hugetlb_sysfs_initialized)
4120 if (nhs->hugepages_kobj)
4121 return; /* already allocated */
4123 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4125 if (!nhs->hugepages_kobj)
4128 for_each_hstate(h) {
4129 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4131 &per_node_hstate_attr_group);
4133 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4134 h->name, node->dev.id);
4135 hugetlb_unregister_node(node);
4142 * hugetlb init time: register hstate attributes for all registered node
4143 * devices of nodes that have memory. All on-line nodes should have
4144 * registered their associated device by this time.
4146 static void __init hugetlb_register_all_nodes(void)
4150 for_each_online_node(nid)
4151 hugetlb_register_node(node_devices[nid]);
4153 #else /* !CONFIG_NUMA */
4155 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4163 static void hugetlb_register_all_nodes(void) { }
4168 static void __init hugetlb_cma_check(void);
4170 static inline __init void hugetlb_cma_check(void)
4175 static void __init hugetlb_sysfs_init(void)
4180 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4181 if (!hugepages_kobj)
4184 for_each_hstate(h) {
4185 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4186 hstate_kobjs, &hstate_attr_group);
4188 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4192 hugetlb_sysfs_initialized = true;
4194 hugetlb_register_all_nodes();
4197 static int __init hugetlb_init(void)
4201 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4204 if (!hugepages_supported()) {
4205 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4206 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4211 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4212 * architectures depend on setup being done here.
4214 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4215 if (!parsed_default_hugepagesz) {
4217 * If we did not parse a default huge page size, set
4218 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4219 * number of huge pages for this default size was implicitly
4220 * specified, set that here as well.
4221 * Note that the implicit setting will overwrite an explicit
4222 * setting. A warning will be printed in this case.
4224 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4225 if (default_hstate_max_huge_pages) {
4226 if (default_hstate.max_huge_pages) {
4229 string_get_size(huge_page_size(&default_hstate),
4230 1, STRING_UNITS_2, buf, 32);
4231 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4232 default_hstate.max_huge_pages, buf);
4233 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4234 default_hstate_max_huge_pages);
4236 default_hstate.max_huge_pages =
4237 default_hstate_max_huge_pages;
4239 for_each_online_node(i)
4240 default_hstate.max_huge_pages_node[i] =
4241 default_hugepages_in_node[i];
4245 hugetlb_cma_check();
4246 hugetlb_init_hstates();
4247 gather_bootmem_prealloc();
4250 hugetlb_sysfs_init();
4251 hugetlb_cgroup_file_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 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 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 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 #endif /* CONFIG_SYSCTL */
4631 void hugetlb_report_meminfo(struct seq_file *m)
4634 unsigned long total = 0;
4636 if (!hugepages_supported())
4639 for_each_hstate(h) {
4640 unsigned long count = h->nr_huge_pages;
4642 total += huge_page_size(h) * count;
4644 if (h == &default_hstate)
4646 "HugePages_Total: %5lu\n"
4647 "HugePages_Free: %5lu\n"
4648 "HugePages_Rsvd: %5lu\n"
4649 "HugePages_Surp: %5lu\n"
4650 "Hugepagesize: %8lu kB\n",
4654 h->surplus_huge_pages,
4655 huge_page_size(h) / SZ_1K);
4658 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4661 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4663 struct hstate *h = &default_hstate;
4665 if (!hugepages_supported())
4668 return sysfs_emit_at(buf, len,
4669 "Node %d HugePages_Total: %5u\n"
4670 "Node %d HugePages_Free: %5u\n"
4671 "Node %d HugePages_Surp: %5u\n",
4672 nid, h->nr_huge_pages_node[nid],
4673 nid, h->free_huge_pages_node[nid],
4674 nid, h->surplus_huge_pages_node[nid]);
4677 void hugetlb_show_meminfo_node(int nid)
4681 if (!hugepages_supported())
4685 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4687 h->nr_huge_pages_node[nid],
4688 h->free_huge_pages_node[nid],
4689 h->surplus_huge_pages_node[nid],
4690 huge_page_size(h) / SZ_1K);
4693 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4695 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4696 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4699 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4700 unsigned long hugetlb_total_pages(void)
4703 unsigned long nr_total_pages = 0;
4706 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4707 return nr_total_pages;
4710 static int hugetlb_acct_memory(struct hstate *h, long delta)
4717 spin_lock_irq(&hugetlb_lock);
4719 * When cpuset is configured, it breaks the strict hugetlb page
4720 * reservation as the accounting is done on a global variable. Such
4721 * reservation is completely rubbish in the presence of cpuset because
4722 * the reservation is not checked against page availability for the
4723 * current cpuset. Application can still potentially OOM'ed by kernel
4724 * with lack of free htlb page in cpuset that the task is in.
4725 * Attempt to enforce strict accounting with cpuset is almost
4726 * impossible (or too ugly) because cpuset is too fluid that
4727 * task or memory node can be dynamically moved between cpusets.
4729 * The change of semantics for shared hugetlb mapping with cpuset is
4730 * undesirable. However, in order to preserve some of the semantics,
4731 * we fall back to check against current free page availability as
4732 * a best attempt and hopefully to minimize the impact of changing
4733 * semantics that cpuset has.
4735 * Apart from cpuset, we also have memory policy mechanism that
4736 * also determines from which node the kernel will allocate memory
4737 * in a NUMA system. So similar to cpuset, we also should consider
4738 * the memory policy of the current task. Similar to the description
4742 if (gather_surplus_pages(h, delta) < 0)
4745 if (delta > allowed_mems_nr(h)) {
4746 return_unused_surplus_pages(h, delta);
4753 return_unused_surplus_pages(h, (unsigned long) -delta);
4756 spin_unlock_irq(&hugetlb_lock);
4760 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4762 struct resv_map *resv = vma_resv_map(vma);
4765 * HPAGE_RESV_OWNER indicates a private mapping.
4766 * This new VMA should share its siblings reservation map if present.
4767 * The VMA will only ever have a valid reservation map pointer where
4768 * it is being copied for another still existing VMA. As that VMA
4769 * has a reference to the reservation map it cannot disappear until
4770 * after this open call completes. It is therefore safe to take a
4771 * new reference here without additional locking.
4773 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4774 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4775 kref_get(&resv->refs);
4779 * vma_lock structure for sharable mappings is vma specific.
4780 * Clear old pointer (if copied via vm_area_dup) and allocate
4781 * new structure. Before clearing, make sure vma_lock is not
4784 if (vma->vm_flags & VM_MAYSHARE) {
4785 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4788 if (vma_lock->vma != vma) {
4789 vma->vm_private_data = NULL;
4790 hugetlb_vma_lock_alloc(vma);
4792 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4794 hugetlb_vma_lock_alloc(vma);
4798 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4800 struct hstate *h = hstate_vma(vma);
4801 struct resv_map *resv;
4802 struct hugepage_subpool *spool = subpool_vma(vma);
4803 unsigned long reserve, start, end;
4806 hugetlb_vma_lock_free(vma);
4808 resv = vma_resv_map(vma);
4809 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4812 start = vma_hugecache_offset(h, vma, vma->vm_start);
4813 end = vma_hugecache_offset(h, vma, vma->vm_end);
4815 reserve = (end - start) - region_count(resv, start, end);
4816 hugetlb_cgroup_uncharge_counter(resv, start, end);
4819 * Decrement reserve counts. The global reserve count may be
4820 * adjusted if the subpool has a minimum size.
4822 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4823 hugetlb_acct_memory(h, -gbl_reserve);
4826 kref_put(&resv->refs, resv_map_release);
4829 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4831 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4835 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4836 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4837 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4839 if (addr & ~PUD_MASK) {
4841 * hugetlb_vm_op_split is called right before we attempt to
4842 * split the VMA. We will need to unshare PMDs in the old and
4843 * new VMAs, so let's unshare before we split.
4845 unsigned long floor = addr & PUD_MASK;
4846 unsigned long ceil = floor + PUD_SIZE;
4848 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4849 hugetlb_unshare_pmds(vma, floor, ceil);
4855 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4857 return huge_page_size(hstate_vma(vma));
4861 * We cannot handle pagefaults against hugetlb pages at all. They cause
4862 * handle_mm_fault() to try to instantiate regular-sized pages in the
4863 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4866 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4873 * When a new function is introduced to vm_operations_struct and added
4874 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4875 * This is because under System V memory model, mappings created via
4876 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4877 * their original vm_ops are overwritten with shm_vm_ops.
4879 const struct vm_operations_struct hugetlb_vm_ops = {
4880 .fault = hugetlb_vm_op_fault,
4881 .open = hugetlb_vm_op_open,
4882 .close = hugetlb_vm_op_close,
4883 .may_split = hugetlb_vm_op_split,
4884 .pagesize = hugetlb_vm_op_pagesize,
4887 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4891 unsigned int shift = huge_page_shift(hstate_vma(vma));
4894 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4895 vma->vm_page_prot)));
4897 entry = huge_pte_wrprotect(mk_huge_pte(page,
4898 vma->vm_page_prot));
4900 entry = pte_mkyoung(entry);
4901 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4906 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4907 unsigned long address, pte_t *ptep)
4911 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4912 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4913 update_mmu_cache(vma, address, ptep);
4916 bool is_hugetlb_entry_migration(pte_t pte)
4920 if (huge_pte_none(pte) || pte_present(pte))
4922 swp = pte_to_swp_entry(pte);
4923 if (is_migration_entry(swp))
4929 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4933 if (huge_pte_none(pte) || pte_present(pte))
4935 swp = pte_to_swp_entry(pte);
4936 if (is_hwpoison_entry(swp))
4943 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4944 struct page *new_page)
4946 __SetPageUptodate(new_page);
4947 hugepage_add_new_anon_rmap(new_page, vma, addr);
4948 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4949 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4950 SetHPageMigratable(new_page);
4953 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4954 struct vm_area_struct *dst_vma,
4955 struct vm_area_struct *src_vma)
4957 pte_t *src_pte, *dst_pte, entry;
4958 struct page *ptepage;
4960 bool cow = is_cow_mapping(src_vma->vm_flags);
4961 struct hstate *h = hstate_vma(src_vma);
4962 unsigned long sz = huge_page_size(h);
4963 unsigned long npages = pages_per_huge_page(h);
4964 struct mmu_notifier_range range;
4965 unsigned long last_addr_mask;
4969 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4972 mmu_notifier_invalidate_range_start(&range);
4973 mmap_assert_write_locked(src);
4974 raw_write_seqcount_begin(&src->write_protect_seq);
4977 * For shared mappings the vma lock must be held before
4978 * calling hugetlb_walk() in the src vma. Otherwise, the
4979 * returned ptep could go away if part of a shared pmd and
4980 * another thread calls huge_pmd_unshare.
4982 hugetlb_vma_lock_read(src_vma);
4985 last_addr_mask = hugetlb_mask_last_page(h);
4986 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4987 spinlock_t *src_ptl, *dst_ptl;
4988 src_pte = hugetlb_walk(src_vma, addr, sz);
4990 addr |= last_addr_mask;
4993 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5000 * If the pagetables are shared don't copy or take references.
5002 * dst_pte == src_pte is the common case of src/dest sharing.
5003 * However, src could have 'unshared' and dst shares with
5004 * another vma. So page_count of ptep page is checked instead
5005 * to reliably determine whether pte is shared.
5007 if (page_count(virt_to_page(dst_pte)) > 1) {
5008 addr |= last_addr_mask;
5012 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5013 src_ptl = huge_pte_lockptr(h, src, src_pte);
5014 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5015 entry = huge_ptep_get(src_pte);
5017 if (huge_pte_none(entry)) {
5019 * Skip if src entry none.
5022 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5023 bool uffd_wp = huge_pte_uffd_wp(entry);
5025 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5026 entry = huge_pte_clear_uffd_wp(entry);
5027 set_huge_pte_at(dst, addr, dst_pte, entry);
5028 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5029 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5030 bool uffd_wp = huge_pte_uffd_wp(entry);
5032 if (!is_readable_migration_entry(swp_entry) && cow) {
5034 * COW mappings require pages in both
5035 * parent and child to be set to read.
5037 swp_entry = make_readable_migration_entry(
5038 swp_offset(swp_entry));
5039 entry = swp_entry_to_pte(swp_entry);
5040 if (userfaultfd_wp(src_vma) && uffd_wp)
5041 entry = huge_pte_mkuffd_wp(entry);
5042 set_huge_pte_at(src, addr, src_pte, entry);
5044 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5045 entry = huge_pte_clear_uffd_wp(entry);
5046 set_huge_pte_at(dst, addr, dst_pte, entry);
5047 } else if (unlikely(is_pte_marker(entry))) {
5048 /* No swap on hugetlb */
5050 is_swapin_error_entry(pte_to_swp_entry(entry)));
5052 * We copy the pte marker only if the dst vma has
5055 if (userfaultfd_wp(dst_vma))
5056 set_huge_pte_at(dst, addr, dst_pte, entry);
5058 entry = huge_ptep_get(src_pte);
5059 ptepage = pte_page(entry);
5063 * Failing to duplicate the anon rmap is a rare case
5064 * where we see pinned hugetlb pages while they're
5065 * prone to COW. We need to do the COW earlier during
5068 * When pre-allocating the page or copying data, we
5069 * need to be without the pgtable locks since we could
5070 * sleep during the process.
5072 if (!PageAnon(ptepage)) {
5073 page_dup_file_rmap(ptepage, true);
5074 } else if (page_try_dup_anon_rmap(ptepage, true,
5076 pte_t src_pte_old = entry;
5079 spin_unlock(src_ptl);
5080 spin_unlock(dst_ptl);
5081 /* Do not use reserve as it's private owned */
5082 new = alloc_huge_page(dst_vma, addr, 1);
5088 copy_user_huge_page(new, ptepage, addr, dst_vma,
5092 /* Install the new huge page if src pte stable */
5093 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5094 src_ptl = huge_pte_lockptr(h, src, src_pte);
5095 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5096 entry = huge_ptep_get(src_pte);
5097 if (!pte_same(src_pte_old, entry)) {
5098 restore_reserve_on_error(h, dst_vma, addr,
5101 /* huge_ptep of dst_pte won't change as in child */
5104 hugetlb_install_page(dst_vma, dst_pte, addr, new);
5105 spin_unlock(src_ptl);
5106 spin_unlock(dst_ptl);
5112 * No need to notify as we are downgrading page
5113 * table protection not changing it to point
5116 * See Documentation/mm/mmu_notifier.rst
5118 huge_ptep_set_wrprotect(src, addr, src_pte);
5119 entry = huge_pte_wrprotect(entry);
5122 set_huge_pte_at(dst, addr, dst_pte, entry);
5123 hugetlb_count_add(npages, dst);
5125 spin_unlock(src_ptl);
5126 spin_unlock(dst_ptl);
5130 raw_write_seqcount_end(&src->write_protect_seq);
5131 mmu_notifier_invalidate_range_end(&range);
5133 hugetlb_vma_unlock_read(src_vma);
5139 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5140 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5142 struct hstate *h = hstate_vma(vma);
5143 struct mm_struct *mm = vma->vm_mm;
5144 spinlock_t *src_ptl, *dst_ptl;
5147 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5148 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5151 * We don't have to worry about the ordering of src and dst ptlocks
5152 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5154 if (src_ptl != dst_ptl)
5155 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5157 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5158 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5160 if (src_ptl != dst_ptl)
5161 spin_unlock(src_ptl);
5162 spin_unlock(dst_ptl);
5165 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5166 struct vm_area_struct *new_vma,
5167 unsigned long old_addr, unsigned long new_addr,
5170 struct hstate *h = hstate_vma(vma);
5171 struct address_space *mapping = vma->vm_file->f_mapping;
5172 unsigned long sz = huge_page_size(h);
5173 struct mm_struct *mm = vma->vm_mm;
5174 unsigned long old_end = old_addr + len;
5175 unsigned long last_addr_mask;
5176 pte_t *src_pte, *dst_pte;
5177 struct mmu_notifier_range range;
5178 bool shared_pmd = false;
5180 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
5182 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5184 * In case of shared PMDs, we should cover the maximum possible
5187 flush_cache_range(vma, range.start, range.end);
5189 mmu_notifier_invalidate_range_start(&range);
5190 last_addr_mask = hugetlb_mask_last_page(h);
5191 /* Prevent race with file truncation */
5192 hugetlb_vma_lock_write(vma);
5193 i_mmap_lock_write(mapping);
5194 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5195 src_pte = hugetlb_walk(vma, old_addr, sz);
5197 old_addr |= last_addr_mask;
5198 new_addr |= last_addr_mask;
5201 if (huge_pte_none(huge_ptep_get(src_pte)))
5204 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5206 old_addr |= last_addr_mask;
5207 new_addr |= last_addr_mask;
5211 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5215 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5219 flush_tlb_range(vma, range.start, range.end);
5221 flush_tlb_range(vma, old_end - len, old_end);
5222 mmu_notifier_invalidate_range_end(&range);
5223 i_mmap_unlock_write(mapping);
5224 hugetlb_vma_unlock_write(vma);
5226 return len + old_addr - old_end;
5229 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5230 unsigned long start, unsigned long end,
5231 struct page *ref_page, zap_flags_t zap_flags)
5233 struct mm_struct *mm = vma->vm_mm;
5234 unsigned long address;
5239 struct hstate *h = hstate_vma(vma);
5240 unsigned long sz = huge_page_size(h);
5241 unsigned long last_addr_mask;
5242 bool force_flush = false;
5244 WARN_ON(!is_vm_hugetlb_page(vma));
5245 BUG_ON(start & ~huge_page_mask(h));
5246 BUG_ON(end & ~huge_page_mask(h));
5249 * This is a hugetlb vma, all the pte entries should point
5252 tlb_change_page_size(tlb, sz);
5253 tlb_start_vma(tlb, vma);
5255 last_addr_mask = hugetlb_mask_last_page(h);
5257 for (; address < end; address += sz) {
5258 ptep = hugetlb_walk(vma, address, sz);
5260 address |= last_addr_mask;
5264 ptl = huge_pte_lock(h, mm, ptep);
5265 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5267 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5269 address |= last_addr_mask;
5273 pte = huge_ptep_get(ptep);
5274 if (huge_pte_none(pte)) {
5280 * Migrating hugepage or HWPoisoned hugepage is already
5281 * unmapped and its refcount is dropped, so just clear pte here.
5283 if (unlikely(!pte_present(pte))) {
5285 * If the pte was wr-protected by uffd-wp in any of the
5286 * swap forms, meanwhile the caller does not want to
5287 * drop the uffd-wp bit in this zap, then replace the
5288 * pte with a marker.
5290 if (pte_swp_uffd_wp_any(pte) &&
5291 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5292 set_huge_pte_at(mm, address, ptep,
5293 make_pte_marker(PTE_MARKER_UFFD_WP));
5295 huge_pte_clear(mm, address, ptep, sz);
5300 page = pte_page(pte);
5302 * If a reference page is supplied, it is because a specific
5303 * page is being unmapped, not a range. Ensure the page we
5304 * are about to unmap is the actual page of interest.
5307 if (page != ref_page) {
5312 * Mark the VMA as having unmapped its page so that
5313 * future faults in this VMA will fail rather than
5314 * looking like data was lost
5316 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5319 pte = huge_ptep_get_and_clear(mm, address, ptep);
5320 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5321 if (huge_pte_dirty(pte))
5322 set_page_dirty(page);
5323 /* Leave a uffd-wp pte marker if needed */
5324 if (huge_pte_uffd_wp(pte) &&
5325 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5326 set_huge_pte_at(mm, address, ptep,
5327 make_pte_marker(PTE_MARKER_UFFD_WP));
5328 hugetlb_count_sub(pages_per_huge_page(h), mm);
5329 page_remove_rmap(page, vma, true);
5332 tlb_remove_page_size(tlb, page, huge_page_size(h));
5334 * Bail out after unmapping reference page if supplied
5339 tlb_end_vma(tlb, vma);
5342 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5343 * could defer the flush until now, since by holding i_mmap_rwsem we
5344 * guaranteed that the last refernece would not be dropped. But we must
5345 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5346 * dropped and the last reference to the shared PMDs page might be
5349 * In theory we could defer the freeing of the PMD pages as well, but
5350 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5351 * detect sharing, so we cannot defer the release of the page either.
5352 * Instead, do flush now.
5355 tlb_flush_mmu_tlbonly(tlb);
5358 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5359 struct vm_area_struct *vma, unsigned long start,
5360 unsigned long end, struct page *ref_page,
5361 zap_flags_t zap_flags)
5363 hugetlb_vma_lock_write(vma);
5364 i_mmap_lock_write(vma->vm_file->f_mapping);
5366 /* mmu notification performed in caller */
5367 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5369 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5371 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5372 * When the vma_lock is freed, this makes the vma ineligible
5373 * for pmd sharing. And, i_mmap_rwsem is required to set up
5374 * pmd sharing. This is important as page tables for this
5375 * unmapped range will be asynchrously deleted. If the page
5376 * tables are shared, there will be issues when accessed by
5379 __hugetlb_vma_unlock_write_free(vma);
5380 i_mmap_unlock_write(vma->vm_file->f_mapping);
5382 i_mmap_unlock_write(vma->vm_file->f_mapping);
5383 hugetlb_vma_unlock_write(vma);
5387 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5388 unsigned long end, struct page *ref_page,
5389 zap_flags_t zap_flags)
5391 struct mmu_notifier_range range;
5392 struct mmu_gather tlb;
5394 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
5396 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5397 mmu_notifier_invalidate_range_start(&range);
5398 tlb_gather_mmu(&tlb, vma->vm_mm);
5400 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5402 mmu_notifier_invalidate_range_end(&range);
5403 tlb_finish_mmu(&tlb);
5407 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5408 * mapping it owns the reserve page for. The intention is to unmap the page
5409 * from other VMAs and let the children be SIGKILLed if they are faulting the
5412 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5413 struct page *page, unsigned long address)
5415 struct hstate *h = hstate_vma(vma);
5416 struct vm_area_struct *iter_vma;
5417 struct address_space *mapping;
5421 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5422 * from page cache lookup which is in HPAGE_SIZE units.
5424 address = address & huge_page_mask(h);
5425 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5427 mapping = vma->vm_file->f_mapping;
5430 * Take the mapping lock for the duration of the table walk. As
5431 * this mapping should be shared between all the VMAs,
5432 * __unmap_hugepage_range() is called as the lock is already held
5434 i_mmap_lock_write(mapping);
5435 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5436 /* Do not unmap the current VMA */
5437 if (iter_vma == vma)
5441 * Shared VMAs have their own reserves and do not affect
5442 * MAP_PRIVATE accounting but it is possible that a shared
5443 * VMA is using the same page so check and skip such VMAs.
5445 if (iter_vma->vm_flags & VM_MAYSHARE)
5449 * Unmap the page from other VMAs without their own reserves.
5450 * They get marked to be SIGKILLed if they fault in these
5451 * areas. This is because a future no-page fault on this VMA
5452 * could insert a zeroed page instead of the data existing
5453 * from the time of fork. This would look like data corruption
5455 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5456 unmap_hugepage_range(iter_vma, address,
5457 address + huge_page_size(h), page, 0);
5459 i_mmap_unlock_write(mapping);
5463 * hugetlb_wp() should be called with page lock of the original hugepage held.
5464 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5465 * cannot race with other handlers or page migration.
5466 * Keep the pte_same checks anyway to make transition from the mutex easier.
5468 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5469 unsigned long address, pte_t *ptep, unsigned int flags,
5470 struct page *pagecache_page, spinlock_t *ptl)
5472 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5474 struct hstate *h = hstate_vma(vma);
5475 struct page *old_page, *new_page;
5476 int outside_reserve = 0;
5478 unsigned long haddr = address & huge_page_mask(h);
5479 struct mmu_notifier_range range;
5482 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5483 * PTE mapped R/O such as maybe_mkwrite() would do.
5485 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5486 return VM_FAULT_SIGSEGV;
5488 /* Let's take out MAP_SHARED mappings first. */
5489 if (vma->vm_flags & VM_MAYSHARE) {
5490 set_huge_ptep_writable(vma, haddr, ptep);
5494 pte = huge_ptep_get(ptep);
5495 old_page = pte_page(pte);
5497 delayacct_wpcopy_start();
5501 * If no-one else is actually using this page, we're the exclusive
5502 * owner and can reuse this page.
5504 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5505 if (!PageAnonExclusive(old_page))
5506 page_move_anon_rmap(old_page, vma);
5507 if (likely(!unshare))
5508 set_huge_ptep_writable(vma, haddr, ptep);
5510 delayacct_wpcopy_end();
5513 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5517 * If the process that created a MAP_PRIVATE mapping is about to
5518 * perform a COW due to a shared page count, attempt to satisfy
5519 * the allocation without using the existing reserves. The pagecache
5520 * page is used to determine if the reserve at this address was
5521 * consumed or not. If reserves were used, a partial faulted mapping
5522 * at the time of fork() could consume its reserves on COW instead
5523 * of the full address range.
5525 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5526 old_page != pagecache_page)
5527 outside_reserve = 1;
5532 * Drop page table lock as buddy allocator may be called. It will
5533 * be acquired again before returning to the caller, as expected.
5536 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5538 if (IS_ERR(new_page)) {
5540 * If a process owning a MAP_PRIVATE mapping fails to COW,
5541 * it is due to references held by a child and an insufficient
5542 * huge page pool. To guarantee the original mappers
5543 * reliability, unmap the page from child processes. The child
5544 * may get SIGKILLed if it later faults.
5546 if (outside_reserve) {
5547 struct address_space *mapping = vma->vm_file->f_mapping;
5553 * Drop hugetlb_fault_mutex and vma_lock before
5554 * unmapping. unmapping needs to hold vma_lock
5555 * in write mode. Dropping vma_lock in read mode
5556 * here is OK as COW mappings do not interact with
5559 * Reacquire both after unmap operation.
5561 idx = vma_hugecache_offset(h, vma, haddr);
5562 hash = hugetlb_fault_mutex_hash(mapping, idx);
5563 hugetlb_vma_unlock_read(vma);
5564 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5566 unmap_ref_private(mm, vma, old_page, haddr);
5568 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5569 hugetlb_vma_lock_read(vma);
5571 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5573 pte_same(huge_ptep_get(ptep), pte)))
5574 goto retry_avoidcopy;
5576 * race occurs while re-acquiring page table
5577 * lock, and our job is done.
5579 delayacct_wpcopy_end();
5583 ret = vmf_error(PTR_ERR(new_page));
5584 goto out_release_old;
5588 * When the original hugepage is shared one, it does not have
5589 * anon_vma prepared.
5591 if (unlikely(anon_vma_prepare(vma))) {
5593 goto out_release_all;
5596 copy_user_huge_page(new_page, old_page, address, vma,
5597 pages_per_huge_page(h));
5598 __SetPageUptodate(new_page);
5600 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5601 haddr + huge_page_size(h));
5602 mmu_notifier_invalidate_range_start(&range);
5605 * Retake the page table lock to check for racing updates
5606 * before the page tables are altered
5609 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5610 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5611 /* Break COW or unshare */
5612 huge_ptep_clear_flush(vma, haddr, ptep);
5613 mmu_notifier_invalidate_range(mm, range.start, range.end);
5614 page_remove_rmap(old_page, vma, true);
5615 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5616 set_huge_pte_at(mm, haddr, ptep,
5617 make_huge_pte(vma, new_page, !unshare));
5618 SetHPageMigratable(new_page);
5619 /* Make the old page be freed below */
5620 new_page = old_page;
5623 mmu_notifier_invalidate_range_end(&range);
5626 * No restore in case of successful pagetable update (Break COW or
5629 if (new_page != old_page)
5630 restore_reserve_on_error(h, vma, haddr, new_page);
5635 spin_lock(ptl); /* Caller expects lock to be held */
5637 delayacct_wpcopy_end();
5642 * Return whether there is a pagecache page to back given address within VMA.
5643 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5645 static bool hugetlbfs_pagecache_present(struct hstate *h,
5646 struct vm_area_struct *vma, unsigned long address)
5648 struct address_space *mapping;
5652 mapping = vma->vm_file->f_mapping;
5653 idx = vma_hugecache_offset(h, vma, address);
5655 page = find_get_page(mapping, idx);
5658 return page != NULL;
5661 int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping,
5664 struct folio *folio = page_folio(page);
5665 struct inode *inode = mapping->host;
5666 struct hstate *h = hstate_inode(inode);
5669 __folio_set_locked(folio);
5670 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5672 if (unlikely(err)) {
5673 __folio_clear_locked(folio);
5676 ClearHPageRestoreReserve(page);
5679 * mark folio dirty so that it will not be removed from cache/file
5680 * by non-hugetlbfs specific code paths.
5682 folio_mark_dirty(folio);
5684 spin_lock(&inode->i_lock);
5685 inode->i_blocks += blocks_per_huge_page(h);
5686 spin_unlock(&inode->i_lock);
5690 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5691 struct address_space *mapping,
5694 unsigned long haddr,
5696 unsigned long reason)
5699 struct vm_fault vmf = {
5702 .real_address = addr,
5706 * Hard to debug if it ends up being
5707 * used by a callee that assumes
5708 * something about the other
5709 * uninitialized fields... same as in
5715 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5716 * userfault. Also mmap_lock could be dropped due to handling
5717 * userfault, any vma operation should be careful from here.
5719 hugetlb_vma_unlock_read(vma);
5720 hash = hugetlb_fault_mutex_hash(mapping, idx);
5721 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5722 return handle_userfault(&vmf, reason);
5726 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5727 * false if pte changed or is changing.
5729 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5730 pte_t *ptep, pte_t old_pte)
5735 ptl = huge_pte_lock(h, mm, ptep);
5736 same = pte_same(huge_ptep_get(ptep), old_pte);
5742 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5743 struct vm_area_struct *vma,
5744 struct address_space *mapping, pgoff_t idx,
5745 unsigned long address, pte_t *ptep,
5746 pte_t old_pte, unsigned int flags)
5748 struct hstate *h = hstate_vma(vma);
5749 vm_fault_t ret = VM_FAULT_SIGBUS;
5755 unsigned long haddr = address & huge_page_mask(h);
5756 bool new_page, new_pagecache_page = false;
5757 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5760 * Currently, we are forced to kill the process in the event the
5761 * original mapper has unmapped pages from the child due to a failed
5762 * COW/unsharing. Warn that such a situation has occurred as it may not
5765 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5766 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5772 * Use page lock to guard against racing truncation
5773 * before we get page_table_lock.
5776 page = find_lock_page(mapping, idx);
5778 size = i_size_read(mapping->host) >> huge_page_shift(h);
5781 /* Check for page in userfault range */
5782 if (userfaultfd_missing(vma)) {
5784 * Since hugetlb_no_page() was examining pte
5785 * without pgtable lock, we need to re-test under
5786 * lock because the pte may not be stable and could
5787 * have changed from under us. Try to detect
5788 * either changed or during-changing ptes and retry
5789 * properly when needed.
5791 * Note that userfaultfd is actually fine with
5792 * false positives (e.g. caused by pte changed),
5793 * but not wrong logical events (e.g. caused by
5794 * reading a pte during changing). The latter can
5795 * confuse the userspace, so the strictness is very
5796 * much preferred. E.g., MISSING event should
5797 * never happen on the page after UFFDIO_COPY has
5798 * correctly installed the page and returned.
5800 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5805 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5810 page = alloc_huge_page(vma, haddr, 0);
5813 * Returning error will result in faulting task being
5814 * sent SIGBUS. The hugetlb fault mutex prevents two
5815 * tasks from racing to fault in the same page which
5816 * could result in false unable to allocate errors.
5817 * Page migration does not take the fault mutex, but
5818 * does a clear then write of pte's under page table
5819 * lock. Page fault code could race with migration,
5820 * notice the clear pte and try to allocate a page
5821 * here. Before returning error, get ptl and make
5822 * sure there really is no pte entry.
5824 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5825 ret = vmf_error(PTR_ERR(page));
5830 clear_huge_page(page, address, pages_per_huge_page(h));
5831 __SetPageUptodate(page);
5834 if (vma->vm_flags & VM_MAYSHARE) {
5835 int err = hugetlb_add_to_page_cache(page, mapping, idx);
5838 * err can't be -EEXIST which implies someone
5839 * else consumed the reservation since hugetlb
5840 * fault mutex is held when add a hugetlb page
5841 * to the page cache. So it's safe to call
5842 * restore_reserve_on_error() here.
5844 restore_reserve_on_error(h, vma, haddr, page);
5848 new_pagecache_page = true;
5851 if (unlikely(anon_vma_prepare(vma))) {
5853 goto backout_unlocked;
5859 * If memory error occurs between mmap() and fault, some process
5860 * don't have hwpoisoned swap entry for errored virtual address.
5861 * So we need to block hugepage fault by PG_hwpoison bit check.
5863 if (unlikely(PageHWPoison(page))) {
5864 ret = VM_FAULT_HWPOISON_LARGE |
5865 VM_FAULT_SET_HINDEX(hstate_index(h));
5866 goto backout_unlocked;
5869 /* Check for page in userfault range. */
5870 if (userfaultfd_minor(vma)) {
5873 /* See comment in userfaultfd_missing() block above */
5874 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5878 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5885 * If we are going to COW a private mapping later, we examine the
5886 * pending reservations for this page now. This will ensure that
5887 * any allocations necessary to record that reservation occur outside
5890 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5891 if (vma_needs_reservation(h, vma, haddr) < 0) {
5893 goto backout_unlocked;
5895 /* Just decrements count, does not deallocate */
5896 vma_end_reservation(h, vma, haddr);
5899 ptl = huge_pte_lock(h, mm, ptep);
5901 /* If pte changed from under us, retry */
5902 if (!pte_same(huge_ptep_get(ptep), old_pte))
5906 hugepage_add_new_anon_rmap(page, vma, haddr);
5908 page_dup_file_rmap(page, true);
5909 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5910 && (vma->vm_flags & VM_SHARED)));
5912 * If this pte was previously wr-protected, keep it wr-protected even
5915 if (unlikely(pte_marker_uffd_wp(old_pte)))
5916 new_pte = huge_pte_mkuffd_wp(new_pte);
5917 set_huge_pte_at(mm, haddr, ptep, new_pte);
5919 hugetlb_count_add(pages_per_huge_page(h), mm);
5920 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5921 /* Optimization, do the COW without a second fault */
5922 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5928 * Only set HPageMigratable in newly allocated pages. Existing pages
5929 * found in the pagecache may not have HPageMigratableset if they have
5930 * been isolated for migration.
5933 SetHPageMigratable(page);
5937 hugetlb_vma_unlock_read(vma);
5938 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5944 if (new_page && !new_pagecache_page)
5945 restore_reserve_on_error(h, vma, haddr, page);
5953 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5955 unsigned long key[2];
5958 key[0] = (unsigned long) mapping;
5961 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5963 return hash & (num_fault_mutexes - 1);
5967 * For uniprocessor systems we always use a single mutex, so just
5968 * return 0 and avoid the hashing overhead.
5970 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5976 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5977 unsigned long address, unsigned int flags)
5984 struct page *page = NULL;
5985 struct page *pagecache_page = NULL;
5986 struct hstate *h = hstate_vma(vma);
5987 struct address_space *mapping;
5988 int need_wait_lock = 0;
5989 unsigned long haddr = address & huge_page_mask(h);
5992 * Serialize hugepage allocation and instantiation, so that we don't
5993 * get spurious allocation failures if two CPUs race to instantiate
5994 * the same page in the page cache.
5996 mapping = vma->vm_file->f_mapping;
5997 idx = vma_hugecache_offset(h, vma, haddr);
5998 hash = hugetlb_fault_mutex_hash(mapping, idx);
5999 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6002 * Acquire vma lock before calling huge_pte_alloc and hold
6003 * until finished with ptep. This prevents huge_pmd_unshare from
6004 * being called elsewhere and making the ptep no longer valid.
6006 hugetlb_vma_lock_read(vma);
6007 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6009 hugetlb_vma_unlock_read(vma);
6010 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6011 return VM_FAULT_OOM;
6014 entry = huge_ptep_get(ptep);
6015 /* PTE markers should be handled the same way as none pte */
6016 if (huge_pte_none_mostly(entry))
6018 * hugetlb_no_page will drop vma lock and hugetlb fault
6019 * mutex internally, which make us return immediately.
6021 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6027 * entry could be a migration/hwpoison entry at this point, so this
6028 * check prevents the kernel from going below assuming that we have
6029 * an active hugepage in pagecache. This goto expects the 2nd page
6030 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6031 * properly handle it.
6033 if (!pte_present(entry)) {
6034 if (unlikely(is_hugetlb_entry_migration(entry))) {
6036 * Release the hugetlb fault lock now, but retain
6037 * the vma lock, because it is needed to guard the
6038 * huge_pte_lockptr() later in
6039 * migration_entry_wait_huge(). The vma lock will
6040 * be released there.
6042 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6043 migration_entry_wait_huge(vma, ptep);
6045 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6046 ret = VM_FAULT_HWPOISON_LARGE |
6047 VM_FAULT_SET_HINDEX(hstate_index(h));
6052 * If we are going to COW/unshare the mapping later, we examine the
6053 * pending reservations for this page now. This will ensure that any
6054 * allocations necessary to record that reservation occur outside the
6055 * spinlock. Also lookup the pagecache page now as it is used to
6056 * determine if a reservation has been consumed.
6058 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6059 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6060 if (vma_needs_reservation(h, vma, haddr) < 0) {
6064 /* Just decrements count, does not deallocate */
6065 vma_end_reservation(h, vma, haddr);
6067 pagecache_page = find_lock_page(mapping, idx);
6070 ptl = huge_pte_lock(h, mm, ptep);
6072 /* Check for a racing update before calling hugetlb_wp() */
6073 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6076 /* Handle userfault-wp first, before trying to lock more pages */
6077 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6078 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6079 struct vm_fault vmf = {
6082 .real_address = address,
6087 if (pagecache_page) {
6088 unlock_page(pagecache_page);
6089 put_page(pagecache_page);
6091 hugetlb_vma_unlock_read(vma);
6092 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6093 return handle_userfault(&vmf, VM_UFFD_WP);
6097 * hugetlb_wp() requires page locks of pte_page(entry) and
6098 * pagecache_page, so here we need take the former one
6099 * when page != pagecache_page or !pagecache_page.
6101 page = pte_page(entry);
6102 if (page != pagecache_page)
6103 if (!trylock_page(page)) {
6110 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6111 if (!huge_pte_write(entry)) {
6112 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6113 pagecache_page, ptl);
6115 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6116 entry = huge_pte_mkdirty(entry);
6119 entry = pte_mkyoung(entry);
6120 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6121 flags & FAULT_FLAG_WRITE))
6122 update_mmu_cache(vma, haddr, ptep);
6124 if (page != pagecache_page)
6130 if (pagecache_page) {
6131 unlock_page(pagecache_page);
6132 put_page(pagecache_page);
6135 hugetlb_vma_unlock_read(vma);
6136 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6138 * Generally it's safe to hold refcount during waiting page lock. But
6139 * here we just wait to defer the next page fault to avoid busy loop and
6140 * the page is not used after unlocked before returning from the current
6141 * page fault. So we are safe from accessing freed page, even if we wait
6142 * here without taking refcount.
6145 wait_on_page_locked(page);
6149 #ifdef CONFIG_USERFAULTFD
6151 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
6152 * modifications for huge pages.
6154 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
6156 struct vm_area_struct *dst_vma,
6157 unsigned long dst_addr,
6158 unsigned long src_addr,
6159 enum mcopy_atomic_mode mode,
6160 struct page **pagep,
6163 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
6164 struct hstate *h = hstate_vma(dst_vma);
6165 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6166 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6168 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6174 bool page_in_pagecache = false;
6178 page = find_lock_page(mapping, idx);
6181 page_in_pagecache = true;
6182 } else if (!*pagep) {
6183 /* If a page already exists, then it's UFFDIO_COPY for
6184 * a non-missing case. Return -EEXIST.
6187 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6192 page = alloc_huge_page(dst_vma, dst_addr, 0);
6198 ret = copy_huge_page_from_user(page,
6199 (const void __user *) src_addr,
6200 pages_per_huge_page(h), false);
6202 /* fallback to copy_from_user outside mmap_lock */
6203 if (unlikely(ret)) {
6205 /* Free the allocated page which may have
6206 * consumed a reservation.
6208 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6211 /* Allocate a temporary page to hold the copied
6214 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
6220 /* Set the outparam pagep and return to the caller to
6221 * copy the contents outside the lock. Don't free the
6228 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6235 page = alloc_huge_page(dst_vma, dst_addr, 0);
6242 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
6243 pages_per_huge_page(h));
6249 * The memory barrier inside __SetPageUptodate makes sure that
6250 * preceding stores to the page contents become visible before
6251 * the set_pte_at() write.
6253 __SetPageUptodate(page);
6255 /* Add shared, newly allocated pages to the page cache. */
6256 if (vm_shared && !is_continue) {
6257 size = i_size_read(mapping->host) >> huge_page_shift(h);
6260 goto out_release_nounlock;
6263 * Serialization between remove_inode_hugepages() and
6264 * hugetlb_add_to_page_cache() below happens through the
6265 * hugetlb_fault_mutex_table that here must be hold by
6268 ret = hugetlb_add_to_page_cache(page, mapping, idx);
6270 goto out_release_nounlock;
6271 page_in_pagecache = true;
6274 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6277 if (PageHWPoison(page))
6278 goto out_release_unlock;
6281 * We allow to overwrite a pte marker: consider when both MISSING|WP
6282 * registered, we firstly wr-protect a none pte which has no page cache
6283 * page backing it, then access the page.
6286 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6287 goto out_release_unlock;
6289 if (page_in_pagecache)
6290 page_dup_file_rmap(page, true);
6292 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6295 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6296 * with wp flag set, don't set pte write bit.
6298 if (wp_copy || (is_continue && !vm_shared))
6301 writable = dst_vma->vm_flags & VM_WRITE;
6303 _dst_pte = make_huge_pte(dst_vma, page, writable);
6305 * Always mark UFFDIO_COPY page dirty; note that this may not be
6306 * extremely important for hugetlbfs for now since swapping is not
6307 * supported, but we should still be clear in that this page cannot be
6308 * thrown away at will, even if write bit not set.
6310 _dst_pte = huge_pte_mkdirty(_dst_pte);
6311 _dst_pte = pte_mkyoung(_dst_pte);
6314 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6316 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6318 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6320 /* No need to invalidate - it was non-present before */
6321 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6325 SetHPageMigratable(page);
6326 if (vm_shared || is_continue)
6333 if (vm_shared || is_continue)
6335 out_release_nounlock:
6336 if (!page_in_pagecache)
6337 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6341 #endif /* CONFIG_USERFAULTFD */
6343 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6344 int refs, struct page **pages,
6345 struct vm_area_struct **vmas)
6349 for (nr = 0; nr < refs; nr++) {
6351 pages[nr] = nth_page(page, nr);
6357 static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma,
6358 unsigned int flags, pte_t *pte,
6361 pte_t pteval = huge_ptep_get(pte);
6364 if (is_swap_pte(pteval))
6366 if (huge_pte_write(pteval))
6368 if (flags & FOLL_WRITE)
6370 if (gup_must_unshare(vma, flags, pte_page(pteval))) {
6377 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6378 unsigned long address, unsigned int flags)
6380 struct hstate *h = hstate_vma(vma);
6381 struct mm_struct *mm = vma->vm_mm;
6382 unsigned long haddr = address & huge_page_mask(h);
6383 struct page *page = NULL;
6388 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
6389 * follow_hugetlb_page().
6391 if (WARN_ON_ONCE(flags & FOLL_PIN))
6394 hugetlb_vma_lock_read(vma);
6395 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6399 ptl = huge_pte_lock(h, mm, pte);
6400 entry = huge_ptep_get(pte);
6401 if (pte_present(entry)) {
6402 page = pte_page(entry) +
6403 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6405 * Note that page may be a sub-page, and with vmemmap
6406 * optimizations the page struct may be read only.
6407 * try_grab_page() will increase the ref count on the
6408 * head page, so this will be OK.
6410 * try_grab_page() should always be able to get the page here,
6411 * because we hold the ptl lock and have verified pte_present().
6413 if (try_grab_page(page, flags)) {
6421 hugetlb_vma_unlock_read(vma);
6425 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6426 struct page **pages, struct vm_area_struct **vmas,
6427 unsigned long *position, unsigned long *nr_pages,
6428 long i, unsigned int flags, int *locked)
6430 unsigned long pfn_offset;
6431 unsigned long vaddr = *position;
6432 unsigned long remainder = *nr_pages;
6433 struct hstate *h = hstate_vma(vma);
6434 int err = -EFAULT, refs;
6436 while (vaddr < vma->vm_end && remainder) {
6438 spinlock_t *ptl = NULL;
6439 bool unshare = false;
6444 * If we have a pending SIGKILL, don't keep faulting pages and
6445 * potentially allocating memory.
6447 if (fatal_signal_pending(current)) {
6452 hugetlb_vma_lock_read(vma);
6454 * Some archs (sparc64, sh*) have multiple pte_ts to
6455 * each hugepage. We have to make sure we get the
6456 * first, for the page indexing below to work.
6458 * Note that page table lock is not held when pte is null.
6460 pte = hugetlb_walk(vma, vaddr & huge_page_mask(h),
6463 ptl = huge_pte_lock(h, mm, pte);
6464 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6467 * When coredumping, it suits get_dump_page if we just return
6468 * an error where there's an empty slot with no huge pagecache
6469 * to back it. This way, we avoid allocating a hugepage, and
6470 * the sparse dumpfile avoids allocating disk blocks, but its
6471 * huge holes still show up with zeroes where they need to be.
6473 if (absent && (flags & FOLL_DUMP) &&
6474 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6477 hugetlb_vma_unlock_read(vma);
6483 * We need call hugetlb_fault for both hugepages under migration
6484 * (in which case hugetlb_fault waits for the migration,) and
6485 * hwpoisoned hugepages (in which case we need to prevent the
6486 * caller from accessing to them.) In order to do this, we use
6487 * here is_swap_pte instead of is_hugetlb_entry_migration and
6488 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6489 * both cases, and because we can't follow correct pages
6490 * directly from any kind of swap entries.
6493 __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) {
6495 unsigned int fault_flags = 0;
6499 hugetlb_vma_unlock_read(vma);
6501 if (flags & FOLL_WRITE)
6502 fault_flags |= FAULT_FLAG_WRITE;
6504 fault_flags |= FAULT_FLAG_UNSHARE;
6506 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6507 FAULT_FLAG_KILLABLE;
6508 if (flags & FOLL_INTERRUPTIBLE)
6509 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
6511 if (flags & FOLL_NOWAIT)
6512 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6513 FAULT_FLAG_RETRY_NOWAIT;
6514 if (flags & FOLL_TRIED) {
6516 * Note: FAULT_FLAG_ALLOW_RETRY and
6517 * FAULT_FLAG_TRIED can co-exist
6519 fault_flags |= FAULT_FLAG_TRIED;
6521 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6522 if (ret & VM_FAULT_ERROR) {
6523 err = vm_fault_to_errno(ret, flags);
6527 if (ret & VM_FAULT_RETRY) {
6529 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6533 * VM_FAULT_RETRY must not return an
6534 * error, it will return zero
6537 * No need to update "position" as the
6538 * caller will not check it after
6539 * *nr_pages is set to 0.
6546 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6547 page = pte_page(huge_ptep_get(pte));
6549 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6550 !PageAnonExclusive(page), page);
6553 * If subpage information not requested, update counters
6554 * and skip the same_page loop below.
6556 if (!pages && !vmas && !pfn_offset &&
6557 (vaddr + huge_page_size(h) < vma->vm_end) &&
6558 (remainder >= pages_per_huge_page(h))) {
6559 vaddr += huge_page_size(h);
6560 remainder -= pages_per_huge_page(h);
6561 i += pages_per_huge_page(h);
6563 hugetlb_vma_unlock_read(vma);
6567 /* vaddr may not be aligned to PAGE_SIZE */
6568 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6569 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6572 record_subpages_vmas(nth_page(page, pfn_offset),
6574 likely(pages) ? pages + i : NULL,
6575 vmas ? vmas + i : NULL);
6579 * try_grab_folio() should always succeed here,
6580 * because: a) we hold the ptl lock, and b) we've just
6581 * checked that the huge page is present in the page
6582 * tables. If the huge page is present, then the tail
6583 * pages must also be present. The ptl prevents the
6584 * head page and tail pages from being rearranged in
6585 * any way. As this is hugetlb, the pages will never
6586 * be p2pdma or not longterm pinable. So this page
6587 * must be available at this point, unless the page
6588 * refcount overflowed:
6590 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6593 hugetlb_vma_unlock_read(vma);
6600 vaddr += (refs << PAGE_SHIFT);
6605 hugetlb_vma_unlock_read(vma);
6607 *nr_pages = remainder;
6609 * setting position is actually required only if remainder is
6610 * not zero but it's faster not to add a "if (remainder)"
6618 long hugetlb_change_protection(struct vm_area_struct *vma,
6619 unsigned long address, unsigned long end,
6620 pgprot_t newprot, unsigned long cp_flags)
6622 struct mm_struct *mm = vma->vm_mm;
6623 unsigned long start = address;
6626 struct hstate *h = hstate_vma(vma);
6627 long pages = 0, psize = huge_page_size(h);
6628 bool shared_pmd = false;
6629 struct mmu_notifier_range range;
6630 unsigned long last_addr_mask;
6631 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6632 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6635 * In the case of shared PMDs, the area to flush could be beyond
6636 * start/end. Set range.start/range.end to cover the maximum possible
6637 * range if PMD sharing is possible.
6639 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6640 0, vma, mm, start, end);
6641 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6643 BUG_ON(address >= end);
6644 flush_cache_range(vma, range.start, range.end);
6646 mmu_notifier_invalidate_range_start(&range);
6647 hugetlb_vma_lock_write(vma);
6648 i_mmap_lock_write(vma->vm_file->f_mapping);
6649 last_addr_mask = hugetlb_mask_last_page(h);
6650 for (; address < end; address += psize) {
6652 ptep = hugetlb_walk(vma, address, psize);
6655 address |= last_addr_mask;
6659 * Userfaultfd wr-protect requires pgtable
6660 * pre-allocations to install pte markers.
6662 ptep = huge_pte_alloc(mm, vma, address, psize);
6666 ptl = huge_pte_lock(h, mm, ptep);
6667 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6669 * When uffd-wp is enabled on the vma, unshare
6670 * shouldn't happen at all. Warn about it if it
6671 * happened due to some reason.
6673 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6677 address |= last_addr_mask;
6680 pte = huge_ptep_get(ptep);
6681 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6682 /* Nothing to do. */
6683 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6684 swp_entry_t entry = pte_to_swp_entry(pte);
6685 struct page *page = pfn_swap_entry_to_page(entry);
6688 if (is_writable_migration_entry(entry)) {
6690 entry = make_readable_exclusive_migration_entry(
6693 entry = make_readable_migration_entry(
6695 newpte = swp_entry_to_pte(entry);
6700 newpte = pte_swp_mkuffd_wp(newpte);
6701 else if (uffd_wp_resolve)
6702 newpte = pte_swp_clear_uffd_wp(newpte);
6703 if (!pte_same(pte, newpte))
6704 set_huge_pte_at(mm, address, ptep, newpte);
6705 } else if (unlikely(is_pte_marker(pte))) {
6706 /* No other markers apply for now. */
6707 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6708 if (uffd_wp_resolve)
6709 /* Safe to modify directly (non-present->none). */
6710 huge_pte_clear(mm, address, ptep, psize);
6711 } else if (!huge_pte_none(pte)) {
6713 unsigned int shift = huge_page_shift(hstate_vma(vma));
6715 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6716 pte = huge_pte_modify(old_pte, newprot);
6717 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6719 pte = huge_pte_mkuffd_wp(pte);
6720 else if (uffd_wp_resolve)
6721 pte = huge_pte_clear_uffd_wp(pte);
6722 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6726 if (unlikely(uffd_wp))
6727 /* Safe to modify directly (none->non-present). */
6728 set_huge_pte_at(mm, address, ptep,
6729 make_pte_marker(PTE_MARKER_UFFD_WP));
6734 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6735 * may have cleared our pud entry and done put_page on the page table:
6736 * once we release i_mmap_rwsem, another task can do the final put_page
6737 * and that page table be reused and filled with junk. If we actually
6738 * did unshare a page of pmds, flush the range corresponding to the pud.
6741 flush_hugetlb_tlb_range(vma, range.start, range.end);
6743 flush_hugetlb_tlb_range(vma, start, end);
6745 * No need to call mmu_notifier_invalidate_range() we are downgrading
6746 * page table protection not changing it to point to a new page.
6748 * See Documentation/mm/mmu_notifier.rst
6750 i_mmap_unlock_write(vma->vm_file->f_mapping);
6751 hugetlb_vma_unlock_write(vma);
6752 mmu_notifier_invalidate_range_end(&range);
6754 return pages << h->order;
6757 /* Return true if reservation was successful, false otherwise. */
6758 bool hugetlb_reserve_pages(struct inode *inode,
6760 struct vm_area_struct *vma,
6761 vm_flags_t vm_flags)
6763 long chg = -1, add = -1;
6764 struct hstate *h = hstate_inode(inode);
6765 struct hugepage_subpool *spool = subpool_inode(inode);
6766 struct resv_map *resv_map;
6767 struct hugetlb_cgroup *h_cg = NULL;
6768 long gbl_reserve, regions_needed = 0;
6770 /* This should never happen */
6772 VM_WARN(1, "%s called with a negative range\n", __func__);
6777 * vma specific semaphore used for pmd sharing and fault/truncation
6780 hugetlb_vma_lock_alloc(vma);
6783 * Only apply hugepage reservation if asked. At fault time, an
6784 * attempt will be made for VM_NORESERVE to allocate a page
6785 * without using reserves
6787 if (vm_flags & VM_NORESERVE)
6791 * Shared mappings base their reservation on the number of pages that
6792 * are already allocated on behalf of the file. Private mappings need
6793 * to reserve the full area even if read-only as mprotect() may be
6794 * called to make the mapping read-write. Assume !vma is a shm mapping
6796 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6798 * resv_map can not be NULL as hugetlb_reserve_pages is only
6799 * called for inodes for which resv_maps were created (see
6800 * hugetlbfs_get_inode).
6802 resv_map = inode_resv_map(inode);
6804 chg = region_chg(resv_map, from, to, ®ions_needed);
6806 /* Private mapping. */
6807 resv_map = resv_map_alloc();
6813 set_vma_resv_map(vma, resv_map);
6814 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6820 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6821 chg * pages_per_huge_page(h), &h_cg) < 0)
6824 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6825 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6828 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6832 * There must be enough pages in the subpool for the mapping. If
6833 * the subpool has a minimum size, there may be some global
6834 * reservations already in place (gbl_reserve).
6836 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6837 if (gbl_reserve < 0)
6838 goto out_uncharge_cgroup;
6841 * Check enough hugepages are available for the reservation.
6842 * Hand the pages back to the subpool if there are not
6844 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6848 * Account for the reservations made. Shared mappings record regions
6849 * that have reservations as they are shared by multiple VMAs.
6850 * When the last VMA disappears, the region map says how much
6851 * the reservation was and the page cache tells how much of
6852 * the reservation was consumed. Private mappings are per-VMA and
6853 * only the consumed reservations are tracked. When the VMA
6854 * disappears, the original reservation is the VMA size and the
6855 * consumed reservations are stored in the map. Hence, nothing
6856 * else has to be done for private mappings here
6858 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6859 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6861 if (unlikely(add < 0)) {
6862 hugetlb_acct_memory(h, -gbl_reserve);
6864 } else if (unlikely(chg > add)) {
6866 * pages in this range were added to the reserve
6867 * map between region_chg and region_add. This
6868 * indicates a race with alloc_huge_page. Adjust
6869 * the subpool and reserve counts modified above
6870 * based on the difference.
6875 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6876 * reference to h_cg->css. See comment below for detail.
6878 hugetlb_cgroup_uncharge_cgroup_rsvd(
6880 (chg - add) * pages_per_huge_page(h), h_cg);
6882 rsv_adjust = hugepage_subpool_put_pages(spool,
6884 hugetlb_acct_memory(h, -rsv_adjust);
6887 * The file_regions will hold their own reference to
6888 * h_cg->css. So we should release the reference held
6889 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6892 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6898 /* put back original number of pages, chg */
6899 (void)hugepage_subpool_put_pages(spool, chg);
6900 out_uncharge_cgroup:
6901 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6902 chg * pages_per_huge_page(h), h_cg);
6904 hugetlb_vma_lock_free(vma);
6905 if (!vma || vma->vm_flags & VM_MAYSHARE)
6906 /* Only call region_abort if the region_chg succeeded but the
6907 * region_add failed or didn't run.
6909 if (chg >= 0 && add < 0)
6910 region_abort(resv_map, from, to, regions_needed);
6911 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6912 kref_put(&resv_map->refs, resv_map_release);
6916 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6919 struct hstate *h = hstate_inode(inode);
6920 struct resv_map *resv_map = inode_resv_map(inode);
6922 struct hugepage_subpool *spool = subpool_inode(inode);
6926 * Since this routine can be called in the evict inode path for all
6927 * hugetlbfs inodes, resv_map could be NULL.
6930 chg = region_del(resv_map, start, end);
6932 * region_del() can fail in the rare case where a region
6933 * must be split and another region descriptor can not be
6934 * allocated. If end == LONG_MAX, it will not fail.
6940 spin_lock(&inode->i_lock);
6941 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6942 spin_unlock(&inode->i_lock);
6945 * If the subpool has a minimum size, the number of global
6946 * reservations to be released may be adjusted.
6948 * Note that !resv_map implies freed == 0. So (chg - freed)
6949 * won't go negative.
6951 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6952 hugetlb_acct_memory(h, -gbl_reserve);
6957 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6958 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6959 struct vm_area_struct *vma,
6960 unsigned long addr, pgoff_t idx)
6962 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6964 unsigned long sbase = saddr & PUD_MASK;
6965 unsigned long s_end = sbase + PUD_SIZE;
6967 /* Allow segments to share if only one is marked locked */
6968 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6969 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6972 * match the virtual addresses, permission and the alignment of the
6975 * Also, vma_lock (vm_private_data) is required for sharing.
6977 if (pmd_index(addr) != pmd_index(saddr) ||
6978 vm_flags != svm_flags ||
6979 !range_in_vma(svma, sbase, s_end) ||
6980 !svma->vm_private_data)
6986 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6988 unsigned long start = addr & PUD_MASK;
6989 unsigned long end = start + PUD_SIZE;
6991 #ifdef CONFIG_USERFAULTFD
6992 if (uffd_disable_huge_pmd_share(vma))
6996 * check on proper vm_flags and page table alignment
6998 if (!(vma->vm_flags & VM_MAYSHARE))
7000 if (!vma->vm_private_data) /* vma lock required for sharing */
7002 if (!range_in_vma(vma, start, end))
7008 * Determine if start,end range within vma could be mapped by shared pmd.
7009 * If yes, adjust start and end to cover range associated with possible
7010 * shared pmd mappings.
7012 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7013 unsigned long *start, unsigned long *end)
7015 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7016 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7019 * vma needs to span at least one aligned PUD size, and the range
7020 * must be at least partially within in.
7022 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7023 (*end <= v_start) || (*start >= v_end))
7026 /* Extend the range to be PUD aligned for a worst case scenario */
7027 if (*start > v_start)
7028 *start = ALIGN_DOWN(*start, PUD_SIZE);
7031 *end = ALIGN(*end, PUD_SIZE);
7035 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7036 * and returns the corresponding pte. While this is not necessary for the
7037 * !shared pmd case because we can allocate the pmd later as well, it makes the
7038 * code much cleaner. pmd allocation is essential for the shared case because
7039 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7040 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7041 * bad pmd for sharing.
7043 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7044 unsigned long addr, pud_t *pud)
7046 struct address_space *mapping = vma->vm_file->f_mapping;
7047 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7049 struct vm_area_struct *svma;
7050 unsigned long saddr;
7055 i_mmap_lock_read(mapping);
7056 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7060 saddr = page_table_shareable(svma, vma, addr, idx);
7062 spte = hugetlb_walk(svma, saddr,
7063 vma_mmu_pagesize(svma));
7065 get_page(virt_to_page(spte));
7074 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
7075 if (pud_none(*pud)) {
7076 pud_populate(mm, pud,
7077 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7080 put_page(virt_to_page(spte));
7084 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7085 i_mmap_unlock_read(mapping);
7090 * unmap huge page backed by shared pte.
7092 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7093 * indicated by page_count > 1, unmap is achieved by clearing pud and
7094 * decrementing the ref count. If count == 1, the pte page is not shared.
7096 * Called with page table lock held.
7098 * returns: 1 successfully unmapped a shared pte page
7099 * 0 the underlying pte page is not shared, or it is the last user
7101 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7102 unsigned long addr, pte_t *ptep)
7104 pgd_t *pgd = pgd_offset(mm, addr);
7105 p4d_t *p4d = p4d_offset(pgd, addr);
7106 pud_t *pud = pud_offset(p4d, addr);
7108 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7109 hugetlb_vma_assert_locked(vma);
7110 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7111 if (page_count(virt_to_page(ptep)) == 1)
7115 put_page(virt_to_page(ptep));
7120 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7122 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7123 unsigned long addr, pud_t *pud)
7128 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7129 unsigned long addr, pte_t *ptep)
7134 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7135 unsigned long *start, unsigned long *end)
7139 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7143 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7145 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7146 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7147 unsigned long addr, unsigned long sz)
7154 pgd = pgd_offset(mm, addr);
7155 p4d = p4d_alloc(mm, pgd, addr);
7158 pud = pud_alloc(mm, p4d, addr);
7160 if (sz == PUD_SIZE) {
7163 BUG_ON(sz != PMD_SIZE);
7164 if (want_pmd_share(vma, addr) && pud_none(*pud))
7165 pte = huge_pmd_share(mm, vma, addr, pud);
7167 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7170 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
7176 * huge_pte_offset() - Walk the page table to resolve the hugepage
7177 * entry at address @addr
7179 * Return: Pointer to page table entry (PUD or PMD) for
7180 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7181 * size @sz doesn't match the hugepage size at this level of the page
7184 pte_t *huge_pte_offset(struct mm_struct *mm,
7185 unsigned long addr, unsigned long sz)
7192 pgd = pgd_offset(mm, addr);
7193 if (!pgd_present(*pgd))
7195 p4d = p4d_offset(pgd, addr);
7196 if (!p4d_present(*p4d))
7199 pud = pud_offset(p4d, addr);
7201 /* must be pud huge, non-present or none */
7202 return (pte_t *)pud;
7203 if (!pud_present(*pud))
7205 /* must have a valid entry and size to go further */
7207 pmd = pmd_offset(pud, addr);
7208 /* must be pmd huge, non-present or none */
7209 return (pte_t *)pmd;
7213 * Return a mask that can be used to update an address to the last huge
7214 * page in a page table page mapping size. Used to skip non-present
7215 * page table entries when linearly scanning address ranges. Architectures
7216 * with unique huge page to page table relationships can define their own
7217 * version of this routine.
7219 unsigned long hugetlb_mask_last_page(struct hstate *h)
7221 unsigned long hp_size = huge_page_size(h);
7223 if (hp_size == PUD_SIZE)
7224 return P4D_SIZE - PUD_SIZE;
7225 else if (hp_size == PMD_SIZE)
7226 return PUD_SIZE - PMD_SIZE;
7233 /* See description above. Architectures can provide their own version. */
7234 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7236 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7237 if (huge_page_size(h) == PMD_SIZE)
7238 return PUD_SIZE - PMD_SIZE;
7243 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7246 * These functions are overwritable if your architecture needs its own
7249 int isolate_hugetlb(struct page *page, struct list_head *list)
7253 spin_lock_irq(&hugetlb_lock);
7254 if (!PageHeadHuge(page) ||
7255 !HPageMigratable(page) ||
7256 !get_page_unless_zero(page)) {
7260 ClearHPageMigratable(page);
7261 list_move_tail(&page->lru, list);
7263 spin_unlock_irq(&hugetlb_lock);
7267 int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison)
7272 spin_lock_irq(&hugetlb_lock);
7273 if (PageHeadHuge(page)) {
7275 if (HPageFreed(page))
7277 else if (HPageMigratable(page) || unpoison)
7278 ret = get_page_unless_zero(page);
7282 spin_unlock_irq(&hugetlb_lock);
7286 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7287 bool *migratable_cleared)
7291 spin_lock_irq(&hugetlb_lock);
7292 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7293 spin_unlock_irq(&hugetlb_lock);
7297 void putback_active_hugepage(struct page *page)
7299 spin_lock_irq(&hugetlb_lock);
7300 SetHPageMigratable(page);
7301 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7302 spin_unlock_irq(&hugetlb_lock);
7306 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7308 struct hstate *h = folio_hstate(old_folio);
7310 hugetlb_cgroup_migrate(old_folio, new_folio);
7311 set_page_owner_migrate_reason(&new_folio->page, reason);
7314 * transfer temporary state of the new hugetlb folio. This is
7315 * reverse to other transitions because the newpage is going to
7316 * be final while the old one will be freed so it takes over
7317 * the temporary status.
7319 * Also note that we have to transfer the per-node surplus state
7320 * here as well otherwise the global surplus count will not match
7323 if (folio_test_hugetlb_temporary(new_folio)) {
7324 int old_nid = folio_nid(old_folio);
7325 int new_nid = folio_nid(new_folio);
7327 folio_set_hugetlb_temporary(old_folio);
7328 folio_clear_hugetlb_temporary(new_folio);
7332 * There is no need to transfer the per-node surplus state
7333 * when we do not cross the node.
7335 if (new_nid == old_nid)
7337 spin_lock_irq(&hugetlb_lock);
7338 if (h->surplus_huge_pages_node[old_nid]) {
7339 h->surplus_huge_pages_node[old_nid]--;
7340 h->surplus_huge_pages_node[new_nid]++;
7342 spin_unlock_irq(&hugetlb_lock);
7346 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7347 unsigned long start,
7350 struct hstate *h = hstate_vma(vma);
7351 unsigned long sz = huge_page_size(h);
7352 struct mm_struct *mm = vma->vm_mm;
7353 struct mmu_notifier_range range;
7354 unsigned long address;
7358 if (!(vma->vm_flags & VM_MAYSHARE))
7364 flush_cache_range(vma, start, end);
7366 * No need to call adjust_range_if_pmd_sharing_possible(), because
7367 * we have already done the PUD_SIZE alignment.
7369 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7371 mmu_notifier_invalidate_range_start(&range);
7372 hugetlb_vma_lock_write(vma);
7373 i_mmap_lock_write(vma->vm_file->f_mapping);
7374 for (address = start; address < end; address += PUD_SIZE) {
7375 ptep = hugetlb_walk(vma, address, sz);
7378 ptl = huge_pte_lock(h, mm, ptep);
7379 huge_pmd_unshare(mm, vma, address, ptep);
7382 flush_hugetlb_tlb_range(vma, start, end);
7383 i_mmap_unlock_write(vma->vm_file->f_mapping);
7384 hugetlb_vma_unlock_write(vma);
7386 * No need to call mmu_notifier_invalidate_range(), see
7387 * Documentation/mm/mmu_notifier.rst.
7389 mmu_notifier_invalidate_range_end(&range);
7393 * This function will unconditionally remove all the shared pmd pgtable entries
7394 * within the specific vma for a hugetlbfs memory range.
7396 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7398 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7399 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7403 static bool cma_reserve_called __initdata;
7405 static int __init cmdline_parse_hugetlb_cma(char *p)
7412 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7415 if (s[count] == ':') {
7416 if (tmp >= MAX_NUMNODES)
7418 nid = array_index_nospec(tmp, MAX_NUMNODES);
7421 tmp = memparse(s, &s);
7422 hugetlb_cma_size_in_node[nid] = tmp;
7423 hugetlb_cma_size += tmp;
7426 * Skip the separator if have one, otherwise
7427 * break the parsing.
7434 hugetlb_cma_size = memparse(p, &p);
7442 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7444 void __init hugetlb_cma_reserve(int order)
7446 unsigned long size, reserved, per_node;
7447 bool node_specific_cma_alloc = false;
7450 cma_reserve_called = true;
7452 if (!hugetlb_cma_size)
7455 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7456 if (hugetlb_cma_size_in_node[nid] == 0)
7459 if (!node_online(nid)) {
7460 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7461 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7462 hugetlb_cma_size_in_node[nid] = 0;
7466 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7467 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7468 nid, (PAGE_SIZE << order) / SZ_1M);
7469 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7470 hugetlb_cma_size_in_node[nid] = 0;
7472 node_specific_cma_alloc = true;
7476 /* Validate the CMA size again in case some invalid nodes specified. */
7477 if (!hugetlb_cma_size)
7480 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7481 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7482 (PAGE_SIZE << order) / SZ_1M);
7483 hugetlb_cma_size = 0;
7487 if (!node_specific_cma_alloc) {
7489 * If 3 GB area is requested on a machine with 4 numa nodes,
7490 * let's allocate 1 GB on first three nodes and ignore the last one.
7492 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7493 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7494 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7498 for_each_online_node(nid) {
7500 char name[CMA_MAX_NAME];
7502 if (node_specific_cma_alloc) {
7503 if (hugetlb_cma_size_in_node[nid] == 0)
7506 size = hugetlb_cma_size_in_node[nid];
7508 size = min(per_node, hugetlb_cma_size - reserved);
7511 size = round_up(size, PAGE_SIZE << order);
7513 snprintf(name, sizeof(name), "hugetlb%d", nid);
7515 * Note that 'order per bit' is based on smallest size that
7516 * may be returned to CMA allocator in the case of
7517 * huge page demotion.
7519 res = cma_declare_contiguous_nid(0, size, 0,
7520 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7522 &hugetlb_cma[nid], nid);
7524 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7530 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7533 if (reserved >= hugetlb_cma_size)
7539 * hugetlb_cma_size is used to determine if allocations from
7540 * cma are possible. Set to zero if no cma regions are set up.
7542 hugetlb_cma_size = 0;
7545 static void __init hugetlb_cma_check(void)
7547 if (!hugetlb_cma_size || cma_reserve_called)
7550 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7553 #endif /* CONFIG_CMA */