1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
40 #include <asm/pgalloc.h>
44 #include <linux/hugetlb.h>
45 #include <linux/hugetlb_cgroup.h>
46 #include <linux/node.h>
47 #include <linux/page_owner.h>
49 #include "hugetlb_vmemmap.h"
51 int hugetlb_max_hstate __read_mostly;
52 unsigned int default_hstate_idx;
53 struct hstate hstates[HUGE_MAX_HSTATE];
56 static struct cma *hugetlb_cma[MAX_NUMNODES];
57 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
58 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
60 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
64 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
69 static unsigned long hugetlb_cma_size __initdata;
71 __initdata LIST_HEAD(huge_boot_pages);
73 /* for command line parsing */
74 static struct hstate * __initdata parsed_hstate;
75 static unsigned long __initdata default_hstate_max_huge_pages;
76 static bool __initdata parsed_valid_hugepagesz = true;
77 static bool __initdata parsed_default_hugepagesz;
78 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
81 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
82 * free_huge_pages, and surplus_huge_pages.
84 DEFINE_SPINLOCK(hugetlb_lock);
87 * Serializes faults on the same logical page. This is used to
88 * prevent spurious OOMs when the hugepage pool is fully utilized.
90 static int num_fault_mutexes;
91 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
93 /* Forward declaration */
94 static int hugetlb_acct_memory(struct hstate *h, long delta);
95 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
96 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
97 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
98 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
99 unsigned long start, unsigned long end);
100 static struct resv_map *vma_resv_map(struct vm_area_struct *vma);
102 static inline bool subpool_is_free(struct hugepage_subpool *spool)
106 if (spool->max_hpages != -1)
107 return spool->used_hpages == 0;
108 if (spool->min_hpages != -1)
109 return spool->rsv_hpages == spool->min_hpages;
114 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
115 unsigned long irq_flags)
117 spin_unlock_irqrestore(&spool->lock, irq_flags);
119 /* If no pages are used, and no other handles to the subpool
120 * remain, give up any reservations based on minimum size and
121 * free the subpool */
122 if (subpool_is_free(spool)) {
123 if (spool->min_hpages != -1)
124 hugetlb_acct_memory(spool->hstate,
130 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
133 struct hugepage_subpool *spool;
135 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
139 spin_lock_init(&spool->lock);
141 spool->max_hpages = max_hpages;
143 spool->min_hpages = min_hpages;
145 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
149 spool->rsv_hpages = min_hpages;
154 void hugepage_put_subpool(struct hugepage_subpool *spool)
158 spin_lock_irqsave(&spool->lock, flags);
159 BUG_ON(!spool->count);
161 unlock_or_release_subpool(spool, flags);
165 * Subpool accounting for allocating and reserving pages.
166 * Return -ENOMEM if there are not enough resources to satisfy the
167 * request. Otherwise, return the number of pages by which the
168 * global pools must be adjusted (upward). The returned value may
169 * only be different than the passed value (delta) in the case where
170 * a subpool minimum size must be maintained.
172 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
180 spin_lock_irq(&spool->lock);
182 if (spool->max_hpages != -1) { /* maximum size accounting */
183 if ((spool->used_hpages + delta) <= spool->max_hpages)
184 spool->used_hpages += delta;
191 /* minimum size accounting */
192 if (spool->min_hpages != -1 && spool->rsv_hpages) {
193 if (delta > spool->rsv_hpages) {
195 * Asking for more reserves than those already taken on
196 * behalf of subpool. Return difference.
198 ret = delta - spool->rsv_hpages;
199 spool->rsv_hpages = 0;
201 ret = 0; /* reserves already accounted for */
202 spool->rsv_hpages -= delta;
207 spin_unlock_irq(&spool->lock);
212 * Subpool accounting for freeing and unreserving pages.
213 * Return the number of global page reservations that must be dropped.
214 * The return value may only be different than the passed value (delta)
215 * in the case where a subpool minimum size must be maintained.
217 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
226 spin_lock_irqsave(&spool->lock, flags);
228 if (spool->max_hpages != -1) /* maximum size accounting */
229 spool->used_hpages -= delta;
231 /* minimum size accounting */
232 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
233 if (spool->rsv_hpages + delta <= spool->min_hpages)
236 ret = spool->rsv_hpages + delta - spool->min_hpages;
238 spool->rsv_hpages += delta;
239 if (spool->rsv_hpages > spool->min_hpages)
240 spool->rsv_hpages = spool->min_hpages;
244 * If hugetlbfs_put_super couldn't free spool due to an outstanding
245 * quota reference, free it now.
247 unlock_or_release_subpool(spool, flags);
252 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
254 return HUGETLBFS_SB(inode->i_sb)->spool;
257 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
259 return subpool_inode(file_inode(vma->vm_file));
263 * hugetlb vma_lock helper routines
265 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
267 if (__vma_shareable_lock(vma)) {
268 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
270 down_read(&vma_lock->rw_sema);
271 } else if (__vma_private_lock(vma)) {
272 struct resv_map *resv_map = vma_resv_map(vma);
274 down_read(&resv_map->rw_sema);
278 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
280 if (__vma_shareable_lock(vma)) {
281 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
283 up_read(&vma_lock->rw_sema);
284 } else if (__vma_private_lock(vma)) {
285 struct resv_map *resv_map = vma_resv_map(vma);
287 up_read(&resv_map->rw_sema);
291 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
293 if (__vma_shareable_lock(vma)) {
294 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
296 down_write(&vma_lock->rw_sema);
297 } else if (__vma_private_lock(vma)) {
298 struct resv_map *resv_map = vma_resv_map(vma);
300 down_write(&resv_map->rw_sema);
304 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
306 if (__vma_shareable_lock(vma)) {
307 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
309 up_write(&vma_lock->rw_sema);
310 } else if (__vma_private_lock(vma)) {
311 struct resv_map *resv_map = vma_resv_map(vma);
313 up_write(&resv_map->rw_sema);
317 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
320 if (__vma_shareable_lock(vma)) {
321 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
323 return down_write_trylock(&vma_lock->rw_sema);
324 } else if (__vma_private_lock(vma)) {
325 struct resv_map *resv_map = vma_resv_map(vma);
327 return down_write_trylock(&resv_map->rw_sema);
333 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
335 if (__vma_shareable_lock(vma)) {
336 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
338 lockdep_assert_held(&vma_lock->rw_sema);
339 } else if (__vma_private_lock(vma)) {
340 struct resv_map *resv_map = vma_resv_map(vma);
342 lockdep_assert_held(&resv_map->rw_sema);
346 void hugetlb_vma_lock_release(struct kref *kref)
348 struct hugetlb_vma_lock *vma_lock = container_of(kref,
349 struct hugetlb_vma_lock, refs);
354 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
356 struct vm_area_struct *vma = vma_lock->vma;
359 * vma_lock structure may or not be released as a result of put,
360 * it certainly will no longer be attached to vma so clear pointer.
361 * Semaphore synchronizes access to vma_lock->vma field.
363 vma_lock->vma = NULL;
364 vma->vm_private_data = NULL;
365 up_write(&vma_lock->rw_sema);
366 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
369 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
371 if (__vma_shareable_lock(vma)) {
372 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
374 __hugetlb_vma_unlock_write_put(vma_lock);
375 } else if (__vma_private_lock(vma)) {
376 struct resv_map *resv_map = vma_resv_map(vma);
378 /* no free for anon vmas, but still need to unlock */
379 up_write(&resv_map->rw_sema);
383 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
386 * Only present in sharable vmas.
388 if (!vma || !__vma_shareable_lock(vma))
391 if (vma->vm_private_data) {
392 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
394 down_write(&vma_lock->rw_sema);
395 __hugetlb_vma_unlock_write_put(vma_lock);
399 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
401 struct hugetlb_vma_lock *vma_lock;
403 /* Only establish in (flags) sharable vmas */
404 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
407 /* Should never get here with non-NULL vm_private_data */
408 if (vma->vm_private_data)
411 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
414 * If we can not allocate structure, then vma can not
415 * participate in pmd sharing. This is only a possible
416 * performance enhancement and memory saving issue.
417 * However, the lock is also used to synchronize page
418 * faults with truncation. If the lock is not present,
419 * unlikely races could leave pages in a file past i_size
420 * until the file is removed. Warn in the unlikely case of
421 * allocation failure.
423 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
427 kref_init(&vma_lock->refs);
428 init_rwsem(&vma_lock->rw_sema);
430 vma->vm_private_data = vma_lock;
433 /* Helper that removes a struct file_region from the resv_map cache and returns
436 static struct file_region *
437 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
439 struct file_region *nrg;
441 VM_BUG_ON(resv->region_cache_count <= 0);
443 resv->region_cache_count--;
444 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
445 list_del(&nrg->link);
453 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
454 struct file_region *rg)
456 #ifdef CONFIG_CGROUP_HUGETLB
457 nrg->reservation_counter = rg->reservation_counter;
464 /* Helper that records hugetlb_cgroup uncharge info. */
465 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
467 struct resv_map *resv,
468 struct file_region *nrg)
470 #ifdef CONFIG_CGROUP_HUGETLB
472 nrg->reservation_counter =
473 &h_cg->rsvd_hugepage[hstate_index(h)];
474 nrg->css = &h_cg->css;
476 * The caller will hold exactly one h_cg->css reference for the
477 * whole contiguous reservation region. But this area might be
478 * scattered when there are already some file_regions reside in
479 * it. As a result, many file_regions may share only one css
480 * reference. In order to ensure that one file_region must hold
481 * exactly one h_cg->css reference, we should do css_get for
482 * each file_region and leave the reference held by caller
486 if (!resv->pages_per_hpage)
487 resv->pages_per_hpage = pages_per_huge_page(h);
488 /* pages_per_hpage should be the same for all entries in
491 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
493 nrg->reservation_counter = NULL;
499 static void put_uncharge_info(struct file_region *rg)
501 #ifdef CONFIG_CGROUP_HUGETLB
507 static bool has_same_uncharge_info(struct file_region *rg,
508 struct file_region *org)
510 #ifdef CONFIG_CGROUP_HUGETLB
511 return rg->reservation_counter == org->reservation_counter &&
519 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
521 struct file_region *nrg, *prg;
523 prg = list_prev_entry(rg, link);
524 if (&prg->link != &resv->regions && prg->to == rg->from &&
525 has_same_uncharge_info(prg, rg)) {
529 put_uncharge_info(rg);
535 nrg = list_next_entry(rg, link);
536 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
537 has_same_uncharge_info(nrg, rg)) {
538 nrg->from = rg->from;
541 put_uncharge_info(rg);
547 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
548 long to, struct hstate *h, struct hugetlb_cgroup *cg,
549 long *regions_needed)
551 struct file_region *nrg;
553 if (!regions_needed) {
554 nrg = get_file_region_entry_from_cache(map, from, to);
555 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
556 list_add(&nrg->link, rg);
557 coalesce_file_region(map, nrg);
559 *regions_needed += 1;
565 * Must be called with resv->lock held.
567 * Calling this with regions_needed != NULL will count the number of pages
568 * to be added but will not modify the linked list. And regions_needed will
569 * indicate the number of file_regions needed in the cache to carry out to add
570 * the regions for this range.
572 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
573 struct hugetlb_cgroup *h_cg,
574 struct hstate *h, long *regions_needed)
577 struct list_head *head = &resv->regions;
578 long last_accounted_offset = f;
579 struct file_region *iter, *trg = NULL;
580 struct list_head *rg = NULL;
585 /* In this loop, we essentially handle an entry for the range
586 * [last_accounted_offset, iter->from), at every iteration, with some
589 list_for_each_entry_safe(iter, trg, head, link) {
590 /* Skip irrelevant regions that start before our range. */
591 if (iter->from < f) {
592 /* If this region ends after the last accounted offset,
593 * then we need to update last_accounted_offset.
595 if (iter->to > last_accounted_offset)
596 last_accounted_offset = iter->to;
600 /* When we find a region that starts beyond our range, we've
603 if (iter->from >= t) {
604 rg = iter->link.prev;
608 /* Add an entry for last_accounted_offset -> iter->from, and
609 * update last_accounted_offset.
611 if (iter->from > last_accounted_offset)
612 add += hugetlb_resv_map_add(resv, iter->link.prev,
613 last_accounted_offset,
617 last_accounted_offset = iter->to;
620 /* Handle the case where our range extends beyond
621 * last_accounted_offset.
625 if (last_accounted_offset < t)
626 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
627 t, h, h_cg, regions_needed);
632 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
634 static int allocate_file_region_entries(struct resv_map *resv,
636 __must_hold(&resv->lock)
638 LIST_HEAD(allocated_regions);
639 int to_allocate = 0, i = 0;
640 struct file_region *trg = NULL, *rg = NULL;
642 VM_BUG_ON(regions_needed < 0);
645 * Check for sufficient descriptors in the cache to accommodate
646 * the number of in progress add operations plus regions_needed.
648 * This is a while loop because when we drop the lock, some other call
649 * to region_add or region_del may have consumed some region_entries,
650 * so we keep looping here until we finally have enough entries for
651 * (adds_in_progress + regions_needed).
653 while (resv->region_cache_count <
654 (resv->adds_in_progress + regions_needed)) {
655 to_allocate = resv->adds_in_progress + regions_needed -
656 resv->region_cache_count;
658 /* At this point, we should have enough entries in the cache
659 * for all the existing adds_in_progress. We should only be
660 * needing to allocate for regions_needed.
662 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
664 spin_unlock(&resv->lock);
665 for (i = 0; i < to_allocate; i++) {
666 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
669 list_add(&trg->link, &allocated_regions);
672 spin_lock(&resv->lock);
674 list_splice(&allocated_regions, &resv->region_cache);
675 resv->region_cache_count += to_allocate;
681 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
689 * Add the huge page range represented by [f, t) to the reserve
690 * map. Regions will be taken from the cache to fill in this range.
691 * Sufficient regions should exist in the cache due to the previous
692 * call to region_chg with the same range, but in some cases the cache will not
693 * have sufficient entries due to races with other code doing region_add or
694 * region_del. The extra needed entries will be allocated.
696 * regions_needed is the out value provided by a previous call to region_chg.
698 * Return the number of new huge pages added to the map. This number is greater
699 * than or equal to zero. If file_region entries needed to be allocated for
700 * this operation and we were not able to allocate, it returns -ENOMEM.
701 * region_add of regions of length 1 never allocate file_regions and cannot
702 * fail; region_chg will always allocate at least 1 entry and a region_add for
703 * 1 page will only require at most 1 entry.
705 static long region_add(struct resv_map *resv, long f, long t,
706 long in_regions_needed, struct hstate *h,
707 struct hugetlb_cgroup *h_cg)
709 long add = 0, actual_regions_needed = 0;
711 spin_lock(&resv->lock);
714 /* Count how many regions are actually needed to execute this add. */
715 add_reservation_in_range(resv, f, t, NULL, NULL,
716 &actual_regions_needed);
719 * Check for sufficient descriptors in the cache to accommodate
720 * this add operation. Note that actual_regions_needed may be greater
721 * than in_regions_needed, as the resv_map may have been modified since
722 * the region_chg call. In this case, we need to make sure that we
723 * allocate extra entries, such that we have enough for all the
724 * existing adds_in_progress, plus the excess needed for this
727 if (actual_regions_needed > in_regions_needed &&
728 resv->region_cache_count <
729 resv->adds_in_progress +
730 (actual_regions_needed - in_regions_needed)) {
731 /* region_add operation of range 1 should never need to
732 * allocate file_region entries.
734 VM_BUG_ON(t - f <= 1);
736 if (allocate_file_region_entries(
737 resv, actual_regions_needed - in_regions_needed)) {
744 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
746 resv->adds_in_progress -= in_regions_needed;
748 spin_unlock(&resv->lock);
753 * Examine the existing reserve map and determine how many
754 * huge pages in the specified range [f, t) are NOT currently
755 * represented. This routine is called before a subsequent
756 * call to region_add that will actually modify the reserve
757 * map to add the specified range [f, t). region_chg does
758 * not change the number of huge pages represented by the
759 * map. A number of new file_region structures is added to the cache as a
760 * placeholder, for the subsequent region_add call to use. At least 1
761 * file_region structure is added.
763 * out_regions_needed is the number of regions added to the
764 * resv->adds_in_progress. This value needs to be provided to a follow up call
765 * to region_add or region_abort for proper accounting.
767 * Returns the number of huge pages that need to be added to the existing
768 * reservation map for the range [f, t). This number is greater or equal to
769 * zero. -ENOMEM is returned if a new file_region structure or cache entry
770 * is needed and can not be allocated.
772 static long region_chg(struct resv_map *resv, long f, long t,
773 long *out_regions_needed)
777 spin_lock(&resv->lock);
779 /* Count how many hugepages in this range are NOT represented. */
780 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
783 if (*out_regions_needed == 0)
784 *out_regions_needed = 1;
786 if (allocate_file_region_entries(resv, *out_regions_needed))
789 resv->adds_in_progress += *out_regions_needed;
791 spin_unlock(&resv->lock);
796 * Abort the in progress add operation. The adds_in_progress field
797 * of the resv_map keeps track of the operations in progress between
798 * calls to region_chg and region_add. Operations are sometimes
799 * aborted after the call to region_chg. In such cases, region_abort
800 * is called to decrement the adds_in_progress counter. regions_needed
801 * is the value returned by the region_chg call, it is used to decrement
802 * the adds_in_progress counter.
804 * NOTE: The range arguments [f, t) are not needed or used in this
805 * routine. They are kept to make reading the calling code easier as
806 * arguments will match the associated region_chg call.
808 static void region_abort(struct resv_map *resv, long f, long t,
811 spin_lock(&resv->lock);
812 VM_BUG_ON(!resv->region_cache_count);
813 resv->adds_in_progress -= regions_needed;
814 spin_unlock(&resv->lock);
818 * Delete the specified range [f, t) from the reserve map. If the
819 * t parameter is LONG_MAX, this indicates that ALL regions after f
820 * should be deleted. Locate the regions which intersect [f, t)
821 * and either trim, delete or split the existing regions.
823 * Returns the number of huge pages deleted from the reserve map.
824 * In the normal case, the return value is zero or more. In the
825 * case where a region must be split, a new region descriptor must
826 * be allocated. If the allocation fails, -ENOMEM will be returned.
827 * NOTE: If the parameter t == LONG_MAX, then we will never split
828 * a region and possibly return -ENOMEM. Callers specifying
829 * t == LONG_MAX do not need to check for -ENOMEM error.
831 static long region_del(struct resv_map *resv, long f, long t)
833 struct list_head *head = &resv->regions;
834 struct file_region *rg, *trg;
835 struct file_region *nrg = NULL;
839 spin_lock(&resv->lock);
840 list_for_each_entry_safe(rg, trg, head, link) {
842 * Skip regions before the range to be deleted. file_region
843 * ranges are normally of the form [from, to). However, there
844 * may be a "placeholder" entry in the map which is of the form
845 * (from, to) with from == to. Check for placeholder entries
846 * at the beginning of the range to be deleted.
848 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
854 if (f > rg->from && t < rg->to) { /* Must split region */
856 * Check for an entry in the cache before dropping
857 * lock and attempting allocation.
860 resv->region_cache_count > resv->adds_in_progress) {
861 nrg = list_first_entry(&resv->region_cache,
864 list_del(&nrg->link);
865 resv->region_cache_count--;
869 spin_unlock(&resv->lock);
870 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
877 hugetlb_cgroup_uncharge_file_region(
878 resv, rg, t - f, false);
880 /* New entry for end of split region */
884 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
886 INIT_LIST_HEAD(&nrg->link);
888 /* Original entry is trimmed */
891 list_add(&nrg->link, &rg->link);
896 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
897 del += rg->to - rg->from;
898 hugetlb_cgroup_uncharge_file_region(resv, rg,
899 rg->to - rg->from, true);
905 if (f <= rg->from) { /* Trim beginning of region */
906 hugetlb_cgroup_uncharge_file_region(resv, rg,
907 t - rg->from, false);
911 } else { /* Trim end of region */
912 hugetlb_cgroup_uncharge_file_region(resv, rg,
920 spin_unlock(&resv->lock);
926 * A rare out of memory error was encountered which prevented removal of
927 * the reserve map region for a page. The huge page itself was free'ed
928 * and removed from the page cache. This routine will adjust the subpool
929 * usage count, and the global reserve count if needed. By incrementing
930 * these counts, the reserve map entry which could not be deleted will
931 * appear as a "reserved" entry instead of simply dangling with incorrect
934 void hugetlb_fix_reserve_counts(struct inode *inode)
936 struct hugepage_subpool *spool = subpool_inode(inode);
938 bool reserved = false;
940 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
941 if (rsv_adjust > 0) {
942 struct hstate *h = hstate_inode(inode);
944 if (!hugetlb_acct_memory(h, 1))
946 } else if (!rsv_adjust) {
951 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
955 * Count and return the number of huge pages in the reserve map
956 * that intersect with the range [f, t).
958 static long region_count(struct resv_map *resv, long f, long t)
960 struct list_head *head = &resv->regions;
961 struct file_region *rg;
964 spin_lock(&resv->lock);
965 /* Locate each segment we overlap with, and count that overlap. */
966 list_for_each_entry(rg, head, link) {
975 seg_from = max(rg->from, f);
976 seg_to = min(rg->to, t);
978 chg += seg_to - seg_from;
980 spin_unlock(&resv->lock);
986 * Convert the address within this vma to the page offset within
987 * the mapping, huge page units here.
989 static pgoff_t vma_hugecache_offset(struct hstate *h,
990 struct vm_area_struct *vma, unsigned long address)
992 return ((address - vma->vm_start) >> huge_page_shift(h)) +
993 (vma->vm_pgoff >> huge_page_order(h));
997 * vma_kernel_pagesize - Page size granularity for this VMA.
998 * @vma: The user mapping.
1000 * Folios in this VMA will be aligned to, and at least the size of the
1001 * number of bytes returned by this function.
1003 * Return: The default size of the folios allocated when backing a VMA.
1005 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
1007 if (vma->vm_ops && vma->vm_ops->pagesize)
1008 return vma->vm_ops->pagesize(vma);
1011 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
1014 * Return the page size being used by the MMU to back a VMA. In the majority
1015 * of cases, the page size used by the kernel matches the MMU size. On
1016 * architectures where it differs, an architecture-specific 'strong'
1017 * version of this symbol is required.
1019 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
1021 return vma_kernel_pagesize(vma);
1025 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1026 * bits of the reservation map pointer, which are always clear due to
1029 #define HPAGE_RESV_OWNER (1UL << 0)
1030 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1031 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1034 * These helpers are used to track how many pages are reserved for
1035 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1036 * is guaranteed to have their future faults succeed.
1038 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1039 * the reserve counters are updated with the hugetlb_lock held. It is safe
1040 * to reset the VMA at fork() time as it is not in use yet and there is no
1041 * chance of the global counters getting corrupted as a result of the values.
1043 * The private mapping reservation is represented in a subtly different
1044 * manner to a shared mapping. A shared mapping has a region map associated
1045 * with the underlying file, this region map represents the backing file
1046 * pages which have ever had a reservation assigned which this persists even
1047 * after the page is instantiated. A private mapping has a region map
1048 * associated with the original mmap which is attached to all VMAs which
1049 * reference it, this region map represents those offsets which have consumed
1050 * reservation ie. where pages have been instantiated.
1052 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1054 return (unsigned long)vma->vm_private_data;
1057 static void set_vma_private_data(struct vm_area_struct *vma,
1058 unsigned long value)
1060 vma->vm_private_data = (void *)value;
1064 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1065 struct hugetlb_cgroup *h_cg,
1068 #ifdef CONFIG_CGROUP_HUGETLB
1070 resv_map->reservation_counter = NULL;
1071 resv_map->pages_per_hpage = 0;
1072 resv_map->css = NULL;
1074 resv_map->reservation_counter =
1075 &h_cg->rsvd_hugepage[hstate_index(h)];
1076 resv_map->pages_per_hpage = pages_per_huge_page(h);
1077 resv_map->css = &h_cg->css;
1082 struct resv_map *resv_map_alloc(void)
1084 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1085 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1087 if (!resv_map || !rg) {
1093 kref_init(&resv_map->refs);
1094 spin_lock_init(&resv_map->lock);
1095 INIT_LIST_HEAD(&resv_map->regions);
1096 init_rwsem(&resv_map->rw_sema);
1098 resv_map->adds_in_progress = 0;
1100 * Initialize these to 0. On shared mappings, 0's here indicate these
1101 * fields don't do cgroup accounting. On private mappings, these will be
1102 * re-initialized to the proper values, to indicate that hugetlb cgroup
1103 * reservations are to be un-charged from here.
1105 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1107 INIT_LIST_HEAD(&resv_map->region_cache);
1108 list_add(&rg->link, &resv_map->region_cache);
1109 resv_map->region_cache_count = 1;
1114 void resv_map_release(struct kref *ref)
1116 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1117 struct list_head *head = &resv_map->region_cache;
1118 struct file_region *rg, *trg;
1120 /* Clear out any active regions before we release the map. */
1121 region_del(resv_map, 0, LONG_MAX);
1123 /* ... and any entries left in the cache */
1124 list_for_each_entry_safe(rg, trg, head, link) {
1125 list_del(&rg->link);
1129 VM_BUG_ON(resv_map->adds_in_progress);
1134 static inline struct resv_map *inode_resv_map(struct inode *inode)
1137 * At inode evict time, i_mapping may not point to the original
1138 * address space within the inode. This original address space
1139 * contains the pointer to the resv_map. So, always use the
1140 * address space embedded within the inode.
1141 * The VERY common case is inode->mapping == &inode->i_data but,
1142 * this may not be true for device special inodes.
1144 return (struct resv_map *)(&inode->i_data)->i_private_data;
1147 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1149 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1150 if (vma->vm_flags & VM_MAYSHARE) {
1151 struct address_space *mapping = vma->vm_file->f_mapping;
1152 struct inode *inode = mapping->host;
1154 return inode_resv_map(inode);
1157 return (struct resv_map *)(get_vma_private_data(vma) &
1162 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1164 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1165 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1167 set_vma_private_data(vma, (unsigned long)map);
1170 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1172 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1173 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1175 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1178 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1180 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1182 return (get_vma_private_data(vma) & flag) != 0;
1185 bool __vma_private_lock(struct vm_area_struct *vma)
1187 return !(vma->vm_flags & VM_MAYSHARE) &&
1188 get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&
1189 is_vma_resv_set(vma, HPAGE_RESV_OWNER);
1192 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1194 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1196 * Clear vm_private_data
1197 * - For shared mappings this is a per-vma semaphore that may be
1198 * allocated in a subsequent call to hugetlb_vm_op_open.
1199 * Before clearing, make sure pointer is not associated with vma
1200 * as this will leak the structure. This is the case when called
1201 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1202 * been called to allocate a new structure.
1203 * - For MAP_PRIVATE mappings, this is the reserve map which does
1204 * not apply to children. Faults generated by the children are
1205 * not guaranteed to succeed, even if read-only.
1207 if (vma->vm_flags & VM_MAYSHARE) {
1208 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1210 if (vma_lock && vma_lock->vma != vma)
1211 vma->vm_private_data = NULL;
1213 vma->vm_private_data = NULL;
1217 * Reset and decrement one ref on hugepage private reservation.
1218 * Called with mm->mmap_lock writer semaphore held.
1219 * This function should be only used by move_vma() and operate on
1220 * same sized vma. It should never come here with last ref on the
1223 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1226 * Clear the old hugetlb private page reservation.
1227 * It has already been transferred to new_vma.
1229 * During a mremap() operation of a hugetlb vma we call move_vma()
1230 * which copies vma into new_vma and unmaps vma. After the copy
1231 * operation both new_vma and vma share a reference to the resv_map
1232 * struct, and at that point vma is about to be unmapped. We don't
1233 * want to return the reservation to the pool at unmap of vma because
1234 * the reservation still lives on in new_vma, so simply decrement the
1235 * ref here and remove the resv_map reference from this vma.
1237 struct resv_map *reservations = vma_resv_map(vma);
1239 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1240 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1241 kref_put(&reservations->refs, resv_map_release);
1244 hugetlb_dup_vma_private(vma);
1247 /* Returns true if the VMA has associated reserve pages */
1248 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1250 if (vma->vm_flags & VM_NORESERVE) {
1252 * This address is already reserved by other process(chg == 0),
1253 * so, we should decrement reserved count. Without decrementing,
1254 * reserve count remains after releasing inode, because this
1255 * allocated page will go into page cache and is regarded as
1256 * coming from reserved pool in releasing step. Currently, we
1257 * don't have any other solution to deal with this situation
1258 * properly, so add work-around here.
1260 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1266 /* Shared mappings always use reserves */
1267 if (vma->vm_flags & VM_MAYSHARE) {
1269 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1270 * be a region map for all pages. The only situation where
1271 * there is no region map is if a hole was punched via
1272 * fallocate. In this case, there really are no reserves to
1273 * use. This situation is indicated if chg != 0.
1282 * Only the process that called mmap() has reserves for
1285 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1287 * Like the shared case above, a hole punch or truncate
1288 * could have been performed on the private mapping.
1289 * Examine the value of chg to determine if reserves
1290 * actually exist or were previously consumed.
1291 * Very Subtle - The value of chg comes from a previous
1292 * call to vma_needs_reserves(). The reserve map for
1293 * private mappings has different (opposite) semantics
1294 * than that of shared mappings. vma_needs_reserves()
1295 * has already taken this difference in semantics into
1296 * account. Therefore, the meaning of chg is the same
1297 * as in the shared case above. Code could easily be
1298 * combined, but keeping it separate draws attention to
1299 * subtle differences.
1310 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1312 int nid = folio_nid(folio);
1314 lockdep_assert_held(&hugetlb_lock);
1315 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1317 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1318 h->free_huge_pages++;
1319 h->free_huge_pages_node[nid]++;
1320 folio_set_hugetlb_freed(folio);
1323 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1326 struct folio *folio;
1327 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1329 lockdep_assert_held(&hugetlb_lock);
1330 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1331 if (pin && !folio_is_longterm_pinnable(folio))
1334 if (folio_test_hwpoison(folio))
1337 list_move(&folio->lru, &h->hugepage_activelist);
1338 folio_ref_unfreeze(folio, 1);
1339 folio_clear_hugetlb_freed(folio);
1340 h->free_huge_pages--;
1341 h->free_huge_pages_node[nid]--;
1348 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1349 int nid, nodemask_t *nmask)
1351 unsigned int cpuset_mems_cookie;
1352 struct zonelist *zonelist;
1355 int node = NUMA_NO_NODE;
1357 zonelist = node_zonelist(nid, gfp_mask);
1360 cpuset_mems_cookie = read_mems_allowed_begin();
1361 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1362 struct folio *folio;
1364 if (!cpuset_zone_allowed(zone, gfp_mask))
1367 * no need to ask again on the same node. Pool is node rather than
1370 if (zone_to_nid(zone) == node)
1372 node = zone_to_nid(zone);
1374 folio = dequeue_hugetlb_folio_node_exact(h, node);
1378 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1384 static unsigned long available_huge_pages(struct hstate *h)
1386 return h->free_huge_pages - h->resv_huge_pages;
1389 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1390 struct vm_area_struct *vma,
1391 unsigned long address, int avoid_reserve,
1394 struct folio *folio = NULL;
1395 struct mempolicy *mpol;
1397 nodemask_t *nodemask;
1401 * A child process with MAP_PRIVATE mappings created by their parent
1402 * have no page reserves. This check ensures that reservations are
1403 * not "stolen". The child may still get SIGKILLed
1405 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1408 /* If reserves cannot be used, ensure enough pages are in the pool */
1409 if (avoid_reserve && !available_huge_pages(h))
1412 gfp_mask = htlb_alloc_mask(h);
1413 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1415 if (mpol_is_preferred_many(mpol)) {
1416 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1419 /* Fallback to all nodes if page==NULL */
1424 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1427 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1428 folio_set_hugetlb_restore_reserve(folio);
1429 h->resv_huge_pages--;
1432 mpol_cond_put(mpol);
1440 * common helper functions for hstate_next_node_to_{alloc|free}.
1441 * We may have allocated or freed a huge page based on a different
1442 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1443 * be outside of *nodes_allowed. Ensure that we use an allowed
1444 * node for alloc or free.
1446 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1448 nid = next_node_in(nid, *nodes_allowed);
1449 VM_BUG_ON(nid >= MAX_NUMNODES);
1454 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1456 if (!node_isset(nid, *nodes_allowed))
1457 nid = next_node_allowed(nid, nodes_allowed);
1462 * returns the previously saved node ["this node"] from which to
1463 * allocate a persistent huge page for the pool and advance the
1464 * next node from which to allocate, handling wrap at end of node
1467 static int hstate_next_node_to_alloc(struct hstate *h,
1468 nodemask_t *nodes_allowed)
1472 VM_BUG_ON(!nodes_allowed);
1474 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1475 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1481 * helper for remove_pool_hugetlb_folio() - return the previously saved
1482 * node ["this node"] from which to free a huge page. Advance the
1483 * next node id whether or not we find a free huge page to free so
1484 * that the next attempt to free addresses the next node.
1486 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1490 VM_BUG_ON(!nodes_allowed);
1492 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1493 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1498 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1499 for (nr_nodes = nodes_weight(*mask); \
1501 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1504 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1505 for (nr_nodes = nodes_weight(*mask); \
1507 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1510 /* used to demote non-gigantic_huge pages as well */
1511 static void __destroy_compound_gigantic_folio(struct folio *folio,
1512 unsigned int order, bool demote)
1515 int nr_pages = 1 << order;
1518 atomic_set(&folio->_entire_mapcount, 0);
1519 atomic_set(&folio->_nr_pages_mapped, 0);
1520 atomic_set(&folio->_pincount, 0);
1522 for (i = 1; i < nr_pages; i++) {
1523 p = folio_page(folio, i);
1524 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1526 clear_compound_head(p);
1528 set_page_refcounted(p);
1531 __folio_clear_head(folio);
1534 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1537 __destroy_compound_gigantic_folio(folio, order, true);
1540 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1541 static void destroy_compound_gigantic_folio(struct folio *folio,
1544 __destroy_compound_gigantic_folio(folio, order, false);
1547 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1550 * If the page isn't allocated using the cma allocator,
1551 * cma_release() returns false.
1554 int nid = folio_nid(folio);
1556 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1560 free_contig_range(folio_pfn(folio), 1 << order);
1563 #ifdef CONFIG_CONTIG_ALLOC
1564 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1565 int nid, nodemask_t *nodemask)
1568 unsigned long nr_pages = pages_per_huge_page(h);
1569 if (nid == NUMA_NO_NODE)
1570 nid = numa_mem_id();
1576 if (hugetlb_cma[nid]) {
1577 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1578 huge_page_order(h), true);
1580 return page_folio(page);
1583 if (!(gfp_mask & __GFP_THISNODE)) {
1584 for_each_node_mask(node, *nodemask) {
1585 if (node == nid || !hugetlb_cma[node])
1588 page = cma_alloc(hugetlb_cma[node], nr_pages,
1589 huge_page_order(h), true);
1591 return page_folio(page);
1597 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1598 return page ? page_folio(page) : NULL;
1601 #else /* !CONFIG_CONTIG_ALLOC */
1602 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1603 int nid, nodemask_t *nodemask)
1607 #endif /* CONFIG_CONTIG_ALLOC */
1609 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1610 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1611 int nid, nodemask_t *nodemask)
1615 static inline void free_gigantic_folio(struct folio *folio,
1616 unsigned int order) { }
1617 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1618 unsigned int order) { }
1621 static inline void __clear_hugetlb_destructor(struct hstate *h,
1622 struct folio *folio)
1624 lockdep_assert_held(&hugetlb_lock);
1626 __folio_clear_hugetlb(folio);
1630 * Remove hugetlb folio from lists.
1631 * If vmemmap exists for the folio, update dtor so that the folio appears
1632 * as just a compound page. Otherwise, wait until after allocating vmemmap
1635 * A reference is held on the folio, except in the case of demote.
1637 * Must be called with hugetlb lock held.
1639 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1640 bool adjust_surplus,
1643 int nid = folio_nid(folio);
1645 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1646 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1648 lockdep_assert_held(&hugetlb_lock);
1649 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1652 list_del(&folio->lru);
1654 if (folio_test_hugetlb_freed(folio)) {
1655 h->free_huge_pages--;
1656 h->free_huge_pages_node[nid]--;
1658 if (adjust_surplus) {
1659 h->surplus_huge_pages--;
1660 h->surplus_huge_pages_node[nid]--;
1664 * We can only clear the hugetlb destructor after allocating vmemmap
1665 * pages. Otherwise, someone (memory error handling) may try to write
1666 * to tail struct pages.
1668 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1669 __clear_hugetlb_destructor(h, folio);
1672 * In the case of demote we do not ref count the page as it will soon
1673 * be turned into a page of smaller size.
1676 folio_ref_unfreeze(folio, 1);
1679 h->nr_huge_pages_node[nid]--;
1682 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1683 bool adjust_surplus)
1685 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1688 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1689 bool adjust_surplus)
1691 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1694 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1695 bool adjust_surplus)
1698 int nid = folio_nid(folio);
1700 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1702 lockdep_assert_held(&hugetlb_lock);
1704 INIT_LIST_HEAD(&folio->lru);
1706 h->nr_huge_pages_node[nid]++;
1708 if (adjust_surplus) {
1709 h->surplus_huge_pages++;
1710 h->surplus_huge_pages_node[nid]++;
1713 __folio_set_hugetlb(folio);
1714 folio_change_private(folio, NULL);
1716 * We have to set hugetlb_vmemmap_optimized again as above
1717 * folio_change_private(folio, NULL) cleared it.
1719 folio_set_hugetlb_vmemmap_optimized(folio);
1722 * This folio is about to be managed by the hugetlb allocator and
1723 * should have no users. Drop our reference, and check for others
1726 zeroed = folio_put_testzero(folio);
1727 if (unlikely(!zeroed))
1729 * It is VERY unlikely soneone else has taken a ref
1730 * on the folio. In this case, we simply return as
1731 * free_huge_folio() will be called when this other ref
1736 arch_clear_hugepage_flags(&folio->page);
1737 enqueue_hugetlb_folio(h, folio);
1740 static void __update_and_free_hugetlb_folio(struct hstate *h,
1741 struct folio *folio)
1743 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1745 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1749 * If we don't know which subpages are hwpoisoned, we can't free
1750 * the hugepage, so it's leaked intentionally.
1752 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1756 * If folio is not vmemmap optimized (!clear_dtor), then the folio
1757 * is no longer identified as a hugetlb page. hugetlb_vmemmap_restore_folio
1758 * can only be passed hugetlb pages and will BUG otherwise.
1760 if (clear_dtor && hugetlb_vmemmap_restore_folio(h, folio)) {
1761 spin_lock_irq(&hugetlb_lock);
1763 * If we cannot allocate vmemmap pages, just refuse to free the
1764 * page and put the page back on the hugetlb free list and treat
1765 * as a surplus page.
1767 add_hugetlb_folio(h, folio, true);
1768 spin_unlock_irq(&hugetlb_lock);
1773 * Move PageHWPoison flag from head page to the raw error pages,
1774 * which makes any healthy subpages reusable.
1776 if (unlikely(folio_test_hwpoison(folio)))
1777 folio_clear_hugetlb_hwpoison(folio);
1780 * If vmemmap pages were allocated above, then we need to clear the
1781 * hugetlb destructor under the hugetlb lock.
1783 if (folio_test_hugetlb(folio)) {
1784 spin_lock_irq(&hugetlb_lock);
1785 __clear_hugetlb_destructor(h, folio);
1786 spin_unlock_irq(&hugetlb_lock);
1790 * Non-gigantic pages demoted from CMA allocated gigantic pages
1791 * need to be given back to CMA in free_gigantic_folio.
1793 if (hstate_is_gigantic(h) ||
1794 hugetlb_cma_folio(folio, huge_page_order(h))) {
1795 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1796 free_gigantic_folio(folio, huge_page_order(h));
1798 __free_pages(&folio->page, huge_page_order(h));
1803 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1804 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1805 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1806 * the vmemmap pages.
1808 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1809 * freed and frees them one-by-one. As the page->mapping pointer is going
1810 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1811 * structure of a lockless linked list of huge pages to be freed.
1813 static LLIST_HEAD(hpage_freelist);
1815 static void free_hpage_workfn(struct work_struct *work)
1817 struct llist_node *node;
1819 node = llist_del_all(&hpage_freelist);
1822 struct folio *folio;
1825 folio = container_of((struct address_space **)node,
1826 struct folio, mapping);
1828 folio->mapping = NULL;
1830 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1831 * folio_hstate() is going to trigger because a previous call to
1832 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1833 * not use folio_hstate() directly.
1835 h = size_to_hstate(folio_size(folio));
1837 __update_and_free_hugetlb_folio(h, folio);
1842 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1844 static inline void flush_free_hpage_work(struct hstate *h)
1846 if (hugetlb_vmemmap_optimizable(h))
1847 flush_work(&free_hpage_work);
1850 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1853 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1854 __update_and_free_hugetlb_folio(h, folio);
1859 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1861 * Only call schedule_work() if hpage_freelist is previously
1862 * empty. Otherwise, schedule_work() had been called but the workfn
1863 * hasn't retrieved the list yet.
1865 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1866 schedule_work(&free_hpage_work);
1869 static void bulk_vmemmap_restore_error(struct hstate *h,
1870 struct list_head *folio_list,
1871 struct list_head *non_hvo_folios)
1873 struct folio *folio, *t_folio;
1875 if (!list_empty(non_hvo_folios)) {
1877 * Free any restored hugetlb pages so that restore of the
1878 * entire list can be retried.
1879 * The idea is that in the common case of ENOMEM errors freeing
1880 * hugetlb pages with vmemmap we will free up memory so that we
1881 * can allocate vmemmap for more hugetlb pages.
1883 list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) {
1884 list_del(&folio->lru);
1885 spin_lock_irq(&hugetlb_lock);
1886 __clear_hugetlb_destructor(h, folio);
1887 spin_unlock_irq(&hugetlb_lock);
1888 update_and_free_hugetlb_folio(h, folio, false);
1893 * In the case where there are no folios which can be
1894 * immediately freed, we loop through the list trying to restore
1895 * vmemmap individually in the hope that someone elsewhere may
1896 * have done something to cause success (such as freeing some
1897 * memory). If unable to restore a hugetlb page, the hugetlb
1898 * page is made a surplus page and removed from the list.
1899 * If are able to restore vmemmap and free one hugetlb page, we
1900 * quit processing the list to retry the bulk operation.
1902 list_for_each_entry_safe(folio, t_folio, folio_list, lru)
1903 if (hugetlb_vmemmap_restore_folio(h, folio)) {
1904 list_del(&folio->lru);
1905 spin_lock_irq(&hugetlb_lock);
1906 add_hugetlb_folio(h, folio, true);
1907 spin_unlock_irq(&hugetlb_lock);
1909 list_del(&folio->lru);
1910 spin_lock_irq(&hugetlb_lock);
1911 __clear_hugetlb_destructor(h, folio);
1912 spin_unlock_irq(&hugetlb_lock);
1913 update_and_free_hugetlb_folio(h, folio, false);
1920 static void update_and_free_pages_bulk(struct hstate *h,
1921 struct list_head *folio_list)
1924 struct folio *folio, *t_folio;
1925 LIST_HEAD(non_hvo_folios);
1928 * First allocate required vmemmmap (if necessary) for all folios.
1929 * Carefully handle errors and free up any available hugetlb pages
1930 * in an effort to make forward progress.
1933 ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios);
1935 bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios);
1940 * At this point, list should be empty, ret should be >= 0 and there
1941 * should only be pages on the non_hvo_folios list.
1942 * Do note that the non_hvo_folios list could be empty.
1943 * Without HVO enabled, ret will be 0 and there is no need to call
1944 * __clear_hugetlb_destructor as this was done previously.
1946 VM_WARN_ON(!list_empty(folio_list));
1947 VM_WARN_ON(ret < 0);
1948 if (!list_empty(&non_hvo_folios) && ret) {
1949 spin_lock_irq(&hugetlb_lock);
1950 list_for_each_entry(folio, &non_hvo_folios, lru)
1951 __clear_hugetlb_destructor(h, folio);
1952 spin_unlock_irq(&hugetlb_lock);
1955 list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) {
1956 update_and_free_hugetlb_folio(h, folio, false);
1961 struct hstate *size_to_hstate(unsigned long size)
1965 for_each_hstate(h) {
1966 if (huge_page_size(h) == size)
1972 void free_huge_folio(struct folio *folio)
1975 * Can't pass hstate in here because it is called from the
1976 * compound page destructor.
1978 struct hstate *h = folio_hstate(folio);
1979 int nid = folio_nid(folio);
1980 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1981 bool restore_reserve;
1982 unsigned long flags;
1984 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1985 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1987 hugetlb_set_folio_subpool(folio, NULL);
1988 if (folio_test_anon(folio))
1989 __ClearPageAnonExclusive(&folio->page);
1990 folio->mapping = NULL;
1991 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1992 folio_clear_hugetlb_restore_reserve(folio);
1995 * If HPageRestoreReserve was set on page, page allocation consumed a
1996 * reservation. If the page was associated with a subpool, there
1997 * would have been a page reserved in the subpool before allocation
1998 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1999 * reservation, do not call hugepage_subpool_put_pages() as this will
2000 * remove the reserved page from the subpool.
2002 if (!restore_reserve) {
2004 * A return code of zero implies that the subpool will be
2005 * under its minimum size if the reservation is not restored
2006 * after page is free. Therefore, force restore_reserve
2009 if (hugepage_subpool_put_pages(spool, 1) == 0)
2010 restore_reserve = true;
2013 spin_lock_irqsave(&hugetlb_lock, flags);
2014 folio_clear_hugetlb_migratable(folio);
2015 hugetlb_cgroup_uncharge_folio(hstate_index(h),
2016 pages_per_huge_page(h), folio);
2017 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
2018 pages_per_huge_page(h), folio);
2019 mem_cgroup_uncharge(folio);
2020 if (restore_reserve)
2021 h->resv_huge_pages++;
2023 if (folio_test_hugetlb_temporary(folio)) {
2024 remove_hugetlb_folio(h, folio, false);
2025 spin_unlock_irqrestore(&hugetlb_lock, flags);
2026 update_and_free_hugetlb_folio(h, folio, true);
2027 } else if (h->surplus_huge_pages_node[nid]) {
2028 /* remove the page from active list */
2029 remove_hugetlb_folio(h, folio, true);
2030 spin_unlock_irqrestore(&hugetlb_lock, flags);
2031 update_and_free_hugetlb_folio(h, folio, true);
2033 arch_clear_hugepage_flags(&folio->page);
2034 enqueue_hugetlb_folio(h, folio);
2035 spin_unlock_irqrestore(&hugetlb_lock, flags);
2040 * Must be called with the hugetlb lock held
2042 static void __prep_account_new_huge_page(struct hstate *h, int nid)
2044 lockdep_assert_held(&hugetlb_lock);
2046 h->nr_huge_pages_node[nid]++;
2049 static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2051 __folio_set_hugetlb(folio);
2052 INIT_LIST_HEAD(&folio->lru);
2053 hugetlb_set_folio_subpool(folio, NULL);
2054 set_hugetlb_cgroup(folio, NULL);
2055 set_hugetlb_cgroup_rsvd(folio, NULL);
2058 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2060 init_new_hugetlb_folio(h, folio);
2061 hugetlb_vmemmap_optimize_folio(h, folio);
2064 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
2066 __prep_new_hugetlb_folio(h, folio);
2067 spin_lock_irq(&hugetlb_lock);
2068 __prep_account_new_huge_page(h, nid);
2069 spin_unlock_irq(&hugetlb_lock);
2072 static bool __prep_compound_gigantic_folio(struct folio *folio,
2073 unsigned int order, bool demote)
2076 int nr_pages = 1 << order;
2079 __folio_clear_reserved(folio);
2080 for (i = 0; i < nr_pages; i++) {
2081 p = folio_page(folio, i);
2084 * For gigantic hugepages allocated through bootmem at
2085 * boot, it's safer to be consistent with the not-gigantic
2086 * hugepages and clear the PG_reserved bit from all tail pages
2087 * too. Otherwise drivers using get_user_pages() to access tail
2088 * pages may get the reference counting wrong if they see
2089 * PG_reserved set on a tail page (despite the head page not
2090 * having PG_reserved set). Enforcing this consistency between
2091 * head and tail pages allows drivers to optimize away a check
2092 * on the head page when they need know if put_page() is needed
2093 * after get_user_pages().
2095 if (i != 0) /* head page cleared above */
2096 __ClearPageReserved(p);
2098 * Subtle and very unlikely
2100 * Gigantic 'page allocators' such as memblock or cma will
2101 * return a set of pages with each page ref counted. We need
2102 * to turn this set of pages into a compound page with tail
2103 * page ref counts set to zero. Code such as speculative page
2104 * cache adding could take a ref on a 'to be' tail page.
2105 * We need to respect any increased ref count, and only set
2106 * the ref count to zero if count is currently 1. If count
2107 * is not 1, we return an error. An error return indicates
2108 * the set of pages can not be converted to a gigantic page.
2109 * The caller who allocated the pages should then discard the
2110 * pages using the appropriate free interface.
2112 * In the case of demote, the ref count will be zero.
2115 if (!page_ref_freeze(p, 1)) {
2116 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2120 VM_BUG_ON_PAGE(page_count(p), p);
2123 set_compound_head(p, &folio->page);
2125 __folio_set_head(folio);
2126 /* we rely on prep_new_hugetlb_folio to set the destructor */
2127 folio_set_order(folio, order);
2128 atomic_set(&folio->_entire_mapcount, -1);
2129 atomic_set(&folio->_nr_pages_mapped, 0);
2130 atomic_set(&folio->_pincount, 0);
2134 /* undo page modifications made above */
2135 for (j = 0; j < i; j++) {
2136 p = folio_page(folio, j);
2138 clear_compound_head(p);
2139 set_page_refcounted(p);
2141 /* need to clear PG_reserved on remaining tail pages */
2142 for (; j < nr_pages; j++) {
2143 p = folio_page(folio, j);
2144 __ClearPageReserved(p);
2149 static bool prep_compound_gigantic_folio(struct folio *folio,
2152 return __prep_compound_gigantic_folio(folio, order, false);
2155 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2158 return __prep_compound_gigantic_folio(folio, order, true);
2162 * Find and lock address space (mapping) in write mode.
2164 * Upon entry, the page is locked which means that page_mapping() is
2165 * stable. Due to locking order, we can only trylock_write. If we can
2166 * not get the lock, simply return NULL to caller.
2168 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2170 struct address_space *mapping = page_mapping(hpage);
2175 if (i_mmap_trylock_write(mapping))
2181 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2182 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2183 nodemask_t *node_alloc_noretry)
2185 int order = huge_page_order(h);
2187 bool alloc_try_hard = true;
2191 * By default we always try hard to allocate the page with
2192 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2193 * a loop (to adjust global huge page counts) and previous allocation
2194 * failed, do not continue to try hard on the same node. Use the
2195 * node_alloc_noretry bitmap to manage this state information.
2197 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2198 alloc_try_hard = false;
2199 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2201 gfp_mask |= __GFP_RETRY_MAYFAIL;
2202 if (nid == NUMA_NO_NODE)
2203 nid = numa_mem_id();
2205 page = __alloc_pages(gfp_mask, order, nid, nmask);
2207 /* Freeze head page */
2208 if (page && !page_ref_freeze(page, 1)) {
2209 __free_pages(page, order);
2210 if (retry) { /* retry once */
2214 /* WOW! twice in a row. */
2215 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2220 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2221 * indicates an overall state change. Clear bit so that we resume
2222 * normal 'try hard' allocations.
2224 if (node_alloc_noretry && page && !alloc_try_hard)
2225 node_clear(nid, *node_alloc_noretry);
2228 * If we tried hard to get a page but failed, set bit so that
2229 * subsequent attempts will not try as hard until there is an
2230 * overall state change.
2232 if (node_alloc_noretry && !page && alloc_try_hard)
2233 node_set(nid, *node_alloc_noretry);
2236 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2240 __count_vm_event(HTLB_BUDDY_PGALLOC);
2241 return page_folio(page);
2244 static struct folio *__alloc_fresh_hugetlb_folio(struct hstate *h,
2245 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2246 nodemask_t *node_alloc_noretry)
2248 struct folio *folio;
2252 if (hstate_is_gigantic(h))
2253 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2255 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2256 nid, nmask, node_alloc_noretry);
2260 if (hstate_is_gigantic(h)) {
2261 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2263 * Rare failure to convert pages to compound page.
2264 * Free pages and try again - ONCE!
2266 free_gigantic_folio(folio, huge_page_order(h));
2278 static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h,
2279 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2280 nodemask_t *node_alloc_noretry)
2282 struct folio *folio;
2284 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2285 node_alloc_noretry);
2287 init_new_hugetlb_folio(h, folio);
2292 * Common helper to allocate a fresh hugetlb page. All specific allocators
2293 * should use this function to get new hugetlb pages
2295 * Note that returned page is 'frozen': ref count of head page and all tail
2298 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2299 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2300 nodemask_t *node_alloc_noretry)
2302 struct folio *folio;
2304 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2305 node_alloc_noretry);
2309 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2313 static void prep_and_add_allocated_folios(struct hstate *h,
2314 struct list_head *folio_list)
2316 unsigned long flags;
2317 struct folio *folio, *tmp_f;
2319 /* Send list for bulk vmemmap optimization processing */
2320 hugetlb_vmemmap_optimize_folios(h, folio_list);
2322 /* Add all new pool pages to free lists in one lock cycle */
2323 spin_lock_irqsave(&hugetlb_lock, flags);
2324 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
2325 __prep_account_new_huge_page(h, folio_nid(folio));
2326 enqueue_hugetlb_folio(h, folio);
2328 spin_unlock_irqrestore(&hugetlb_lock, flags);
2332 * Allocates a fresh hugetlb page in a node interleaved manner. The page
2333 * will later be added to the appropriate hugetlb pool.
2335 static struct folio *alloc_pool_huge_folio(struct hstate *h,
2336 nodemask_t *nodes_allowed,
2337 nodemask_t *node_alloc_noretry)
2339 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2342 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2343 struct folio *folio;
2345 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2346 nodes_allowed, node_alloc_noretry);
2355 * Remove huge page from pool from next node to free. Attempt to keep
2356 * persistent huge pages more or less balanced over allowed nodes.
2357 * This routine only 'removes' the hugetlb page. The caller must make
2358 * an additional call to free the page to low level allocators.
2359 * Called with hugetlb_lock locked.
2361 static struct folio *remove_pool_hugetlb_folio(struct hstate *h,
2362 nodemask_t *nodes_allowed, bool acct_surplus)
2365 struct folio *folio = NULL;
2367 lockdep_assert_held(&hugetlb_lock);
2368 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2370 * If we're returning unused surplus pages, only examine
2371 * nodes with surplus pages.
2373 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2374 !list_empty(&h->hugepage_freelists[node])) {
2375 folio = list_entry(h->hugepage_freelists[node].next,
2377 remove_hugetlb_folio(h, folio, acct_surplus);
2386 * Dissolve a given free hugepage into free buddy pages. This function does
2387 * nothing for in-use hugepages and non-hugepages.
2388 * This function returns values like below:
2390 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2391 * when the system is under memory pressure and the feature of
2392 * freeing unused vmemmap pages associated with each hugetlb page
2394 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2395 * (allocated or reserved.)
2396 * 0: successfully dissolved free hugepages or the page is not a
2397 * hugepage (considered as already dissolved)
2399 int dissolve_free_huge_page(struct page *page)
2402 struct folio *folio = page_folio(page);
2405 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2406 if (!folio_test_hugetlb(folio))
2409 spin_lock_irq(&hugetlb_lock);
2410 if (!folio_test_hugetlb(folio)) {
2415 if (!folio_ref_count(folio)) {
2416 struct hstate *h = folio_hstate(folio);
2417 if (!available_huge_pages(h))
2421 * We should make sure that the page is already on the free list
2422 * when it is dissolved.
2424 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2425 spin_unlock_irq(&hugetlb_lock);
2429 * Theoretically, we should return -EBUSY when we
2430 * encounter this race. In fact, we have a chance
2431 * to successfully dissolve the page if we do a
2432 * retry. Because the race window is quite small.
2433 * If we seize this opportunity, it is an optimization
2434 * for increasing the success rate of dissolving page.
2439 remove_hugetlb_folio(h, folio, false);
2440 h->max_huge_pages--;
2441 spin_unlock_irq(&hugetlb_lock);
2444 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2445 * before freeing the page. update_and_free_hugtlb_folio will fail to
2446 * free the page if it can not allocate required vmemmap. We
2447 * need to adjust max_huge_pages if the page is not freed.
2448 * Attempt to allocate vmemmmap here so that we can take
2449 * appropriate action on failure.
2451 * The folio_test_hugetlb check here is because
2452 * remove_hugetlb_folio will clear hugetlb folio flag for
2453 * non-vmemmap optimized hugetlb folios.
2455 if (folio_test_hugetlb(folio)) {
2456 rc = hugetlb_vmemmap_restore_folio(h, folio);
2458 spin_lock_irq(&hugetlb_lock);
2459 add_hugetlb_folio(h, folio, false);
2460 h->max_huge_pages++;
2466 update_and_free_hugetlb_folio(h, folio, false);
2470 spin_unlock_irq(&hugetlb_lock);
2475 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2476 * make specified memory blocks removable from the system.
2477 * Note that this will dissolve a free gigantic hugepage completely, if any
2478 * part of it lies within the given range.
2479 * Also note that if dissolve_free_huge_page() returns with an error, all
2480 * free hugepages that were dissolved before that error are lost.
2482 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2490 if (!hugepages_supported())
2493 order = huge_page_order(&default_hstate);
2495 order = min(order, huge_page_order(h));
2497 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2498 page = pfn_to_page(pfn);
2499 rc = dissolve_free_huge_page(page);
2508 * Allocates a fresh surplus page from the page allocator.
2510 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2511 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2513 struct folio *folio = NULL;
2515 if (hstate_is_gigantic(h))
2518 spin_lock_irq(&hugetlb_lock);
2519 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2521 spin_unlock_irq(&hugetlb_lock);
2523 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2527 spin_lock_irq(&hugetlb_lock);
2529 * We could have raced with the pool size change.
2530 * Double check that and simply deallocate the new page
2531 * if we would end up overcommiting the surpluses. Abuse
2532 * temporary page to workaround the nasty free_huge_folio
2535 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2536 folio_set_hugetlb_temporary(folio);
2537 spin_unlock_irq(&hugetlb_lock);
2538 free_huge_folio(folio);
2542 h->surplus_huge_pages++;
2543 h->surplus_huge_pages_node[folio_nid(folio)]++;
2546 spin_unlock_irq(&hugetlb_lock);
2551 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2552 int nid, nodemask_t *nmask)
2554 struct folio *folio;
2556 if (hstate_is_gigantic(h))
2559 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2563 /* fresh huge pages are frozen */
2564 folio_ref_unfreeze(folio, 1);
2566 * We do not account these pages as surplus because they are only
2567 * temporary and will be released properly on the last reference
2569 folio_set_hugetlb_temporary(folio);
2575 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2578 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2579 struct vm_area_struct *vma, unsigned long addr)
2581 struct folio *folio = NULL;
2582 struct mempolicy *mpol;
2583 gfp_t gfp_mask = htlb_alloc_mask(h);
2585 nodemask_t *nodemask;
2587 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2588 if (mpol_is_preferred_many(mpol)) {
2589 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2591 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2592 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2594 /* Fallback to all nodes if page==NULL */
2599 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2600 mpol_cond_put(mpol);
2604 /* folio migration callback function */
2605 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2606 nodemask_t *nmask, gfp_t gfp_mask)
2608 spin_lock_irq(&hugetlb_lock);
2609 if (available_huge_pages(h)) {
2610 struct folio *folio;
2612 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2613 preferred_nid, nmask);
2615 spin_unlock_irq(&hugetlb_lock);
2619 spin_unlock_irq(&hugetlb_lock);
2621 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2625 * Increase the hugetlb pool such that it can accommodate a reservation
2628 static int gather_surplus_pages(struct hstate *h, long delta)
2629 __must_hold(&hugetlb_lock)
2631 LIST_HEAD(surplus_list);
2632 struct folio *folio, *tmp;
2635 long needed, allocated;
2636 bool alloc_ok = true;
2638 lockdep_assert_held(&hugetlb_lock);
2639 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2641 h->resv_huge_pages += delta;
2649 spin_unlock_irq(&hugetlb_lock);
2650 for (i = 0; i < needed; i++) {
2651 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2652 NUMA_NO_NODE, NULL);
2657 list_add(&folio->lru, &surplus_list);
2663 * After retaking hugetlb_lock, we need to recalculate 'needed'
2664 * because either resv_huge_pages or free_huge_pages may have changed.
2666 spin_lock_irq(&hugetlb_lock);
2667 needed = (h->resv_huge_pages + delta) -
2668 (h->free_huge_pages + allocated);
2673 * We were not able to allocate enough pages to
2674 * satisfy the entire reservation so we free what
2675 * we've allocated so far.
2680 * The surplus_list now contains _at_least_ the number of extra pages
2681 * needed to accommodate the reservation. Add the appropriate number
2682 * of pages to the hugetlb pool and free the extras back to the buddy
2683 * allocator. Commit the entire reservation here to prevent another
2684 * process from stealing the pages as they are added to the pool but
2685 * before they are reserved.
2687 needed += allocated;
2688 h->resv_huge_pages += delta;
2691 /* Free the needed pages to the hugetlb pool */
2692 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2695 /* Add the page to the hugetlb allocator */
2696 enqueue_hugetlb_folio(h, folio);
2699 spin_unlock_irq(&hugetlb_lock);
2702 * Free unnecessary surplus pages to the buddy allocator.
2703 * Pages have no ref count, call free_huge_folio directly.
2705 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2706 free_huge_folio(folio);
2707 spin_lock_irq(&hugetlb_lock);
2713 * This routine has two main purposes:
2714 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2715 * in unused_resv_pages. This corresponds to the prior adjustments made
2716 * to the associated reservation map.
2717 * 2) Free any unused surplus pages that may have been allocated to satisfy
2718 * the reservation. As many as unused_resv_pages may be freed.
2720 static void return_unused_surplus_pages(struct hstate *h,
2721 unsigned long unused_resv_pages)
2723 unsigned long nr_pages;
2724 LIST_HEAD(page_list);
2726 lockdep_assert_held(&hugetlb_lock);
2727 /* Uncommit the reservation */
2728 h->resv_huge_pages -= unused_resv_pages;
2730 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2734 * Part (or even all) of the reservation could have been backed
2735 * by pre-allocated pages. Only free surplus pages.
2737 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2740 * We want to release as many surplus pages as possible, spread
2741 * evenly across all nodes with memory. Iterate across these nodes
2742 * until we can no longer free unreserved surplus pages. This occurs
2743 * when the nodes with surplus pages have no free pages.
2744 * remove_pool_hugetlb_folio() will balance the freed pages across the
2745 * on-line nodes with memory and will handle the hstate accounting.
2747 while (nr_pages--) {
2748 struct folio *folio;
2750 folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1);
2754 list_add(&folio->lru, &page_list);
2758 spin_unlock_irq(&hugetlb_lock);
2759 update_and_free_pages_bulk(h, &page_list);
2760 spin_lock_irq(&hugetlb_lock);
2765 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2766 * are used by the huge page allocation routines to manage reservations.
2768 * vma_needs_reservation is called to determine if the huge page at addr
2769 * within the vma has an associated reservation. If a reservation is
2770 * needed, the value 1 is returned. The caller is then responsible for
2771 * managing the global reservation and subpool usage counts. After
2772 * the huge page has been allocated, vma_commit_reservation is called
2773 * to add the page to the reservation map. If the page allocation fails,
2774 * the reservation must be ended instead of committed. vma_end_reservation
2775 * is called in such cases.
2777 * In the normal case, vma_commit_reservation returns the same value
2778 * as the preceding vma_needs_reservation call. The only time this
2779 * is not the case is if a reserve map was changed between calls. It
2780 * is the responsibility of the caller to notice the difference and
2781 * take appropriate action.
2783 * vma_add_reservation is used in error paths where a reservation must
2784 * be restored when a newly allocated huge page must be freed. It is
2785 * to be called after calling vma_needs_reservation to determine if a
2786 * reservation exists.
2788 * vma_del_reservation is used in error paths where an entry in the reserve
2789 * map was created during huge page allocation and must be removed. It is to
2790 * be called after calling vma_needs_reservation to determine if a reservation
2793 enum vma_resv_mode {
2800 static long __vma_reservation_common(struct hstate *h,
2801 struct vm_area_struct *vma, unsigned long addr,
2802 enum vma_resv_mode mode)
2804 struct resv_map *resv;
2807 long dummy_out_regions_needed;
2809 resv = vma_resv_map(vma);
2813 idx = vma_hugecache_offset(h, vma, addr);
2815 case VMA_NEEDS_RESV:
2816 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2817 /* We assume that vma_reservation_* routines always operate on
2818 * 1 page, and that adding to resv map a 1 page entry can only
2819 * ever require 1 region.
2821 VM_BUG_ON(dummy_out_regions_needed != 1);
2823 case VMA_COMMIT_RESV:
2824 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2825 /* region_add calls of range 1 should never fail. */
2829 region_abort(resv, idx, idx + 1, 1);
2833 if (vma->vm_flags & VM_MAYSHARE) {
2834 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2835 /* region_add calls of range 1 should never fail. */
2838 region_abort(resv, idx, idx + 1, 1);
2839 ret = region_del(resv, idx, idx + 1);
2843 if (vma->vm_flags & VM_MAYSHARE) {
2844 region_abort(resv, idx, idx + 1, 1);
2845 ret = region_del(resv, idx, idx + 1);
2847 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2848 /* region_add calls of range 1 should never fail. */
2856 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2859 * We know private mapping must have HPAGE_RESV_OWNER set.
2861 * In most cases, reserves always exist for private mappings.
2862 * However, a file associated with mapping could have been
2863 * hole punched or truncated after reserves were consumed.
2864 * As subsequent fault on such a range will not use reserves.
2865 * Subtle - The reserve map for private mappings has the
2866 * opposite meaning than that of shared mappings. If NO
2867 * entry is in the reserve map, it means a reservation exists.
2868 * If an entry exists in the reserve map, it means the
2869 * reservation has already been consumed. As a result, the
2870 * return value of this routine is the opposite of the
2871 * value returned from reserve map manipulation routines above.
2880 static long vma_needs_reservation(struct hstate *h,
2881 struct vm_area_struct *vma, unsigned long addr)
2883 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2886 static long vma_commit_reservation(struct hstate *h,
2887 struct vm_area_struct *vma, unsigned long addr)
2889 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2892 static void vma_end_reservation(struct hstate *h,
2893 struct vm_area_struct *vma, unsigned long addr)
2895 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2898 static long vma_add_reservation(struct hstate *h,
2899 struct vm_area_struct *vma, unsigned long addr)
2901 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2904 static long vma_del_reservation(struct hstate *h,
2905 struct vm_area_struct *vma, unsigned long addr)
2907 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2911 * This routine is called to restore reservation information on error paths.
2912 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2913 * and the hugetlb mutex should remain held when calling this routine.
2915 * It handles two specific cases:
2916 * 1) A reservation was in place and the folio consumed the reservation.
2917 * hugetlb_restore_reserve is set in the folio.
2918 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2919 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2921 * In case 1, free_huge_folio later in the error path will increment the
2922 * global reserve count. But, free_huge_folio does not have enough context
2923 * to adjust the reservation map. This case deals primarily with private
2924 * mappings. Adjust the reserve map here to be consistent with global
2925 * reserve count adjustments to be made by free_huge_folio. Make sure the
2926 * reserve map indicates there is a reservation present.
2928 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2930 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2931 unsigned long address, struct folio *folio)
2933 long rc = vma_needs_reservation(h, vma, address);
2935 if (folio_test_hugetlb_restore_reserve(folio)) {
2936 if (unlikely(rc < 0))
2938 * Rare out of memory condition in reserve map
2939 * manipulation. Clear hugetlb_restore_reserve so
2940 * that global reserve count will not be incremented
2941 * by free_huge_folio. This will make it appear
2942 * as though the reservation for this folio was
2943 * consumed. This may prevent the task from
2944 * faulting in the folio at a later time. This
2945 * is better than inconsistent global huge page
2946 * accounting of reserve counts.
2948 folio_clear_hugetlb_restore_reserve(folio);
2950 (void)vma_add_reservation(h, vma, address);
2952 vma_end_reservation(h, vma, address);
2956 * This indicates there is an entry in the reserve map
2957 * not added by alloc_hugetlb_folio. We know it was added
2958 * before the alloc_hugetlb_folio call, otherwise
2959 * hugetlb_restore_reserve would be set on the folio.
2960 * Remove the entry so that a subsequent allocation
2961 * does not consume a reservation.
2963 rc = vma_del_reservation(h, vma, address);
2966 * VERY rare out of memory condition. Since
2967 * we can not delete the entry, set
2968 * hugetlb_restore_reserve so that the reserve
2969 * count will be incremented when the folio
2970 * is freed. This reserve will be consumed
2971 * on a subsequent allocation.
2973 folio_set_hugetlb_restore_reserve(folio);
2974 } else if (rc < 0) {
2976 * Rare out of memory condition from
2977 * vma_needs_reservation call. Memory allocation is
2978 * only attempted if a new entry is needed. Therefore,
2979 * this implies there is not an entry in the
2982 * For shared mappings, no entry in the map indicates
2983 * no reservation. We are done.
2985 if (!(vma->vm_flags & VM_MAYSHARE))
2987 * For private mappings, no entry indicates
2988 * a reservation is present. Since we can
2989 * not add an entry, set hugetlb_restore_reserve
2990 * on the folio so reserve count will be
2991 * incremented when freed. This reserve will
2992 * be consumed on a subsequent allocation.
2994 folio_set_hugetlb_restore_reserve(folio);
2997 * No reservation present, do nothing
2999 vma_end_reservation(h, vma, address);
3004 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
3006 * @h: struct hstate old page belongs to
3007 * @old_folio: Old folio to dissolve
3008 * @list: List to isolate the page in case we need to
3009 * Returns 0 on success, otherwise negated error.
3011 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
3012 struct folio *old_folio, struct list_head *list)
3014 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3015 int nid = folio_nid(old_folio);
3016 struct folio *new_folio;
3020 * Before dissolving the folio, we need to allocate a new one for the
3021 * pool to remain stable. Here, we allocate the folio and 'prep' it
3022 * by doing everything but actually updating counters and adding to
3023 * the pool. This simplifies and let us do most of the processing
3026 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
3029 __prep_new_hugetlb_folio(h, new_folio);
3032 spin_lock_irq(&hugetlb_lock);
3033 if (!folio_test_hugetlb(old_folio)) {
3035 * Freed from under us. Drop new_folio too.
3038 } else if (folio_ref_count(old_folio)) {
3042 * Someone has grabbed the folio, try to isolate it here.
3043 * Fail with -EBUSY if not possible.
3045 spin_unlock_irq(&hugetlb_lock);
3046 isolated = isolate_hugetlb(old_folio, list);
3047 ret = isolated ? 0 : -EBUSY;
3048 spin_lock_irq(&hugetlb_lock);
3050 } else if (!folio_test_hugetlb_freed(old_folio)) {
3052 * Folio's refcount is 0 but it has not been enqueued in the
3053 * freelist yet. Race window is small, so we can succeed here if
3056 spin_unlock_irq(&hugetlb_lock);
3061 * Ok, old_folio is still a genuine free hugepage. Remove it from
3062 * the freelist and decrease the counters. These will be
3063 * incremented again when calling __prep_account_new_huge_page()
3064 * and enqueue_hugetlb_folio() for new_folio. The counters will
3065 * remain stable since this happens under the lock.
3067 remove_hugetlb_folio(h, old_folio, false);
3070 * Ref count on new_folio is already zero as it was dropped
3071 * earlier. It can be directly added to the pool free list.
3073 __prep_account_new_huge_page(h, nid);
3074 enqueue_hugetlb_folio(h, new_folio);
3077 * Folio has been replaced, we can safely free the old one.
3079 spin_unlock_irq(&hugetlb_lock);
3080 update_and_free_hugetlb_folio(h, old_folio, false);
3086 spin_unlock_irq(&hugetlb_lock);
3087 /* Folio has a zero ref count, but needs a ref to be freed */
3088 folio_ref_unfreeze(new_folio, 1);
3089 update_and_free_hugetlb_folio(h, new_folio, false);
3094 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3097 struct folio *folio = page_folio(page);
3101 * The page might have been dissolved from under our feet, so make sure
3102 * to carefully check the state under the lock.
3103 * Return success when racing as if we dissolved the page ourselves.
3105 spin_lock_irq(&hugetlb_lock);
3106 if (folio_test_hugetlb(folio)) {
3107 h = folio_hstate(folio);
3109 spin_unlock_irq(&hugetlb_lock);
3112 spin_unlock_irq(&hugetlb_lock);
3115 * Fence off gigantic pages as there is a cyclic dependency between
3116 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3117 * of bailing out right away without further retrying.
3119 if (hstate_is_gigantic(h))
3122 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3124 else if (!folio_ref_count(folio))
3125 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3130 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3131 unsigned long addr, int avoid_reserve)
3133 struct hugepage_subpool *spool = subpool_vma(vma);
3134 struct hstate *h = hstate_vma(vma);
3135 struct folio *folio;
3136 long map_chg, map_commit, nr_pages = pages_per_huge_page(h);
3138 int memcg_charge_ret, ret, idx;
3139 struct hugetlb_cgroup *h_cg = NULL;
3140 struct mem_cgroup *memcg;
3141 bool deferred_reserve;
3142 gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL;
3144 memcg = get_mem_cgroup_from_current();
3145 memcg_charge_ret = mem_cgroup_hugetlb_try_charge(memcg, gfp, nr_pages);
3146 if (memcg_charge_ret == -ENOMEM) {
3147 mem_cgroup_put(memcg);
3148 return ERR_PTR(-ENOMEM);
3151 idx = hstate_index(h);
3153 * Examine the region/reserve map to determine if the process
3154 * has a reservation for the page to be allocated. A return
3155 * code of zero indicates a reservation exists (no change).
3157 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3159 if (!memcg_charge_ret)
3160 mem_cgroup_cancel_charge(memcg, nr_pages);
3161 mem_cgroup_put(memcg);
3162 return ERR_PTR(-ENOMEM);
3166 * Processes that did not create the mapping will have no
3167 * reserves as indicated by the region/reserve map. Check
3168 * that the allocation will not exceed the subpool limit.
3169 * Allocations for MAP_NORESERVE mappings also need to be
3170 * checked against any subpool limit.
3172 if (map_chg || avoid_reserve) {
3173 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3175 goto out_end_reservation;
3178 * Even though there was no reservation in the region/reserve
3179 * map, there could be reservations associated with the
3180 * subpool that can be used. This would be indicated if the
3181 * return value of hugepage_subpool_get_pages() is zero.
3182 * However, if avoid_reserve is specified we still avoid even
3183 * the subpool reservations.
3189 /* If this allocation is not consuming a reservation, charge it now.
3191 deferred_reserve = map_chg || avoid_reserve;
3192 if (deferred_reserve) {
3193 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3194 idx, pages_per_huge_page(h), &h_cg);
3196 goto out_subpool_put;
3199 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3201 goto out_uncharge_cgroup_reservation;
3203 spin_lock_irq(&hugetlb_lock);
3205 * glb_chg is passed to indicate whether or not a page must be taken
3206 * from the global free pool (global change). gbl_chg == 0 indicates
3207 * a reservation exists for the allocation.
3209 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3211 spin_unlock_irq(&hugetlb_lock);
3212 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3214 goto out_uncharge_cgroup;
3215 spin_lock_irq(&hugetlb_lock);
3216 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3217 folio_set_hugetlb_restore_reserve(folio);
3218 h->resv_huge_pages--;
3220 list_add(&folio->lru, &h->hugepage_activelist);
3221 folio_ref_unfreeze(folio, 1);
3225 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3226 /* If allocation is not consuming a reservation, also store the
3227 * hugetlb_cgroup pointer on the page.
3229 if (deferred_reserve) {
3230 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3234 spin_unlock_irq(&hugetlb_lock);
3236 hugetlb_set_folio_subpool(folio, spool);
3238 map_commit = vma_commit_reservation(h, vma, addr);
3239 if (unlikely(map_chg > map_commit)) {
3241 * The page was added to the reservation map between
3242 * vma_needs_reservation and vma_commit_reservation.
3243 * This indicates a race with hugetlb_reserve_pages.
3244 * Adjust for the subpool count incremented above AND
3245 * in hugetlb_reserve_pages for the same page. Also,
3246 * the reservation count added in hugetlb_reserve_pages
3247 * no longer applies.
3251 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3252 hugetlb_acct_memory(h, -rsv_adjust);
3253 if (deferred_reserve) {
3254 spin_lock_irq(&hugetlb_lock);
3255 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3256 pages_per_huge_page(h), folio);
3257 spin_unlock_irq(&hugetlb_lock);
3261 if (!memcg_charge_ret)
3262 mem_cgroup_commit_charge(folio, memcg);
3263 mem_cgroup_put(memcg);
3267 out_uncharge_cgroup:
3268 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3269 out_uncharge_cgroup_reservation:
3270 if (deferred_reserve)
3271 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3274 if (map_chg || avoid_reserve)
3275 hugepage_subpool_put_pages(spool, 1);
3276 out_end_reservation:
3277 vma_end_reservation(h, vma, addr);
3278 if (!memcg_charge_ret)
3279 mem_cgroup_cancel_charge(memcg, nr_pages);
3280 mem_cgroup_put(memcg);
3281 return ERR_PTR(-ENOSPC);
3284 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3285 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3286 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3288 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3291 /* do node specific alloc */
3292 if (nid != NUMA_NO_NODE) {
3293 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3294 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3299 /* allocate from next node when distributing huge pages */
3300 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3301 m = memblock_alloc_try_nid_raw(
3302 huge_page_size(h), huge_page_size(h),
3303 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3305 * Use the beginning of the huge page to store the
3306 * huge_bootmem_page struct (until gather_bootmem
3307 * puts them into the mem_map).
3317 * Only initialize the head struct page in memmap_init_reserved_pages,
3318 * rest of the struct pages will be initialized by the HugeTLB
3320 * The head struct page is used to get folio information by the HugeTLB
3321 * subsystem like zone id and node id.
3323 memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE),
3324 huge_page_size(h) - PAGE_SIZE);
3325 /* Put them into a private list first because mem_map is not up yet */
3326 INIT_LIST_HEAD(&m->list);
3327 list_add(&m->list, &huge_boot_pages);
3332 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
3333 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
3334 unsigned long start_page_number,
3335 unsigned long end_page_number)
3337 enum zone_type zone = zone_idx(folio_zone(folio));
3338 int nid = folio_nid(folio);
3339 unsigned long head_pfn = folio_pfn(folio);
3340 unsigned long pfn, end_pfn = head_pfn + end_page_number;
3343 for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) {
3344 struct page *page = pfn_to_page(pfn);
3346 __init_single_page(page, pfn, zone, nid);
3347 prep_compound_tail((struct page *)folio, pfn - head_pfn);
3348 ret = page_ref_freeze(page, 1);
3353 static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
3355 unsigned long nr_pages)
3359 /* Prepare folio head */
3360 __folio_clear_reserved(folio);
3361 __folio_set_head(folio);
3362 ret = folio_ref_freeze(folio, 1);
3364 /* Initialize the necessary tail struct pages */
3365 hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);
3366 prep_compound_head((struct page *)folio, huge_page_order(h));
3369 static void __init prep_and_add_bootmem_folios(struct hstate *h,
3370 struct list_head *folio_list)
3372 unsigned long flags;
3373 struct folio *folio, *tmp_f;
3375 /* Send list for bulk vmemmap optimization processing */
3376 hugetlb_vmemmap_optimize_folios(h, folio_list);
3378 /* Add all new pool pages to free lists in one lock cycle */
3379 spin_lock_irqsave(&hugetlb_lock, flags);
3380 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
3381 if (!folio_test_hugetlb_vmemmap_optimized(folio)) {
3383 * If HVO fails, initialize all tail struct pages
3384 * We do not worry about potential long lock hold
3385 * time as this is early in boot and there should
3388 hugetlb_folio_init_tail_vmemmap(folio,
3389 HUGETLB_VMEMMAP_RESERVE_PAGES,
3390 pages_per_huge_page(h));
3392 __prep_account_new_huge_page(h, folio_nid(folio));
3393 enqueue_hugetlb_folio(h, folio);
3395 spin_unlock_irqrestore(&hugetlb_lock, flags);
3399 * Put bootmem huge pages into the standard lists after mem_map is up.
3400 * Note: This only applies to gigantic (order > MAX_PAGE_ORDER) pages.
3402 static void __init gather_bootmem_prealloc(void)
3404 LIST_HEAD(folio_list);
3405 struct huge_bootmem_page *m;
3406 struct hstate *h = NULL, *prev_h = NULL;
3408 list_for_each_entry(m, &huge_boot_pages, list) {
3409 struct page *page = virt_to_page(m);
3410 struct folio *folio = (void *)page;
3414 * It is possible to have multiple huge page sizes (hstates)
3415 * in this list. If so, process each size separately.
3417 if (h != prev_h && prev_h != NULL)
3418 prep_and_add_bootmem_folios(prev_h, &folio_list);
3421 VM_BUG_ON(!hstate_is_gigantic(h));
3422 WARN_ON(folio_ref_count(folio) != 1);
3424 hugetlb_folio_init_vmemmap(folio, h,
3425 HUGETLB_VMEMMAP_RESERVE_PAGES);
3426 init_new_hugetlb_folio(h, folio);
3427 list_add(&folio->lru, &folio_list);
3430 * We need to restore the 'stolen' pages to totalram_pages
3431 * in order to fix confusing memory reports from free(1) and
3432 * other side-effects, like CommitLimit going negative.
3434 adjust_managed_page_count(page, pages_per_huge_page(h));
3438 prep_and_add_bootmem_folios(h, &folio_list);
3441 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3446 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3447 if (hstate_is_gigantic(h)) {
3448 if (!alloc_bootmem_huge_page(h, nid))
3451 struct folio *folio;
3452 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3454 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3455 &node_states[N_MEMORY], NULL);
3458 free_huge_folio(folio); /* free it into the hugepage allocator */
3462 if (i == h->max_huge_pages_node[nid])
3465 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3466 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3467 h->max_huge_pages_node[nid], buf, nid, i);
3468 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3469 h->max_huge_pages_node[nid] = i;
3473 * NOTE: this routine is called in different contexts for gigantic and
3474 * non-gigantic pages.
3475 * - For gigantic pages, this is called early in the boot process and
3476 * pages are allocated from memblock allocated or something similar.
3477 * Gigantic pages are actually added to pools later with the routine
3478 * gather_bootmem_prealloc.
3479 * - For non-gigantic pages, this is called later in the boot process after
3480 * all of mm is up and functional. Pages are allocated from buddy and
3481 * then added to hugetlb pools.
3483 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3486 struct folio *folio;
3487 LIST_HEAD(folio_list);
3488 nodemask_t *node_alloc_noretry;
3489 bool node_specific_alloc = false;
3491 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3492 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3493 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3497 /* do node specific alloc */
3498 for_each_online_node(i) {
3499 if (h->max_huge_pages_node[i] > 0) {
3500 hugetlb_hstate_alloc_pages_onenode(h, i);
3501 node_specific_alloc = true;
3505 if (node_specific_alloc)
3508 /* below will do all node balanced alloc */
3509 if (!hstate_is_gigantic(h)) {
3511 * Bit mask controlling how hard we retry per-node allocations.
3512 * Ignore errors as lower level routines can deal with
3513 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3514 * time, we are likely in bigger trouble.
3516 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3519 /* allocations done at boot time */
3520 node_alloc_noretry = NULL;
3523 /* bit mask controlling how hard we retry per-node allocations */
3524 if (node_alloc_noretry)
3525 nodes_clear(*node_alloc_noretry);
3527 for (i = 0; i < h->max_huge_pages; ++i) {
3528 if (hstate_is_gigantic(h)) {
3530 * gigantic pages not added to list as they are not
3531 * added to pools now.
3533 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3536 folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
3537 node_alloc_noretry);
3540 list_add(&folio->lru, &folio_list);
3545 /* list will be empty if hstate_is_gigantic */
3546 prep_and_add_allocated_folios(h, &folio_list);
3548 if (i < h->max_huge_pages) {
3551 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3552 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3553 h->max_huge_pages, buf, i);
3554 h->max_huge_pages = i;
3556 kfree(node_alloc_noretry);
3559 static void __init hugetlb_init_hstates(void)
3561 struct hstate *h, *h2;
3563 for_each_hstate(h) {
3564 /* oversize hugepages were init'ed in early boot */
3565 if (!hstate_is_gigantic(h))
3566 hugetlb_hstate_alloc_pages(h);
3569 * Set demote order for each hstate. Note that
3570 * h->demote_order is initially 0.
3571 * - We can not demote gigantic pages if runtime freeing
3572 * is not supported, so skip this.
3573 * - If CMA allocation is possible, we can not demote
3574 * HUGETLB_PAGE_ORDER or smaller size pages.
3576 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3578 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3580 for_each_hstate(h2) {
3583 if (h2->order < h->order &&
3584 h2->order > h->demote_order)
3585 h->demote_order = h2->order;
3590 static void __init report_hugepages(void)
3594 for_each_hstate(h) {
3597 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3598 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3599 buf, h->free_huge_pages);
3600 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3601 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3605 #ifdef CONFIG_HIGHMEM
3606 static void try_to_free_low(struct hstate *h, unsigned long count,
3607 nodemask_t *nodes_allowed)
3610 LIST_HEAD(page_list);
3612 lockdep_assert_held(&hugetlb_lock);
3613 if (hstate_is_gigantic(h))
3617 * Collect pages to be freed on a list, and free after dropping lock
3619 for_each_node_mask(i, *nodes_allowed) {
3620 struct folio *folio, *next;
3621 struct list_head *freel = &h->hugepage_freelists[i];
3622 list_for_each_entry_safe(folio, next, freel, lru) {
3623 if (count >= h->nr_huge_pages)
3625 if (folio_test_highmem(folio))
3627 remove_hugetlb_folio(h, folio, false);
3628 list_add(&folio->lru, &page_list);
3633 spin_unlock_irq(&hugetlb_lock);
3634 update_and_free_pages_bulk(h, &page_list);
3635 spin_lock_irq(&hugetlb_lock);
3638 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3639 nodemask_t *nodes_allowed)
3645 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3646 * balanced by operating on them in a round-robin fashion.
3647 * Returns 1 if an adjustment was made.
3649 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3654 lockdep_assert_held(&hugetlb_lock);
3655 VM_BUG_ON(delta != -1 && delta != 1);
3658 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3659 if (h->surplus_huge_pages_node[node])
3663 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3664 if (h->surplus_huge_pages_node[node] <
3665 h->nr_huge_pages_node[node])
3672 h->surplus_huge_pages += delta;
3673 h->surplus_huge_pages_node[node] += delta;
3677 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3678 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3679 nodemask_t *nodes_allowed)
3681 unsigned long min_count;
3682 unsigned long allocated;
3683 struct folio *folio;
3684 LIST_HEAD(page_list);
3685 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3688 * Bit mask controlling how hard we retry per-node allocations.
3689 * If we can not allocate the bit mask, do not attempt to allocate
3690 * the requested huge pages.
3692 if (node_alloc_noretry)
3693 nodes_clear(*node_alloc_noretry);
3698 * resize_lock mutex prevents concurrent adjustments to number of
3699 * pages in hstate via the proc/sysfs interfaces.
3701 mutex_lock(&h->resize_lock);
3702 flush_free_hpage_work(h);
3703 spin_lock_irq(&hugetlb_lock);
3706 * Check for a node specific request.
3707 * Changing node specific huge page count may require a corresponding
3708 * change to the global count. In any case, the passed node mask
3709 * (nodes_allowed) will restrict alloc/free to the specified node.
3711 if (nid != NUMA_NO_NODE) {
3712 unsigned long old_count = count;
3714 count += persistent_huge_pages(h) -
3715 (h->nr_huge_pages_node[nid] -
3716 h->surplus_huge_pages_node[nid]);
3718 * User may have specified a large count value which caused the
3719 * above calculation to overflow. In this case, they wanted
3720 * to allocate as many huge pages as possible. Set count to
3721 * largest possible value to align with their intention.
3723 if (count < old_count)
3728 * Gigantic pages runtime allocation depend on the capability for large
3729 * page range allocation.
3730 * If the system does not provide this feature, return an error when
3731 * the user tries to allocate gigantic pages but let the user free the
3732 * boottime allocated gigantic pages.
3734 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3735 if (count > persistent_huge_pages(h)) {
3736 spin_unlock_irq(&hugetlb_lock);
3737 mutex_unlock(&h->resize_lock);
3738 NODEMASK_FREE(node_alloc_noretry);
3741 /* Fall through to decrease pool */
3745 * Increase the pool size
3746 * First take pages out of surplus state. Then make up the
3747 * remaining difference by allocating fresh huge pages.
3749 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3750 * to convert a surplus huge page to a normal huge page. That is
3751 * not critical, though, it just means the overall size of the
3752 * pool might be one hugepage larger than it needs to be, but
3753 * within all the constraints specified by the sysctls.
3755 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3756 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3761 while (count > (persistent_huge_pages(h) + allocated)) {
3763 * If this allocation races such that we no longer need the
3764 * page, free_huge_folio will handle it by freeing the page
3765 * and reducing the surplus.
3767 spin_unlock_irq(&hugetlb_lock);
3769 /* yield cpu to avoid soft lockup */
3772 folio = alloc_pool_huge_folio(h, nodes_allowed,
3773 node_alloc_noretry);
3775 prep_and_add_allocated_folios(h, &page_list);
3776 spin_lock_irq(&hugetlb_lock);
3780 list_add(&folio->lru, &page_list);
3783 /* Bail for signals. Probably ctrl-c from user */
3784 if (signal_pending(current)) {
3785 prep_and_add_allocated_folios(h, &page_list);
3786 spin_lock_irq(&hugetlb_lock);
3790 spin_lock_irq(&hugetlb_lock);
3793 /* Add allocated pages to the pool */
3794 if (!list_empty(&page_list)) {
3795 spin_unlock_irq(&hugetlb_lock);
3796 prep_and_add_allocated_folios(h, &page_list);
3797 spin_lock_irq(&hugetlb_lock);
3801 * Decrease the pool size
3802 * First return free pages to the buddy allocator (being careful
3803 * to keep enough around to satisfy reservations). Then place
3804 * pages into surplus state as needed so the pool will shrink
3805 * to the desired size as pages become free.
3807 * By placing pages into the surplus state independent of the
3808 * overcommit value, we are allowing the surplus pool size to
3809 * exceed overcommit. There are few sane options here. Since
3810 * alloc_surplus_hugetlb_folio() is checking the global counter,
3811 * though, we'll note that we're not allowed to exceed surplus
3812 * and won't grow the pool anywhere else. Not until one of the
3813 * sysctls are changed, or the surplus pages go out of use.
3815 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3816 min_count = max(count, min_count);
3817 try_to_free_low(h, min_count, nodes_allowed);
3820 * Collect pages to be removed on list without dropping lock
3822 while (min_count < persistent_huge_pages(h)) {
3823 folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0);
3827 list_add(&folio->lru, &page_list);
3829 /* free the pages after dropping lock */
3830 spin_unlock_irq(&hugetlb_lock);
3831 update_and_free_pages_bulk(h, &page_list);
3832 flush_free_hpage_work(h);
3833 spin_lock_irq(&hugetlb_lock);
3835 while (count < persistent_huge_pages(h)) {
3836 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3840 h->max_huge_pages = persistent_huge_pages(h);
3841 spin_unlock_irq(&hugetlb_lock);
3842 mutex_unlock(&h->resize_lock);
3844 NODEMASK_FREE(node_alloc_noretry);
3849 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3851 int i, nid = folio_nid(folio);
3852 struct hstate *target_hstate;
3853 struct page *subpage;
3854 struct folio *inner_folio;
3857 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3859 remove_hugetlb_folio_for_demote(h, folio, false);
3860 spin_unlock_irq(&hugetlb_lock);
3863 * If vmemmap already existed for folio, the remove routine above would
3864 * have cleared the hugetlb folio flag. Hence the folio is technically
3865 * no longer a hugetlb folio. hugetlb_vmemmap_restore_folio can only be
3866 * passed hugetlb folios and will BUG otherwise.
3868 if (folio_test_hugetlb(folio)) {
3869 rc = hugetlb_vmemmap_restore_folio(h, folio);
3871 /* Allocation of vmemmmap failed, we can not demote folio */
3872 spin_lock_irq(&hugetlb_lock);
3873 folio_ref_unfreeze(folio, 1);
3874 add_hugetlb_folio(h, folio, false);
3880 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3881 * sizes as it will not ref count folios.
3883 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3886 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3887 * Without the mutex, pages added to target hstate could be marked
3890 * Note that we already hold h->resize_lock. To prevent deadlock,
3891 * use the convention of always taking larger size hstate mutex first.
3893 mutex_lock(&target_hstate->resize_lock);
3894 for (i = 0; i < pages_per_huge_page(h);
3895 i += pages_per_huge_page(target_hstate)) {
3896 subpage = folio_page(folio, i);
3897 inner_folio = page_folio(subpage);
3898 if (hstate_is_gigantic(target_hstate))
3899 prep_compound_gigantic_folio_for_demote(inner_folio,
3900 target_hstate->order);
3902 prep_compound_page(subpage, target_hstate->order);
3903 folio_change_private(inner_folio, NULL);
3904 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3905 free_huge_folio(inner_folio);
3907 mutex_unlock(&target_hstate->resize_lock);
3909 spin_lock_irq(&hugetlb_lock);
3912 * Not absolutely necessary, but for consistency update max_huge_pages
3913 * based on pool changes for the demoted page.
3915 h->max_huge_pages--;
3916 target_hstate->max_huge_pages +=
3917 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3922 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3923 __must_hold(&hugetlb_lock)
3926 struct folio *folio;
3928 lockdep_assert_held(&hugetlb_lock);
3930 /* We should never get here if no demote order */
3931 if (!h->demote_order) {
3932 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3933 return -EINVAL; /* internal error */
3936 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3937 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3938 if (folio_test_hwpoison(folio))
3940 return demote_free_hugetlb_folio(h, folio);
3945 * Only way to get here is if all pages on free lists are poisoned.
3946 * Return -EBUSY so that caller will not retry.
3951 #define HSTATE_ATTR_RO(_name) \
3952 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3954 #define HSTATE_ATTR_WO(_name) \
3955 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3957 #define HSTATE_ATTR(_name) \
3958 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3960 static struct kobject *hugepages_kobj;
3961 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3963 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3965 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3969 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3970 if (hstate_kobjs[i] == kobj) {
3972 *nidp = NUMA_NO_NODE;
3976 return kobj_to_node_hstate(kobj, nidp);
3979 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3980 struct kobj_attribute *attr, char *buf)
3983 unsigned long nr_huge_pages;
3986 h = kobj_to_hstate(kobj, &nid);
3987 if (nid == NUMA_NO_NODE)
3988 nr_huge_pages = h->nr_huge_pages;
3990 nr_huge_pages = h->nr_huge_pages_node[nid];
3992 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3995 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3996 struct hstate *h, int nid,
3997 unsigned long count, size_t len)
4000 nodemask_t nodes_allowed, *n_mask;
4002 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
4005 if (nid == NUMA_NO_NODE) {
4007 * global hstate attribute
4009 if (!(obey_mempolicy &&
4010 init_nodemask_of_mempolicy(&nodes_allowed)))
4011 n_mask = &node_states[N_MEMORY];
4013 n_mask = &nodes_allowed;
4016 * Node specific request. count adjustment happens in
4017 * set_max_huge_pages() after acquiring hugetlb_lock.
4019 init_nodemask_of_node(&nodes_allowed, nid);
4020 n_mask = &nodes_allowed;
4023 err = set_max_huge_pages(h, count, nid, n_mask);
4025 return err ? err : len;
4028 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
4029 struct kobject *kobj, const char *buf,
4033 unsigned long count;
4037 err = kstrtoul(buf, 10, &count);
4041 h = kobj_to_hstate(kobj, &nid);
4042 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
4045 static ssize_t nr_hugepages_show(struct kobject *kobj,
4046 struct kobj_attribute *attr, char *buf)
4048 return nr_hugepages_show_common(kobj, attr, buf);
4051 static ssize_t nr_hugepages_store(struct kobject *kobj,
4052 struct kobj_attribute *attr, const char *buf, size_t len)
4054 return nr_hugepages_store_common(false, kobj, buf, len);
4056 HSTATE_ATTR(nr_hugepages);
4061 * hstate attribute for optionally mempolicy-based constraint on persistent
4062 * huge page alloc/free.
4064 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
4065 struct kobj_attribute *attr,
4068 return nr_hugepages_show_common(kobj, attr, buf);
4071 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
4072 struct kobj_attribute *attr, const char *buf, size_t len)
4074 return nr_hugepages_store_common(true, kobj, buf, len);
4076 HSTATE_ATTR(nr_hugepages_mempolicy);
4080 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
4081 struct kobj_attribute *attr, char *buf)
4083 struct hstate *h = kobj_to_hstate(kobj, NULL);
4084 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
4087 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
4088 struct kobj_attribute *attr, const char *buf, size_t count)
4091 unsigned long input;
4092 struct hstate *h = kobj_to_hstate(kobj, NULL);
4094 if (hstate_is_gigantic(h))
4097 err = kstrtoul(buf, 10, &input);
4101 spin_lock_irq(&hugetlb_lock);
4102 h->nr_overcommit_huge_pages = input;
4103 spin_unlock_irq(&hugetlb_lock);
4107 HSTATE_ATTR(nr_overcommit_hugepages);
4109 static ssize_t free_hugepages_show(struct kobject *kobj,
4110 struct kobj_attribute *attr, char *buf)
4113 unsigned long free_huge_pages;
4116 h = kobj_to_hstate(kobj, &nid);
4117 if (nid == NUMA_NO_NODE)
4118 free_huge_pages = h->free_huge_pages;
4120 free_huge_pages = h->free_huge_pages_node[nid];
4122 return sysfs_emit(buf, "%lu\n", free_huge_pages);
4124 HSTATE_ATTR_RO(free_hugepages);
4126 static ssize_t resv_hugepages_show(struct kobject *kobj,
4127 struct kobj_attribute *attr, char *buf)
4129 struct hstate *h = kobj_to_hstate(kobj, NULL);
4130 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
4132 HSTATE_ATTR_RO(resv_hugepages);
4134 static ssize_t surplus_hugepages_show(struct kobject *kobj,
4135 struct kobj_attribute *attr, char *buf)
4138 unsigned long surplus_huge_pages;
4141 h = kobj_to_hstate(kobj, &nid);
4142 if (nid == NUMA_NO_NODE)
4143 surplus_huge_pages = h->surplus_huge_pages;
4145 surplus_huge_pages = h->surplus_huge_pages_node[nid];
4147 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
4149 HSTATE_ATTR_RO(surplus_hugepages);
4151 static ssize_t demote_store(struct kobject *kobj,
4152 struct kobj_attribute *attr, const char *buf, size_t len)
4154 unsigned long nr_demote;
4155 unsigned long nr_available;
4156 nodemask_t nodes_allowed, *n_mask;
4161 err = kstrtoul(buf, 10, &nr_demote);
4164 h = kobj_to_hstate(kobj, &nid);
4166 if (nid != NUMA_NO_NODE) {
4167 init_nodemask_of_node(&nodes_allowed, nid);
4168 n_mask = &nodes_allowed;
4170 n_mask = &node_states[N_MEMORY];
4173 /* Synchronize with other sysfs operations modifying huge pages */
4174 mutex_lock(&h->resize_lock);
4175 spin_lock_irq(&hugetlb_lock);
4179 * Check for available pages to demote each time thorough the
4180 * loop as demote_pool_huge_page will drop hugetlb_lock.
4182 if (nid != NUMA_NO_NODE)
4183 nr_available = h->free_huge_pages_node[nid];
4185 nr_available = h->free_huge_pages;
4186 nr_available -= h->resv_huge_pages;
4190 err = demote_pool_huge_page(h, n_mask);
4197 spin_unlock_irq(&hugetlb_lock);
4198 mutex_unlock(&h->resize_lock);
4204 HSTATE_ATTR_WO(demote);
4206 static ssize_t demote_size_show(struct kobject *kobj,
4207 struct kobj_attribute *attr, char *buf)
4209 struct hstate *h = kobj_to_hstate(kobj, NULL);
4210 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
4212 return sysfs_emit(buf, "%lukB\n", demote_size);
4215 static ssize_t demote_size_store(struct kobject *kobj,
4216 struct kobj_attribute *attr,
4217 const char *buf, size_t count)
4219 struct hstate *h, *demote_hstate;
4220 unsigned long demote_size;
4221 unsigned int demote_order;
4223 demote_size = (unsigned long)memparse(buf, NULL);
4225 demote_hstate = size_to_hstate(demote_size);
4228 demote_order = demote_hstate->order;
4229 if (demote_order < HUGETLB_PAGE_ORDER)
4232 /* demote order must be smaller than hstate order */
4233 h = kobj_to_hstate(kobj, NULL);
4234 if (demote_order >= h->order)
4237 /* resize_lock synchronizes access to demote size and writes */
4238 mutex_lock(&h->resize_lock);
4239 h->demote_order = demote_order;
4240 mutex_unlock(&h->resize_lock);
4244 HSTATE_ATTR(demote_size);
4246 static struct attribute *hstate_attrs[] = {
4247 &nr_hugepages_attr.attr,
4248 &nr_overcommit_hugepages_attr.attr,
4249 &free_hugepages_attr.attr,
4250 &resv_hugepages_attr.attr,
4251 &surplus_hugepages_attr.attr,
4253 &nr_hugepages_mempolicy_attr.attr,
4258 static const struct attribute_group hstate_attr_group = {
4259 .attrs = hstate_attrs,
4262 static struct attribute *hstate_demote_attrs[] = {
4263 &demote_size_attr.attr,
4268 static const struct attribute_group hstate_demote_attr_group = {
4269 .attrs = hstate_demote_attrs,
4272 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4273 struct kobject **hstate_kobjs,
4274 const struct attribute_group *hstate_attr_group)
4277 int hi = hstate_index(h);
4279 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4280 if (!hstate_kobjs[hi])
4283 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4285 kobject_put(hstate_kobjs[hi]);
4286 hstate_kobjs[hi] = NULL;
4290 if (h->demote_order) {
4291 retval = sysfs_create_group(hstate_kobjs[hi],
4292 &hstate_demote_attr_group);
4294 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4295 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4296 kobject_put(hstate_kobjs[hi]);
4297 hstate_kobjs[hi] = NULL;
4306 static bool hugetlb_sysfs_initialized __ro_after_init;
4309 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4310 * with node devices in node_devices[] using a parallel array. The array
4311 * index of a node device or _hstate == node id.
4312 * This is here to avoid any static dependency of the node device driver, in
4313 * the base kernel, on the hugetlb module.
4315 struct node_hstate {
4316 struct kobject *hugepages_kobj;
4317 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4319 static struct node_hstate node_hstates[MAX_NUMNODES];
4322 * A subset of global hstate attributes for node devices
4324 static struct attribute *per_node_hstate_attrs[] = {
4325 &nr_hugepages_attr.attr,
4326 &free_hugepages_attr.attr,
4327 &surplus_hugepages_attr.attr,
4331 static const struct attribute_group per_node_hstate_attr_group = {
4332 .attrs = per_node_hstate_attrs,
4336 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4337 * Returns node id via non-NULL nidp.
4339 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4343 for (nid = 0; nid < nr_node_ids; nid++) {
4344 struct node_hstate *nhs = &node_hstates[nid];
4346 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4347 if (nhs->hstate_kobjs[i] == kobj) {
4359 * Unregister hstate attributes from a single node device.
4360 * No-op if no hstate attributes attached.
4362 void hugetlb_unregister_node(struct node *node)
4365 struct node_hstate *nhs = &node_hstates[node->dev.id];
4367 if (!nhs->hugepages_kobj)
4368 return; /* no hstate attributes */
4370 for_each_hstate(h) {
4371 int idx = hstate_index(h);
4372 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4376 if (h->demote_order)
4377 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4378 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4379 kobject_put(hstate_kobj);
4380 nhs->hstate_kobjs[idx] = NULL;
4383 kobject_put(nhs->hugepages_kobj);
4384 nhs->hugepages_kobj = NULL;
4389 * Register hstate attributes for a single node device.
4390 * No-op if attributes already registered.
4392 void hugetlb_register_node(struct node *node)
4395 struct node_hstate *nhs = &node_hstates[node->dev.id];
4398 if (!hugetlb_sysfs_initialized)
4401 if (nhs->hugepages_kobj)
4402 return; /* already allocated */
4404 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4406 if (!nhs->hugepages_kobj)
4409 for_each_hstate(h) {
4410 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4412 &per_node_hstate_attr_group);
4414 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4415 h->name, node->dev.id);
4416 hugetlb_unregister_node(node);
4423 * hugetlb init time: register hstate attributes for all registered node
4424 * devices of nodes that have memory. All on-line nodes should have
4425 * registered their associated device by this time.
4427 static void __init hugetlb_register_all_nodes(void)
4431 for_each_online_node(nid)
4432 hugetlb_register_node(node_devices[nid]);
4434 #else /* !CONFIG_NUMA */
4436 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4444 static void hugetlb_register_all_nodes(void) { }
4449 static void __init hugetlb_cma_check(void);
4451 static inline __init void hugetlb_cma_check(void)
4456 static void __init hugetlb_sysfs_init(void)
4461 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4462 if (!hugepages_kobj)
4465 for_each_hstate(h) {
4466 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4467 hstate_kobjs, &hstate_attr_group);
4469 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4473 hugetlb_sysfs_initialized = true;
4475 hugetlb_register_all_nodes();
4478 #ifdef CONFIG_SYSCTL
4479 static void hugetlb_sysctl_init(void);
4481 static inline void hugetlb_sysctl_init(void) { }
4484 static int __init hugetlb_init(void)
4488 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4491 if (!hugepages_supported()) {
4492 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4493 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4498 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4499 * architectures depend on setup being done here.
4501 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4502 if (!parsed_default_hugepagesz) {
4504 * If we did not parse a default huge page size, set
4505 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4506 * number of huge pages for this default size was implicitly
4507 * specified, set that here as well.
4508 * Note that the implicit setting will overwrite an explicit
4509 * setting. A warning will be printed in this case.
4511 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4512 if (default_hstate_max_huge_pages) {
4513 if (default_hstate.max_huge_pages) {
4516 string_get_size(huge_page_size(&default_hstate),
4517 1, STRING_UNITS_2, buf, 32);
4518 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4519 default_hstate.max_huge_pages, buf);
4520 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4521 default_hstate_max_huge_pages);
4523 default_hstate.max_huge_pages =
4524 default_hstate_max_huge_pages;
4526 for_each_online_node(i)
4527 default_hstate.max_huge_pages_node[i] =
4528 default_hugepages_in_node[i];
4532 hugetlb_cma_check();
4533 hugetlb_init_hstates();
4534 gather_bootmem_prealloc();
4537 hugetlb_sysfs_init();
4538 hugetlb_cgroup_file_init();
4539 hugetlb_sysctl_init();
4542 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4544 num_fault_mutexes = 1;
4546 hugetlb_fault_mutex_table =
4547 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4549 BUG_ON(!hugetlb_fault_mutex_table);
4551 for (i = 0; i < num_fault_mutexes; i++)
4552 mutex_init(&hugetlb_fault_mutex_table[i]);
4555 subsys_initcall(hugetlb_init);
4557 /* Overwritten by architectures with more huge page sizes */
4558 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4560 return size == HPAGE_SIZE;
4563 void __init hugetlb_add_hstate(unsigned int order)
4568 if (size_to_hstate(PAGE_SIZE << order)) {
4571 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4572 BUG_ON(order < order_base_2(__NR_USED_SUBPAGE));
4573 h = &hstates[hugetlb_max_hstate++];
4574 mutex_init(&h->resize_lock);
4576 h->mask = ~(huge_page_size(h) - 1);
4577 for (i = 0; i < MAX_NUMNODES; ++i)
4578 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4579 INIT_LIST_HEAD(&h->hugepage_activelist);
4580 h->next_nid_to_alloc = first_memory_node;
4581 h->next_nid_to_free = first_memory_node;
4582 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4583 huge_page_size(h)/SZ_1K);
4588 bool __init __weak hugetlb_node_alloc_supported(void)
4593 static void __init hugepages_clear_pages_in_node(void)
4595 if (!hugetlb_max_hstate) {
4596 default_hstate_max_huge_pages = 0;
4597 memset(default_hugepages_in_node, 0,
4598 sizeof(default_hugepages_in_node));
4600 parsed_hstate->max_huge_pages = 0;
4601 memset(parsed_hstate->max_huge_pages_node, 0,
4602 sizeof(parsed_hstate->max_huge_pages_node));
4607 * hugepages command line processing
4608 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4609 * specification. If not, ignore the hugepages value. hugepages can also
4610 * be the first huge page command line option in which case it implicitly
4611 * specifies the number of huge pages for the default size.
4613 static int __init hugepages_setup(char *s)
4616 static unsigned long *last_mhp;
4617 int node = NUMA_NO_NODE;
4622 if (!parsed_valid_hugepagesz) {
4623 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4624 parsed_valid_hugepagesz = true;
4629 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4630 * yet, so this hugepages= parameter goes to the "default hstate".
4631 * Otherwise, it goes with the previously parsed hugepagesz or
4632 * default_hugepagesz.
4634 else if (!hugetlb_max_hstate)
4635 mhp = &default_hstate_max_huge_pages;
4637 mhp = &parsed_hstate->max_huge_pages;
4639 if (mhp == last_mhp) {
4640 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4646 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4648 /* Parameter is node format */
4649 if (p[count] == ':') {
4650 if (!hugetlb_node_alloc_supported()) {
4651 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4654 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4656 node = array_index_nospec(tmp, MAX_NUMNODES);
4658 /* Parse hugepages */
4659 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4661 if (!hugetlb_max_hstate)
4662 default_hugepages_in_node[node] = tmp;
4664 parsed_hstate->max_huge_pages_node[node] = tmp;
4666 /* Go to parse next node*/
4667 if (p[count] == ',')
4680 * Global state is always initialized later in hugetlb_init.
4681 * But we need to allocate gigantic hstates here early to still
4682 * use the bootmem allocator.
4684 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4685 hugetlb_hstate_alloc_pages(parsed_hstate);
4692 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4693 hugepages_clear_pages_in_node();
4696 __setup("hugepages=", hugepages_setup);
4699 * hugepagesz command line processing
4700 * A specific huge page size can only be specified once with hugepagesz.
4701 * hugepagesz is followed by hugepages on the command line. The global
4702 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4703 * hugepagesz argument was valid.
4705 static int __init hugepagesz_setup(char *s)
4710 parsed_valid_hugepagesz = false;
4711 size = (unsigned long)memparse(s, NULL);
4713 if (!arch_hugetlb_valid_size(size)) {
4714 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4718 h = size_to_hstate(size);
4721 * hstate for this size already exists. This is normally
4722 * an error, but is allowed if the existing hstate is the
4723 * default hstate. More specifically, it is only allowed if
4724 * the number of huge pages for the default hstate was not
4725 * previously specified.
4727 if (!parsed_default_hugepagesz || h != &default_hstate ||
4728 default_hstate.max_huge_pages) {
4729 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4734 * No need to call hugetlb_add_hstate() as hstate already
4735 * exists. But, do set parsed_hstate so that a following
4736 * hugepages= parameter will be applied to this hstate.
4739 parsed_valid_hugepagesz = true;
4743 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4744 parsed_valid_hugepagesz = true;
4747 __setup("hugepagesz=", hugepagesz_setup);
4750 * default_hugepagesz command line input
4751 * Only one instance of default_hugepagesz allowed on command line.
4753 static int __init default_hugepagesz_setup(char *s)
4758 parsed_valid_hugepagesz = false;
4759 if (parsed_default_hugepagesz) {
4760 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4764 size = (unsigned long)memparse(s, NULL);
4766 if (!arch_hugetlb_valid_size(size)) {
4767 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4771 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4772 parsed_valid_hugepagesz = true;
4773 parsed_default_hugepagesz = true;
4774 default_hstate_idx = hstate_index(size_to_hstate(size));
4777 * The number of default huge pages (for this size) could have been
4778 * specified as the first hugetlb parameter: hugepages=X. If so,
4779 * then default_hstate_max_huge_pages is set. If the default huge
4780 * page size is gigantic (> MAX_PAGE_ORDER), then the pages must be
4781 * allocated here from bootmem allocator.
4783 if (default_hstate_max_huge_pages) {
4784 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4785 for_each_online_node(i)
4786 default_hstate.max_huge_pages_node[i] =
4787 default_hugepages_in_node[i];
4788 if (hstate_is_gigantic(&default_hstate))
4789 hugetlb_hstate_alloc_pages(&default_hstate);
4790 default_hstate_max_huge_pages = 0;
4795 __setup("default_hugepagesz=", default_hugepagesz_setup);
4797 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4800 struct mempolicy *mpol = get_task_policy(current);
4803 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4804 * (from policy_nodemask) specifically for hugetlb case
4806 if (mpol->mode == MPOL_BIND &&
4807 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4808 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4809 return &mpol->nodes;
4814 static unsigned int allowed_mems_nr(struct hstate *h)
4817 unsigned int nr = 0;
4818 nodemask_t *mbind_nodemask;
4819 unsigned int *array = h->free_huge_pages_node;
4820 gfp_t gfp_mask = htlb_alloc_mask(h);
4822 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4823 for_each_node_mask(node, cpuset_current_mems_allowed) {
4824 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4831 #ifdef CONFIG_SYSCTL
4832 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4833 void *buffer, size_t *length,
4834 loff_t *ppos, unsigned long *out)
4836 struct ctl_table dup_table;
4839 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4840 * can duplicate the @table and alter the duplicate of it.
4843 dup_table.data = out;
4845 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4848 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4849 struct ctl_table *table, int write,
4850 void *buffer, size_t *length, loff_t *ppos)
4852 struct hstate *h = &default_hstate;
4853 unsigned long tmp = h->max_huge_pages;
4856 if (!hugepages_supported())
4859 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4865 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4866 NUMA_NO_NODE, tmp, *length);
4871 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4872 void *buffer, size_t *length, loff_t *ppos)
4875 return hugetlb_sysctl_handler_common(false, table, write,
4876 buffer, length, ppos);
4880 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4881 void *buffer, size_t *length, loff_t *ppos)
4883 return hugetlb_sysctl_handler_common(true, table, write,
4884 buffer, length, ppos);
4886 #endif /* CONFIG_NUMA */
4888 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4889 void *buffer, size_t *length, loff_t *ppos)
4891 struct hstate *h = &default_hstate;
4895 if (!hugepages_supported())
4898 tmp = h->nr_overcommit_huge_pages;
4900 if (write && hstate_is_gigantic(h))
4903 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4909 spin_lock_irq(&hugetlb_lock);
4910 h->nr_overcommit_huge_pages = tmp;
4911 spin_unlock_irq(&hugetlb_lock);
4917 static struct ctl_table hugetlb_table[] = {
4919 .procname = "nr_hugepages",
4921 .maxlen = sizeof(unsigned long),
4923 .proc_handler = hugetlb_sysctl_handler,
4927 .procname = "nr_hugepages_mempolicy",
4929 .maxlen = sizeof(unsigned long),
4931 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4935 .procname = "hugetlb_shm_group",
4936 .data = &sysctl_hugetlb_shm_group,
4937 .maxlen = sizeof(gid_t),
4939 .proc_handler = proc_dointvec,
4942 .procname = "nr_overcommit_hugepages",
4944 .maxlen = sizeof(unsigned long),
4946 .proc_handler = hugetlb_overcommit_handler,
4951 static void hugetlb_sysctl_init(void)
4953 register_sysctl_init("vm", hugetlb_table);
4955 #endif /* CONFIG_SYSCTL */
4957 void hugetlb_report_meminfo(struct seq_file *m)
4960 unsigned long total = 0;
4962 if (!hugepages_supported())
4965 for_each_hstate(h) {
4966 unsigned long count = h->nr_huge_pages;
4968 total += huge_page_size(h) * count;
4970 if (h == &default_hstate)
4972 "HugePages_Total: %5lu\n"
4973 "HugePages_Free: %5lu\n"
4974 "HugePages_Rsvd: %5lu\n"
4975 "HugePages_Surp: %5lu\n"
4976 "Hugepagesize: %8lu kB\n",
4980 h->surplus_huge_pages,
4981 huge_page_size(h) / SZ_1K);
4984 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4987 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4989 struct hstate *h = &default_hstate;
4991 if (!hugepages_supported())
4994 return sysfs_emit_at(buf, len,
4995 "Node %d HugePages_Total: %5u\n"
4996 "Node %d HugePages_Free: %5u\n"
4997 "Node %d HugePages_Surp: %5u\n",
4998 nid, h->nr_huge_pages_node[nid],
4999 nid, h->free_huge_pages_node[nid],
5000 nid, h->surplus_huge_pages_node[nid]);
5003 void hugetlb_show_meminfo_node(int nid)
5007 if (!hugepages_supported())
5011 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
5013 h->nr_huge_pages_node[nid],
5014 h->free_huge_pages_node[nid],
5015 h->surplus_huge_pages_node[nid],
5016 huge_page_size(h) / SZ_1K);
5019 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
5021 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
5022 K(atomic_long_read(&mm->hugetlb_usage)));
5025 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
5026 unsigned long hugetlb_total_pages(void)
5029 unsigned long nr_total_pages = 0;
5032 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
5033 return nr_total_pages;
5036 static int hugetlb_acct_memory(struct hstate *h, long delta)
5043 spin_lock_irq(&hugetlb_lock);
5045 * When cpuset is configured, it breaks the strict hugetlb page
5046 * reservation as the accounting is done on a global variable. Such
5047 * reservation is completely rubbish in the presence of cpuset because
5048 * the reservation is not checked against page availability for the
5049 * current cpuset. Application can still potentially OOM'ed by kernel
5050 * with lack of free htlb page in cpuset that the task is in.
5051 * Attempt to enforce strict accounting with cpuset is almost
5052 * impossible (or too ugly) because cpuset is too fluid that
5053 * task or memory node can be dynamically moved between cpusets.
5055 * The change of semantics for shared hugetlb mapping with cpuset is
5056 * undesirable. However, in order to preserve some of the semantics,
5057 * we fall back to check against current free page availability as
5058 * a best attempt and hopefully to minimize the impact of changing
5059 * semantics that cpuset has.
5061 * Apart from cpuset, we also have memory policy mechanism that
5062 * also determines from which node the kernel will allocate memory
5063 * in a NUMA system. So similar to cpuset, we also should consider
5064 * the memory policy of the current task. Similar to the description
5068 if (gather_surplus_pages(h, delta) < 0)
5071 if (delta > allowed_mems_nr(h)) {
5072 return_unused_surplus_pages(h, delta);
5079 return_unused_surplus_pages(h, (unsigned long) -delta);
5082 spin_unlock_irq(&hugetlb_lock);
5086 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
5088 struct resv_map *resv = vma_resv_map(vma);
5091 * HPAGE_RESV_OWNER indicates a private mapping.
5092 * This new VMA should share its siblings reservation map if present.
5093 * The VMA will only ever have a valid reservation map pointer where
5094 * it is being copied for another still existing VMA. As that VMA
5095 * has a reference to the reservation map it cannot disappear until
5096 * after this open call completes. It is therefore safe to take a
5097 * new reference here without additional locking.
5099 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
5100 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
5101 kref_get(&resv->refs);
5105 * vma_lock structure for sharable mappings is vma specific.
5106 * Clear old pointer (if copied via vm_area_dup) and allocate
5107 * new structure. Before clearing, make sure vma_lock is not
5110 if (vma->vm_flags & VM_MAYSHARE) {
5111 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
5114 if (vma_lock->vma != vma) {
5115 vma->vm_private_data = NULL;
5116 hugetlb_vma_lock_alloc(vma);
5118 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
5120 hugetlb_vma_lock_alloc(vma);
5124 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
5126 struct hstate *h = hstate_vma(vma);
5127 struct resv_map *resv;
5128 struct hugepage_subpool *spool = subpool_vma(vma);
5129 unsigned long reserve, start, end;
5132 hugetlb_vma_lock_free(vma);
5134 resv = vma_resv_map(vma);
5135 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5138 start = vma_hugecache_offset(h, vma, vma->vm_start);
5139 end = vma_hugecache_offset(h, vma, vma->vm_end);
5141 reserve = (end - start) - region_count(resv, start, end);
5142 hugetlb_cgroup_uncharge_counter(resv, start, end);
5145 * Decrement reserve counts. The global reserve count may be
5146 * adjusted if the subpool has a minimum size.
5148 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
5149 hugetlb_acct_memory(h, -gbl_reserve);
5152 kref_put(&resv->refs, resv_map_release);
5155 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
5157 if (addr & ~(huge_page_mask(hstate_vma(vma))))
5161 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
5162 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
5163 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
5165 if (addr & ~PUD_MASK) {
5167 * hugetlb_vm_op_split is called right before we attempt to
5168 * split the VMA. We will need to unshare PMDs in the old and
5169 * new VMAs, so let's unshare before we split.
5171 unsigned long floor = addr & PUD_MASK;
5172 unsigned long ceil = floor + PUD_SIZE;
5174 if (floor >= vma->vm_start && ceil <= vma->vm_end)
5175 hugetlb_unshare_pmds(vma, floor, ceil);
5181 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
5183 return huge_page_size(hstate_vma(vma));
5187 * We cannot handle pagefaults against hugetlb pages at all. They cause
5188 * handle_mm_fault() to try to instantiate regular-sized pages in the
5189 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
5192 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
5199 * When a new function is introduced to vm_operations_struct and added
5200 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
5201 * This is because under System V memory model, mappings created via
5202 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
5203 * their original vm_ops are overwritten with shm_vm_ops.
5205 const struct vm_operations_struct hugetlb_vm_ops = {
5206 .fault = hugetlb_vm_op_fault,
5207 .open = hugetlb_vm_op_open,
5208 .close = hugetlb_vm_op_close,
5209 .may_split = hugetlb_vm_op_split,
5210 .pagesize = hugetlb_vm_op_pagesize,
5213 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
5217 unsigned int shift = huge_page_shift(hstate_vma(vma));
5220 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
5221 vma->vm_page_prot)));
5223 entry = huge_pte_wrprotect(mk_huge_pte(page,
5224 vma->vm_page_prot));
5226 entry = pte_mkyoung(entry);
5227 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
5232 static void set_huge_ptep_writable(struct vm_area_struct *vma,
5233 unsigned long address, pte_t *ptep)
5237 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
5238 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
5239 update_mmu_cache(vma, address, ptep);
5242 bool is_hugetlb_entry_migration(pte_t pte)
5246 if (huge_pte_none(pte) || pte_present(pte))
5248 swp = pte_to_swp_entry(pte);
5249 if (is_migration_entry(swp))
5255 bool is_hugetlb_entry_hwpoisoned(pte_t pte)
5259 if (huge_pte_none(pte) || pte_present(pte))
5261 swp = pte_to_swp_entry(pte);
5262 if (is_hwpoison_entry(swp))
5269 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5270 struct folio *new_folio, pte_t old, unsigned long sz)
5272 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5274 __folio_mark_uptodate(new_folio);
5275 hugetlb_add_new_anon_rmap(new_folio, vma, addr);
5276 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5277 newpte = huge_pte_mkuffd_wp(newpte);
5278 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5279 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5280 folio_set_hugetlb_migratable(new_folio);
5283 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5284 struct vm_area_struct *dst_vma,
5285 struct vm_area_struct *src_vma)
5287 pte_t *src_pte, *dst_pte, entry;
5288 struct folio *pte_folio;
5290 bool cow = is_cow_mapping(src_vma->vm_flags);
5291 struct hstate *h = hstate_vma(src_vma);
5292 unsigned long sz = huge_page_size(h);
5293 unsigned long npages = pages_per_huge_page(h);
5294 struct mmu_notifier_range range;
5295 unsigned long last_addr_mask;
5299 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5302 mmu_notifier_invalidate_range_start(&range);
5303 vma_assert_write_locked(src_vma);
5304 raw_write_seqcount_begin(&src->write_protect_seq);
5307 * For shared mappings the vma lock must be held before
5308 * calling hugetlb_walk() in the src vma. Otherwise, the
5309 * returned ptep could go away if part of a shared pmd and
5310 * another thread calls huge_pmd_unshare.
5312 hugetlb_vma_lock_read(src_vma);
5315 last_addr_mask = hugetlb_mask_last_page(h);
5316 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5317 spinlock_t *src_ptl, *dst_ptl;
5318 src_pte = hugetlb_walk(src_vma, addr, sz);
5320 addr |= last_addr_mask;
5323 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5330 * If the pagetables are shared don't copy or take references.
5332 * dst_pte == src_pte is the common case of src/dest sharing.
5333 * However, src could have 'unshared' and dst shares with
5334 * another vma. So page_count of ptep page is checked instead
5335 * to reliably determine whether pte is shared.
5337 if (page_count(virt_to_page(dst_pte)) > 1) {
5338 addr |= last_addr_mask;
5342 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5343 src_ptl = huge_pte_lockptr(h, src, src_pte);
5344 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5345 entry = huge_ptep_get(src_pte);
5347 if (huge_pte_none(entry)) {
5349 * Skip if src entry none.
5352 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5353 if (!userfaultfd_wp(dst_vma))
5354 entry = huge_pte_clear_uffd_wp(entry);
5355 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5356 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5357 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5358 bool uffd_wp = pte_swp_uffd_wp(entry);
5360 if (!is_readable_migration_entry(swp_entry) && cow) {
5362 * COW mappings require pages in both
5363 * parent and child to be set to read.
5365 swp_entry = make_readable_migration_entry(
5366 swp_offset(swp_entry));
5367 entry = swp_entry_to_pte(swp_entry);
5368 if (userfaultfd_wp(src_vma) && uffd_wp)
5369 entry = pte_swp_mkuffd_wp(entry);
5370 set_huge_pte_at(src, addr, src_pte, entry, sz);
5372 if (!userfaultfd_wp(dst_vma))
5373 entry = huge_pte_clear_uffd_wp(entry);
5374 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5375 } else if (unlikely(is_pte_marker(entry))) {
5376 pte_marker marker = copy_pte_marker(
5377 pte_to_swp_entry(entry), dst_vma);
5380 set_huge_pte_at(dst, addr, dst_pte,
5381 make_pte_marker(marker), sz);
5383 entry = huge_ptep_get(src_pte);
5384 pte_folio = page_folio(pte_page(entry));
5385 folio_get(pte_folio);
5388 * Failing to duplicate the anon rmap is a rare case
5389 * where we see pinned hugetlb pages while they're
5390 * prone to COW. We need to do the COW earlier during
5393 * When pre-allocating the page or copying data, we
5394 * need to be without the pgtable locks since we could
5395 * sleep during the process.
5397 if (!folio_test_anon(pte_folio)) {
5398 hugetlb_add_file_rmap(pte_folio);
5399 } else if (hugetlb_try_dup_anon_rmap(pte_folio, src_vma)) {
5400 pte_t src_pte_old = entry;
5401 struct folio *new_folio;
5403 spin_unlock(src_ptl);
5404 spin_unlock(dst_ptl);
5405 /* Do not use reserve as it's private owned */
5406 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5407 if (IS_ERR(new_folio)) {
5408 folio_put(pte_folio);
5409 ret = PTR_ERR(new_folio);
5412 ret = copy_user_large_folio(new_folio,
5415 folio_put(pte_folio);
5417 folio_put(new_folio);
5421 /* Install the new hugetlb folio if src pte stable */
5422 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5423 src_ptl = huge_pte_lockptr(h, src, src_pte);
5424 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5425 entry = huge_ptep_get(src_pte);
5426 if (!pte_same(src_pte_old, entry)) {
5427 restore_reserve_on_error(h, dst_vma, addr,
5429 folio_put(new_folio);
5430 /* huge_ptep of dst_pte won't change as in child */
5433 hugetlb_install_folio(dst_vma, dst_pte, addr,
5434 new_folio, src_pte_old, sz);
5435 spin_unlock(src_ptl);
5436 spin_unlock(dst_ptl);
5442 * No need to notify as we are downgrading page
5443 * table protection not changing it to point
5446 * See Documentation/mm/mmu_notifier.rst
5448 huge_ptep_set_wrprotect(src, addr, src_pte);
5449 entry = huge_pte_wrprotect(entry);
5452 if (!userfaultfd_wp(dst_vma))
5453 entry = huge_pte_clear_uffd_wp(entry);
5455 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5456 hugetlb_count_add(npages, dst);
5458 spin_unlock(src_ptl);
5459 spin_unlock(dst_ptl);
5463 raw_write_seqcount_end(&src->write_protect_seq);
5464 mmu_notifier_invalidate_range_end(&range);
5466 hugetlb_vma_unlock_read(src_vma);
5472 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5473 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5476 struct hstate *h = hstate_vma(vma);
5477 struct mm_struct *mm = vma->vm_mm;
5478 spinlock_t *src_ptl, *dst_ptl;
5481 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5482 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5485 * We don't have to worry about the ordering of src and dst ptlocks
5486 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5488 if (src_ptl != dst_ptl)
5489 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5491 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5492 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5494 if (src_ptl != dst_ptl)
5495 spin_unlock(src_ptl);
5496 spin_unlock(dst_ptl);
5499 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5500 struct vm_area_struct *new_vma,
5501 unsigned long old_addr, unsigned long new_addr,
5504 struct hstate *h = hstate_vma(vma);
5505 struct address_space *mapping = vma->vm_file->f_mapping;
5506 unsigned long sz = huge_page_size(h);
5507 struct mm_struct *mm = vma->vm_mm;
5508 unsigned long old_end = old_addr + len;
5509 unsigned long last_addr_mask;
5510 pte_t *src_pte, *dst_pte;
5511 struct mmu_notifier_range range;
5512 bool shared_pmd = false;
5514 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5516 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5518 * In case of shared PMDs, we should cover the maximum possible
5521 flush_cache_range(vma, range.start, range.end);
5523 mmu_notifier_invalidate_range_start(&range);
5524 last_addr_mask = hugetlb_mask_last_page(h);
5525 /* Prevent race with file truncation */
5526 hugetlb_vma_lock_write(vma);
5527 i_mmap_lock_write(mapping);
5528 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5529 src_pte = hugetlb_walk(vma, old_addr, sz);
5531 old_addr |= last_addr_mask;
5532 new_addr |= last_addr_mask;
5535 if (huge_pte_none(huge_ptep_get(src_pte)))
5538 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5540 old_addr |= last_addr_mask;
5541 new_addr |= last_addr_mask;
5545 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5549 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5553 flush_hugetlb_tlb_range(vma, range.start, range.end);
5555 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5556 mmu_notifier_invalidate_range_end(&range);
5557 i_mmap_unlock_write(mapping);
5558 hugetlb_vma_unlock_write(vma);
5560 return len + old_addr - old_end;
5563 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5564 unsigned long start, unsigned long end,
5565 struct page *ref_page, zap_flags_t zap_flags)
5567 struct mm_struct *mm = vma->vm_mm;
5568 unsigned long address;
5573 struct hstate *h = hstate_vma(vma);
5574 unsigned long sz = huge_page_size(h);
5575 unsigned long last_addr_mask;
5576 bool force_flush = false;
5578 WARN_ON(!is_vm_hugetlb_page(vma));
5579 BUG_ON(start & ~huge_page_mask(h));
5580 BUG_ON(end & ~huge_page_mask(h));
5583 * This is a hugetlb vma, all the pte entries should point
5586 tlb_change_page_size(tlb, sz);
5587 tlb_start_vma(tlb, vma);
5589 last_addr_mask = hugetlb_mask_last_page(h);
5591 for (; address < end; address += sz) {
5592 ptep = hugetlb_walk(vma, address, sz);
5594 address |= last_addr_mask;
5598 ptl = huge_pte_lock(h, mm, ptep);
5599 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5601 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5603 address |= last_addr_mask;
5607 pte = huge_ptep_get(ptep);
5608 if (huge_pte_none(pte)) {
5614 * Migrating hugepage or HWPoisoned hugepage is already
5615 * unmapped and its refcount is dropped, so just clear pte here.
5617 if (unlikely(!pte_present(pte))) {
5619 * If the pte was wr-protected by uffd-wp in any of the
5620 * swap forms, meanwhile the caller does not want to
5621 * drop the uffd-wp bit in this zap, then replace the
5622 * pte with a marker.
5624 if (pte_swp_uffd_wp_any(pte) &&
5625 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5626 set_huge_pte_at(mm, address, ptep,
5627 make_pte_marker(PTE_MARKER_UFFD_WP),
5630 huge_pte_clear(mm, address, ptep, sz);
5635 page = pte_page(pte);
5637 * If a reference page is supplied, it is because a specific
5638 * page is being unmapped, not a range. Ensure the page we
5639 * are about to unmap is the actual page of interest.
5642 if (page != ref_page) {
5647 * Mark the VMA as having unmapped its page so that
5648 * future faults in this VMA will fail rather than
5649 * looking like data was lost
5651 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5654 pte = huge_ptep_get_and_clear(mm, address, ptep);
5655 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5656 if (huge_pte_dirty(pte))
5657 set_page_dirty(page);
5658 /* Leave a uffd-wp pte marker if needed */
5659 if (huge_pte_uffd_wp(pte) &&
5660 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5661 set_huge_pte_at(mm, address, ptep,
5662 make_pte_marker(PTE_MARKER_UFFD_WP),
5664 hugetlb_count_sub(pages_per_huge_page(h), mm);
5665 hugetlb_remove_rmap(page_folio(page));
5668 tlb_remove_page_size(tlb, page, huge_page_size(h));
5670 * Bail out after unmapping reference page if supplied
5675 tlb_end_vma(tlb, vma);
5678 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5679 * could defer the flush until now, since by holding i_mmap_rwsem we
5680 * guaranteed that the last refernece would not be dropped. But we must
5681 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5682 * dropped and the last reference to the shared PMDs page might be
5685 * In theory we could defer the freeing of the PMD pages as well, but
5686 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5687 * detect sharing, so we cannot defer the release of the page either.
5688 * Instead, do flush now.
5691 tlb_flush_mmu_tlbonly(tlb);
5694 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5695 unsigned long *start, unsigned long *end)
5697 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5700 adjust_range_if_pmd_sharing_possible(vma, start, end);
5701 hugetlb_vma_lock_write(vma);
5703 i_mmap_lock_write(vma->vm_file->f_mapping);
5706 void __hugetlb_zap_end(struct vm_area_struct *vma,
5707 struct zap_details *details)
5709 zap_flags_t zap_flags = details ? details->zap_flags : 0;
5711 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5714 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5716 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5717 * When the vma_lock is freed, this makes the vma ineligible
5718 * for pmd sharing. And, i_mmap_rwsem is required to set up
5719 * pmd sharing. This is important as page tables for this
5720 * unmapped range will be asynchrously deleted. If the page
5721 * tables are shared, there will be issues when accessed by
5724 __hugetlb_vma_unlock_write_free(vma);
5726 hugetlb_vma_unlock_write(vma);
5730 i_mmap_unlock_write(vma->vm_file->f_mapping);
5733 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5734 unsigned long end, struct page *ref_page,
5735 zap_flags_t zap_flags)
5737 struct mmu_notifier_range range;
5738 struct mmu_gather tlb;
5740 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5742 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5743 mmu_notifier_invalidate_range_start(&range);
5744 tlb_gather_mmu(&tlb, vma->vm_mm);
5746 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5748 mmu_notifier_invalidate_range_end(&range);
5749 tlb_finish_mmu(&tlb);
5753 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5754 * mapping it owns the reserve page for. The intention is to unmap the page
5755 * from other VMAs and let the children be SIGKILLed if they are faulting the
5758 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5759 struct page *page, unsigned long address)
5761 struct hstate *h = hstate_vma(vma);
5762 struct vm_area_struct *iter_vma;
5763 struct address_space *mapping;
5767 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5768 * from page cache lookup which is in HPAGE_SIZE units.
5770 address = address & huge_page_mask(h);
5771 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5773 mapping = vma->vm_file->f_mapping;
5776 * Take the mapping lock for the duration of the table walk. As
5777 * this mapping should be shared between all the VMAs,
5778 * __unmap_hugepage_range() is called as the lock is already held
5780 i_mmap_lock_write(mapping);
5781 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5782 /* Do not unmap the current VMA */
5783 if (iter_vma == vma)
5787 * Shared VMAs have their own reserves and do not affect
5788 * MAP_PRIVATE accounting but it is possible that a shared
5789 * VMA is using the same page so check and skip such VMAs.
5791 if (iter_vma->vm_flags & VM_MAYSHARE)
5795 * Unmap the page from other VMAs without their own reserves.
5796 * They get marked to be SIGKILLed if they fault in these
5797 * areas. This is because a future no-page fault on this VMA
5798 * could insert a zeroed page instead of the data existing
5799 * from the time of fork. This would look like data corruption
5801 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5802 unmap_hugepage_range(iter_vma, address,
5803 address + huge_page_size(h), page, 0);
5805 i_mmap_unlock_write(mapping);
5809 * hugetlb_wp() should be called with page lock of the original hugepage held.
5810 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5811 * cannot race with other handlers or page migration.
5812 * Keep the pte_same checks anyway to make transition from the mutex easier.
5814 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5815 unsigned long address, pte_t *ptep, unsigned int flags,
5816 struct folio *pagecache_folio, spinlock_t *ptl)
5818 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5819 pte_t pte = huge_ptep_get(ptep);
5820 struct hstate *h = hstate_vma(vma);
5821 struct folio *old_folio;
5822 struct folio *new_folio;
5823 int outside_reserve = 0;
5825 unsigned long haddr = address & huge_page_mask(h);
5826 struct mmu_notifier_range range;
5829 * Never handle CoW for uffd-wp protected pages. It should be only
5830 * handled when the uffd-wp protection is removed.
5832 * Note that only the CoW optimization path (in hugetlb_no_page())
5833 * can trigger this, because hugetlb_fault() will always resolve
5834 * uffd-wp bit first.
5836 if (!unshare && huge_pte_uffd_wp(pte))
5840 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5841 * PTE mapped R/O such as maybe_mkwrite() would do.
5843 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5844 return VM_FAULT_SIGSEGV;
5846 /* Let's take out MAP_SHARED mappings first. */
5847 if (vma->vm_flags & VM_MAYSHARE) {
5848 set_huge_ptep_writable(vma, haddr, ptep);
5852 old_folio = page_folio(pte_page(pte));
5854 delayacct_wpcopy_start();
5858 * If no-one else is actually using this page, we're the exclusive
5859 * owner and can reuse this page.
5861 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5862 if (!PageAnonExclusive(&old_folio->page)) {
5863 folio_move_anon_rmap(old_folio, vma);
5864 SetPageAnonExclusive(&old_folio->page);
5866 if (likely(!unshare))
5867 set_huge_ptep_writable(vma, haddr, ptep);
5869 delayacct_wpcopy_end();
5872 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5873 PageAnonExclusive(&old_folio->page), &old_folio->page);
5876 * If the process that created a MAP_PRIVATE mapping is about to
5877 * perform a COW due to a shared page count, attempt to satisfy
5878 * the allocation without using the existing reserves. The pagecache
5879 * page is used to determine if the reserve at this address was
5880 * consumed or not. If reserves were used, a partial faulted mapping
5881 * at the time of fork() could consume its reserves on COW instead
5882 * of the full address range.
5884 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5885 old_folio != pagecache_folio)
5886 outside_reserve = 1;
5888 folio_get(old_folio);
5891 * Drop page table lock as buddy allocator may be called. It will
5892 * be acquired again before returning to the caller, as expected.
5895 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5897 if (IS_ERR(new_folio)) {
5899 * If a process owning a MAP_PRIVATE mapping fails to COW,
5900 * it is due to references held by a child and an insufficient
5901 * huge page pool. To guarantee the original mappers
5902 * reliability, unmap the page from child processes. The child
5903 * may get SIGKILLed if it later faults.
5905 if (outside_reserve) {
5906 struct address_space *mapping = vma->vm_file->f_mapping;
5910 folio_put(old_folio);
5912 * Drop hugetlb_fault_mutex and vma_lock before
5913 * unmapping. unmapping needs to hold vma_lock
5914 * in write mode. Dropping vma_lock in read mode
5915 * here is OK as COW mappings do not interact with
5918 * Reacquire both after unmap operation.
5920 idx = vma_hugecache_offset(h, vma, haddr);
5921 hash = hugetlb_fault_mutex_hash(mapping, idx);
5922 hugetlb_vma_unlock_read(vma);
5923 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5925 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5927 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5928 hugetlb_vma_lock_read(vma);
5930 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5932 pte_same(huge_ptep_get(ptep), pte)))
5933 goto retry_avoidcopy;
5935 * race occurs while re-acquiring page table
5936 * lock, and our job is done.
5938 delayacct_wpcopy_end();
5942 ret = vmf_error(PTR_ERR(new_folio));
5943 goto out_release_old;
5947 * When the original hugepage is shared one, it does not have
5948 * anon_vma prepared.
5950 if (unlikely(anon_vma_prepare(vma))) {
5952 goto out_release_all;
5955 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5956 ret = VM_FAULT_HWPOISON_LARGE;
5957 goto out_release_all;
5959 __folio_mark_uptodate(new_folio);
5961 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5962 haddr + huge_page_size(h));
5963 mmu_notifier_invalidate_range_start(&range);
5966 * Retake the page table lock to check for racing updates
5967 * before the page tables are altered
5970 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5971 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5972 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5974 /* Break COW or unshare */
5975 huge_ptep_clear_flush(vma, haddr, ptep);
5976 hugetlb_remove_rmap(old_folio);
5977 hugetlb_add_new_anon_rmap(new_folio, vma, haddr);
5978 if (huge_pte_uffd_wp(pte))
5979 newpte = huge_pte_mkuffd_wp(newpte);
5980 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5981 folio_set_hugetlb_migratable(new_folio);
5982 /* Make the old page be freed below */
5983 new_folio = old_folio;
5986 mmu_notifier_invalidate_range_end(&range);
5989 * No restore in case of successful pagetable update (Break COW or
5992 if (new_folio != old_folio)
5993 restore_reserve_on_error(h, vma, haddr, new_folio);
5994 folio_put(new_folio);
5996 folio_put(old_folio);
5998 spin_lock(ptl); /* Caller expects lock to be held */
6000 delayacct_wpcopy_end();
6005 * Return whether there is a pagecache page to back given address within VMA.
6007 static bool hugetlbfs_pagecache_present(struct hstate *h,
6008 struct vm_area_struct *vma, unsigned long address)
6010 struct address_space *mapping = vma->vm_file->f_mapping;
6011 pgoff_t idx = linear_page_index(vma, address);
6012 struct folio *folio;
6014 folio = filemap_get_folio(mapping, idx);
6021 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
6024 struct inode *inode = mapping->host;
6025 struct hstate *h = hstate_inode(inode);
6028 idx <<= huge_page_order(h);
6029 __folio_set_locked(folio);
6030 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
6032 if (unlikely(err)) {
6033 __folio_clear_locked(folio);
6036 folio_clear_hugetlb_restore_reserve(folio);
6039 * mark folio dirty so that it will not be removed from cache/file
6040 * by non-hugetlbfs specific code paths.
6042 folio_mark_dirty(folio);
6044 spin_lock(&inode->i_lock);
6045 inode->i_blocks += blocks_per_huge_page(h);
6046 spin_unlock(&inode->i_lock);
6050 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
6051 struct address_space *mapping,
6054 unsigned long haddr,
6056 unsigned long reason)
6059 struct vm_fault vmf = {
6062 .real_address = addr,
6066 * Hard to debug if it ends up being
6067 * used by a callee that assumes
6068 * something about the other
6069 * uninitialized fields... same as in
6075 * vma_lock and hugetlb_fault_mutex must be dropped before handling
6076 * userfault. Also mmap_lock could be dropped due to handling
6077 * userfault, any vma operation should be careful from here.
6079 hugetlb_vma_unlock_read(vma);
6080 hash = hugetlb_fault_mutex_hash(mapping, idx);
6081 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6082 return handle_userfault(&vmf, reason);
6086 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
6087 * false if pte changed or is changing.
6089 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
6090 pte_t *ptep, pte_t old_pte)
6095 ptl = huge_pte_lock(h, mm, ptep);
6096 same = pte_same(huge_ptep_get(ptep), old_pte);
6102 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
6103 struct vm_area_struct *vma,
6104 struct address_space *mapping, pgoff_t idx,
6105 unsigned long address, pte_t *ptep,
6106 pte_t old_pte, unsigned int flags)
6108 struct hstate *h = hstate_vma(vma);
6109 vm_fault_t ret = VM_FAULT_SIGBUS;
6112 struct folio *folio;
6115 unsigned long haddr = address & huge_page_mask(h);
6116 bool new_folio, new_pagecache_folio = false;
6117 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
6120 * Currently, we are forced to kill the process in the event the
6121 * original mapper has unmapped pages from the child due to a failed
6122 * COW/unsharing. Warn that such a situation has occurred as it may not
6125 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
6126 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
6132 * Use page lock to guard against racing truncation
6133 * before we get page_table_lock.
6136 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6137 if (IS_ERR(folio)) {
6138 size = i_size_read(mapping->host) >> huge_page_shift(h);
6141 /* Check for page in userfault range */
6142 if (userfaultfd_missing(vma)) {
6144 * Since hugetlb_no_page() was examining pte
6145 * without pgtable lock, we need to re-test under
6146 * lock because the pte may not be stable and could
6147 * have changed from under us. Try to detect
6148 * either changed or during-changing ptes and retry
6149 * properly when needed.
6151 * Note that userfaultfd is actually fine with
6152 * false positives (e.g. caused by pte changed),
6153 * but not wrong logical events (e.g. caused by
6154 * reading a pte during changing). The latter can
6155 * confuse the userspace, so the strictness is very
6156 * much preferred. E.g., MISSING event should
6157 * never happen on the page after UFFDIO_COPY has
6158 * correctly installed the page and returned.
6160 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6165 return hugetlb_handle_userfault(vma, mapping, idx, flags,
6170 folio = alloc_hugetlb_folio(vma, haddr, 0);
6171 if (IS_ERR(folio)) {
6173 * Returning error will result in faulting task being
6174 * sent SIGBUS. The hugetlb fault mutex prevents two
6175 * tasks from racing to fault in the same page which
6176 * could result in false unable to allocate errors.
6177 * Page migration does not take the fault mutex, but
6178 * does a clear then write of pte's under page table
6179 * lock. Page fault code could race with migration,
6180 * notice the clear pte and try to allocate a page
6181 * here. Before returning error, get ptl and make
6182 * sure there really is no pte entry.
6184 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
6185 ret = vmf_error(PTR_ERR(folio));
6190 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
6191 __folio_mark_uptodate(folio);
6194 if (vma->vm_flags & VM_MAYSHARE) {
6195 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
6198 * err can't be -EEXIST which implies someone
6199 * else consumed the reservation since hugetlb
6200 * fault mutex is held when add a hugetlb page
6201 * to the page cache. So it's safe to call
6202 * restore_reserve_on_error() here.
6204 restore_reserve_on_error(h, vma, haddr, folio);
6208 new_pagecache_folio = true;
6211 if (unlikely(anon_vma_prepare(vma))) {
6213 goto backout_unlocked;
6219 * If memory error occurs between mmap() and fault, some process
6220 * don't have hwpoisoned swap entry for errored virtual address.
6221 * So we need to block hugepage fault by PG_hwpoison bit check.
6223 if (unlikely(folio_test_hwpoison(folio))) {
6224 ret = VM_FAULT_HWPOISON_LARGE |
6225 VM_FAULT_SET_HINDEX(hstate_index(h));
6226 goto backout_unlocked;
6229 /* Check for page in userfault range. */
6230 if (userfaultfd_minor(vma)) {
6231 folio_unlock(folio);
6233 /* See comment in userfaultfd_missing() block above */
6234 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6238 return hugetlb_handle_userfault(vma, mapping, idx, flags,
6245 * If we are going to COW a private mapping later, we examine the
6246 * pending reservations for this page now. This will ensure that
6247 * any allocations necessary to record that reservation occur outside
6250 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6251 if (vma_needs_reservation(h, vma, haddr) < 0) {
6253 goto backout_unlocked;
6255 /* Just decrements count, does not deallocate */
6256 vma_end_reservation(h, vma, haddr);
6259 ptl = huge_pte_lock(h, mm, ptep);
6261 /* If pte changed from under us, retry */
6262 if (!pte_same(huge_ptep_get(ptep), old_pte))
6266 hugetlb_add_new_anon_rmap(folio, vma, haddr);
6268 hugetlb_add_file_rmap(folio);
6269 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6270 && (vma->vm_flags & VM_SHARED)));
6272 * If this pte was previously wr-protected, keep it wr-protected even
6275 if (unlikely(pte_marker_uffd_wp(old_pte)))
6276 new_pte = huge_pte_mkuffd_wp(new_pte);
6277 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6279 hugetlb_count_add(pages_per_huge_page(h), mm);
6280 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6281 /* Optimization, do the COW without a second fault */
6282 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
6288 * Only set hugetlb_migratable in newly allocated pages. Existing pages
6289 * found in the pagecache may not have hugetlb_migratable if they have
6290 * been isolated for migration.
6293 folio_set_hugetlb_migratable(folio);
6295 folio_unlock(folio);
6297 hugetlb_vma_unlock_read(vma);
6298 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6304 if (new_folio && !new_pagecache_folio)
6305 restore_reserve_on_error(h, vma, haddr, folio);
6307 folio_unlock(folio);
6313 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6315 unsigned long key[2];
6318 key[0] = (unsigned long) mapping;
6321 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6323 return hash & (num_fault_mutexes - 1);
6327 * For uniprocessor systems we always use a single mutex, so just
6328 * return 0 and avoid the hashing overhead.
6330 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6336 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6337 unsigned long address, unsigned int flags)
6344 struct folio *folio = NULL;
6345 struct folio *pagecache_folio = NULL;
6346 struct hstate *h = hstate_vma(vma);
6347 struct address_space *mapping;
6348 int need_wait_lock = 0;
6349 unsigned long haddr = address & huge_page_mask(h);
6351 /* TODO: Handle faults under the VMA lock */
6352 if (flags & FAULT_FLAG_VMA_LOCK) {
6354 return VM_FAULT_RETRY;
6358 * Serialize hugepage allocation and instantiation, so that we don't
6359 * get spurious allocation failures if two CPUs race to instantiate
6360 * the same page in the page cache.
6362 mapping = vma->vm_file->f_mapping;
6363 idx = vma_hugecache_offset(h, vma, haddr);
6364 hash = hugetlb_fault_mutex_hash(mapping, idx);
6365 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6368 * Acquire vma lock before calling huge_pte_alloc and hold
6369 * until finished with ptep. This prevents huge_pmd_unshare from
6370 * being called elsewhere and making the ptep no longer valid.
6372 hugetlb_vma_lock_read(vma);
6373 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6375 hugetlb_vma_unlock_read(vma);
6376 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6377 return VM_FAULT_OOM;
6380 entry = huge_ptep_get(ptep);
6381 if (huge_pte_none_mostly(entry)) {
6382 if (is_pte_marker(entry)) {
6384 pte_marker_get(pte_to_swp_entry(entry));
6386 if (marker & PTE_MARKER_POISONED) {
6387 ret = VM_FAULT_HWPOISON_LARGE;
6393 * Other PTE markers should be handled the same way as none PTE.
6395 * hugetlb_no_page will drop vma lock and hugetlb fault
6396 * mutex internally, which make us return immediately.
6398 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6405 * entry could be a migration/hwpoison entry at this point, so this
6406 * check prevents the kernel from going below assuming that we have
6407 * an active hugepage in pagecache. This goto expects the 2nd page
6408 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6409 * properly handle it.
6411 if (!pte_present(entry)) {
6412 if (unlikely(is_hugetlb_entry_migration(entry))) {
6414 * Release the hugetlb fault lock now, but retain
6415 * the vma lock, because it is needed to guard the
6416 * huge_pte_lockptr() later in
6417 * migration_entry_wait_huge(). The vma lock will
6418 * be released there.
6420 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6421 migration_entry_wait_huge(vma, ptep);
6423 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6424 ret = VM_FAULT_HWPOISON_LARGE |
6425 VM_FAULT_SET_HINDEX(hstate_index(h));
6430 * If we are going to COW/unshare the mapping later, we examine the
6431 * pending reservations for this page now. This will ensure that any
6432 * allocations necessary to record that reservation occur outside the
6433 * spinlock. Also lookup the pagecache page now as it is used to
6434 * determine if a reservation has been consumed.
6436 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6437 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6438 if (vma_needs_reservation(h, vma, haddr) < 0) {
6442 /* Just decrements count, does not deallocate */
6443 vma_end_reservation(h, vma, haddr);
6445 pagecache_folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6446 if (IS_ERR(pagecache_folio))
6447 pagecache_folio = NULL;
6450 ptl = huge_pte_lock(h, mm, ptep);
6452 /* Check for a racing update before calling hugetlb_wp() */
6453 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6456 /* Handle userfault-wp first, before trying to lock more pages */
6457 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6458 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6459 if (!userfaultfd_wp_async(vma)) {
6460 struct vm_fault vmf = {
6463 .real_address = address,
6468 if (pagecache_folio) {
6469 folio_unlock(pagecache_folio);
6470 folio_put(pagecache_folio);
6472 hugetlb_vma_unlock_read(vma);
6473 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6474 return handle_userfault(&vmf, VM_UFFD_WP);
6477 entry = huge_pte_clear_uffd_wp(entry);
6478 set_huge_pte_at(mm, haddr, ptep, entry,
6479 huge_page_size(hstate_vma(vma)));
6480 /* Fallthrough to CoW */
6484 * hugetlb_wp() requires page locks of pte_page(entry) and
6485 * pagecache_folio, so here we need take the former one
6486 * when folio != pagecache_folio or !pagecache_folio.
6488 folio = page_folio(pte_page(entry));
6489 if (folio != pagecache_folio)
6490 if (!folio_trylock(folio)) {
6497 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6498 if (!huge_pte_write(entry)) {
6499 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6500 pagecache_folio, ptl);
6502 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6503 entry = huge_pte_mkdirty(entry);
6506 entry = pte_mkyoung(entry);
6507 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6508 flags & FAULT_FLAG_WRITE))
6509 update_mmu_cache(vma, haddr, ptep);
6511 if (folio != pagecache_folio)
6512 folio_unlock(folio);
6517 if (pagecache_folio) {
6518 folio_unlock(pagecache_folio);
6519 folio_put(pagecache_folio);
6522 hugetlb_vma_unlock_read(vma);
6523 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6525 * Generally it's safe to hold refcount during waiting page lock. But
6526 * here we just wait to defer the next page fault to avoid busy loop and
6527 * the page is not used after unlocked before returning from the current
6528 * page fault. So we are safe from accessing freed page, even if we wait
6529 * here without taking refcount.
6532 folio_wait_locked(folio);
6536 #ifdef CONFIG_USERFAULTFD
6538 * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte().
6540 static struct folio *alloc_hugetlb_folio_vma(struct hstate *h,
6541 struct vm_area_struct *vma, unsigned long address)
6543 struct mempolicy *mpol;
6544 nodemask_t *nodemask;
6545 struct folio *folio;
6549 gfp_mask = htlb_alloc_mask(h);
6550 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
6551 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
6552 mpol_cond_put(mpol);
6558 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6559 * with modifications for hugetlb pages.
6561 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6562 struct vm_area_struct *dst_vma,
6563 unsigned long dst_addr,
6564 unsigned long src_addr,
6566 struct folio **foliop)
6568 struct mm_struct *dst_mm = dst_vma->vm_mm;
6569 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6570 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6571 struct hstate *h = hstate_vma(dst_vma);
6572 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6573 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6575 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6579 struct folio *folio;
6581 bool folio_in_pagecache = false;
6583 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6584 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6586 /* Don't overwrite any existing PTEs (even markers) */
6587 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6592 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6593 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6596 /* No need to invalidate - it was non-present before */
6597 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6605 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6608 folio_in_pagecache = true;
6609 } else if (!*foliop) {
6610 /* If a folio already exists, then it's UFFDIO_COPY for
6611 * a non-missing case. Return -EEXIST.
6614 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6619 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6620 if (IS_ERR(folio)) {
6625 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6628 /* fallback to copy_from_user outside mmap_lock */
6629 if (unlikely(ret)) {
6631 /* Free the allocated folio which may have
6632 * consumed a reservation.
6634 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6637 /* Allocate a temporary folio to hold the copied
6640 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6646 /* Set the outparam foliop and return to the caller to
6647 * copy the contents outside the lock. Don't free the
6654 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6661 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6662 if (IS_ERR(folio)) {
6668 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6678 * The memory barrier inside __folio_mark_uptodate makes sure that
6679 * preceding stores to the page contents become visible before
6680 * the set_pte_at() write.
6682 __folio_mark_uptodate(folio);
6684 /* Add shared, newly allocated pages to the page cache. */
6685 if (vm_shared && !is_continue) {
6686 size = i_size_read(mapping->host) >> huge_page_shift(h);
6689 goto out_release_nounlock;
6692 * Serialization between remove_inode_hugepages() and
6693 * hugetlb_add_to_page_cache() below happens through the
6694 * hugetlb_fault_mutex_table that here must be hold by
6697 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6699 goto out_release_nounlock;
6700 folio_in_pagecache = true;
6703 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6706 if (folio_test_hwpoison(folio))
6707 goto out_release_unlock;
6710 * We allow to overwrite a pte marker: consider when both MISSING|WP
6711 * registered, we firstly wr-protect a none pte which has no page cache
6712 * page backing it, then access the page.
6715 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6716 goto out_release_unlock;
6718 if (folio_in_pagecache)
6719 hugetlb_add_file_rmap(folio);
6721 hugetlb_add_new_anon_rmap(folio, dst_vma, dst_addr);
6724 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6725 * with wp flag set, don't set pte write bit.
6727 if (wp_enabled || (is_continue && !vm_shared))
6730 writable = dst_vma->vm_flags & VM_WRITE;
6732 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6734 * Always mark UFFDIO_COPY page dirty; note that this may not be
6735 * extremely important for hugetlbfs for now since swapping is not
6736 * supported, but we should still be clear in that this page cannot be
6737 * thrown away at will, even if write bit not set.
6739 _dst_pte = huge_pte_mkdirty(_dst_pte);
6740 _dst_pte = pte_mkyoung(_dst_pte);
6743 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6745 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6747 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6749 /* No need to invalidate - it was non-present before */
6750 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6754 folio_set_hugetlb_migratable(folio);
6755 if (vm_shared || is_continue)
6756 folio_unlock(folio);
6762 if (vm_shared || is_continue)
6763 folio_unlock(folio);
6764 out_release_nounlock:
6765 if (!folio_in_pagecache)
6766 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6770 #endif /* CONFIG_USERFAULTFD */
6772 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6773 unsigned long address, unsigned int flags,
6774 unsigned int *page_mask)
6776 struct hstate *h = hstate_vma(vma);
6777 struct mm_struct *mm = vma->vm_mm;
6778 unsigned long haddr = address & huge_page_mask(h);
6779 struct page *page = NULL;
6784 hugetlb_vma_lock_read(vma);
6785 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6789 ptl = huge_pte_lock(h, mm, pte);
6790 entry = huge_ptep_get(pte);
6791 if (pte_present(entry)) {
6792 page = pte_page(entry);
6794 if (!huge_pte_write(entry)) {
6795 if (flags & FOLL_WRITE) {
6800 if (gup_must_unshare(vma, flags, page)) {
6801 /* Tell the caller to do unsharing */
6802 page = ERR_PTR(-EMLINK);
6807 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6810 * Note that page may be a sub-page, and with vmemmap
6811 * optimizations the page struct may be read only.
6812 * try_grab_page() will increase the ref count on the
6813 * head page, so this will be OK.
6815 * try_grab_page() should always be able to get the page here,
6816 * because we hold the ptl lock and have verified pte_present().
6818 ret = try_grab_page(page, flags);
6820 if (WARN_ON_ONCE(ret)) {
6821 page = ERR_PTR(ret);
6825 *page_mask = (1U << huge_page_order(h)) - 1;
6830 hugetlb_vma_unlock_read(vma);
6833 * Fixup retval for dump requests: if pagecache doesn't exist,
6834 * don't try to allocate a new page but just skip it.
6836 if (!page && (flags & FOLL_DUMP) &&
6837 !hugetlbfs_pagecache_present(h, vma, address))
6838 page = ERR_PTR(-EFAULT);
6843 long hugetlb_change_protection(struct vm_area_struct *vma,
6844 unsigned long address, unsigned long end,
6845 pgprot_t newprot, unsigned long cp_flags)
6847 struct mm_struct *mm = vma->vm_mm;
6848 unsigned long start = address;
6851 struct hstate *h = hstate_vma(vma);
6852 long pages = 0, psize = huge_page_size(h);
6853 bool shared_pmd = false;
6854 struct mmu_notifier_range range;
6855 unsigned long last_addr_mask;
6856 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6857 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6860 * In the case of shared PMDs, the area to flush could be beyond
6861 * start/end. Set range.start/range.end to cover the maximum possible
6862 * range if PMD sharing is possible.
6864 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6866 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6868 BUG_ON(address >= end);
6869 flush_cache_range(vma, range.start, range.end);
6871 mmu_notifier_invalidate_range_start(&range);
6872 hugetlb_vma_lock_write(vma);
6873 i_mmap_lock_write(vma->vm_file->f_mapping);
6874 last_addr_mask = hugetlb_mask_last_page(h);
6875 for (; address < end; address += psize) {
6877 ptep = hugetlb_walk(vma, address, psize);
6880 address |= last_addr_mask;
6884 * Userfaultfd wr-protect requires pgtable
6885 * pre-allocations to install pte markers.
6887 ptep = huge_pte_alloc(mm, vma, address, psize);
6893 ptl = huge_pte_lock(h, mm, ptep);
6894 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6896 * When uffd-wp is enabled on the vma, unshare
6897 * shouldn't happen at all. Warn about it if it
6898 * happened due to some reason.
6900 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6904 address |= last_addr_mask;
6907 pte = huge_ptep_get(ptep);
6908 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6909 /* Nothing to do. */
6910 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6911 swp_entry_t entry = pte_to_swp_entry(pte);
6912 struct page *page = pfn_swap_entry_to_page(entry);
6915 if (is_writable_migration_entry(entry)) {
6917 entry = make_readable_exclusive_migration_entry(
6920 entry = make_readable_migration_entry(
6922 newpte = swp_entry_to_pte(entry);
6927 newpte = pte_swp_mkuffd_wp(newpte);
6928 else if (uffd_wp_resolve)
6929 newpte = pte_swp_clear_uffd_wp(newpte);
6930 if (!pte_same(pte, newpte))
6931 set_huge_pte_at(mm, address, ptep, newpte, psize);
6932 } else if (unlikely(is_pte_marker(pte))) {
6934 * Do nothing on a poison marker; page is
6935 * corrupted, permissons do not apply. Here
6936 * pte_marker_uffd_wp()==true implies !poison
6937 * because they're mutual exclusive.
6939 if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)
6940 /* Safe to modify directly (non-present->none). */
6941 huge_pte_clear(mm, address, ptep, psize);
6942 } else if (!huge_pte_none(pte)) {
6944 unsigned int shift = huge_page_shift(hstate_vma(vma));
6946 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6947 pte = huge_pte_modify(old_pte, newprot);
6948 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6950 pte = huge_pte_mkuffd_wp(pte);
6951 else if (uffd_wp_resolve)
6952 pte = huge_pte_clear_uffd_wp(pte);
6953 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6957 if (unlikely(uffd_wp))
6958 /* Safe to modify directly (none->non-present). */
6959 set_huge_pte_at(mm, address, ptep,
6960 make_pte_marker(PTE_MARKER_UFFD_WP),
6966 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6967 * may have cleared our pud entry and done put_page on the page table:
6968 * once we release i_mmap_rwsem, another task can do the final put_page
6969 * and that page table be reused and filled with junk. If we actually
6970 * did unshare a page of pmds, flush the range corresponding to the pud.
6973 flush_hugetlb_tlb_range(vma, range.start, range.end);
6975 flush_hugetlb_tlb_range(vma, start, end);
6977 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6978 * downgrading page table protection not changing it to point to a new
6981 * See Documentation/mm/mmu_notifier.rst
6983 i_mmap_unlock_write(vma->vm_file->f_mapping);
6984 hugetlb_vma_unlock_write(vma);
6985 mmu_notifier_invalidate_range_end(&range);
6987 return pages > 0 ? (pages << h->order) : pages;
6990 /* Return true if reservation was successful, false otherwise. */
6991 bool hugetlb_reserve_pages(struct inode *inode,
6993 struct vm_area_struct *vma,
6994 vm_flags_t vm_flags)
6996 long chg = -1, add = -1;
6997 struct hstate *h = hstate_inode(inode);
6998 struct hugepage_subpool *spool = subpool_inode(inode);
6999 struct resv_map *resv_map;
7000 struct hugetlb_cgroup *h_cg = NULL;
7001 long gbl_reserve, regions_needed = 0;
7003 /* This should never happen */
7005 VM_WARN(1, "%s called with a negative range\n", __func__);
7010 * vma specific semaphore used for pmd sharing and fault/truncation
7013 hugetlb_vma_lock_alloc(vma);
7016 * Only apply hugepage reservation if asked. At fault time, an
7017 * attempt will be made for VM_NORESERVE to allocate a page
7018 * without using reserves
7020 if (vm_flags & VM_NORESERVE)
7024 * Shared mappings base their reservation on the number of pages that
7025 * are already allocated on behalf of the file. Private mappings need
7026 * to reserve the full area even if read-only as mprotect() may be
7027 * called to make the mapping read-write. Assume !vma is a shm mapping
7029 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7031 * resv_map can not be NULL as hugetlb_reserve_pages is only
7032 * called for inodes for which resv_maps were created (see
7033 * hugetlbfs_get_inode).
7035 resv_map = inode_resv_map(inode);
7037 chg = region_chg(resv_map, from, to, ®ions_needed);
7039 /* Private mapping. */
7040 resv_map = resv_map_alloc();
7046 set_vma_resv_map(vma, resv_map);
7047 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
7053 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
7054 chg * pages_per_huge_page(h), &h_cg) < 0)
7057 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
7058 /* For private mappings, the hugetlb_cgroup uncharge info hangs
7061 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
7065 * There must be enough pages in the subpool for the mapping. If
7066 * the subpool has a minimum size, there may be some global
7067 * reservations already in place (gbl_reserve).
7069 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
7070 if (gbl_reserve < 0)
7071 goto out_uncharge_cgroup;
7074 * Check enough hugepages are available for the reservation.
7075 * Hand the pages back to the subpool if there are not
7077 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
7081 * Account for the reservations made. Shared mappings record regions
7082 * that have reservations as they are shared by multiple VMAs.
7083 * When the last VMA disappears, the region map says how much
7084 * the reservation was and the page cache tells how much of
7085 * the reservation was consumed. Private mappings are per-VMA and
7086 * only the consumed reservations are tracked. When the VMA
7087 * disappears, the original reservation is the VMA size and the
7088 * consumed reservations are stored in the map. Hence, nothing
7089 * else has to be done for private mappings here
7091 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7092 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
7094 if (unlikely(add < 0)) {
7095 hugetlb_acct_memory(h, -gbl_reserve);
7097 } else if (unlikely(chg > add)) {
7099 * pages in this range were added to the reserve
7100 * map between region_chg and region_add. This
7101 * indicates a race with alloc_hugetlb_folio. Adjust
7102 * the subpool and reserve counts modified above
7103 * based on the difference.
7108 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
7109 * reference to h_cg->css. See comment below for detail.
7111 hugetlb_cgroup_uncharge_cgroup_rsvd(
7113 (chg - add) * pages_per_huge_page(h), h_cg);
7115 rsv_adjust = hugepage_subpool_put_pages(spool,
7117 hugetlb_acct_memory(h, -rsv_adjust);
7120 * The file_regions will hold their own reference to
7121 * h_cg->css. So we should release the reference held
7122 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
7125 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
7131 /* put back original number of pages, chg */
7132 (void)hugepage_subpool_put_pages(spool, chg);
7133 out_uncharge_cgroup:
7134 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
7135 chg * pages_per_huge_page(h), h_cg);
7137 hugetlb_vma_lock_free(vma);
7138 if (!vma || vma->vm_flags & VM_MAYSHARE)
7139 /* Only call region_abort if the region_chg succeeded but the
7140 * region_add failed or didn't run.
7142 if (chg >= 0 && add < 0)
7143 region_abort(resv_map, from, to, regions_needed);
7144 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
7145 kref_put(&resv_map->refs, resv_map_release);
7146 set_vma_resv_map(vma, NULL);
7151 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
7154 struct hstate *h = hstate_inode(inode);
7155 struct resv_map *resv_map = inode_resv_map(inode);
7157 struct hugepage_subpool *spool = subpool_inode(inode);
7161 * Since this routine can be called in the evict inode path for all
7162 * hugetlbfs inodes, resv_map could be NULL.
7165 chg = region_del(resv_map, start, end);
7167 * region_del() can fail in the rare case where a region
7168 * must be split and another region descriptor can not be
7169 * allocated. If end == LONG_MAX, it will not fail.
7175 spin_lock(&inode->i_lock);
7176 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
7177 spin_unlock(&inode->i_lock);
7180 * If the subpool has a minimum size, the number of global
7181 * reservations to be released may be adjusted.
7183 * Note that !resv_map implies freed == 0. So (chg - freed)
7184 * won't go negative.
7186 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
7187 hugetlb_acct_memory(h, -gbl_reserve);
7192 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7193 static unsigned long page_table_shareable(struct vm_area_struct *svma,
7194 struct vm_area_struct *vma,
7195 unsigned long addr, pgoff_t idx)
7197 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
7199 unsigned long sbase = saddr & PUD_MASK;
7200 unsigned long s_end = sbase + PUD_SIZE;
7202 /* Allow segments to share if only one is marked locked */
7203 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
7204 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
7207 * match the virtual addresses, permission and the alignment of the
7210 * Also, vma_lock (vm_private_data) is required for sharing.
7212 if (pmd_index(addr) != pmd_index(saddr) ||
7213 vm_flags != svm_flags ||
7214 !range_in_vma(svma, sbase, s_end) ||
7215 !svma->vm_private_data)
7221 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7223 unsigned long start = addr & PUD_MASK;
7224 unsigned long end = start + PUD_SIZE;
7226 #ifdef CONFIG_USERFAULTFD
7227 if (uffd_disable_huge_pmd_share(vma))
7231 * check on proper vm_flags and page table alignment
7233 if (!(vma->vm_flags & VM_MAYSHARE))
7235 if (!vma->vm_private_data) /* vma lock required for sharing */
7237 if (!range_in_vma(vma, start, end))
7243 * Determine if start,end range within vma could be mapped by shared pmd.
7244 * If yes, adjust start and end to cover range associated with possible
7245 * shared pmd mappings.
7247 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7248 unsigned long *start, unsigned long *end)
7250 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7251 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7254 * vma needs to span at least one aligned PUD size, and the range
7255 * must be at least partially within in.
7257 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7258 (*end <= v_start) || (*start >= v_end))
7261 /* Extend the range to be PUD aligned for a worst case scenario */
7262 if (*start > v_start)
7263 *start = ALIGN_DOWN(*start, PUD_SIZE);
7266 *end = ALIGN(*end, PUD_SIZE);
7270 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7271 * and returns the corresponding pte. While this is not necessary for the
7272 * !shared pmd case because we can allocate the pmd later as well, it makes the
7273 * code much cleaner. pmd allocation is essential for the shared case because
7274 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7275 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7276 * bad pmd for sharing.
7278 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7279 unsigned long addr, pud_t *pud)
7281 struct address_space *mapping = vma->vm_file->f_mapping;
7282 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7284 struct vm_area_struct *svma;
7285 unsigned long saddr;
7289 i_mmap_lock_read(mapping);
7290 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7294 saddr = page_table_shareable(svma, vma, addr, idx);
7296 spte = hugetlb_walk(svma, saddr,
7297 vma_mmu_pagesize(svma));
7299 get_page(virt_to_page(spte));
7308 spin_lock(&mm->page_table_lock);
7309 if (pud_none(*pud)) {
7310 pud_populate(mm, pud,
7311 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7314 put_page(virt_to_page(spte));
7316 spin_unlock(&mm->page_table_lock);
7318 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7319 i_mmap_unlock_read(mapping);
7324 * unmap huge page backed by shared pte.
7326 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7327 * indicated by page_count > 1, unmap is achieved by clearing pud and
7328 * decrementing the ref count. If count == 1, the pte page is not shared.
7330 * Called with page table lock held.
7332 * returns: 1 successfully unmapped a shared pte page
7333 * 0 the underlying pte page is not shared, or it is the last user
7335 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7336 unsigned long addr, pte_t *ptep)
7338 pgd_t *pgd = pgd_offset(mm, addr);
7339 p4d_t *p4d = p4d_offset(pgd, addr);
7340 pud_t *pud = pud_offset(p4d, addr);
7342 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7343 hugetlb_vma_assert_locked(vma);
7344 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7345 if (page_count(virt_to_page(ptep)) == 1)
7349 put_page(virt_to_page(ptep));
7354 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7356 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7357 unsigned long addr, pud_t *pud)
7362 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7363 unsigned long addr, pte_t *ptep)
7368 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7369 unsigned long *start, unsigned long *end)
7373 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7377 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7379 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7380 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7381 unsigned long addr, unsigned long sz)
7388 pgd = pgd_offset(mm, addr);
7389 p4d = p4d_alloc(mm, pgd, addr);
7392 pud = pud_alloc(mm, p4d, addr);
7394 if (sz == PUD_SIZE) {
7397 BUG_ON(sz != PMD_SIZE);
7398 if (want_pmd_share(vma, addr) && pud_none(*pud))
7399 pte = huge_pmd_share(mm, vma, addr, pud);
7401 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7406 pte_t pteval = ptep_get_lockless(pte);
7408 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7415 * huge_pte_offset() - Walk the page table to resolve the hugepage
7416 * entry at address @addr
7418 * Return: Pointer to page table entry (PUD or PMD) for
7419 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7420 * size @sz doesn't match the hugepage size at this level of the page
7423 pte_t *huge_pte_offset(struct mm_struct *mm,
7424 unsigned long addr, unsigned long sz)
7431 pgd = pgd_offset(mm, addr);
7432 if (!pgd_present(*pgd))
7434 p4d = p4d_offset(pgd, addr);
7435 if (!p4d_present(*p4d))
7438 pud = pud_offset(p4d, addr);
7440 /* must be pud huge, non-present or none */
7441 return (pte_t *)pud;
7442 if (!pud_present(*pud))
7444 /* must have a valid entry and size to go further */
7446 pmd = pmd_offset(pud, addr);
7447 /* must be pmd huge, non-present or none */
7448 return (pte_t *)pmd;
7452 * Return a mask that can be used to update an address to the last huge
7453 * page in a page table page mapping size. Used to skip non-present
7454 * page table entries when linearly scanning address ranges. Architectures
7455 * with unique huge page to page table relationships can define their own
7456 * version of this routine.
7458 unsigned long hugetlb_mask_last_page(struct hstate *h)
7460 unsigned long hp_size = huge_page_size(h);
7462 if (hp_size == PUD_SIZE)
7463 return P4D_SIZE - PUD_SIZE;
7464 else if (hp_size == PMD_SIZE)
7465 return PUD_SIZE - PMD_SIZE;
7472 /* See description above. Architectures can provide their own version. */
7473 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7475 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7476 if (huge_page_size(h) == PMD_SIZE)
7477 return PUD_SIZE - PMD_SIZE;
7482 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7485 * These functions are overwritable if your architecture needs its own
7488 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7492 spin_lock_irq(&hugetlb_lock);
7493 if (!folio_test_hugetlb(folio) ||
7494 !folio_test_hugetlb_migratable(folio) ||
7495 !folio_try_get(folio)) {
7499 folio_clear_hugetlb_migratable(folio);
7500 list_move_tail(&folio->lru, list);
7502 spin_unlock_irq(&hugetlb_lock);
7506 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7511 spin_lock_irq(&hugetlb_lock);
7512 if (folio_test_hugetlb(folio)) {
7514 if (folio_test_hugetlb_freed(folio))
7516 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7517 ret = folio_try_get(folio);
7521 spin_unlock_irq(&hugetlb_lock);
7525 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7526 bool *migratable_cleared)
7530 spin_lock_irq(&hugetlb_lock);
7531 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7532 spin_unlock_irq(&hugetlb_lock);
7536 void folio_putback_active_hugetlb(struct folio *folio)
7538 spin_lock_irq(&hugetlb_lock);
7539 folio_set_hugetlb_migratable(folio);
7540 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7541 spin_unlock_irq(&hugetlb_lock);
7545 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7547 struct hstate *h = folio_hstate(old_folio);
7549 hugetlb_cgroup_migrate(old_folio, new_folio);
7550 set_page_owner_migrate_reason(&new_folio->page, reason);
7553 * transfer temporary state of the new hugetlb folio. This is
7554 * reverse to other transitions because the newpage is going to
7555 * be final while the old one will be freed so it takes over
7556 * the temporary status.
7558 * Also note that we have to transfer the per-node surplus state
7559 * here as well otherwise the global surplus count will not match
7562 if (folio_test_hugetlb_temporary(new_folio)) {
7563 int old_nid = folio_nid(old_folio);
7564 int new_nid = folio_nid(new_folio);
7566 folio_set_hugetlb_temporary(old_folio);
7567 folio_clear_hugetlb_temporary(new_folio);
7571 * There is no need to transfer the per-node surplus state
7572 * when we do not cross the node.
7574 if (new_nid == old_nid)
7576 spin_lock_irq(&hugetlb_lock);
7577 if (h->surplus_huge_pages_node[old_nid]) {
7578 h->surplus_huge_pages_node[old_nid]--;
7579 h->surplus_huge_pages_node[new_nid]++;
7581 spin_unlock_irq(&hugetlb_lock);
7585 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7586 unsigned long start,
7589 struct hstate *h = hstate_vma(vma);
7590 unsigned long sz = huge_page_size(h);
7591 struct mm_struct *mm = vma->vm_mm;
7592 struct mmu_notifier_range range;
7593 unsigned long address;
7597 if (!(vma->vm_flags & VM_MAYSHARE))
7603 flush_cache_range(vma, start, end);
7605 * No need to call adjust_range_if_pmd_sharing_possible(), because
7606 * we have already done the PUD_SIZE alignment.
7608 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7610 mmu_notifier_invalidate_range_start(&range);
7611 hugetlb_vma_lock_write(vma);
7612 i_mmap_lock_write(vma->vm_file->f_mapping);
7613 for (address = start; address < end; address += PUD_SIZE) {
7614 ptep = hugetlb_walk(vma, address, sz);
7617 ptl = huge_pte_lock(h, mm, ptep);
7618 huge_pmd_unshare(mm, vma, address, ptep);
7621 flush_hugetlb_tlb_range(vma, start, end);
7622 i_mmap_unlock_write(vma->vm_file->f_mapping);
7623 hugetlb_vma_unlock_write(vma);
7625 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7626 * Documentation/mm/mmu_notifier.rst.
7628 mmu_notifier_invalidate_range_end(&range);
7632 * This function will unconditionally remove all the shared pmd pgtable entries
7633 * within the specific vma for a hugetlbfs memory range.
7635 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7637 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7638 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7642 static bool cma_reserve_called __initdata;
7644 static int __init cmdline_parse_hugetlb_cma(char *p)
7651 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7654 if (s[count] == ':') {
7655 if (tmp >= MAX_NUMNODES)
7657 nid = array_index_nospec(tmp, MAX_NUMNODES);
7660 tmp = memparse(s, &s);
7661 hugetlb_cma_size_in_node[nid] = tmp;
7662 hugetlb_cma_size += tmp;
7665 * Skip the separator if have one, otherwise
7666 * break the parsing.
7673 hugetlb_cma_size = memparse(p, &p);
7681 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7683 void __init hugetlb_cma_reserve(int order)
7685 unsigned long size, reserved, per_node;
7686 bool node_specific_cma_alloc = false;
7689 cma_reserve_called = true;
7691 if (!hugetlb_cma_size)
7694 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7695 if (hugetlb_cma_size_in_node[nid] == 0)
7698 if (!node_online(nid)) {
7699 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7700 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7701 hugetlb_cma_size_in_node[nid] = 0;
7705 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7706 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7707 nid, (PAGE_SIZE << order) / SZ_1M);
7708 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7709 hugetlb_cma_size_in_node[nid] = 0;
7711 node_specific_cma_alloc = true;
7715 /* Validate the CMA size again in case some invalid nodes specified. */
7716 if (!hugetlb_cma_size)
7719 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7720 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7721 (PAGE_SIZE << order) / SZ_1M);
7722 hugetlb_cma_size = 0;
7726 if (!node_specific_cma_alloc) {
7728 * If 3 GB area is requested on a machine with 4 numa nodes,
7729 * let's allocate 1 GB on first three nodes and ignore the last one.
7731 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7732 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7733 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7737 for_each_online_node(nid) {
7739 char name[CMA_MAX_NAME];
7741 if (node_specific_cma_alloc) {
7742 if (hugetlb_cma_size_in_node[nid] == 0)
7745 size = hugetlb_cma_size_in_node[nid];
7747 size = min(per_node, hugetlb_cma_size - reserved);
7750 size = round_up(size, PAGE_SIZE << order);
7752 snprintf(name, sizeof(name), "hugetlb%d", nid);
7754 * Note that 'order per bit' is based on smallest size that
7755 * may be returned to CMA allocator in the case of
7756 * huge page demotion.
7758 res = cma_declare_contiguous_nid(0, size, 0,
7759 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7761 &hugetlb_cma[nid], nid);
7763 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7769 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7772 if (reserved >= hugetlb_cma_size)
7778 * hugetlb_cma_size is used to determine if allocations from
7779 * cma are possible. Set to zero if no cma regions are set up.
7781 hugetlb_cma_size = 0;
7784 static void __init hugetlb_cma_check(void)
7786 if (!hugetlb_cma_size || cma_reserve_called)
7789 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7792 #endif /* CONFIG_CMA */