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)->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 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1187 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1189 * Clear vm_private_data
1190 * - For shared mappings this is a per-vma semaphore that may be
1191 * allocated in a subsequent call to hugetlb_vm_op_open.
1192 * Before clearing, make sure pointer is not associated with vma
1193 * as this will leak the structure. This is the case when called
1194 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1195 * been called to allocate a new structure.
1196 * - For MAP_PRIVATE mappings, this is the reserve map which does
1197 * not apply to children. Faults generated by the children are
1198 * not guaranteed to succeed, even if read-only.
1200 if (vma->vm_flags & VM_MAYSHARE) {
1201 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1203 if (vma_lock && vma_lock->vma != vma)
1204 vma->vm_private_data = NULL;
1206 vma->vm_private_data = NULL;
1210 * Reset and decrement one ref on hugepage private reservation.
1211 * Called with mm->mmap_lock writer semaphore held.
1212 * This function should be only used by move_vma() and operate on
1213 * same sized vma. It should never come here with last ref on the
1216 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1219 * Clear the old hugetlb private page reservation.
1220 * It has already been transferred to new_vma.
1222 * During a mremap() operation of a hugetlb vma we call move_vma()
1223 * which copies vma into new_vma and unmaps vma. After the copy
1224 * operation both new_vma and vma share a reference to the resv_map
1225 * struct, and at that point vma is about to be unmapped. We don't
1226 * want to return the reservation to the pool at unmap of vma because
1227 * the reservation still lives on in new_vma, so simply decrement the
1228 * ref here and remove the resv_map reference from this vma.
1230 struct resv_map *reservations = vma_resv_map(vma);
1232 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1233 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1234 kref_put(&reservations->refs, resv_map_release);
1237 hugetlb_dup_vma_private(vma);
1240 /* Returns true if the VMA has associated reserve pages */
1241 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1243 if (vma->vm_flags & VM_NORESERVE) {
1245 * This address is already reserved by other process(chg == 0),
1246 * so, we should decrement reserved count. Without decrementing,
1247 * reserve count remains after releasing inode, because this
1248 * allocated page will go into page cache and is regarded as
1249 * coming from reserved pool in releasing step. Currently, we
1250 * don't have any other solution to deal with this situation
1251 * properly, so add work-around here.
1253 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1259 /* Shared mappings always use reserves */
1260 if (vma->vm_flags & VM_MAYSHARE) {
1262 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1263 * be a region map for all pages. The only situation where
1264 * there is no region map is if a hole was punched via
1265 * fallocate. In this case, there really are no reserves to
1266 * use. This situation is indicated if chg != 0.
1275 * Only the process that called mmap() has reserves for
1278 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1280 * Like the shared case above, a hole punch or truncate
1281 * could have been performed on the private mapping.
1282 * Examine the value of chg to determine if reserves
1283 * actually exist or were previously consumed.
1284 * Very Subtle - The value of chg comes from a previous
1285 * call to vma_needs_reserves(). The reserve map for
1286 * private mappings has different (opposite) semantics
1287 * than that of shared mappings. vma_needs_reserves()
1288 * has already taken this difference in semantics into
1289 * account. Therefore, the meaning of chg is the same
1290 * as in the shared case above. Code could easily be
1291 * combined, but keeping it separate draws attention to
1292 * subtle differences.
1303 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1305 int nid = folio_nid(folio);
1307 lockdep_assert_held(&hugetlb_lock);
1308 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1310 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1311 h->free_huge_pages++;
1312 h->free_huge_pages_node[nid]++;
1313 folio_set_hugetlb_freed(folio);
1316 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1319 struct folio *folio;
1320 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1322 lockdep_assert_held(&hugetlb_lock);
1323 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1324 if (pin && !folio_is_longterm_pinnable(folio))
1327 if (folio_test_hwpoison(folio))
1330 list_move(&folio->lru, &h->hugepage_activelist);
1331 folio_ref_unfreeze(folio, 1);
1332 folio_clear_hugetlb_freed(folio);
1333 h->free_huge_pages--;
1334 h->free_huge_pages_node[nid]--;
1341 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1342 int nid, nodemask_t *nmask)
1344 unsigned int cpuset_mems_cookie;
1345 struct zonelist *zonelist;
1348 int node = NUMA_NO_NODE;
1350 zonelist = node_zonelist(nid, gfp_mask);
1353 cpuset_mems_cookie = read_mems_allowed_begin();
1354 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1355 struct folio *folio;
1357 if (!cpuset_zone_allowed(zone, gfp_mask))
1360 * no need to ask again on the same node. Pool is node rather than
1363 if (zone_to_nid(zone) == node)
1365 node = zone_to_nid(zone);
1367 folio = dequeue_hugetlb_folio_node_exact(h, node);
1371 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1377 static unsigned long available_huge_pages(struct hstate *h)
1379 return h->free_huge_pages - h->resv_huge_pages;
1382 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1383 struct vm_area_struct *vma,
1384 unsigned long address, int avoid_reserve,
1387 struct folio *folio = NULL;
1388 struct mempolicy *mpol;
1390 nodemask_t *nodemask;
1394 * A child process with MAP_PRIVATE mappings created by their parent
1395 * have no page reserves. This check ensures that reservations are
1396 * not "stolen". The child may still get SIGKILLed
1398 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1401 /* If reserves cannot be used, ensure enough pages are in the pool */
1402 if (avoid_reserve && !available_huge_pages(h))
1405 gfp_mask = htlb_alloc_mask(h);
1406 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1408 if (mpol_is_preferred_many(mpol)) {
1409 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1412 /* Fallback to all nodes if page==NULL */
1417 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1420 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1421 folio_set_hugetlb_restore_reserve(folio);
1422 h->resv_huge_pages--;
1425 mpol_cond_put(mpol);
1433 * common helper functions for hstate_next_node_to_{alloc|free}.
1434 * We may have allocated or freed a huge page based on a different
1435 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1436 * be outside of *nodes_allowed. Ensure that we use an allowed
1437 * node for alloc or free.
1439 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1441 nid = next_node_in(nid, *nodes_allowed);
1442 VM_BUG_ON(nid >= MAX_NUMNODES);
1447 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1449 if (!node_isset(nid, *nodes_allowed))
1450 nid = next_node_allowed(nid, nodes_allowed);
1455 * returns the previously saved node ["this node"] from which to
1456 * allocate a persistent huge page for the pool and advance the
1457 * next node from which to allocate, handling wrap at end of node
1460 static int hstate_next_node_to_alloc(struct hstate *h,
1461 nodemask_t *nodes_allowed)
1465 VM_BUG_ON(!nodes_allowed);
1467 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1468 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1474 * helper for remove_pool_hugetlb_folio() - return the previously saved
1475 * node ["this node"] from which to free a huge page. Advance the
1476 * next node id whether or not we find a free huge page to free so
1477 * that the next attempt to free addresses the next node.
1479 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1483 VM_BUG_ON(!nodes_allowed);
1485 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1486 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1491 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1492 for (nr_nodes = nodes_weight(*mask); \
1494 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1497 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1498 for (nr_nodes = nodes_weight(*mask); \
1500 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1503 /* used to demote non-gigantic_huge pages as well */
1504 static void __destroy_compound_gigantic_folio(struct folio *folio,
1505 unsigned int order, bool demote)
1508 int nr_pages = 1 << order;
1511 atomic_set(&folio->_entire_mapcount, 0);
1512 atomic_set(&folio->_nr_pages_mapped, 0);
1513 atomic_set(&folio->_pincount, 0);
1515 for (i = 1; i < nr_pages; i++) {
1516 p = folio_page(folio, i);
1517 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1519 clear_compound_head(p);
1521 set_page_refcounted(p);
1524 __folio_clear_head(folio);
1527 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1530 __destroy_compound_gigantic_folio(folio, order, true);
1533 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1534 static void destroy_compound_gigantic_folio(struct folio *folio,
1537 __destroy_compound_gigantic_folio(folio, order, false);
1540 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1543 * If the page isn't allocated using the cma allocator,
1544 * cma_release() returns false.
1547 int nid = folio_nid(folio);
1549 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1553 free_contig_range(folio_pfn(folio), 1 << order);
1556 #ifdef CONFIG_CONTIG_ALLOC
1557 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1558 int nid, nodemask_t *nodemask)
1561 unsigned long nr_pages = pages_per_huge_page(h);
1562 if (nid == NUMA_NO_NODE)
1563 nid = numa_mem_id();
1569 if (hugetlb_cma[nid]) {
1570 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1571 huge_page_order(h), true);
1573 return page_folio(page);
1576 if (!(gfp_mask & __GFP_THISNODE)) {
1577 for_each_node_mask(node, *nodemask) {
1578 if (node == nid || !hugetlb_cma[node])
1581 page = cma_alloc(hugetlb_cma[node], nr_pages,
1582 huge_page_order(h), true);
1584 return page_folio(page);
1590 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1591 return page ? page_folio(page) : NULL;
1594 #else /* !CONFIG_CONTIG_ALLOC */
1595 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1596 int nid, nodemask_t *nodemask)
1600 #endif /* CONFIG_CONTIG_ALLOC */
1602 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1603 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1604 int nid, nodemask_t *nodemask)
1608 static inline void free_gigantic_folio(struct folio *folio,
1609 unsigned int order) { }
1610 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1611 unsigned int order) { }
1614 static inline void __clear_hugetlb_destructor(struct hstate *h,
1615 struct folio *folio)
1617 lockdep_assert_held(&hugetlb_lock);
1619 folio_clear_hugetlb(folio);
1623 * Remove hugetlb folio from lists.
1624 * If vmemmap exists for the folio, update dtor so that the folio appears
1625 * as just a compound page. Otherwise, wait until after allocating vmemmap
1628 * A reference is held on the folio, except in the case of demote.
1630 * Must be called with hugetlb lock held.
1632 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1633 bool adjust_surplus,
1636 int nid = folio_nid(folio);
1638 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1639 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1641 lockdep_assert_held(&hugetlb_lock);
1642 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1645 list_del(&folio->lru);
1647 if (folio_test_hugetlb_freed(folio)) {
1648 h->free_huge_pages--;
1649 h->free_huge_pages_node[nid]--;
1651 if (adjust_surplus) {
1652 h->surplus_huge_pages--;
1653 h->surplus_huge_pages_node[nid]--;
1657 * We can only clear the hugetlb destructor after allocating vmemmap
1658 * pages. Otherwise, someone (memory error handling) may try to write
1659 * to tail struct pages.
1661 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1662 __clear_hugetlb_destructor(h, folio);
1665 * In the case of demote we do not ref count the page as it will soon
1666 * be turned into a page of smaller size.
1669 folio_ref_unfreeze(folio, 1);
1672 h->nr_huge_pages_node[nid]--;
1675 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1676 bool adjust_surplus)
1678 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1681 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1682 bool adjust_surplus)
1684 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1687 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1688 bool adjust_surplus)
1691 int nid = folio_nid(folio);
1693 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1695 lockdep_assert_held(&hugetlb_lock);
1697 INIT_LIST_HEAD(&folio->lru);
1699 h->nr_huge_pages_node[nid]++;
1701 if (adjust_surplus) {
1702 h->surplus_huge_pages++;
1703 h->surplus_huge_pages_node[nid]++;
1706 folio_set_hugetlb(folio);
1707 folio_change_private(folio, NULL);
1709 * We have to set hugetlb_vmemmap_optimized again as above
1710 * folio_change_private(folio, NULL) cleared it.
1712 folio_set_hugetlb_vmemmap_optimized(folio);
1715 * This folio is about to be managed by the hugetlb allocator and
1716 * should have no users. Drop our reference, and check for others
1719 zeroed = folio_put_testzero(folio);
1720 if (unlikely(!zeroed))
1722 * It is VERY unlikely soneone else has taken a ref
1723 * on the folio. In this case, we simply return as
1724 * free_huge_folio() will be called when this other ref
1729 arch_clear_hugepage_flags(&folio->page);
1730 enqueue_hugetlb_folio(h, folio);
1733 static void __update_and_free_hugetlb_folio(struct hstate *h,
1734 struct folio *folio)
1736 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1738 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1742 * If we don't know which subpages are hwpoisoned, we can't free
1743 * the hugepage, so it's leaked intentionally.
1745 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1749 * If folio is not vmemmap optimized (!clear_dtor), then the folio
1750 * is no longer identified as a hugetlb page. hugetlb_vmemmap_restore_folio
1751 * can only be passed hugetlb pages and will BUG otherwise.
1753 if (clear_dtor && hugetlb_vmemmap_restore_folio(h, folio)) {
1754 spin_lock_irq(&hugetlb_lock);
1756 * If we cannot allocate vmemmap pages, just refuse to free the
1757 * page and put the page back on the hugetlb free list and treat
1758 * as a surplus page.
1760 add_hugetlb_folio(h, folio, true);
1761 spin_unlock_irq(&hugetlb_lock);
1766 * Move PageHWPoison flag from head page to the raw error pages,
1767 * which makes any healthy subpages reusable.
1769 if (unlikely(folio_test_hwpoison(folio)))
1770 folio_clear_hugetlb_hwpoison(folio);
1773 * If vmemmap pages were allocated above, then we need to clear the
1774 * hugetlb destructor under the hugetlb lock.
1777 spin_lock_irq(&hugetlb_lock);
1778 __clear_hugetlb_destructor(h, folio);
1779 spin_unlock_irq(&hugetlb_lock);
1783 * Non-gigantic pages demoted from CMA allocated gigantic pages
1784 * need to be given back to CMA in free_gigantic_folio.
1786 if (hstate_is_gigantic(h) ||
1787 hugetlb_cma_folio(folio, huge_page_order(h))) {
1788 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1789 free_gigantic_folio(folio, huge_page_order(h));
1791 __free_pages(&folio->page, huge_page_order(h));
1796 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1797 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1798 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1799 * the vmemmap pages.
1801 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1802 * freed and frees them one-by-one. As the page->mapping pointer is going
1803 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1804 * structure of a lockless linked list of huge pages to be freed.
1806 static LLIST_HEAD(hpage_freelist);
1808 static void free_hpage_workfn(struct work_struct *work)
1810 struct llist_node *node;
1812 node = llist_del_all(&hpage_freelist);
1815 struct folio *folio;
1818 folio = container_of((struct address_space **)node,
1819 struct folio, mapping);
1821 folio->mapping = NULL;
1823 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1824 * folio_hstate() is going to trigger because a previous call to
1825 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1826 * not use folio_hstate() directly.
1828 h = size_to_hstate(folio_size(folio));
1830 __update_and_free_hugetlb_folio(h, folio);
1835 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1837 static inline void flush_free_hpage_work(struct hstate *h)
1839 if (hugetlb_vmemmap_optimizable(h))
1840 flush_work(&free_hpage_work);
1843 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1846 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1847 __update_and_free_hugetlb_folio(h, folio);
1852 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1854 * Only call schedule_work() if hpage_freelist is previously
1855 * empty. Otherwise, schedule_work() had been called but the workfn
1856 * hasn't retrieved the list yet.
1858 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1859 schedule_work(&free_hpage_work);
1862 static void bulk_vmemmap_restore_error(struct hstate *h,
1863 struct list_head *folio_list,
1864 struct list_head *non_hvo_folios)
1866 struct folio *folio, *t_folio;
1868 if (!list_empty(non_hvo_folios)) {
1870 * Free any restored hugetlb pages so that restore of the
1871 * entire list can be retried.
1872 * The idea is that in the common case of ENOMEM errors freeing
1873 * hugetlb pages with vmemmap we will free up memory so that we
1874 * can allocate vmemmap for more hugetlb pages.
1876 list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) {
1877 list_del(&folio->lru);
1878 spin_lock_irq(&hugetlb_lock);
1879 __clear_hugetlb_destructor(h, folio);
1880 spin_unlock_irq(&hugetlb_lock);
1881 update_and_free_hugetlb_folio(h, folio, false);
1886 * In the case where there are no folios which can be
1887 * immediately freed, we loop through the list trying to restore
1888 * vmemmap individually in the hope that someone elsewhere may
1889 * have done something to cause success (such as freeing some
1890 * memory). If unable to restore a hugetlb page, the hugetlb
1891 * page is made a surplus page and removed from the list.
1892 * If are able to restore vmemmap and free one hugetlb page, we
1893 * quit processing the list to retry the bulk operation.
1895 list_for_each_entry_safe(folio, t_folio, folio_list, lru)
1896 if (hugetlb_vmemmap_restore_folio(h, folio)) {
1897 list_del(&folio->lru);
1898 spin_lock_irq(&hugetlb_lock);
1899 add_hugetlb_folio(h, folio, true);
1900 spin_unlock_irq(&hugetlb_lock);
1902 list_del(&folio->lru);
1903 spin_lock_irq(&hugetlb_lock);
1904 __clear_hugetlb_destructor(h, folio);
1905 spin_unlock_irq(&hugetlb_lock);
1906 update_and_free_hugetlb_folio(h, folio, false);
1913 static void update_and_free_pages_bulk(struct hstate *h,
1914 struct list_head *folio_list)
1917 struct folio *folio, *t_folio;
1918 LIST_HEAD(non_hvo_folios);
1921 * First allocate required vmemmmap (if necessary) for all folios.
1922 * Carefully handle errors and free up any available hugetlb pages
1923 * in an effort to make forward progress.
1926 ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios);
1928 bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios);
1933 * At this point, list should be empty, ret should be >= 0 and there
1934 * should only be pages on the non_hvo_folios list.
1935 * Do note that the non_hvo_folios list could be empty.
1936 * Without HVO enabled, ret will be 0 and there is no need to call
1937 * __clear_hugetlb_destructor as this was done previously.
1939 VM_WARN_ON(!list_empty(folio_list));
1940 VM_WARN_ON(ret < 0);
1941 if (!list_empty(&non_hvo_folios) && ret) {
1942 spin_lock_irq(&hugetlb_lock);
1943 list_for_each_entry(folio, &non_hvo_folios, lru)
1944 __clear_hugetlb_destructor(h, folio);
1945 spin_unlock_irq(&hugetlb_lock);
1948 list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) {
1949 update_and_free_hugetlb_folio(h, folio, false);
1954 struct hstate *size_to_hstate(unsigned long size)
1958 for_each_hstate(h) {
1959 if (huge_page_size(h) == size)
1965 void free_huge_folio(struct folio *folio)
1968 * Can't pass hstate in here because it is called from the
1969 * compound page destructor.
1971 struct hstate *h = folio_hstate(folio);
1972 int nid = folio_nid(folio);
1973 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1974 bool restore_reserve;
1975 unsigned long flags;
1977 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1978 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1980 hugetlb_set_folio_subpool(folio, NULL);
1981 if (folio_test_anon(folio))
1982 __ClearPageAnonExclusive(&folio->page);
1983 folio->mapping = NULL;
1984 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1985 folio_clear_hugetlb_restore_reserve(folio);
1988 * If HPageRestoreReserve was set on page, page allocation consumed a
1989 * reservation. If the page was associated with a subpool, there
1990 * would have been a page reserved in the subpool before allocation
1991 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1992 * reservation, do not call hugepage_subpool_put_pages() as this will
1993 * remove the reserved page from the subpool.
1995 if (!restore_reserve) {
1997 * A return code of zero implies that the subpool will be
1998 * under its minimum size if the reservation is not restored
1999 * after page is free. Therefore, force restore_reserve
2002 if (hugepage_subpool_put_pages(spool, 1) == 0)
2003 restore_reserve = true;
2006 spin_lock_irqsave(&hugetlb_lock, flags);
2007 folio_clear_hugetlb_migratable(folio);
2008 hugetlb_cgroup_uncharge_folio(hstate_index(h),
2009 pages_per_huge_page(h), folio);
2010 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
2011 pages_per_huge_page(h), folio);
2012 mem_cgroup_uncharge(folio);
2013 if (restore_reserve)
2014 h->resv_huge_pages++;
2016 if (folio_test_hugetlb_temporary(folio)) {
2017 remove_hugetlb_folio(h, folio, false);
2018 spin_unlock_irqrestore(&hugetlb_lock, flags);
2019 update_and_free_hugetlb_folio(h, folio, true);
2020 } else if (h->surplus_huge_pages_node[nid]) {
2021 /* remove the page from active list */
2022 remove_hugetlb_folio(h, folio, true);
2023 spin_unlock_irqrestore(&hugetlb_lock, flags);
2024 update_and_free_hugetlb_folio(h, folio, true);
2026 arch_clear_hugepage_flags(&folio->page);
2027 enqueue_hugetlb_folio(h, folio);
2028 spin_unlock_irqrestore(&hugetlb_lock, flags);
2033 * Must be called with the hugetlb lock held
2035 static void __prep_account_new_huge_page(struct hstate *h, int nid)
2037 lockdep_assert_held(&hugetlb_lock);
2039 h->nr_huge_pages_node[nid]++;
2042 static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2044 folio_set_hugetlb(folio);
2045 INIT_LIST_HEAD(&folio->lru);
2046 hugetlb_set_folio_subpool(folio, NULL);
2047 set_hugetlb_cgroup(folio, NULL);
2048 set_hugetlb_cgroup_rsvd(folio, NULL);
2051 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2053 init_new_hugetlb_folio(h, folio);
2054 hugetlb_vmemmap_optimize_folio(h, folio);
2057 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
2059 __prep_new_hugetlb_folio(h, folio);
2060 spin_lock_irq(&hugetlb_lock);
2061 __prep_account_new_huge_page(h, nid);
2062 spin_unlock_irq(&hugetlb_lock);
2065 static bool __prep_compound_gigantic_folio(struct folio *folio,
2066 unsigned int order, bool demote)
2069 int nr_pages = 1 << order;
2072 __folio_clear_reserved(folio);
2073 for (i = 0; i < nr_pages; i++) {
2074 p = folio_page(folio, i);
2077 * For gigantic hugepages allocated through bootmem at
2078 * boot, it's safer to be consistent with the not-gigantic
2079 * hugepages and clear the PG_reserved bit from all tail pages
2080 * too. Otherwise drivers using get_user_pages() to access tail
2081 * pages may get the reference counting wrong if they see
2082 * PG_reserved set on a tail page (despite the head page not
2083 * having PG_reserved set). Enforcing this consistency between
2084 * head and tail pages allows drivers to optimize away a check
2085 * on the head page when they need know if put_page() is needed
2086 * after get_user_pages().
2088 if (i != 0) /* head page cleared above */
2089 __ClearPageReserved(p);
2091 * Subtle and very unlikely
2093 * Gigantic 'page allocators' such as memblock or cma will
2094 * return a set of pages with each page ref counted. We need
2095 * to turn this set of pages into a compound page with tail
2096 * page ref counts set to zero. Code such as speculative page
2097 * cache adding could take a ref on a 'to be' tail page.
2098 * We need to respect any increased ref count, and only set
2099 * the ref count to zero if count is currently 1. If count
2100 * is not 1, we return an error. An error return indicates
2101 * the set of pages can not be converted to a gigantic page.
2102 * The caller who allocated the pages should then discard the
2103 * pages using the appropriate free interface.
2105 * In the case of demote, the ref count will be zero.
2108 if (!page_ref_freeze(p, 1)) {
2109 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2113 VM_BUG_ON_PAGE(page_count(p), p);
2116 set_compound_head(p, &folio->page);
2118 __folio_set_head(folio);
2119 /* we rely on prep_new_hugetlb_folio to set the destructor */
2120 folio_set_order(folio, order);
2121 atomic_set(&folio->_entire_mapcount, -1);
2122 atomic_set(&folio->_nr_pages_mapped, 0);
2123 atomic_set(&folio->_pincount, 0);
2127 /* undo page modifications made above */
2128 for (j = 0; j < i; j++) {
2129 p = folio_page(folio, j);
2131 clear_compound_head(p);
2132 set_page_refcounted(p);
2134 /* need to clear PG_reserved on remaining tail pages */
2135 for (; j < nr_pages; j++) {
2136 p = folio_page(folio, j);
2137 __ClearPageReserved(p);
2142 static bool prep_compound_gigantic_folio(struct folio *folio,
2145 return __prep_compound_gigantic_folio(folio, order, false);
2148 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2151 return __prep_compound_gigantic_folio(folio, order, true);
2155 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2156 * transparent huge pages. See the PageTransHuge() documentation for more
2159 int PageHuge(struct page *page)
2161 struct folio *folio;
2163 if (!PageCompound(page))
2165 folio = page_folio(page);
2166 return folio_test_hugetlb(folio);
2168 EXPORT_SYMBOL_GPL(PageHuge);
2171 * Find and lock address space (mapping) in write mode.
2173 * Upon entry, the page is locked which means that page_mapping() is
2174 * stable. Due to locking order, we can only trylock_write. If we can
2175 * not get the lock, simply return NULL to caller.
2177 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2179 struct address_space *mapping = page_mapping(hpage);
2184 if (i_mmap_trylock_write(mapping))
2190 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2191 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2192 nodemask_t *node_alloc_noretry)
2194 int order = huge_page_order(h);
2196 bool alloc_try_hard = true;
2200 * By default we always try hard to allocate the page with
2201 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2202 * a loop (to adjust global huge page counts) and previous allocation
2203 * failed, do not continue to try hard on the same node. Use the
2204 * node_alloc_noretry bitmap to manage this state information.
2206 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2207 alloc_try_hard = false;
2208 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2210 gfp_mask |= __GFP_RETRY_MAYFAIL;
2211 if (nid == NUMA_NO_NODE)
2212 nid = numa_mem_id();
2214 page = __alloc_pages(gfp_mask, order, nid, nmask);
2216 /* Freeze head page */
2217 if (page && !page_ref_freeze(page, 1)) {
2218 __free_pages(page, order);
2219 if (retry) { /* retry once */
2223 /* WOW! twice in a row. */
2224 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2229 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2230 * indicates an overall state change. Clear bit so that we resume
2231 * normal 'try hard' allocations.
2233 if (node_alloc_noretry && page && !alloc_try_hard)
2234 node_clear(nid, *node_alloc_noretry);
2237 * If we tried hard to get a page but failed, set bit so that
2238 * subsequent attempts will not try as hard until there is an
2239 * overall state change.
2241 if (node_alloc_noretry && !page && alloc_try_hard)
2242 node_set(nid, *node_alloc_noretry);
2245 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2249 __count_vm_event(HTLB_BUDDY_PGALLOC);
2250 return page_folio(page);
2253 static struct folio *__alloc_fresh_hugetlb_folio(struct hstate *h,
2254 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2255 nodemask_t *node_alloc_noretry)
2257 struct folio *folio;
2261 if (hstate_is_gigantic(h))
2262 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2264 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2265 nid, nmask, node_alloc_noretry);
2269 if (hstate_is_gigantic(h)) {
2270 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2272 * Rare failure to convert pages to compound page.
2273 * Free pages and try again - ONCE!
2275 free_gigantic_folio(folio, huge_page_order(h));
2287 static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h,
2288 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2289 nodemask_t *node_alloc_noretry)
2291 struct folio *folio;
2293 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2294 node_alloc_noretry);
2296 init_new_hugetlb_folio(h, folio);
2301 * Common helper to allocate a fresh hugetlb page. All specific allocators
2302 * should use this function to get new hugetlb pages
2304 * Note that returned page is 'frozen': ref count of head page and all tail
2307 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2308 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2309 nodemask_t *node_alloc_noretry)
2311 struct folio *folio;
2313 folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2314 node_alloc_noretry);
2318 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2322 static void prep_and_add_allocated_folios(struct hstate *h,
2323 struct list_head *folio_list)
2325 unsigned long flags;
2326 struct folio *folio, *tmp_f;
2328 /* Send list for bulk vmemmap optimization processing */
2329 hugetlb_vmemmap_optimize_folios(h, folio_list);
2331 /* Add all new pool pages to free lists in one lock cycle */
2332 spin_lock_irqsave(&hugetlb_lock, flags);
2333 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
2334 __prep_account_new_huge_page(h, folio_nid(folio));
2335 enqueue_hugetlb_folio(h, folio);
2337 spin_unlock_irqrestore(&hugetlb_lock, flags);
2341 * Allocates a fresh hugetlb page in a node interleaved manner. The page
2342 * will later be added to the appropriate hugetlb pool.
2344 static struct folio *alloc_pool_huge_folio(struct hstate *h,
2345 nodemask_t *nodes_allowed,
2346 nodemask_t *node_alloc_noretry)
2348 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2351 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2352 struct folio *folio;
2354 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2355 nodes_allowed, node_alloc_noretry);
2364 * Remove huge page from pool from next node to free. Attempt to keep
2365 * persistent huge pages more or less balanced over allowed nodes.
2366 * This routine only 'removes' the hugetlb page. The caller must make
2367 * an additional call to free the page to low level allocators.
2368 * Called with hugetlb_lock locked.
2370 static struct folio *remove_pool_hugetlb_folio(struct hstate *h,
2371 nodemask_t *nodes_allowed, bool acct_surplus)
2374 struct folio *folio = NULL;
2376 lockdep_assert_held(&hugetlb_lock);
2377 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2379 * If we're returning unused surplus pages, only examine
2380 * nodes with surplus pages.
2382 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2383 !list_empty(&h->hugepage_freelists[node])) {
2384 folio = list_entry(h->hugepage_freelists[node].next,
2386 remove_hugetlb_folio(h, folio, acct_surplus);
2395 * Dissolve a given free hugepage into free buddy pages. This function does
2396 * nothing for in-use hugepages and non-hugepages.
2397 * This function returns values like below:
2399 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2400 * when the system is under memory pressure and the feature of
2401 * freeing unused vmemmap pages associated with each hugetlb page
2403 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2404 * (allocated or reserved.)
2405 * 0: successfully dissolved free hugepages or the page is not a
2406 * hugepage (considered as already dissolved)
2408 int dissolve_free_huge_page(struct page *page)
2411 struct folio *folio = page_folio(page);
2414 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2415 if (!folio_test_hugetlb(folio))
2418 spin_lock_irq(&hugetlb_lock);
2419 if (!folio_test_hugetlb(folio)) {
2424 if (!folio_ref_count(folio)) {
2425 struct hstate *h = folio_hstate(folio);
2426 if (!available_huge_pages(h))
2430 * We should make sure that the page is already on the free list
2431 * when it is dissolved.
2433 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2434 spin_unlock_irq(&hugetlb_lock);
2438 * Theoretically, we should return -EBUSY when we
2439 * encounter this race. In fact, we have a chance
2440 * to successfully dissolve the page if we do a
2441 * retry. Because the race window is quite small.
2442 * If we seize this opportunity, it is an optimization
2443 * for increasing the success rate of dissolving page.
2448 remove_hugetlb_folio(h, folio, false);
2449 h->max_huge_pages--;
2450 spin_unlock_irq(&hugetlb_lock);
2453 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2454 * before freeing the page. update_and_free_hugtlb_folio will fail to
2455 * free the page if it can not allocate required vmemmap. We
2456 * need to adjust max_huge_pages if the page is not freed.
2457 * Attempt to allocate vmemmmap here so that we can take
2458 * appropriate action on failure.
2460 * The folio_test_hugetlb check here is because
2461 * remove_hugetlb_folio will clear hugetlb folio flag for
2462 * non-vmemmap optimized hugetlb folios.
2464 if (folio_test_hugetlb(folio)) {
2465 rc = hugetlb_vmemmap_restore_folio(h, folio);
2467 spin_lock_irq(&hugetlb_lock);
2468 add_hugetlb_folio(h, folio, false);
2469 h->max_huge_pages++;
2475 update_and_free_hugetlb_folio(h, folio, false);
2479 spin_unlock_irq(&hugetlb_lock);
2484 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2485 * make specified memory blocks removable from the system.
2486 * Note that this will dissolve a free gigantic hugepage completely, if any
2487 * part of it lies within the given range.
2488 * Also note that if dissolve_free_huge_page() returns with an error, all
2489 * free hugepages that were dissolved before that error are lost.
2491 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2499 if (!hugepages_supported())
2502 order = huge_page_order(&default_hstate);
2504 order = min(order, huge_page_order(h));
2506 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2507 page = pfn_to_page(pfn);
2508 rc = dissolve_free_huge_page(page);
2517 * Allocates a fresh surplus page from the page allocator.
2519 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2520 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2522 struct folio *folio = NULL;
2524 if (hstate_is_gigantic(h))
2527 spin_lock_irq(&hugetlb_lock);
2528 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2530 spin_unlock_irq(&hugetlb_lock);
2532 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2536 spin_lock_irq(&hugetlb_lock);
2538 * We could have raced with the pool size change.
2539 * Double check that and simply deallocate the new page
2540 * if we would end up overcommiting the surpluses. Abuse
2541 * temporary page to workaround the nasty free_huge_folio
2544 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2545 folio_set_hugetlb_temporary(folio);
2546 spin_unlock_irq(&hugetlb_lock);
2547 free_huge_folio(folio);
2551 h->surplus_huge_pages++;
2552 h->surplus_huge_pages_node[folio_nid(folio)]++;
2555 spin_unlock_irq(&hugetlb_lock);
2560 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2561 int nid, nodemask_t *nmask)
2563 struct folio *folio;
2565 if (hstate_is_gigantic(h))
2568 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2572 /* fresh huge pages are frozen */
2573 folio_ref_unfreeze(folio, 1);
2575 * We do not account these pages as surplus because they are only
2576 * temporary and will be released properly on the last reference
2578 folio_set_hugetlb_temporary(folio);
2584 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2587 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2588 struct vm_area_struct *vma, unsigned long addr)
2590 struct folio *folio = NULL;
2591 struct mempolicy *mpol;
2592 gfp_t gfp_mask = htlb_alloc_mask(h);
2594 nodemask_t *nodemask;
2596 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2597 if (mpol_is_preferred_many(mpol)) {
2598 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2600 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2601 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2603 /* Fallback to all nodes if page==NULL */
2608 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2609 mpol_cond_put(mpol);
2613 /* folio migration callback function */
2614 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2615 nodemask_t *nmask, gfp_t gfp_mask)
2617 spin_lock_irq(&hugetlb_lock);
2618 if (available_huge_pages(h)) {
2619 struct folio *folio;
2621 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2622 preferred_nid, nmask);
2624 spin_unlock_irq(&hugetlb_lock);
2628 spin_unlock_irq(&hugetlb_lock);
2630 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2634 * Increase the hugetlb pool such that it can accommodate a reservation
2637 static int gather_surplus_pages(struct hstate *h, long delta)
2638 __must_hold(&hugetlb_lock)
2640 LIST_HEAD(surplus_list);
2641 struct folio *folio, *tmp;
2644 long needed, allocated;
2645 bool alloc_ok = true;
2647 lockdep_assert_held(&hugetlb_lock);
2648 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2650 h->resv_huge_pages += delta;
2658 spin_unlock_irq(&hugetlb_lock);
2659 for (i = 0; i < needed; i++) {
2660 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2661 NUMA_NO_NODE, NULL);
2666 list_add(&folio->lru, &surplus_list);
2672 * After retaking hugetlb_lock, we need to recalculate 'needed'
2673 * because either resv_huge_pages or free_huge_pages may have changed.
2675 spin_lock_irq(&hugetlb_lock);
2676 needed = (h->resv_huge_pages + delta) -
2677 (h->free_huge_pages + allocated);
2682 * We were not able to allocate enough pages to
2683 * satisfy the entire reservation so we free what
2684 * we've allocated so far.
2689 * The surplus_list now contains _at_least_ the number of extra pages
2690 * needed to accommodate the reservation. Add the appropriate number
2691 * of pages to the hugetlb pool and free the extras back to the buddy
2692 * allocator. Commit the entire reservation here to prevent another
2693 * process from stealing the pages as they are added to the pool but
2694 * before they are reserved.
2696 needed += allocated;
2697 h->resv_huge_pages += delta;
2700 /* Free the needed pages to the hugetlb pool */
2701 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2704 /* Add the page to the hugetlb allocator */
2705 enqueue_hugetlb_folio(h, folio);
2708 spin_unlock_irq(&hugetlb_lock);
2711 * Free unnecessary surplus pages to the buddy allocator.
2712 * Pages have no ref count, call free_huge_folio directly.
2714 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2715 free_huge_folio(folio);
2716 spin_lock_irq(&hugetlb_lock);
2722 * This routine has two main purposes:
2723 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2724 * in unused_resv_pages. This corresponds to the prior adjustments made
2725 * to the associated reservation map.
2726 * 2) Free any unused surplus pages that may have been allocated to satisfy
2727 * the reservation. As many as unused_resv_pages may be freed.
2729 static void return_unused_surplus_pages(struct hstate *h,
2730 unsigned long unused_resv_pages)
2732 unsigned long nr_pages;
2733 LIST_HEAD(page_list);
2735 lockdep_assert_held(&hugetlb_lock);
2736 /* Uncommit the reservation */
2737 h->resv_huge_pages -= unused_resv_pages;
2739 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2743 * Part (or even all) of the reservation could have been backed
2744 * by pre-allocated pages. Only free surplus pages.
2746 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2749 * We want to release as many surplus pages as possible, spread
2750 * evenly across all nodes with memory. Iterate across these nodes
2751 * until we can no longer free unreserved surplus pages. This occurs
2752 * when the nodes with surplus pages have no free pages.
2753 * remove_pool_hugetlb_folio() will balance the freed pages across the
2754 * on-line nodes with memory and will handle the hstate accounting.
2756 while (nr_pages--) {
2757 struct folio *folio;
2759 folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1);
2763 list_add(&folio->lru, &page_list);
2767 spin_unlock_irq(&hugetlb_lock);
2768 update_and_free_pages_bulk(h, &page_list);
2769 spin_lock_irq(&hugetlb_lock);
2774 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2775 * are used by the huge page allocation routines to manage reservations.
2777 * vma_needs_reservation is called to determine if the huge page at addr
2778 * within the vma has an associated reservation. If a reservation is
2779 * needed, the value 1 is returned. The caller is then responsible for
2780 * managing the global reservation and subpool usage counts. After
2781 * the huge page has been allocated, vma_commit_reservation is called
2782 * to add the page to the reservation map. If the page allocation fails,
2783 * the reservation must be ended instead of committed. vma_end_reservation
2784 * is called in such cases.
2786 * In the normal case, vma_commit_reservation returns the same value
2787 * as the preceding vma_needs_reservation call. The only time this
2788 * is not the case is if a reserve map was changed between calls. It
2789 * is the responsibility of the caller to notice the difference and
2790 * take appropriate action.
2792 * vma_add_reservation is used in error paths where a reservation must
2793 * be restored when a newly allocated huge page must be freed. It is
2794 * to be called after calling vma_needs_reservation to determine if a
2795 * reservation exists.
2797 * vma_del_reservation is used in error paths where an entry in the reserve
2798 * map was created during huge page allocation and must be removed. It is to
2799 * be called after calling vma_needs_reservation to determine if a reservation
2802 enum vma_resv_mode {
2809 static long __vma_reservation_common(struct hstate *h,
2810 struct vm_area_struct *vma, unsigned long addr,
2811 enum vma_resv_mode mode)
2813 struct resv_map *resv;
2816 long dummy_out_regions_needed;
2818 resv = vma_resv_map(vma);
2822 idx = vma_hugecache_offset(h, vma, addr);
2824 case VMA_NEEDS_RESV:
2825 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2826 /* We assume that vma_reservation_* routines always operate on
2827 * 1 page, and that adding to resv map a 1 page entry can only
2828 * ever require 1 region.
2830 VM_BUG_ON(dummy_out_regions_needed != 1);
2832 case VMA_COMMIT_RESV:
2833 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2834 /* region_add calls of range 1 should never fail. */
2838 region_abort(resv, idx, idx + 1, 1);
2842 if (vma->vm_flags & VM_MAYSHARE) {
2843 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2844 /* region_add calls of range 1 should never fail. */
2847 region_abort(resv, idx, idx + 1, 1);
2848 ret = region_del(resv, idx, idx + 1);
2852 if (vma->vm_flags & VM_MAYSHARE) {
2853 region_abort(resv, idx, idx + 1, 1);
2854 ret = region_del(resv, idx, idx + 1);
2856 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2857 /* region_add calls of range 1 should never fail. */
2865 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2868 * We know private mapping must have HPAGE_RESV_OWNER set.
2870 * In most cases, reserves always exist for private mappings.
2871 * However, a file associated with mapping could have been
2872 * hole punched or truncated after reserves were consumed.
2873 * As subsequent fault on such a range will not use reserves.
2874 * Subtle - The reserve map for private mappings has the
2875 * opposite meaning than that of shared mappings. If NO
2876 * entry is in the reserve map, it means a reservation exists.
2877 * If an entry exists in the reserve map, it means the
2878 * reservation has already been consumed. As a result, the
2879 * return value of this routine is the opposite of the
2880 * value returned from reserve map manipulation routines above.
2889 static long vma_needs_reservation(struct hstate *h,
2890 struct vm_area_struct *vma, unsigned long addr)
2892 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2895 static long vma_commit_reservation(struct hstate *h,
2896 struct vm_area_struct *vma, unsigned long addr)
2898 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2901 static void vma_end_reservation(struct hstate *h,
2902 struct vm_area_struct *vma, unsigned long addr)
2904 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2907 static long vma_add_reservation(struct hstate *h,
2908 struct vm_area_struct *vma, unsigned long addr)
2910 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2913 static long vma_del_reservation(struct hstate *h,
2914 struct vm_area_struct *vma, unsigned long addr)
2916 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2920 * This routine is called to restore reservation information on error paths.
2921 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2922 * and the hugetlb mutex should remain held when calling this routine.
2924 * It handles two specific cases:
2925 * 1) A reservation was in place and the folio consumed the reservation.
2926 * hugetlb_restore_reserve is set in the folio.
2927 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2928 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2930 * In case 1, free_huge_folio later in the error path will increment the
2931 * global reserve count. But, free_huge_folio does not have enough context
2932 * to adjust the reservation map. This case deals primarily with private
2933 * mappings. Adjust the reserve map here to be consistent with global
2934 * reserve count adjustments to be made by free_huge_folio. Make sure the
2935 * reserve map indicates there is a reservation present.
2937 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2939 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2940 unsigned long address, struct folio *folio)
2942 long rc = vma_needs_reservation(h, vma, address);
2944 if (folio_test_hugetlb_restore_reserve(folio)) {
2945 if (unlikely(rc < 0))
2947 * Rare out of memory condition in reserve map
2948 * manipulation. Clear hugetlb_restore_reserve so
2949 * that global reserve count will not be incremented
2950 * by free_huge_folio. This will make it appear
2951 * as though the reservation for this folio was
2952 * consumed. This may prevent the task from
2953 * faulting in the folio at a later time. This
2954 * is better than inconsistent global huge page
2955 * accounting of reserve counts.
2957 folio_clear_hugetlb_restore_reserve(folio);
2959 (void)vma_add_reservation(h, vma, address);
2961 vma_end_reservation(h, vma, address);
2965 * This indicates there is an entry in the reserve map
2966 * not added by alloc_hugetlb_folio. We know it was added
2967 * before the alloc_hugetlb_folio call, otherwise
2968 * hugetlb_restore_reserve would be set on the folio.
2969 * Remove the entry so that a subsequent allocation
2970 * does not consume a reservation.
2972 rc = vma_del_reservation(h, vma, address);
2975 * VERY rare out of memory condition. Since
2976 * we can not delete the entry, set
2977 * hugetlb_restore_reserve so that the reserve
2978 * count will be incremented when the folio
2979 * is freed. This reserve will be consumed
2980 * on a subsequent allocation.
2982 folio_set_hugetlb_restore_reserve(folio);
2983 } else if (rc < 0) {
2985 * Rare out of memory condition from
2986 * vma_needs_reservation call. Memory allocation is
2987 * only attempted if a new entry is needed. Therefore,
2988 * this implies there is not an entry in the
2991 * For shared mappings, no entry in the map indicates
2992 * no reservation. We are done.
2994 if (!(vma->vm_flags & VM_MAYSHARE))
2996 * For private mappings, no entry indicates
2997 * a reservation is present. Since we can
2998 * not add an entry, set hugetlb_restore_reserve
2999 * on the folio so reserve count will be
3000 * incremented when freed. This reserve will
3001 * be consumed on a subsequent allocation.
3003 folio_set_hugetlb_restore_reserve(folio);
3006 * No reservation present, do nothing
3008 vma_end_reservation(h, vma, address);
3013 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
3015 * @h: struct hstate old page belongs to
3016 * @old_folio: Old folio to dissolve
3017 * @list: List to isolate the page in case we need to
3018 * Returns 0 on success, otherwise negated error.
3020 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
3021 struct folio *old_folio, struct list_head *list)
3023 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3024 int nid = folio_nid(old_folio);
3025 struct folio *new_folio;
3029 * Before dissolving the folio, we need to allocate a new one for the
3030 * pool to remain stable. Here, we allocate the folio and 'prep' it
3031 * by doing everything but actually updating counters and adding to
3032 * the pool. This simplifies and let us do most of the processing
3035 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
3038 __prep_new_hugetlb_folio(h, new_folio);
3041 spin_lock_irq(&hugetlb_lock);
3042 if (!folio_test_hugetlb(old_folio)) {
3044 * Freed from under us. Drop new_folio too.
3047 } else if (folio_ref_count(old_folio)) {
3051 * Someone has grabbed the folio, try to isolate it here.
3052 * Fail with -EBUSY if not possible.
3054 spin_unlock_irq(&hugetlb_lock);
3055 isolated = isolate_hugetlb(old_folio, list);
3056 ret = isolated ? 0 : -EBUSY;
3057 spin_lock_irq(&hugetlb_lock);
3059 } else if (!folio_test_hugetlb_freed(old_folio)) {
3061 * Folio's refcount is 0 but it has not been enqueued in the
3062 * freelist yet. Race window is small, so we can succeed here if
3065 spin_unlock_irq(&hugetlb_lock);
3070 * Ok, old_folio is still a genuine free hugepage. Remove it from
3071 * the freelist and decrease the counters. These will be
3072 * incremented again when calling __prep_account_new_huge_page()
3073 * and enqueue_hugetlb_folio() for new_folio. The counters will
3074 * remain stable since this happens under the lock.
3076 remove_hugetlb_folio(h, old_folio, false);
3079 * Ref count on new_folio is already zero as it was dropped
3080 * earlier. It can be directly added to the pool free list.
3082 __prep_account_new_huge_page(h, nid);
3083 enqueue_hugetlb_folio(h, new_folio);
3086 * Folio has been replaced, we can safely free the old one.
3088 spin_unlock_irq(&hugetlb_lock);
3089 update_and_free_hugetlb_folio(h, old_folio, false);
3095 spin_unlock_irq(&hugetlb_lock);
3096 /* Folio has a zero ref count, but needs a ref to be freed */
3097 folio_ref_unfreeze(new_folio, 1);
3098 update_and_free_hugetlb_folio(h, new_folio, false);
3103 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3106 struct folio *folio = page_folio(page);
3110 * The page might have been dissolved from under our feet, so make sure
3111 * to carefully check the state under the lock.
3112 * Return success when racing as if we dissolved the page ourselves.
3114 spin_lock_irq(&hugetlb_lock);
3115 if (folio_test_hugetlb(folio)) {
3116 h = folio_hstate(folio);
3118 spin_unlock_irq(&hugetlb_lock);
3121 spin_unlock_irq(&hugetlb_lock);
3124 * Fence off gigantic pages as there is a cyclic dependency between
3125 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3126 * of bailing out right away without further retrying.
3128 if (hstate_is_gigantic(h))
3131 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3133 else if (!folio_ref_count(folio))
3134 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3139 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3140 unsigned long addr, int avoid_reserve)
3142 struct hugepage_subpool *spool = subpool_vma(vma);
3143 struct hstate *h = hstate_vma(vma);
3144 struct folio *folio;
3145 long map_chg, map_commit, nr_pages = pages_per_huge_page(h);
3147 int memcg_charge_ret, ret, idx;
3148 struct hugetlb_cgroup *h_cg = NULL;
3149 struct mem_cgroup *memcg;
3150 bool deferred_reserve;
3151 gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL;
3153 memcg = get_mem_cgroup_from_current();
3154 memcg_charge_ret = mem_cgroup_hugetlb_try_charge(memcg, gfp, nr_pages);
3155 if (memcg_charge_ret == -ENOMEM) {
3156 mem_cgroup_put(memcg);
3157 return ERR_PTR(-ENOMEM);
3160 idx = hstate_index(h);
3162 * Examine the region/reserve map to determine if the process
3163 * has a reservation for the page to be allocated. A return
3164 * code of zero indicates a reservation exists (no change).
3166 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3168 if (!memcg_charge_ret)
3169 mem_cgroup_cancel_charge(memcg, nr_pages);
3170 mem_cgroup_put(memcg);
3171 return ERR_PTR(-ENOMEM);
3175 * Processes that did not create the mapping will have no
3176 * reserves as indicated by the region/reserve map. Check
3177 * that the allocation will not exceed the subpool limit.
3178 * Allocations for MAP_NORESERVE mappings also need to be
3179 * checked against any subpool limit.
3181 if (map_chg || avoid_reserve) {
3182 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3184 goto out_end_reservation;
3187 * Even though there was no reservation in the region/reserve
3188 * map, there could be reservations associated with the
3189 * subpool that can be used. This would be indicated if the
3190 * return value of hugepage_subpool_get_pages() is zero.
3191 * However, if avoid_reserve is specified we still avoid even
3192 * the subpool reservations.
3198 /* If this allocation is not consuming a reservation, charge it now.
3200 deferred_reserve = map_chg || avoid_reserve;
3201 if (deferred_reserve) {
3202 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3203 idx, pages_per_huge_page(h), &h_cg);
3205 goto out_subpool_put;
3208 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3210 goto out_uncharge_cgroup_reservation;
3212 spin_lock_irq(&hugetlb_lock);
3214 * glb_chg is passed to indicate whether or not a page must be taken
3215 * from the global free pool (global change). gbl_chg == 0 indicates
3216 * a reservation exists for the allocation.
3218 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3220 spin_unlock_irq(&hugetlb_lock);
3221 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3223 goto out_uncharge_cgroup;
3224 spin_lock_irq(&hugetlb_lock);
3225 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3226 folio_set_hugetlb_restore_reserve(folio);
3227 h->resv_huge_pages--;
3229 list_add(&folio->lru, &h->hugepage_activelist);
3230 folio_ref_unfreeze(folio, 1);
3234 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3235 /* If allocation is not consuming a reservation, also store the
3236 * hugetlb_cgroup pointer on the page.
3238 if (deferred_reserve) {
3239 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3243 spin_unlock_irq(&hugetlb_lock);
3245 hugetlb_set_folio_subpool(folio, spool);
3247 map_commit = vma_commit_reservation(h, vma, addr);
3248 if (unlikely(map_chg > map_commit)) {
3250 * The page was added to the reservation map between
3251 * vma_needs_reservation and vma_commit_reservation.
3252 * This indicates a race with hugetlb_reserve_pages.
3253 * Adjust for the subpool count incremented above AND
3254 * in hugetlb_reserve_pages for the same page. Also,
3255 * the reservation count added in hugetlb_reserve_pages
3256 * no longer applies.
3260 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3261 hugetlb_acct_memory(h, -rsv_adjust);
3262 if (deferred_reserve)
3263 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3264 pages_per_huge_page(h), folio);
3267 if (!memcg_charge_ret)
3268 mem_cgroup_commit_charge(folio, memcg);
3269 mem_cgroup_put(memcg);
3273 out_uncharge_cgroup:
3274 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3275 out_uncharge_cgroup_reservation:
3276 if (deferred_reserve)
3277 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3280 if (map_chg || avoid_reserve)
3281 hugepage_subpool_put_pages(spool, 1);
3282 out_end_reservation:
3283 vma_end_reservation(h, vma, addr);
3284 if (!memcg_charge_ret)
3285 mem_cgroup_cancel_charge(memcg, nr_pages);
3286 mem_cgroup_put(memcg);
3287 return ERR_PTR(-ENOSPC);
3290 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3291 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3292 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3294 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3297 /* do node specific alloc */
3298 if (nid != NUMA_NO_NODE) {
3299 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3300 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3305 /* allocate from next node when distributing huge pages */
3306 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3307 m = memblock_alloc_try_nid_raw(
3308 huge_page_size(h), huge_page_size(h),
3309 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3311 * Use the beginning of the huge page to store the
3312 * huge_bootmem_page struct (until gather_bootmem
3313 * puts them into the mem_map).
3323 * Only initialize the head struct page in memmap_init_reserved_pages,
3324 * rest of the struct pages will be initialized by the HugeTLB
3326 * The head struct page is used to get folio information by the HugeTLB
3327 * subsystem like zone id and node id.
3329 memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE),
3330 huge_page_size(h) - PAGE_SIZE);
3331 /* Put them into a private list first because mem_map is not up yet */
3332 INIT_LIST_HEAD(&m->list);
3333 list_add(&m->list, &huge_boot_pages);
3338 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
3339 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
3340 unsigned long start_page_number,
3341 unsigned long end_page_number)
3343 enum zone_type zone = zone_idx(folio_zone(folio));
3344 int nid = folio_nid(folio);
3345 unsigned long head_pfn = folio_pfn(folio);
3346 unsigned long pfn, end_pfn = head_pfn + end_page_number;
3349 for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) {
3350 struct page *page = pfn_to_page(pfn);
3352 __init_single_page(page, pfn, zone, nid);
3353 prep_compound_tail((struct page *)folio, pfn - head_pfn);
3354 ret = page_ref_freeze(page, 1);
3359 static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
3361 unsigned long nr_pages)
3365 /* Prepare folio head */
3366 __folio_clear_reserved(folio);
3367 __folio_set_head(folio);
3368 ret = folio_ref_freeze(folio, 1);
3370 /* Initialize the necessary tail struct pages */
3371 hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);
3372 prep_compound_head((struct page *)folio, huge_page_order(h));
3375 static void __init prep_and_add_bootmem_folios(struct hstate *h,
3376 struct list_head *folio_list)
3378 unsigned long flags;
3379 struct folio *folio, *tmp_f;
3381 /* Send list for bulk vmemmap optimization processing */
3382 hugetlb_vmemmap_optimize_folios(h, folio_list);
3384 /* Add all new pool pages to free lists in one lock cycle */
3385 spin_lock_irqsave(&hugetlb_lock, flags);
3386 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
3387 if (!folio_test_hugetlb_vmemmap_optimized(folio)) {
3389 * If HVO fails, initialize all tail struct pages
3390 * We do not worry about potential long lock hold
3391 * time as this is early in boot and there should
3394 hugetlb_folio_init_tail_vmemmap(folio,
3395 HUGETLB_VMEMMAP_RESERVE_PAGES,
3396 pages_per_huge_page(h));
3398 __prep_account_new_huge_page(h, folio_nid(folio));
3399 enqueue_hugetlb_folio(h, folio);
3401 spin_unlock_irqrestore(&hugetlb_lock, flags);
3405 * Put bootmem huge pages into the standard lists after mem_map is up.
3406 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3408 static void __init gather_bootmem_prealloc(void)
3410 LIST_HEAD(folio_list);
3411 struct huge_bootmem_page *m;
3412 struct hstate *h = NULL, *prev_h = NULL;
3414 list_for_each_entry(m, &huge_boot_pages, list) {
3415 struct page *page = virt_to_page(m);
3416 struct folio *folio = (void *)page;
3420 * It is possible to have multiple huge page sizes (hstates)
3421 * in this list. If so, process each size separately.
3423 if (h != prev_h && prev_h != NULL)
3424 prep_and_add_bootmem_folios(prev_h, &folio_list);
3427 VM_BUG_ON(!hstate_is_gigantic(h));
3428 WARN_ON(folio_ref_count(folio) != 1);
3430 hugetlb_folio_init_vmemmap(folio, h,
3431 HUGETLB_VMEMMAP_RESERVE_PAGES);
3432 init_new_hugetlb_folio(h, folio);
3433 list_add(&folio->lru, &folio_list);
3436 * We need to restore the 'stolen' pages to totalram_pages
3437 * in order to fix confusing memory reports from free(1) and
3438 * other side-effects, like CommitLimit going negative.
3440 adjust_managed_page_count(page, pages_per_huge_page(h));
3444 prep_and_add_bootmem_folios(h, &folio_list);
3447 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3452 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3453 if (hstate_is_gigantic(h)) {
3454 if (!alloc_bootmem_huge_page(h, nid))
3457 struct folio *folio;
3458 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3460 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3461 &node_states[N_MEMORY], NULL);
3464 free_huge_folio(folio); /* free it into the hugepage allocator */
3468 if (i == h->max_huge_pages_node[nid])
3471 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3472 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3473 h->max_huge_pages_node[nid], buf, nid, i);
3474 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3475 h->max_huge_pages_node[nid] = i;
3479 * NOTE: this routine is called in different contexts for gigantic and
3480 * non-gigantic pages.
3481 * - For gigantic pages, this is called early in the boot process and
3482 * pages are allocated from memblock allocated or something similar.
3483 * Gigantic pages are actually added to pools later with the routine
3484 * gather_bootmem_prealloc.
3485 * - For non-gigantic pages, this is called later in the boot process after
3486 * all of mm is up and functional. Pages are allocated from buddy and
3487 * then added to hugetlb pools.
3489 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3492 struct folio *folio;
3493 LIST_HEAD(folio_list);
3494 nodemask_t *node_alloc_noretry;
3495 bool node_specific_alloc = false;
3497 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3498 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3499 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3503 /* do node specific alloc */
3504 for_each_online_node(i) {
3505 if (h->max_huge_pages_node[i] > 0) {
3506 hugetlb_hstate_alloc_pages_onenode(h, i);
3507 node_specific_alloc = true;
3511 if (node_specific_alloc)
3514 /* below will do all node balanced alloc */
3515 if (!hstate_is_gigantic(h)) {
3517 * Bit mask controlling how hard we retry per-node allocations.
3518 * Ignore errors as lower level routines can deal with
3519 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3520 * time, we are likely in bigger trouble.
3522 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3525 /* allocations done at boot time */
3526 node_alloc_noretry = NULL;
3529 /* bit mask controlling how hard we retry per-node allocations */
3530 if (node_alloc_noretry)
3531 nodes_clear(*node_alloc_noretry);
3533 for (i = 0; i < h->max_huge_pages; ++i) {
3534 if (hstate_is_gigantic(h)) {
3536 * gigantic pages not added to list as they are not
3537 * added to pools now.
3539 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3542 folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
3543 node_alloc_noretry);
3546 list_add(&folio->lru, &folio_list);
3551 /* list will be empty if hstate_is_gigantic */
3552 prep_and_add_allocated_folios(h, &folio_list);
3554 if (i < h->max_huge_pages) {
3557 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3558 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3559 h->max_huge_pages, buf, i);
3560 h->max_huge_pages = i;
3562 kfree(node_alloc_noretry);
3565 static void __init hugetlb_init_hstates(void)
3567 struct hstate *h, *h2;
3569 for_each_hstate(h) {
3570 /* oversize hugepages were init'ed in early boot */
3571 if (!hstate_is_gigantic(h))
3572 hugetlb_hstate_alloc_pages(h);
3575 * Set demote order for each hstate. Note that
3576 * h->demote_order is initially 0.
3577 * - We can not demote gigantic pages if runtime freeing
3578 * is not supported, so skip this.
3579 * - If CMA allocation is possible, we can not demote
3580 * HUGETLB_PAGE_ORDER or smaller size pages.
3582 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3584 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3586 for_each_hstate(h2) {
3589 if (h2->order < h->order &&
3590 h2->order > h->demote_order)
3591 h->demote_order = h2->order;
3596 static void __init report_hugepages(void)
3600 for_each_hstate(h) {
3603 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3604 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3605 buf, h->free_huge_pages);
3606 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3607 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3611 #ifdef CONFIG_HIGHMEM
3612 static void try_to_free_low(struct hstate *h, unsigned long count,
3613 nodemask_t *nodes_allowed)
3616 LIST_HEAD(page_list);
3618 lockdep_assert_held(&hugetlb_lock);
3619 if (hstate_is_gigantic(h))
3623 * Collect pages to be freed on a list, and free after dropping lock
3625 for_each_node_mask(i, *nodes_allowed) {
3626 struct folio *folio, *next;
3627 struct list_head *freel = &h->hugepage_freelists[i];
3628 list_for_each_entry_safe(folio, next, freel, lru) {
3629 if (count >= h->nr_huge_pages)
3631 if (folio_test_highmem(folio))
3633 remove_hugetlb_folio(h, folio, false);
3634 list_add(&folio->lru, &page_list);
3639 spin_unlock_irq(&hugetlb_lock);
3640 update_and_free_pages_bulk(h, &page_list);
3641 spin_lock_irq(&hugetlb_lock);
3644 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3645 nodemask_t *nodes_allowed)
3651 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3652 * balanced by operating on them in a round-robin fashion.
3653 * Returns 1 if an adjustment was made.
3655 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3660 lockdep_assert_held(&hugetlb_lock);
3661 VM_BUG_ON(delta != -1 && delta != 1);
3664 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3665 if (h->surplus_huge_pages_node[node])
3669 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3670 if (h->surplus_huge_pages_node[node] <
3671 h->nr_huge_pages_node[node])
3678 h->surplus_huge_pages += delta;
3679 h->surplus_huge_pages_node[node] += delta;
3683 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3684 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3685 nodemask_t *nodes_allowed)
3687 unsigned long min_count;
3688 unsigned long allocated;
3689 struct folio *folio;
3690 LIST_HEAD(page_list);
3691 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3694 * Bit mask controlling how hard we retry per-node allocations.
3695 * If we can not allocate the bit mask, do not attempt to allocate
3696 * the requested huge pages.
3698 if (node_alloc_noretry)
3699 nodes_clear(*node_alloc_noretry);
3704 * resize_lock mutex prevents concurrent adjustments to number of
3705 * pages in hstate via the proc/sysfs interfaces.
3707 mutex_lock(&h->resize_lock);
3708 flush_free_hpage_work(h);
3709 spin_lock_irq(&hugetlb_lock);
3712 * Check for a node specific request.
3713 * Changing node specific huge page count may require a corresponding
3714 * change to the global count. In any case, the passed node mask
3715 * (nodes_allowed) will restrict alloc/free to the specified node.
3717 if (nid != NUMA_NO_NODE) {
3718 unsigned long old_count = count;
3720 count += persistent_huge_pages(h) -
3721 (h->nr_huge_pages_node[nid] -
3722 h->surplus_huge_pages_node[nid]);
3724 * User may have specified a large count value which caused the
3725 * above calculation to overflow. In this case, they wanted
3726 * to allocate as many huge pages as possible. Set count to
3727 * largest possible value to align with their intention.
3729 if (count < old_count)
3734 * Gigantic pages runtime allocation depend on the capability for large
3735 * page range allocation.
3736 * If the system does not provide this feature, return an error when
3737 * the user tries to allocate gigantic pages but let the user free the
3738 * boottime allocated gigantic pages.
3740 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3741 if (count > persistent_huge_pages(h)) {
3742 spin_unlock_irq(&hugetlb_lock);
3743 mutex_unlock(&h->resize_lock);
3744 NODEMASK_FREE(node_alloc_noretry);
3747 /* Fall through to decrease pool */
3751 * Increase the pool size
3752 * First take pages out of surplus state. Then make up the
3753 * remaining difference by allocating fresh huge pages.
3755 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3756 * to convert a surplus huge page to a normal huge page. That is
3757 * not critical, though, it just means the overall size of the
3758 * pool might be one hugepage larger than it needs to be, but
3759 * within all the constraints specified by the sysctls.
3761 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3762 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3767 while (count > (persistent_huge_pages(h) + allocated)) {
3769 * If this allocation races such that we no longer need the
3770 * page, free_huge_folio will handle it by freeing the page
3771 * and reducing the surplus.
3773 spin_unlock_irq(&hugetlb_lock);
3775 /* yield cpu to avoid soft lockup */
3778 folio = alloc_pool_huge_folio(h, nodes_allowed,
3779 node_alloc_noretry);
3781 prep_and_add_allocated_folios(h, &page_list);
3782 spin_lock_irq(&hugetlb_lock);
3786 list_add(&folio->lru, &page_list);
3789 /* Bail for signals. Probably ctrl-c from user */
3790 if (signal_pending(current)) {
3791 prep_and_add_allocated_folios(h, &page_list);
3792 spin_lock_irq(&hugetlb_lock);
3796 spin_lock_irq(&hugetlb_lock);
3799 /* Add allocated pages to the pool */
3800 if (!list_empty(&page_list)) {
3801 spin_unlock_irq(&hugetlb_lock);
3802 prep_and_add_allocated_folios(h, &page_list);
3803 spin_lock_irq(&hugetlb_lock);
3807 * Decrease the pool size
3808 * First return free pages to the buddy allocator (being careful
3809 * to keep enough around to satisfy reservations). Then place
3810 * pages into surplus state as needed so the pool will shrink
3811 * to the desired size as pages become free.
3813 * By placing pages into the surplus state independent of the
3814 * overcommit value, we are allowing the surplus pool size to
3815 * exceed overcommit. There are few sane options here. Since
3816 * alloc_surplus_hugetlb_folio() is checking the global counter,
3817 * though, we'll note that we're not allowed to exceed surplus
3818 * and won't grow the pool anywhere else. Not until one of the
3819 * sysctls are changed, or the surplus pages go out of use.
3821 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3822 min_count = max(count, min_count);
3823 try_to_free_low(h, min_count, nodes_allowed);
3826 * Collect pages to be removed on list without dropping lock
3828 while (min_count < persistent_huge_pages(h)) {
3829 folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0);
3833 list_add(&folio->lru, &page_list);
3835 /* free the pages after dropping lock */
3836 spin_unlock_irq(&hugetlb_lock);
3837 update_and_free_pages_bulk(h, &page_list);
3838 flush_free_hpage_work(h);
3839 spin_lock_irq(&hugetlb_lock);
3841 while (count < persistent_huge_pages(h)) {
3842 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3846 h->max_huge_pages = persistent_huge_pages(h);
3847 spin_unlock_irq(&hugetlb_lock);
3848 mutex_unlock(&h->resize_lock);
3850 NODEMASK_FREE(node_alloc_noretry);
3855 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3857 int i, nid = folio_nid(folio);
3858 struct hstate *target_hstate;
3859 struct page *subpage;
3860 struct folio *inner_folio;
3863 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3865 remove_hugetlb_folio_for_demote(h, folio, false);
3866 spin_unlock_irq(&hugetlb_lock);
3869 * If vmemmap already existed for folio, the remove routine above would
3870 * have cleared the hugetlb folio flag. Hence the folio is technically
3871 * no longer a hugetlb folio. hugetlb_vmemmap_restore_folio can only be
3872 * passed hugetlb folios and will BUG otherwise.
3874 if (folio_test_hugetlb(folio)) {
3875 rc = hugetlb_vmemmap_restore_folio(h, folio);
3877 /* Allocation of vmemmmap failed, we can not demote folio */
3878 spin_lock_irq(&hugetlb_lock);
3879 folio_ref_unfreeze(folio, 1);
3880 add_hugetlb_folio(h, folio, false);
3886 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3887 * sizes as it will not ref count folios.
3889 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3892 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3893 * Without the mutex, pages added to target hstate could be marked
3896 * Note that we already hold h->resize_lock. To prevent deadlock,
3897 * use the convention of always taking larger size hstate mutex first.
3899 mutex_lock(&target_hstate->resize_lock);
3900 for (i = 0; i < pages_per_huge_page(h);
3901 i += pages_per_huge_page(target_hstate)) {
3902 subpage = folio_page(folio, i);
3903 inner_folio = page_folio(subpage);
3904 if (hstate_is_gigantic(target_hstate))
3905 prep_compound_gigantic_folio_for_demote(inner_folio,
3906 target_hstate->order);
3908 prep_compound_page(subpage, target_hstate->order);
3909 folio_change_private(inner_folio, NULL);
3910 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3911 free_huge_folio(inner_folio);
3913 mutex_unlock(&target_hstate->resize_lock);
3915 spin_lock_irq(&hugetlb_lock);
3918 * Not absolutely necessary, but for consistency update max_huge_pages
3919 * based on pool changes for the demoted page.
3921 h->max_huge_pages--;
3922 target_hstate->max_huge_pages +=
3923 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3928 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3929 __must_hold(&hugetlb_lock)
3932 struct folio *folio;
3934 lockdep_assert_held(&hugetlb_lock);
3936 /* We should never get here if no demote order */
3937 if (!h->demote_order) {
3938 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3939 return -EINVAL; /* internal error */
3942 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3943 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3944 if (folio_test_hwpoison(folio))
3946 return demote_free_hugetlb_folio(h, folio);
3951 * Only way to get here is if all pages on free lists are poisoned.
3952 * Return -EBUSY so that caller will not retry.
3957 #define HSTATE_ATTR_RO(_name) \
3958 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3960 #define HSTATE_ATTR_WO(_name) \
3961 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3963 #define HSTATE_ATTR(_name) \
3964 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3966 static struct kobject *hugepages_kobj;
3967 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3969 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3971 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3975 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3976 if (hstate_kobjs[i] == kobj) {
3978 *nidp = NUMA_NO_NODE;
3982 return kobj_to_node_hstate(kobj, nidp);
3985 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3986 struct kobj_attribute *attr, char *buf)
3989 unsigned long nr_huge_pages;
3992 h = kobj_to_hstate(kobj, &nid);
3993 if (nid == NUMA_NO_NODE)
3994 nr_huge_pages = h->nr_huge_pages;
3996 nr_huge_pages = h->nr_huge_pages_node[nid];
3998 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
4001 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
4002 struct hstate *h, int nid,
4003 unsigned long count, size_t len)
4006 nodemask_t nodes_allowed, *n_mask;
4008 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
4011 if (nid == NUMA_NO_NODE) {
4013 * global hstate attribute
4015 if (!(obey_mempolicy &&
4016 init_nodemask_of_mempolicy(&nodes_allowed)))
4017 n_mask = &node_states[N_MEMORY];
4019 n_mask = &nodes_allowed;
4022 * Node specific request. count adjustment happens in
4023 * set_max_huge_pages() after acquiring hugetlb_lock.
4025 init_nodemask_of_node(&nodes_allowed, nid);
4026 n_mask = &nodes_allowed;
4029 err = set_max_huge_pages(h, count, nid, n_mask);
4031 return err ? err : len;
4034 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
4035 struct kobject *kobj, const char *buf,
4039 unsigned long count;
4043 err = kstrtoul(buf, 10, &count);
4047 h = kobj_to_hstate(kobj, &nid);
4048 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
4051 static ssize_t nr_hugepages_show(struct kobject *kobj,
4052 struct kobj_attribute *attr, char *buf)
4054 return nr_hugepages_show_common(kobj, attr, buf);
4057 static ssize_t nr_hugepages_store(struct kobject *kobj,
4058 struct kobj_attribute *attr, const char *buf, size_t len)
4060 return nr_hugepages_store_common(false, kobj, buf, len);
4062 HSTATE_ATTR(nr_hugepages);
4067 * hstate attribute for optionally mempolicy-based constraint on persistent
4068 * huge page alloc/free.
4070 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
4071 struct kobj_attribute *attr,
4074 return nr_hugepages_show_common(kobj, attr, buf);
4077 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
4078 struct kobj_attribute *attr, const char *buf, size_t len)
4080 return nr_hugepages_store_common(true, kobj, buf, len);
4082 HSTATE_ATTR(nr_hugepages_mempolicy);
4086 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
4087 struct kobj_attribute *attr, char *buf)
4089 struct hstate *h = kobj_to_hstate(kobj, NULL);
4090 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
4093 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
4094 struct kobj_attribute *attr, const char *buf, size_t count)
4097 unsigned long input;
4098 struct hstate *h = kobj_to_hstate(kobj, NULL);
4100 if (hstate_is_gigantic(h))
4103 err = kstrtoul(buf, 10, &input);
4107 spin_lock_irq(&hugetlb_lock);
4108 h->nr_overcommit_huge_pages = input;
4109 spin_unlock_irq(&hugetlb_lock);
4113 HSTATE_ATTR(nr_overcommit_hugepages);
4115 static ssize_t free_hugepages_show(struct kobject *kobj,
4116 struct kobj_attribute *attr, char *buf)
4119 unsigned long free_huge_pages;
4122 h = kobj_to_hstate(kobj, &nid);
4123 if (nid == NUMA_NO_NODE)
4124 free_huge_pages = h->free_huge_pages;
4126 free_huge_pages = h->free_huge_pages_node[nid];
4128 return sysfs_emit(buf, "%lu\n", free_huge_pages);
4130 HSTATE_ATTR_RO(free_hugepages);
4132 static ssize_t resv_hugepages_show(struct kobject *kobj,
4133 struct kobj_attribute *attr, char *buf)
4135 struct hstate *h = kobj_to_hstate(kobj, NULL);
4136 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
4138 HSTATE_ATTR_RO(resv_hugepages);
4140 static ssize_t surplus_hugepages_show(struct kobject *kobj,
4141 struct kobj_attribute *attr, char *buf)
4144 unsigned long surplus_huge_pages;
4147 h = kobj_to_hstate(kobj, &nid);
4148 if (nid == NUMA_NO_NODE)
4149 surplus_huge_pages = h->surplus_huge_pages;
4151 surplus_huge_pages = h->surplus_huge_pages_node[nid];
4153 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
4155 HSTATE_ATTR_RO(surplus_hugepages);
4157 static ssize_t demote_store(struct kobject *kobj,
4158 struct kobj_attribute *attr, const char *buf, size_t len)
4160 unsigned long nr_demote;
4161 unsigned long nr_available;
4162 nodemask_t nodes_allowed, *n_mask;
4167 err = kstrtoul(buf, 10, &nr_demote);
4170 h = kobj_to_hstate(kobj, &nid);
4172 if (nid != NUMA_NO_NODE) {
4173 init_nodemask_of_node(&nodes_allowed, nid);
4174 n_mask = &nodes_allowed;
4176 n_mask = &node_states[N_MEMORY];
4179 /* Synchronize with other sysfs operations modifying huge pages */
4180 mutex_lock(&h->resize_lock);
4181 spin_lock_irq(&hugetlb_lock);
4185 * Check for available pages to demote each time thorough the
4186 * loop as demote_pool_huge_page will drop hugetlb_lock.
4188 if (nid != NUMA_NO_NODE)
4189 nr_available = h->free_huge_pages_node[nid];
4191 nr_available = h->free_huge_pages;
4192 nr_available -= h->resv_huge_pages;
4196 err = demote_pool_huge_page(h, n_mask);
4203 spin_unlock_irq(&hugetlb_lock);
4204 mutex_unlock(&h->resize_lock);
4210 HSTATE_ATTR_WO(demote);
4212 static ssize_t demote_size_show(struct kobject *kobj,
4213 struct kobj_attribute *attr, char *buf)
4215 struct hstate *h = kobj_to_hstate(kobj, NULL);
4216 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
4218 return sysfs_emit(buf, "%lukB\n", demote_size);
4221 static ssize_t demote_size_store(struct kobject *kobj,
4222 struct kobj_attribute *attr,
4223 const char *buf, size_t count)
4225 struct hstate *h, *demote_hstate;
4226 unsigned long demote_size;
4227 unsigned int demote_order;
4229 demote_size = (unsigned long)memparse(buf, NULL);
4231 demote_hstate = size_to_hstate(demote_size);
4234 demote_order = demote_hstate->order;
4235 if (demote_order < HUGETLB_PAGE_ORDER)
4238 /* demote order must be smaller than hstate order */
4239 h = kobj_to_hstate(kobj, NULL);
4240 if (demote_order >= h->order)
4243 /* resize_lock synchronizes access to demote size and writes */
4244 mutex_lock(&h->resize_lock);
4245 h->demote_order = demote_order;
4246 mutex_unlock(&h->resize_lock);
4250 HSTATE_ATTR(demote_size);
4252 static struct attribute *hstate_attrs[] = {
4253 &nr_hugepages_attr.attr,
4254 &nr_overcommit_hugepages_attr.attr,
4255 &free_hugepages_attr.attr,
4256 &resv_hugepages_attr.attr,
4257 &surplus_hugepages_attr.attr,
4259 &nr_hugepages_mempolicy_attr.attr,
4264 static const struct attribute_group hstate_attr_group = {
4265 .attrs = hstate_attrs,
4268 static struct attribute *hstate_demote_attrs[] = {
4269 &demote_size_attr.attr,
4274 static const struct attribute_group hstate_demote_attr_group = {
4275 .attrs = hstate_demote_attrs,
4278 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4279 struct kobject **hstate_kobjs,
4280 const struct attribute_group *hstate_attr_group)
4283 int hi = hstate_index(h);
4285 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4286 if (!hstate_kobjs[hi])
4289 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4291 kobject_put(hstate_kobjs[hi]);
4292 hstate_kobjs[hi] = NULL;
4296 if (h->demote_order) {
4297 retval = sysfs_create_group(hstate_kobjs[hi],
4298 &hstate_demote_attr_group);
4300 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4301 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4302 kobject_put(hstate_kobjs[hi]);
4303 hstate_kobjs[hi] = NULL;
4312 static bool hugetlb_sysfs_initialized __ro_after_init;
4315 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4316 * with node devices in node_devices[] using a parallel array. The array
4317 * index of a node device or _hstate == node id.
4318 * This is here to avoid any static dependency of the node device driver, in
4319 * the base kernel, on the hugetlb module.
4321 struct node_hstate {
4322 struct kobject *hugepages_kobj;
4323 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4325 static struct node_hstate node_hstates[MAX_NUMNODES];
4328 * A subset of global hstate attributes for node devices
4330 static struct attribute *per_node_hstate_attrs[] = {
4331 &nr_hugepages_attr.attr,
4332 &free_hugepages_attr.attr,
4333 &surplus_hugepages_attr.attr,
4337 static const struct attribute_group per_node_hstate_attr_group = {
4338 .attrs = per_node_hstate_attrs,
4342 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4343 * Returns node id via non-NULL nidp.
4345 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4349 for (nid = 0; nid < nr_node_ids; nid++) {
4350 struct node_hstate *nhs = &node_hstates[nid];
4352 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4353 if (nhs->hstate_kobjs[i] == kobj) {
4365 * Unregister hstate attributes from a single node device.
4366 * No-op if no hstate attributes attached.
4368 void hugetlb_unregister_node(struct node *node)
4371 struct node_hstate *nhs = &node_hstates[node->dev.id];
4373 if (!nhs->hugepages_kobj)
4374 return; /* no hstate attributes */
4376 for_each_hstate(h) {
4377 int idx = hstate_index(h);
4378 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4382 if (h->demote_order)
4383 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4384 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4385 kobject_put(hstate_kobj);
4386 nhs->hstate_kobjs[idx] = NULL;
4389 kobject_put(nhs->hugepages_kobj);
4390 nhs->hugepages_kobj = NULL;
4395 * Register hstate attributes for a single node device.
4396 * No-op if attributes already registered.
4398 void hugetlb_register_node(struct node *node)
4401 struct node_hstate *nhs = &node_hstates[node->dev.id];
4404 if (!hugetlb_sysfs_initialized)
4407 if (nhs->hugepages_kobj)
4408 return; /* already allocated */
4410 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4412 if (!nhs->hugepages_kobj)
4415 for_each_hstate(h) {
4416 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4418 &per_node_hstate_attr_group);
4420 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4421 h->name, node->dev.id);
4422 hugetlb_unregister_node(node);
4429 * hugetlb init time: register hstate attributes for all registered node
4430 * devices of nodes that have memory. All on-line nodes should have
4431 * registered their associated device by this time.
4433 static void __init hugetlb_register_all_nodes(void)
4437 for_each_online_node(nid)
4438 hugetlb_register_node(node_devices[nid]);
4440 #else /* !CONFIG_NUMA */
4442 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4450 static void hugetlb_register_all_nodes(void) { }
4455 static void __init hugetlb_cma_check(void);
4457 static inline __init void hugetlb_cma_check(void)
4462 static void __init hugetlb_sysfs_init(void)
4467 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4468 if (!hugepages_kobj)
4471 for_each_hstate(h) {
4472 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4473 hstate_kobjs, &hstate_attr_group);
4475 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4479 hugetlb_sysfs_initialized = true;
4481 hugetlb_register_all_nodes();
4484 #ifdef CONFIG_SYSCTL
4485 static void hugetlb_sysctl_init(void);
4487 static inline void hugetlb_sysctl_init(void) { }
4490 static int __init hugetlb_init(void)
4494 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4497 if (!hugepages_supported()) {
4498 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4499 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4504 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4505 * architectures depend on setup being done here.
4507 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4508 if (!parsed_default_hugepagesz) {
4510 * If we did not parse a default huge page size, set
4511 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4512 * number of huge pages for this default size was implicitly
4513 * specified, set that here as well.
4514 * Note that the implicit setting will overwrite an explicit
4515 * setting. A warning will be printed in this case.
4517 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4518 if (default_hstate_max_huge_pages) {
4519 if (default_hstate.max_huge_pages) {
4522 string_get_size(huge_page_size(&default_hstate),
4523 1, STRING_UNITS_2, buf, 32);
4524 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4525 default_hstate.max_huge_pages, buf);
4526 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4527 default_hstate_max_huge_pages);
4529 default_hstate.max_huge_pages =
4530 default_hstate_max_huge_pages;
4532 for_each_online_node(i)
4533 default_hstate.max_huge_pages_node[i] =
4534 default_hugepages_in_node[i];
4538 hugetlb_cma_check();
4539 hugetlb_init_hstates();
4540 gather_bootmem_prealloc();
4543 hugetlb_sysfs_init();
4544 hugetlb_cgroup_file_init();
4545 hugetlb_sysctl_init();
4548 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4550 num_fault_mutexes = 1;
4552 hugetlb_fault_mutex_table =
4553 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4555 BUG_ON(!hugetlb_fault_mutex_table);
4557 for (i = 0; i < num_fault_mutexes; i++)
4558 mutex_init(&hugetlb_fault_mutex_table[i]);
4561 subsys_initcall(hugetlb_init);
4563 /* Overwritten by architectures with more huge page sizes */
4564 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4566 return size == HPAGE_SIZE;
4569 void __init hugetlb_add_hstate(unsigned int order)
4574 if (size_to_hstate(PAGE_SIZE << order)) {
4577 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4578 BUG_ON(order < order_base_2(__NR_USED_SUBPAGE));
4579 h = &hstates[hugetlb_max_hstate++];
4580 mutex_init(&h->resize_lock);
4582 h->mask = ~(huge_page_size(h) - 1);
4583 for (i = 0; i < MAX_NUMNODES; ++i)
4584 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4585 INIT_LIST_HEAD(&h->hugepage_activelist);
4586 h->next_nid_to_alloc = first_memory_node;
4587 h->next_nid_to_free = first_memory_node;
4588 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4589 huge_page_size(h)/SZ_1K);
4594 bool __init __weak hugetlb_node_alloc_supported(void)
4599 static void __init hugepages_clear_pages_in_node(void)
4601 if (!hugetlb_max_hstate) {
4602 default_hstate_max_huge_pages = 0;
4603 memset(default_hugepages_in_node, 0,
4604 sizeof(default_hugepages_in_node));
4606 parsed_hstate->max_huge_pages = 0;
4607 memset(parsed_hstate->max_huge_pages_node, 0,
4608 sizeof(parsed_hstate->max_huge_pages_node));
4613 * hugepages command line processing
4614 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4615 * specification. If not, ignore the hugepages value. hugepages can also
4616 * be the first huge page command line option in which case it implicitly
4617 * specifies the number of huge pages for the default size.
4619 static int __init hugepages_setup(char *s)
4622 static unsigned long *last_mhp;
4623 int node = NUMA_NO_NODE;
4628 if (!parsed_valid_hugepagesz) {
4629 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4630 parsed_valid_hugepagesz = true;
4635 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4636 * yet, so this hugepages= parameter goes to the "default hstate".
4637 * Otherwise, it goes with the previously parsed hugepagesz or
4638 * default_hugepagesz.
4640 else if (!hugetlb_max_hstate)
4641 mhp = &default_hstate_max_huge_pages;
4643 mhp = &parsed_hstate->max_huge_pages;
4645 if (mhp == last_mhp) {
4646 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4652 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4654 /* Parameter is node format */
4655 if (p[count] == ':') {
4656 if (!hugetlb_node_alloc_supported()) {
4657 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4660 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4662 node = array_index_nospec(tmp, MAX_NUMNODES);
4664 /* Parse hugepages */
4665 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4667 if (!hugetlb_max_hstate)
4668 default_hugepages_in_node[node] = tmp;
4670 parsed_hstate->max_huge_pages_node[node] = tmp;
4672 /* Go to parse next node*/
4673 if (p[count] == ',')
4686 * Global state is always initialized later in hugetlb_init.
4687 * But we need to allocate gigantic hstates here early to still
4688 * use the bootmem allocator.
4690 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4691 hugetlb_hstate_alloc_pages(parsed_hstate);
4698 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4699 hugepages_clear_pages_in_node();
4702 __setup("hugepages=", hugepages_setup);
4705 * hugepagesz command line processing
4706 * A specific huge page size can only be specified once with hugepagesz.
4707 * hugepagesz is followed by hugepages on the command line. The global
4708 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4709 * hugepagesz argument was valid.
4711 static int __init hugepagesz_setup(char *s)
4716 parsed_valid_hugepagesz = false;
4717 size = (unsigned long)memparse(s, NULL);
4719 if (!arch_hugetlb_valid_size(size)) {
4720 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4724 h = size_to_hstate(size);
4727 * hstate for this size already exists. This is normally
4728 * an error, but is allowed if the existing hstate is the
4729 * default hstate. More specifically, it is only allowed if
4730 * the number of huge pages for the default hstate was not
4731 * previously specified.
4733 if (!parsed_default_hugepagesz || h != &default_hstate ||
4734 default_hstate.max_huge_pages) {
4735 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4740 * No need to call hugetlb_add_hstate() as hstate already
4741 * exists. But, do set parsed_hstate so that a following
4742 * hugepages= parameter will be applied to this hstate.
4745 parsed_valid_hugepagesz = true;
4749 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4750 parsed_valid_hugepagesz = true;
4753 __setup("hugepagesz=", hugepagesz_setup);
4756 * default_hugepagesz command line input
4757 * Only one instance of default_hugepagesz allowed on command line.
4759 static int __init default_hugepagesz_setup(char *s)
4764 parsed_valid_hugepagesz = false;
4765 if (parsed_default_hugepagesz) {
4766 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4770 size = (unsigned long)memparse(s, NULL);
4772 if (!arch_hugetlb_valid_size(size)) {
4773 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4777 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4778 parsed_valid_hugepagesz = true;
4779 parsed_default_hugepagesz = true;
4780 default_hstate_idx = hstate_index(size_to_hstate(size));
4783 * The number of default huge pages (for this size) could have been
4784 * specified as the first hugetlb parameter: hugepages=X. If so,
4785 * then default_hstate_max_huge_pages is set. If the default huge
4786 * page size is gigantic (> MAX_ORDER), then the pages must be
4787 * allocated here from bootmem allocator.
4789 if (default_hstate_max_huge_pages) {
4790 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4791 for_each_online_node(i)
4792 default_hstate.max_huge_pages_node[i] =
4793 default_hugepages_in_node[i];
4794 if (hstate_is_gigantic(&default_hstate))
4795 hugetlb_hstate_alloc_pages(&default_hstate);
4796 default_hstate_max_huge_pages = 0;
4801 __setup("default_hugepagesz=", default_hugepagesz_setup);
4803 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4806 struct mempolicy *mpol = get_task_policy(current);
4809 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4810 * (from policy_nodemask) specifically for hugetlb case
4812 if (mpol->mode == MPOL_BIND &&
4813 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4814 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4815 return &mpol->nodes;
4820 static unsigned int allowed_mems_nr(struct hstate *h)
4823 unsigned int nr = 0;
4824 nodemask_t *mbind_nodemask;
4825 unsigned int *array = h->free_huge_pages_node;
4826 gfp_t gfp_mask = htlb_alloc_mask(h);
4828 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4829 for_each_node_mask(node, cpuset_current_mems_allowed) {
4830 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4837 #ifdef CONFIG_SYSCTL
4838 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4839 void *buffer, size_t *length,
4840 loff_t *ppos, unsigned long *out)
4842 struct ctl_table dup_table;
4845 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4846 * can duplicate the @table and alter the duplicate of it.
4849 dup_table.data = out;
4851 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4854 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4855 struct ctl_table *table, int write,
4856 void *buffer, size_t *length, loff_t *ppos)
4858 struct hstate *h = &default_hstate;
4859 unsigned long tmp = h->max_huge_pages;
4862 if (!hugepages_supported())
4865 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4871 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4872 NUMA_NO_NODE, tmp, *length);
4877 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4878 void *buffer, size_t *length, loff_t *ppos)
4881 return hugetlb_sysctl_handler_common(false, table, write,
4882 buffer, length, ppos);
4886 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4887 void *buffer, size_t *length, loff_t *ppos)
4889 return hugetlb_sysctl_handler_common(true, table, write,
4890 buffer, length, ppos);
4892 #endif /* CONFIG_NUMA */
4894 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4895 void *buffer, size_t *length, loff_t *ppos)
4897 struct hstate *h = &default_hstate;
4901 if (!hugepages_supported())
4904 tmp = h->nr_overcommit_huge_pages;
4906 if (write && hstate_is_gigantic(h))
4909 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4915 spin_lock_irq(&hugetlb_lock);
4916 h->nr_overcommit_huge_pages = tmp;
4917 spin_unlock_irq(&hugetlb_lock);
4923 static struct ctl_table hugetlb_table[] = {
4925 .procname = "nr_hugepages",
4927 .maxlen = sizeof(unsigned long),
4929 .proc_handler = hugetlb_sysctl_handler,
4933 .procname = "nr_hugepages_mempolicy",
4935 .maxlen = sizeof(unsigned long),
4937 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4941 .procname = "hugetlb_shm_group",
4942 .data = &sysctl_hugetlb_shm_group,
4943 .maxlen = sizeof(gid_t),
4945 .proc_handler = proc_dointvec,
4948 .procname = "nr_overcommit_hugepages",
4950 .maxlen = sizeof(unsigned long),
4952 .proc_handler = hugetlb_overcommit_handler,
4957 static void hugetlb_sysctl_init(void)
4959 register_sysctl_init("vm", hugetlb_table);
4961 #endif /* CONFIG_SYSCTL */
4963 void hugetlb_report_meminfo(struct seq_file *m)
4966 unsigned long total = 0;
4968 if (!hugepages_supported())
4971 for_each_hstate(h) {
4972 unsigned long count = h->nr_huge_pages;
4974 total += huge_page_size(h) * count;
4976 if (h == &default_hstate)
4978 "HugePages_Total: %5lu\n"
4979 "HugePages_Free: %5lu\n"
4980 "HugePages_Rsvd: %5lu\n"
4981 "HugePages_Surp: %5lu\n"
4982 "Hugepagesize: %8lu kB\n",
4986 h->surplus_huge_pages,
4987 huge_page_size(h) / SZ_1K);
4990 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4993 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4995 struct hstate *h = &default_hstate;
4997 if (!hugepages_supported())
5000 return sysfs_emit_at(buf, len,
5001 "Node %d HugePages_Total: %5u\n"
5002 "Node %d HugePages_Free: %5u\n"
5003 "Node %d HugePages_Surp: %5u\n",
5004 nid, h->nr_huge_pages_node[nid],
5005 nid, h->free_huge_pages_node[nid],
5006 nid, h->surplus_huge_pages_node[nid]);
5009 void hugetlb_show_meminfo_node(int nid)
5013 if (!hugepages_supported())
5017 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
5019 h->nr_huge_pages_node[nid],
5020 h->free_huge_pages_node[nid],
5021 h->surplus_huge_pages_node[nid],
5022 huge_page_size(h) / SZ_1K);
5025 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
5027 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
5028 K(atomic_long_read(&mm->hugetlb_usage)));
5031 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
5032 unsigned long hugetlb_total_pages(void)
5035 unsigned long nr_total_pages = 0;
5038 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
5039 return nr_total_pages;
5042 static int hugetlb_acct_memory(struct hstate *h, long delta)
5049 spin_lock_irq(&hugetlb_lock);
5051 * When cpuset is configured, it breaks the strict hugetlb page
5052 * reservation as the accounting is done on a global variable. Such
5053 * reservation is completely rubbish in the presence of cpuset because
5054 * the reservation is not checked against page availability for the
5055 * current cpuset. Application can still potentially OOM'ed by kernel
5056 * with lack of free htlb page in cpuset that the task is in.
5057 * Attempt to enforce strict accounting with cpuset is almost
5058 * impossible (or too ugly) because cpuset is too fluid that
5059 * task or memory node can be dynamically moved between cpusets.
5061 * The change of semantics for shared hugetlb mapping with cpuset is
5062 * undesirable. However, in order to preserve some of the semantics,
5063 * we fall back to check against current free page availability as
5064 * a best attempt and hopefully to minimize the impact of changing
5065 * semantics that cpuset has.
5067 * Apart from cpuset, we also have memory policy mechanism that
5068 * also determines from which node the kernel will allocate memory
5069 * in a NUMA system. So similar to cpuset, we also should consider
5070 * the memory policy of the current task. Similar to the description
5074 if (gather_surplus_pages(h, delta) < 0)
5077 if (delta > allowed_mems_nr(h)) {
5078 return_unused_surplus_pages(h, delta);
5085 return_unused_surplus_pages(h, (unsigned long) -delta);
5088 spin_unlock_irq(&hugetlb_lock);
5092 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
5094 struct resv_map *resv = vma_resv_map(vma);
5097 * HPAGE_RESV_OWNER indicates a private mapping.
5098 * This new VMA should share its siblings reservation map if present.
5099 * The VMA will only ever have a valid reservation map pointer where
5100 * it is being copied for another still existing VMA. As that VMA
5101 * has a reference to the reservation map it cannot disappear until
5102 * after this open call completes. It is therefore safe to take a
5103 * new reference here without additional locking.
5105 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
5106 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
5107 kref_get(&resv->refs);
5111 * vma_lock structure for sharable mappings is vma specific.
5112 * Clear old pointer (if copied via vm_area_dup) and allocate
5113 * new structure. Before clearing, make sure vma_lock is not
5116 if (vma->vm_flags & VM_MAYSHARE) {
5117 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
5120 if (vma_lock->vma != vma) {
5121 vma->vm_private_data = NULL;
5122 hugetlb_vma_lock_alloc(vma);
5124 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
5126 hugetlb_vma_lock_alloc(vma);
5130 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
5132 struct hstate *h = hstate_vma(vma);
5133 struct resv_map *resv;
5134 struct hugepage_subpool *spool = subpool_vma(vma);
5135 unsigned long reserve, start, end;
5138 hugetlb_vma_lock_free(vma);
5140 resv = vma_resv_map(vma);
5141 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5144 start = vma_hugecache_offset(h, vma, vma->vm_start);
5145 end = vma_hugecache_offset(h, vma, vma->vm_end);
5147 reserve = (end - start) - region_count(resv, start, end);
5148 hugetlb_cgroup_uncharge_counter(resv, start, end);
5151 * Decrement reserve counts. The global reserve count may be
5152 * adjusted if the subpool has a minimum size.
5154 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
5155 hugetlb_acct_memory(h, -gbl_reserve);
5158 kref_put(&resv->refs, resv_map_release);
5161 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
5163 if (addr & ~(huge_page_mask(hstate_vma(vma))))
5167 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
5168 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
5169 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
5171 if (addr & ~PUD_MASK) {
5173 * hugetlb_vm_op_split is called right before we attempt to
5174 * split the VMA. We will need to unshare PMDs in the old and
5175 * new VMAs, so let's unshare before we split.
5177 unsigned long floor = addr & PUD_MASK;
5178 unsigned long ceil = floor + PUD_SIZE;
5180 if (floor >= vma->vm_start && ceil <= vma->vm_end)
5181 hugetlb_unshare_pmds(vma, floor, ceil);
5187 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
5189 return huge_page_size(hstate_vma(vma));
5193 * We cannot handle pagefaults against hugetlb pages at all. They cause
5194 * handle_mm_fault() to try to instantiate regular-sized pages in the
5195 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
5198 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
5205 * When a new function is introduced to vm_operations_struct and added
5206 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
5207 * This is because under System V memory model, mappings created via
5208 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
5209 * their original vm_ops are overwritten with shm_vm_ops.
5211 const struct vm_operations_struct hugetlb_vm_ops = {
5212 .fault = hugetlb_vm_op_fault,
5213 .open = hugetlb_vm_op_open,
5214 .close = hugetlb_vm_op_close,
5215 .may_split = hugetlb_vm_op_split,
5216 .pagesize = hugetlb_vm_op_pagesize,
5219 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
5223 unsigned int shift = huge_page_shift(hstate_vma(vma));
5226 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
5227 vma->vm_page_prot)));
5229 entry = huge_pte_wrprotect(mk_huge_pte(page,
5230 vma->vm_page_prot));
5232 entry = pte_mkyoung(entry);
5233 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
5238 static void set_huge_ptep_writable(struct vm_area_struct *vma,
5239 unsigned long address, pte_t *ptep)
5243 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
5244 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
5245 update_mmu_cache(vma, address, ptep);
5248 bool is_hugetlb_entry_migration(pte_t pte)
5252 if (huge_pte_none(pte) || pte_present(pte))
5254 swp = pte_to_swp_entry(pte);
5255 if (is_migration_entry(swp))
5261 bool is_hugetlb_entry_hwpoisoned(pte_t pte)
5265 if (huge_pte_none(pte) || pte_present(pte))
5267 swp = pte_to_swp_entry(pte);
5268 if (is_hwpoison_entry(swp))
5275 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5276 struct folio *new_folio, pte_t old, unsigned long sz)
5278 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5280 __folio_mark_uptodate(new_folio);
5281 hugepage_add_new_anon_rmap(new_folio, vma, addr);
5282 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5283 newpte = huge_pte_mkuffd_wp(newpte);
5284 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5285 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5286 folio_set_hugetlb_migratable(new_folio);
5289 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5290 struct vm_area_struct *dst_vma,
5291 struct vm_area_struct *src_vma)
5293 pte_t *src_pte, *dst_pte, entry;
5294 struct folio *pte_folio;
5296 bool cow = is_cow_mapping(src_vma->vm_flags);
5297 struct hstate *h = hstate_vma(src_vma);
5298 unsigned long sz = huge_page_size(h);
5299 unsigned long npages = pages_per_huge_page(h);
5300 struct mmu_notifier_range range;
5301 unsigned long last_addr_mask;
5305 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5308 mmu_notifier_invalidate_range_start(&range);
5309 vma_assert_write_locked(src_vma);
5310 raw_write_seqcount_begin(&src->write_protect_seq);
5313 * For shared mappings the vma lock must be held before
5314 * calling hugetlb_walk() in the src vma. Otherwise, the
5315 * returned ptep could go away if part of a shared pmd and
5316 * another thread calls huge_pmd_unshare.
5318 hugetlb_vma_lock_read(src_vma);
5321 last_addr_mask = hugetlb_mask_last_page(h);
5322 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5323 spinlock_t *src_ptl, *dst_ptl;
5324 src_pte = hugetlb_walk(src_vma, addr, sz);
5326 addr |= last_addr_mask;
5329 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5336 * If the pagetables are shared don't copy or take references.
5338 * dst_pte == src_pte is the common case of src/dest sharing.
5339 * However, src could have 'unshared' and dst shares with
5340 * another vma. So page_count of ptep page is checked instead
5341 * to reliably determine whether pte is shared.
5343 if (page_count(virt_to_page(dst_pte)) > 1) {
5344 addr |= last_addr_mask;
5348 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5349 src_ptl = huge_pte_lockptr(h, src, src_pte);
5350 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5351 entry = huge_ptep_get(src_pte);
5353 if (huge_pte_none(entry)) {
5355 * Skip if src entry none.
5358 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5359 if (!userfaultfd_wp(dst_vma))
5360 entry = huge_pte_clear_uffd_wp(entry);
5361 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5362 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5363 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5364 bool uffd_wp = pte_swp_uffd_wp(entry);
5366 if (!is_readable_migration_entry(swp_entry) && cow) {
5368 * COW mappings require pages in both
5369 * parent and child to be set to read.
5371 swp_entry = make_readable_migration_entry(
5372 swp_offset(swp_entry));
5373 entry = swp_entry_to_pte(swp_entry);
5374 if (userfaultfd_wp(src_vma) && uffd_wp)
5375 entry = pte_swp_mkuffd_wp(entry);
5376 set_huge_pte_at(src, addr, src_pte, entry, sz);
5378 if (!userfaultfd_wp(dst_vma))
5379 entry = huge_pte_clear_uffd_wp(entry);
5380 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5381 } else if (unlikely(is_pte_marker(entry))) {
5382 pte_marker marker = copy_pte_marker(
5383 pte_to_swp_entry(entry), dst_vma);
5386 set_huge_pte_at(dst, addr, dst_pte,
5387 make_pte_marker(marker), sz);
5389 entry = huge_ptep_get(src_pte);
5390 pte_folio = page_folio(pte_page(entry));
5391 folio_get(pte_folio);
5394 * Failing to duplicate the anon rmap is a rare case
5395 * where we see pinned hugetlb pages while they're
5396 * prone to COW. We need to do the COW earlier during
5399 * When pre-allocating the page or copying data, we
5400 * need to be without the pgtable locks since we could
5401 * sleep during the process.
5403 if (!folio_test_anon(pte_folio)) {
5404 page_dup_file_rmap(&pte_folio->page, true);
5405 } else if (page_try_dup_anon_rmap(&pte_folio->page,
5407 pte_t src_pte_old = entry;
5408 struct folio *new_folio;
5410 spin_unlock(src_ptl);
5411 spin_unlock(dst_ptl);
5412 /* Do not use reserve as it's private owned */
5413 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5414 if (IS_ERR(new_folio)) {
5415 folio_put(pte_folio);
5416 ret = PTR_ERR(new_folio);
5419 ret = copy_user_large_folio(new_folio,
5422 folio_put(pte_folio);
5424 folio_put(new_folio);
5428 /* Install the new hugetlb folio if src pte stable */
5429 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5430 src_ptl = huge_pte_lockptr(h, src, src_pte);
5431 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5432 entry = huge_ptep_get(src_pte);
5433 if (!pte_same(src_pte_old, entry)) {
5434 restore_reserve_on_error(h, dst_vma, addr,
5436 folio_put(new_folio);
5437 /* huge_ptep of dst_pte won't change as in child */
5440 hugetlb_install_folio(dst_vma, dst_pte, addr,
5441 new_folio, src_pte_old, sz);
5442 spin_unlock(src_ptl);
5443 spin_unlock(dst_ptl);
5449 * No need to notify as we are downgrading page
5450 * table protection not changing it to point
5453 * See Documentation/mm/mmu_notifier.rst
5455 huge_ptep_set_wrprotect(src, addr, src_pte);
5456 entry = huge_pte_wrprotect(entry);
5459 if (!userfaultfd_wp(dst_vma))
5460 entry = huge_pte_clear_uffd_wp(entry);
5462 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5463 hugetlb_count_add(npages, dst);
5465 spin_unlock(src_ptl);
5466 spin_unlock(dst_ptl);
5470 raw_write_seqcount_end(&src->write_protect_seq);
5471 mmu_notifier_invalidate_range_end(&range);
5473 hugetlb_vma_unlock_read(src_vma);
5479 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5480 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5483 struct hstate *h = hstate_vma(vma);
5484 struct mm_struct *mm = vma->vm_mm;
5485 spinlock_t *src_ptl, *dst_ptl;
5488 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5489 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5492 * We don't have to worry about the ordering of src and dst ptlocks
5493 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5495 if (src_ptl != dst_ptl)
5496 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5498 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5499 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5501 if (src_ptl != dst_ptl)
5502 spin_unlock(src_ptl);
5503 spin_unlock(dst_ptl);
5506 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5507 struct vm_area_struct *new_vma,
5508 unsigned long old_addr, unsigned long new_addr,
5511 struct hstate *h = hstate_vma(vma);
5512 struct address_space *mapping = vma->vm_file->f_mapping;
5513 unsigned long sz = huge_page_size(h);
5514 struct mm_struct *mm = vma->vm_mm;
5515 unsigned long old_end = old_addr + len;
5516 unsigned long last_addr_mask;
5517 pte_t *src_pte, *dst_pte;
5518 struct mmu_notifier_range range;
5519 bool shared_pmd = false;
5521 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5523 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5525 * In case of shared PMDs, we should cover the maximum possible
5528 flush_cache_range(vma, range.start, range.end);
5530 mmu_notifier_invalidate_range_start(&range);
5531 last_addr_mask = hugetlb_mask_last_page(h);
5532 /* Prevent race with file truncation */
5533 hugetlb_vma_lock_write(vma);
5534 i_mmap_lock_write(mapping);
5535 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5536 src_pte = hugetlb_walk(vma, old_addr, sz);
5538 old_addr |= last_addr_mask;
5539 new_addr |= last_addr_mask;
5542 if (huge_pte_none(huge_ptep_get(src_pte)))
5545 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5547 old_addr |= last_addr_mask;
5548 new_addr |= last_addr_mask;
5552 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5556 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5560 flush_hugetlb_tlb_range(vma, range.start, range.end);
5562 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5563 mmu_notifier_invalidate_range_end(&range);
5564 i_mmap_unlock_write(mapping);
5565 hugetlb_vma_unlock_write(vma);
5567 return len + old_addr - old_end;
5570 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5571 unsigned long start, unsigned long end,
5572 struct page *ref_page, zap_flags_t zap_flags)
5574 struct mm_struct *mm = vma->vm_mm;
5575 unsigned long address;
5580 struct hstate *h = hstate_vma(vma);
5581 unsigned long sz = huge_page_size(h);
5582 unsigned long last_addr_mask;
5583 bool force_flush = false;
5585 WARN_ON(!is_vm_hugetlb_page(vma));
5586 BUG_ON(start & ~huge_page_mask(h));
5587 BUG_ON(end & ~huge_page_mask(h));
5590 * This is a hugetlb vma, all the pte entries should point
5593 tlb_change_page_size(tlb, sz);
5594 tlb_start_vma(tlb, vma);
5596 last_addr_mask = hugetlb_mask_last_page(h);
5598 for (; address < end; address += sz) {
5599 ptep = hugetlb_walk(vma, address, sz);
5601 address |= last_addr_mask;
5605 ptl = huge_pte_lock(h, mm, ptep);
5606 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5608 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5610 address |= last_addr_mask;
5614 pte = huge_ptep_get(ptep);
5615 if (huge_pte_none(pte)) {
5621 * Migrating hugepage or HWPoisoned hugepage is already
5622 * unmapped and its refcount is dropped, so just clear pte here.
5624 if (unlikely(!pte_present(pte))) {
5626 * If the pte was wr-protected by uffd-wp in any of the
5627 * swap forms, meanwhile the caller does not want to
5628 * drop the uffd-wp bit in this zap, then replace the
5629 * pte with a marker.
5631 if (pte_swp_uffd_wp_any(pte) &&
5632 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5633 set_huge_pte_at(mm, address, ptep,
5634 make_pte_marker(PTE_MARKER_UFFD_WP),
5637 huge_pte_clear(mm, address, ptep, sz);
5642 page = pte_page(pte);
5644 * If a reference page is supplied, it is because a specific
5645 * page is being unmapped, not a range. Ensure the page we
5646 * are about to unmap is the actual page of interest.
5649 if (page != ref_page) {
5654 * Mark the VMA as having unmapped its page so that
5655 * future faults in this VMA will fail rather than
5656 * looking like data was lost
5658 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5661 pte = huge_ptep_get_and_clear(mm, address, ptep);
5662 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5663 if (huge_pte_dirty(pte))
5664 set_page_dirty(page);
5665 /* Leave a uffd-wp pte marker if needed */
5666 if (huge_pte_uffd_wp(pte) &&
5667 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5668 set_huge_pte_at(mm, address, ptep,
5669 make_pte_marker(PTE_MARKER_UFFD_WP),
5671 hugetlb_count_sub(pages_per_huge_page(h), mm);
5672 page_remove_rmap(page, vma, true);
5675 tlb_remove_page_size(tlb, page, huge_page_size(h));
5677 * Bail out after unmapping reference page if supplied
5682 tlb_end_vma(tlb, vma);
5685 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5686 * could defer the flush until now, since by holding i_mmap_rwsem we
5687 * guaranteed that the last refernece would not be dropped. But we must
5688 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5689 * dropped and the last reference to the shared PMDs page might be
5692 * In theory we could defer the freeing of the PMD pages as well, but
5693 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5694 * detect sharing, so we cannot defer the release of the page either.
5695 * Instead, do flush now.
5698 tlb_flush_mmu_tlbonly(tlb);
5701 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5702 unsigned long *start, unsigned long *end)
5704 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5707 adjust_range_if_pmd_sharing_possible(vma, start, end);
5708 hugetlb_vma_lock_write(vma);
5710 i_mmap_lock_write(vma->vm_file->f_mapping);
5713 void __hugetlb_zap_end(struct vm_area_struct *vma,
5714 struct zap_details *details)
5716 zap_flags_t zap_flags = details ? details->zap_flags : 0;
5718 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5721 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5723 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5724 * When the vma_lock is freed, this makes the vma ineligible
5725 * for pmd sharing. And, i_mmap_rwsem is required to set up
5726 * pmd sharing. This is important as page tables for this
5727 * unmapped range will be asynchrously deleted. If the page
5728 * tables are shared, there will be issues when accessed by
5731 __hugetlb_vma_unlock_write_free(vma);
5733 hugetlb_vma_unlock_write(vma);
5737 i_mmap_unlock_write(vma->vm_file->f_mapping);
5740 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5741 unsigned long end, struct page *ref_page,
5742 zap_flags_t zap_flags)
5744 struct mmu_notifier_range range;
5745 struct mmu_gather tlb;
5747 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5749 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5750 mmu_notifier_invalidate_range_start(&range);
5751 tlb_gather_mmu(&tlb, vma->vm_mm);
5753 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5755 mmu_notifier_invalidate_range_end(&range);
5756 tlb_finish_mmu(&tlb);
5760 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5761 * mapping it owns the reserve page for. The intention is to unmap the page
5762 * from other VMAs and let the children be SIGKILLed if they are faulting the
5765 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5766 struct page *page, unsigned long address)
5768 struct hstate *h = hstate_vma(vma);
5769 struct vm_area_struct *iter_vma;
5770 struct address_space *mapping;
5774 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5775 * from page cache lookup which is in HPAGE_SIZE units.
5777 address = address & huge_page_mask(h);
5778 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5780 mapping = vma->vm_file->f_mapping;
5783 * Take the mapping lock for the duration of the table walk. As
5784 * this mapping should be shared between all the VMAs,
5785 * __unmap_hugepage_range() is called as the lock is already held
5787 i_mmap_lock_write(mapping);
5788 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5789 /* Do not unmap the current VMA */
5790 if (iter_vma == vma)
5794 * Shared VMAs have their own reserves and do not affect
5795 * MAP_PRIVATE accounting but it is possible that a shared
5796 * VMA is using the same page so check and skip such VMAs.
5798 if (iter_vma->vm_flags & VM_MAYSHARE)
5802 * Unmap the page from other VMAs without their own reserves.
5803 * They get marked to be SIGKILLed if they fault in these
5804 * areas. This is because a future no-page fault on this VMA
5805 * could insert a zeroed page instead of the data existing
5806 * from the time of fork. This would look like data corruption
5808 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5809 unmap_hugepage_range(iter_vma, address,
5810 address + huge_page_size(h), page, 0);
5812 i_mmap_unlock_write(mapping);
5816 * hugetlb_wp() should be called with page lock of the original hugepage held.
5817 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5818 * cannot race with other handlers or page migration.
5819 * Keep the pte_same checks anyway to make transition from the mutex easier.
5821 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5822 unsigned long address, pte_t *ptep, unsigned int flags,
5823 struct folio *pagecache_folio, spinlock_t *ptl)
5825 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5826 pte_t pte = huge_ptep_get(ptep);
5827 struct hstate *h = hstate_vma(vma);
5828 struct folio *old_folio;
5829 struct folio *new_folio;
5830 int outside_reserve = 0;
5832 unsigned long haddr = address & huge_page_mask(h);
5833 struct mmu_notifier_range range;
5836 * Never handle CoW for uffd-wp protected pages. It should be only
5837 * handled when the uffd-wp protection is removed.
5839 * Note that only the CoW optimization path (in hugetlb_no_page())
5840 * can trigger this, because hugetlb_fault() will always resolve
5841 * uffd-wp bit first.
5843 if (!unshare && huge_pte_uffd_wp(pte))
5847 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5848 * PTE mapped R/O such as maybe_mkwrite() would do.
5850 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5851 return VM_FAULT_SIGSEGV;
5853 /* Let's take out MAP_SHARED mappings first. */
5854 if (vma->vm_flags & VM_MAYSHARE) {
5855 set_huge_ptep_writable(vma, haddr, ptep);
5859 old_folio = page_folio(pte_page(pte));
5861 delayacct_wpcopy_start();
5865 * If no-one else is actually using this page, we're the exclusive
5866 * owner and can reuse this page.
5868 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5869 if (!PageAnonExclusive(&old_folio->page)) {
5870 folio_move_anon_rmap(old_folio, vma);
5871 SetPageAnonExclusive(&old_folio->page);
5873 if (likely(!unshare))
5874 set_huge_ptep_writable(vma, haddr, ptep);
5876 delayacct_wpcopy_end();
5879 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5880 PageAnonExclusive(&old_folio->page), &old_folio->page);
5883 * If the process that created a MAP_PRIVATE mapping is about to
5884 * perform a COW due to a shared page count, attempt to satisfy
5885 * the allocation without using the existing reserves. The pagecache
5886 * page is used to determine if the reserve at this address was
5887 * consumed or not. If reserves were used, a partial faulted mapping
5888 * at the time of fork() could consume its reserves on COW instead
5889 * of the full address range.
5891 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5892 old_folio != pagecache_folio)
5893 outside_reserve = 1;
5895 folio_get(old_folio);
5898 * Drop page table lock as buddy allocator may be called. It will
5899 * be acquired again before returning to the caller, as expected.
5902 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5904 if (IS_ERR(new_folio)) {
5906 * If a process owning a MAP_PRIVATE mapping fails to COW,
5907 * it is due to references held by a child and an insufficient
5908 * huge page pool. To guarantee the original mappers
5909 * reliability, unmap the page from child processes. The child
5910 * may get SIGKILLed if it later faults.
5912 if (outside_reserve) {
5913 struct address_space *mapping = vma->vm_file->f_mapping;
5917 folio_put(old_folio);
5919 * Drop hugetlb_fault_mutex and vma_lock before
5920 * unmapping. unmapping needs to hold vma_lock
5921 * in write mode. Dropping vma_lock in read mode
5922 * here is OK as COW mappings do not interact with
5925 * Reacquire both after unmap operation.
5927 idx = vma_hugecache_offset(h, vma, haddr);
5928 hash = hugetlb_fault_mutex_hash(mapping, idx);
5929 hugetlb_vma_unlock_read(vma);
5930 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5932 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5934 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5935 hugetlb_vma_lock_read(vma);
5937 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5939 pte_same(huge_ptep_get(ptep), pte)))
5940 goto retry_avoidcopy;
5942 * race occurs while re-acquiring page table
5943 * lock, and our job is done.
5945 delayacct_wpcopy_end();
5949 ret = vmf_error(PTR_ERR(new_folio));
5950 goto out_release_old;
5954 * When the original hugepage is shared one, it does not have
5955 * anon_vma prepared.
5957 if (unlikely(anon_vma_prepare(vma))) {
5959 goto out_release_all;
5962 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5963 ret = VM_FAULT_HWPOISON_LARGE;
5964 goto out_release_all;
5966 __folio_mark_uptodate(new_folio);
5968 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5969 haddr + huge_page_size(h));
5970 mmu_notifier_invalidate_range_start(&range);
5973 * Retake the page table lock to check for racing updates
5974 * before the page tables are altered
5977 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5978 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5979 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5981 /* Break COW or unshare */
5982 huge_ptep_clear_flush(vma, haddr, ptep);
5983 page_remove_rmap(&old_folio->page, vma, true);
5984 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5985 if (huge_pte_uffd_wp(pte))
5986 newpte = huge_pte_mkuffd_wp(newpte);
5987 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5988 folio_set_hugetlb_migratable(new_folio);
5989 /* Make the old page be freed below */
5990 new_folio = old_folio;
5993 mmu_notifier_invalidate_range_end(&range);
5996 * No restore in case of successful pagetable update (Break COW or
5999 if (new_folio != old_folio)
6000 restore_reserve_on_error(h, vma, haddr, new_folio);
6001 folio_put(new_folio);
6003 folio_put(old_folio);
6005 spin_lock(ptl); /* Caller expects lock to be held */
6007 delayacct_wpcopy_end();
6012 * Return whether there is a pagecache page to back given address within VMA.
6014 static bool hugetlbfs_pagecache_present(struct hstate *h,
6015 struct vm_area_struct *vma, unsigned long address)
6017 struct address_space *mapping = vma->vm_file->f_mapping;
6018 pgoff_t idx = linear_page_index(vma, address);
6019 struct folio *folio;
6021 folio = filemap_get_folio(mapping, idx);
6028 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
6031 struct inode *inode = mapping->host;
6032 struct hstate *h = hstate_inode(inode);
6035 idx <<= huge_page_order(h);
6036 __folio_set_locked(folio);
6037 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
6039 if (unlikely(err)) {
6040 __folio_clear_locked(folio);
6043 folio_clear_hugetlb_restore_reserve(folio);
6046 * mark folio dirty so that it will not be removed from cache/file
6047 * by non-hugetlbfs specific code paths.
6049 folio_mark_dirty(folio);
6051 spin_lock(&inode->i_lock);
6052 inode->i_blocks += blocks_per_huge_page(h);
6053 spin_unlock(&inode->i_lock);
6057 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
6058 struct address_space *mapping,
6061 unsigned long haddr,
6063 unsigned long reason)
6066 struct vm_fault vmf = {
6069 .real_address = addr,
6073 * Hard to debug if it ends up being
6074 * used by a callee that assumes
6075 * something about the other
6076 * uninitialized fields... same as in
6082 * vma_lock and hugetlb_fault_mutex must be dropped before handling
6083 * userfault. Also mmap_lock could be dropped due to handling
6084 * userfault, any vma operation should be careful from here.
6086 hugetlb_vma_unlock_read(vma);
6087 hash = hugetlb_fault_mutex_hash(mapping, idx);
6088 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6089 return handle_userfault(&vmf, reason);
6093 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
6094 * false if pte changed or is changing.
6096 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
6097 pte_t *ptep, pte_t old_pte)
6102 ptl = huge_pte_lock(h, mm, ptep);
6103 same = pte_same(huge_ptep_get(ptep), old_pte);
6109 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
6110 struct vm_area_struct *vma,
6111 struct address_space *mapping, pgoff_t idx,
6112 unsigned long address, pte_t *ptep,
6113 pte_t old_pte, unsigned int flags)
6115 struct hstate *h = hstate_vma(vma);
6116 vm_fault_t ret = VM_FAULT_SIGBUS;
6119 struct folio *folio;
6122 unsigned long haddr = address & huge_page_mask(h);
6123 bool new_folio, new_pagecache_folio = false;
6124 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
6127 * Currently, we are forced to kill the process in the event the
6128 * original mapper has unmapped pages from the child due to a failed
6129 * COW/unsharing. Warn that such a situation has occurred as it may not
6132 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
6133 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
6139 * Use page lock to guard against racing truncation
6140 * before we get page_table_lock.
6143 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6144 if (IS_ERR(folio)) {
6145 size = i_size_read(mapping->host) >> huge_page_shift(h);
6148 /* Check for page in userfault range */
6149 if (userfaultfd_missing(vma)) {
6151 * Since hugetlb_no_page() was examining pte
6152 * without pgtable lock, we need to re-test under
6153 * lock because the pte may not be stable and could
6154 * have changed from under us. Try to detect
6155 * either changed or during-changing ptes and retry
6156 * properly when needed.
6158 * Note that userfaultfd is actually fine with
6159 * false positives (e.g. caused by pte changed),
6160 * but not wrong logical events (e.g. caused by
6161 * reading a pte during changing). The latter can
6162 * confuse the userspace, so the strictness is very
6163 * much preferred. E.g., MISSING event should
6164 * never happen on the page after UFFDIO_COPY has
6165 * correctly installed the page and returned.
6167 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6172 return hugetlb_handle_userfault(vma, mapping, idx, flags,
6177 folio = alloc_hugetlb_folio(vma, haddr, 0);
6178 if (IS_ERR(folio)) {
6180 * Returning error will result in faulting task being
6181 * sent SIGBUS. The hugetlb fault mutex prevents two
6182 * tasks from racing to fault in the same page which
6183 * could result in false unable to allocate errors.
6184 * Page migration does not take the fault mutex, but
6185 * does a clear then write of pte's under page table
6186 * lock. Page fault code could race with migration,
6187 * notice the clear pte and try to allocate a page
6188 * here. Before returning error, get ptl and make
6189 * sure there really is no pte entry.
6191 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
6192 ret = vmf_error(PTR_ERR(folio));
6197 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
6198 __folio_mark_uptodate(folio);
6201 if (vma->vm_flags & VM_MAYSHARE) {
6202 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
6205 * err can't be -EEXIST which implies someone
6206 * else consumed the reservation since hugetlb
6207 * fault mutex is held when add a hugetlb page
6208 * to the page cache. So it's safe to call
6209 * restore_reserve_on_error() here.
6211 restore_reserve_on_error(h, vma, haddr, folio);
6215 new_pagecache_folio = true;
6218 if (unlikely(anon_vma_prepare(vma))) {
6220 goto backout_unlocked;
6226 * If memory error occurs between mmap() and fault, some process
6227 * don't have hwpoisoned swap entry for errored virtual address.
6228 * So we need to block hugepage fault by PG_hwpoison bit check.
6230 if (unlikely(folio_test_hwpoison(folio))) {
6231 ret = VM_FAULT_HWPOISON_LARGE |
6232 VM_FAULT_SET_HINDEX(hstate_index(h));
6233 goto backout_unlocked;
6236 /* Check for page in userfault range. */
6237 if (userfaultfd_minor(vma)) {
6238 folio_unlock(folio);
6240 /* See comment in userfaultfd_missing() block above */
6241 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6245 return hugetlb_handle_userfault(vma, mapping, idx, flags,
6252 * If we are going to COW a private mapping later, we examine the
6253 * pending reservations for this page now. This will ensure that
6254 * any allocations necessary to record that reservation occur outside
6257 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6258 if (vma_needs_reservation(h, vma, haddr) < 0) {
6260 goto backout_unlocked;
6262 /* Just decrements count, does not deallocate */
6263 vma_end_reservation(h, vma, haddr);
6266 ptl = huge_pte_lock(h, mm, ptep);
6268 /* If pte changed from under us, retry */
6269 if (!pte_same(huge_ptep_get(ptep), old_pte))
6273 hugepage_add_new_anon_rmap(folio, vma, haddr);
6275 page_dup_file_rmap(&folio->page, true);
6276 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6277 && (vma->vm_flags & VM_SHARED)));
6279 * If this pte was previously wr-protected, keep it wr-protected even
6282 if (unlikely(pte_marker_uffd_wp(old_pte)))
6283 new_pte = huge_pte_mkuffd_wp(new_pte);
6284 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6286 hugetlb_count_add(pages_per_huge_page(h), mm);
6287 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6288 /* Optimization, do the COW without a second fault */
6289 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
6295 * Only set hugetlb_migratable in newly allocated pages. Existing pages
6296 * found in the pagecache may not have hugetlb_migratable if they have
6297 * been isolated for migration.
6300 folio_set_hugetlb_migratable(folio);
6302 folio_unlock(folio);
6304 hugetlb_vma_unlock_read(vma);
6305 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6311 if (new_folio && !new_pagecache_folio)
6312 restore_reserve_on_error(h, vma, haddr, folio);
6314 folio_unlock(folio);
6320 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6322 unsigned long key[2];
6325 key[0] = (unsigned long) mapping;
6328 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6330 return hash & (num_fault_mutexes - 1);
6334 * For uniprocessor systems we always use a single mutex, so just
6335 * return 0 and avoid the hashing overhead.
6337 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6343 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6344 unsigned long address, unsigned int flags)
6351 struct folio *folio = NULL;
6352 struct folio *pagecache_folio = NULL;
6353 struct hstate *h = hstate_vma(vma);
6354 struct address_space *mapping;
6355 int need_wait_lock = 0;
6356 unsigned long haddr = address & huge_page_mask(h);
6358 /* TODO: Handle faults under the VMA lock */
6359 if (flags & FAULT_FLAG_VMA_LOCK) {
6361 return VM_FAULT_RETRY;
6365 * Serialize hugepage allocation and instantiation, so that we don't
6366 * get spurious allocation failures if two CPUs race to instantiate
6367 * the same page in the page cache.
6369 mapping = vma->vm_file->f_mapping;
6370 idx = vma_hugecache_offset(h, vma, haddr);
6371 hash = hugetlb_fault_mutex_hash(mapping, idx);
6372 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6375 * Acquire vma lock before calling huge_pte_alloc and hold
6376 * until finished with ptep. This prevents huge_pmd_unshare from
6377 * being called elsewhere and making the ptep no longer valid.
6379 hugetlb_vma_lock_read(vma);
6380 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6382 hugetlb_vma_unlock_read(vma);
6383 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6384 return VM_FAULT_OOM;
6387 entry = huge_ptep_get(ptep);
6388 if (huge_pte_none_mostly(entry)) {
6389 if (is_pte_marker(entry)) {
6391 pte_marker_get(pte_to_swp_entry(entry));
6393 if (marker & PTE_MARKER_POISONED) {
6394 ret = VM_FAULT_HWPOISON_LARGE;
6400 * Other PTE markers should be handled the same way as none PTE.
6402 * hugetlb_no_page will drop vma lock and hugetlb fault
6403 * mutex internally, which make us return immediately.
6405 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6412 * entry could be a migration/hwpoison entry at this point, so this
6413 * check prevents the kernel from going below assuming that we have
6414 * an active hugepage in pagecache. This goto expects the 2nd page
6415 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6416 * properly handle it.
6418 if (!pte_present(entry)) {
6419 if (unlikely(is_hugetlb_entry_migration(entry))) {
6421 * Release the hugetlb fault lock now, but retain
6422 * the vma lock, because it is needed to guard the
6423 * huge_pte_lockptr() later in
6424 * migration_entry_wait_huge(). The vma lock will
6425 * be released there.
6427 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6428 migration_entry_wait_huge(vma, ptep);
6430 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6431 ret = VM_FAULT_HWPOISON_LARGE |
6432 VM_FAULT_SET_HINDEX(hstate_index(h));
6437 * If we are going to COW/unshare the mapping later, we examine the
6438 * pending reservations for this page now. This will ensure that any
6439 * allocations necessary to record that reservation occur outside the
6440 * spinlock. Also lookup the pagecache page now as it is used to
6441 * determine if a reservation has been consumed.
6443 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6444 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6445 if (vma_needs_reservation(h, vma, haddr) < 0) {
6449 /* Just decrements count, does not deallocate */
6450 vma_end_reservation(h, vma, haddr);
6452 pagecache_folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6453 if (IS_ERR(pagecache_folio))
6454 pagecache_folio = NULL;
6457 ptl = huge_pte_lock(h, mm, ptep);
6459 /* Check for a racing update before calling hugetlb_wp() */
6460 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6463 /* Handle userfault-wp first, before trying to lock more pages */
6464 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6465 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6466 if (!userfaultfd_wp_async(vma)) {
6467 struct vm_fault vmf = {
6470 .real_address = address,
6475 if (pagecache_folio) {
6476 folio_unlock(pagecache_folio);
6477 folio_put(pagecache_folio);
6479 hugetlb_vma_unlock_read(vma);
6480 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6481 return handle_userfault(&vmf, VM_UFFD_WP);
6484 entry = huge_pte_clear_uffd_wp(entry);
6485 set_huge_pte_at(mm, haddr, ptep, entry,
6486 huge_page_size(hstate_vma(vma)));
6487 /* Fallthrough to CoW */
6491 * hugetlb_wp() requires page locks of pte_page(entry) and
6492 * pagecache_folio, so here we need take the former one
6493 * when folio != pagecache_folio or !pagecache_folio.
6495 folio = page_folio(pte_page(entry));
6496 if (folio != pagecache_folio)
6497 if (!folio_trylock(folio)) {
6504 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6505 if (!huge_pte_write(entry)) {
6506 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6507 pagecache_folio, ptl);
6509 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6510 entry = huge_pte_mkdirty(entry);
6513 entry = pte_mkyoung(entry);
6514 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6515 flags & FAULT_FLAG_WRITE))
6516 update_mmu_cache(vma, haddr, ptep);
6518 if (folio != pagecache_folio)
6519 folio_unlock(folio);
6524 if (pagecache_folio) {
6525 folio_unlock(pagecache_folio);
6526 folio_put(pagecache_folio);
6529 hugetlb_vma_unlock_read(vma);
6530 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6532 * Generally it's safe to hold refcount during waiting page lock. But
6533 * here we just wait to defer the next page fault to avoid busy loop and
6534 * the page is not used after unlocked before returning from the current
6535 * page fault. So we are safe from accessing freed page, even if we wait
6536 * here without taking refcount.
6539 folio_wait_locked(folio);
6543 #ifdef CONFIG_USERFAULTFD
6545 * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte().
6547 static struct folio *alloc_hugetlb_folio_vma(struct hstate *h,
6548 struct vm_area_struct *vma, unsigned long address)
6550 struct mempolicy *mpol;
6551 nodemask_t *nodemask;
6552 struct folio *folio;
6556 gfp_mask = htlb_alloc_mask(h);
6557 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
6558 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
6559 mpol_cond_put(mpol);
6565 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6566 * with modifications for hugetlb pages.
6568 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6569 struct vm_area_struct *dst_vma,
6570 unsigned long dst_addr,
6571 unsigned long src_addr,
6573 struct folio **foliop)
6575 struct mm_struct *dst_mm = dst_vma->vm_mm;
6576 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6577 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6578 struct hstate *h = hstate_vma(dst_vma);
6579 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6580 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6582 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6586 struct folio *folio;
6588 bool folio_in_pagecache = false;
6590 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6591 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6593 /* Don't overwrite any existing PTEs (even markers) */
6594 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6599 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6600 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6603 /* No need to invalidate - it was non-present before */
6604 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6612 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6615 folio_in_pagecache = true;
6616 } else if (!*foliop) {
6617 /* If a folio already exists, then it's UFFDIO_COPY for
6618 * a non-missing case. Return -EEXIST.
6621 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6626 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6627 if (IS_ERR(folio)) {
6632 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6635 /* fallback to copy_from_user outside mmap_lock */
6636 if (unlikely(ret)) {
6638 /* Free the allocated folio which may have
6639 * consumed a reservation.
6641 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6644 /* Allocate a temporary folio to hold the copied
6647 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6653 /* Set the outparam foliop and return to the caller to
6654 * copy the contents outside the lock. Don't free the
6661 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6668 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6669 if (IS_ERR(folio)) {
6675 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6685 * The memory barrier inside __folio_mark_uptodate makes sure that
6686 * preceding stores to the page contents become visible before
6687 * the set_pte_at() write.
6689 __folio_mark_uptodate(folio);
6691 /* Add shared, newly allocated pages to the page cache. */
6692 if (vm_shared && !is_continue) {
6693 size = i_size_read(mapping->host) >> huge_page_shift(h);
6696 goto out_release_nounlock;
6699 * Serialization between remove_inode_hugepages() and
6700 * hugetlb_add_to_page_cache() below happens through the
6701 * hugetlb_fault_mutex_table that here must be hold by
6704 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6706 goto out_release_nounlock;
6707 folio_in_pagecache = true;
6710 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6713 if (folio_test_hwpoison(folio))
6714 goto out_release_unlock;
6717 * We allow to overwrite a pte marker: consider when both MISSING|WP
6718 * registered, we firstly wr-protect a none pte which has no page cache
6719 * page backing it, then access the page.
6722 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6723 goto out_release_unlock;
6725 if (folio_in_pagecache)
6726 page_dup_file_rmap(&folio->page, true);
6728 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6731 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6732 * with wp flag set, don't set pte write bit.
6734 if (wp_enabled || (is_continue && !vm_shared))
6737 writable = dst_vma->vm_flags & VM_WRITE;
6739 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6741 * Always mark UFFDIO_COPY page dirty; note that this may not be
6742 * extremely important for hugetlbfs for now since swapping is not
6743 * supported, but we should still be clear in that this page cannot be
6744 * thrown away at will, even if write bit not set.
6746 _dst_pte = huge_pte_mkdirty(_dst_pte);
6747 _dst_pte = pte_mkyoung(_dst_pte);
6750 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6752 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6754 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6756 /* No need to invalidate - it was non-present before */
6757 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6761 folio_set_hugetlb_migratable(folio);
6762 if (vm_shared || is_continue)
6763 folio_unlock(folio);
6769 if (vm_shared || is_continue)
6770 folio_unlock(folio);
6771 out_release_nounlock:
6772 if (!folio_in_pagecache)
6773 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6777 #endif /* CONFIG_USERFAULTFD */
6779 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6780 unsigned long address, unsigned int flags,
6781 unsigned int *page_mask)
6783 struct hstate *h = hstate_vma(vma);
6784 struct mm_struct *mm = vma->vm_mm;
6785 unsigned long haddr = address & huge_page_mask(h);
6786 struct page *page = NULL;
6791 hugetlb_vma_lock_read(vma);
6792 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6796 ptl = huge_pte_lock(h, mm, pte);
6797 entry = huge_ptep_get(pte);
6798 if (pte_present(entry)) {
6799 page = pte_page(entry);
6801 if (!huge_pte_write(entry)) {
6802 if (flags & FOLL_WRITE) {
6807 if (gup_must_unshare(vma, flags, page)) {
6808 /* Tell the caller to do unsharing */
6809 page = ERR_PTR(-EMLINK);
6814 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6817 * Note that page may be a sub-page, and with vmemmap
6818 * optimizations the page struct may be read only.
6819 * try_grab_page() will increase the ref count on the
6820 * head page, so this will be OK.
6822 * try_grab_page() should always be able to get the page here,
6823 * because we hold the ptl lock and have verified pte_present().
6825 ret = try_grab_page(page, flags);
6827 if (WARN_ON_ONCE(ret)) {
6828 page = ERR_PTR(ret);
6832 *page_mask = (1U << huge_page_order(h)) - 1;
6837 hugetlb_vma_unlock_read(vma);
6840 * Fixup retval for dump requests: if pagecache doesn't exist,
6841 * don't try to allocate a new page but just skip it.
6843 if (!page && (flags & FOLL_DUMP) &&
6844 !hugetlbfs_pagecache_present(h, vma, address))
6845 page = ERR_PTR(-EFAULT);
6850 long hugetlb_change_protection(struct vm_area_struct *vma,
6851 unsigned long address, unsigned long end,
6852 pgprot_t newprot, unsigned long cp_flags)
6854 struct mm_struct *mm = vma->vm_mm;
6855 unsigned long start = address;
6858 struct hstate *h = hstate_vma(vma);
6859 long pages = 0, psize = huge_page_size(h);
6860 bool shared_pmd = false;
6861 struct mmu_notifier_range range;
6862 unsigned long last_addr_mask;
6863 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6864 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6867 * In the case of shared PMDs, the area to flush could be beyond
6868 * start/end. Set range.start/range.end to cover the maximum possible
6869 * range if PMD sharing is possible.
6871 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6873 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6875 BUG_ON(address >= end);
6876 flush_cache_range(vma, range.start, range.end);
6878 mmu_notifier_invalidate_range_start(&range);
6879 hugetlb_vma_lock_write(vma);
6880 i_mmap_lock_write(vma->vm_file->f_mapping);
6881 last_addr_mask = hugetlb_mask_last_page(h);
6882 for (; address < end; address += psize) {
6884 ptep = hugetlb_walk(vma, address, psize);
6887 address |= last_addr_mask;
6891 * Userfaultfd wr-protect requires pgtable
6892 * pre-allocations to install pte markers.
6894 ptep = huge_pte_alloc(mm, vma, address, psize);
6900 ptl = huge_pte_lock(h, mm, ptep);
6901 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6903 * When uffd-wp is enabled on the vma, unshare
6904 * shouldn't happen at all. Warn about it if it
6905 * happened due to some reason.
6907 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6911 address |= last_addr_mask;
6914 pte = huge_ptep_get(ptep);
6915 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6916 /* Nothing to do. */
6917 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6918 swp_entry_t entry = pte_to_swp_entry(pte);
6919 struct page *page = pfn_swap_entry_to_page(entry);
6922 if (is_writable_migration_entry(entry)) {
6924 entry = make_readable_exclusive_migration_entry(
6927 entry = make_readable_migration_entry(
6929 newpte = swp_entry_to_pte(entry);
6934 newpte = pte_swp_mkuffd_wp(newpte);
6935 else if (uffd_wp_resolve)
6936 newpte = pte_swp_clear_uffd_wp(newpte);
6937 if (!pte_same(pte, newpte))
6938 set_huge_pte_at(mm, address, ptep, newpte, psize);
6939 } else if (unlikely(is_pte_marker(pte))) {
6940 /* No other markers apply for now. */
6941 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6942 if (uffd_wp_resolve)
6943 /* Safe to modify directly (non-present->none). */
6944 huge_pte_clear(mm, address, ptep, psize);
6945 } else if (!huge_pte_none(pte)) {
6947 unsigned int shift = huge_page_shift(hstate_vma(vma));
6949 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6950 pte = huge_pte_modify(old_pte, newprot);
6951 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6953 pte = huge_pte_mkuffd_wp(pte);
6954 else if (uffd_wp_resolve)
6955 pte = huge_pte_clear_uffd_wp(pte);
6956 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6960 if (unlikely(uffd_wp))
6961 /* Safe to modify directly (none->non-present). */
6962 set_huge_pte_at(mm, address, ptep,
6963 make_pte_marker(PTE_MARKER_UFFD_WP),
6969 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6970 * may have cleared our pud entry and done put_page on the page table:
6971 * once we release i_mmap_rwsem, another task can do the final put_page
6972 * and that page table be reused and filled with junk. If we actually
6973 * did unshare a page of pmds, flush the range corresponding to the pud.
6976 flush_hugetlb_tlb_range(vma, range.start, range.end);
6978 flush_hugetlb_tlb_range(vma, start, end);
6980 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6981 * downgrading page table protection not changing it to point to a new
6984 * See Documentation/mm/mmu_notifier.rst
6986 i_mmap_unlock_write(vma->vm_file->f_mapping);
6987 hugetlb_vma_unlock_write(vma);
6988 mmu_notifier_invalidate_range_end(&range);
6990 return pages > 0 ? (pages << h->order) : pages;
6993 /* Return true if reservation was successful, false otherwise. */
6994 bool hugetlb_reserve_pages(struct inode *inode,
6996 struct vm_area_struct *vma,
6997 vm_flags_t vm_flags)
6999 long chg = -1, add = -1;
7000 struct hstate *h = hstate_inode(inode);
7001 struct hugepage_subpool *spool = subpool_inode(inode);
7002 struct resv_map *resv_map;
7003 struct hugetlb_cgroup *h_cg = NULL;
7004 long gbl_reserve, regions_needed = 0;
7006 /* This should never happen */
7008 VM_WARN(1, "%s called with a negative range\n", __func__);
7013 * vma specific semaphore used for pmd sharing and fault/truncation
7016 hugetlb_vma_lock_alloc(vma);
7019 * Only apply hugepage reservation if asked. At fault time, an
7020 * attempt will be made for VM_NORESERVE to allocate a page
7021 * without using reserves
7023 if (vm_flags & VM_NORESERVE)
7027 * Shared mappings base their reservation on the number of pages that
7028 * are already allocated on behalf of the file. Private mappings need
7029 * to reserve the full area even if read-only as mprotect() may be
7030 * called to make the mapping read-write. Assume !vma is a shm mapping
7032 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7034 * resv_map can not be NULL as hugetlb_reserve_pages is only
7035 * called for inodes for which resv_maps were created (see
7036 * hugetlbfs_get_inode).
7038 resv_map = inode_resv_map(inode);
7040 chg = region_chg(resv_map, from, to, ®ions_needed);
7042 /* Private mapping. */
7043 resv_map = resv_map_alloc();
7049 set_vma_resv_map(vma, resv_map);
7050 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
7056 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
7057 chg * pages_per_huge_page(h), &h_cg) < 0)
7060 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
7061 /* For private mappings, the hugetlb_cgroup uncharge info hangs
7064 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
7068 * There must be enough pages in the subpool for the mapping. If
7069 * the subpool has a minimum size, there may be some global
7070 * reservations already in place (gbl_reserve).
7072 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
7073 if (gbl_reserve < 0)
7074 goto out_uncharge_cgroup;
7077 * Check enough hugepages are available for the reservation.
7078 * Hand the pages back to the subpool if there are not
7080 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
7084 * Account for the reservations made. Shared mappings record regions
7085 * that have reservations as they are shared by multiple VMAs.
7086 * When the last VMA disappears, the region map says how much
7087 * the reservation was and the page cache tells how much of
7088 * the reservation was consumed. Private mappings are per-VMA and
7089 * only the consumed reservations are tracked. When the VMA
7090 * disappears, the original reservation is the VMA size and the
7091 * consumed reservations are stored in the map. Hence, nothing
7092 * else has to be done for private mappings here
7094 if (!vma || vma->vm_flags & VM_MAYSHARE) {
7095 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
7097 if (unlikely(add < 0)) {
7098 hugetlb_acct_memory(h, -gbl_reserve);
7100 } else if (unlikely(chg > add)) {
7102 * pages in this range were added to the reserve
7103 * map between region_chg and region_add. This
7104 * indicates a race with alloc_hugetlb_folio. Adjust
7105 * the subpool and reserve counts modified above
7106 * based on the difference.
7111 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
7112 * reference to h_cg->css. See comment below for detail.
7114 hugetlb_cgroup_uncharge_cgroup_rsvd(
7116 (chg - add) * pages_per_huge_page(h), h_cg);
7118 rsv_adjust = hugepage_subpool_put_pages(spool,
7120 hugetlb_acct_memory(h, -rsv_adjust);
7123 * The file_regions will hold their own reference to
7124 * h_cg->css. So we should release the reference held
7125 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
7128 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
7134 /* put back original number of pages, chg */
7135 (void)hugepage_subpool_put_pages(spool, chg);
7136 out_uncharge_cgroup:
7137 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
7138 chg * pages_per_huge_page(h), h_cg);
7140 hugetlb_vma_lock_free(vma);
7141 if (!vma || vma->vm_flags & VM_MAYSHARE)
7142 /* Only call region_abort if the region_chg succeeded but the
7143 * region_add failed or didn't run.
7145 if (chg >= 0 && add < 0)
7146 region_abort(resv_map, from, to, regions_needed);
7147 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
7148 kref_put(&resv_map->refs, resv_map_release);
7149 set_vma_resv_map(vma, NULL);
7154 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
7157 struct hstate *h = hstate_inode(inode);
7158 struct resv_map *resv_map = inode_resv_map(inode);
7160 struct hugepage_subpool *spool = subpool_inode(inode);
7164 * Since this routine can be called in the evict inode path for all
7165 * hugetlbfs inodes, resv_map could be NULL.
7168 chg = region_del(resv_map, start, end);
7170 * region_del() can fail in the rare case where a region
7171 * must be split and another region descriptor can not be
7172 * allocated. If end == LONG_MAX, it will not fail.
7178 spin_lock(&inode->i_lock);
7179 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
7180 spin_unlock(&inode->i_lock);
7183 * If the subpool has a minimum size, the number of global
7184 * reservations to be released may be adjusted.
7186 * Note that !resv_map implies freed == 0. So (chg - freed)
7187 * won't go negative.
7189 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
7190 hugetlb_acct_memory(h, -gbl_reserve);
7195 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7196 static unsigned long page_table_shareable(struct vm_area_struct *svma,
7197 struct vm_area_struct *vma,
7198 unsigned long addr, pgoff_t idx)
7200 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
7202 unsigned long sbase = saddr & PUD_MASK;
7203 unsigned long s_end = sbase + PUD_SIZE;
7205 /* Allow segments to share if only one is marked locked */
7206 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
7207 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
7210 * match the virtual addresses, permission and the alignment of the
7213 * Also, vma_lock (vm_private_data) is required for sharing.
7215 if (pmd_index(addr) != pmd_index(saddr) ||
7216 vm_flags != svm_flags ||
7217 !range_in_vma(svma, sbase, s_end) ||
7218 !svma->vm_private_data)
7224 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7226 unsigned long start = addr & PUD_MASK;
7227 unsigned long end = start + PUD_SIZE;
7229 #ifdef CONFIG_USERFAULTFD
7230 if (uffd_disable_huge_pmd_share(vma))
7234 * check on proper vm_flags and page table alignment
7236 if (!(vma->vm_flags & VM_MAYSHARE))
7238 if (!vma->vm_private_data) /* vma lock required for sharing */
7240 if (!range_in_vma(vma, start, end))
7246 * Determine if start,end range within vma could be mapped by shared pmd.
7247 * If yes, adjust start and end to cover range associated with possible
7248 * shared pmd mappings.
7250 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7251 unsigned long *start, unsigned long *end)
7253 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7254 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7257 * vma needs to span at least one aligned PUD size, and the range
7258 * must be at least partially within in.
7260 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7261 (*end <= v_start) || (*start >= v_end))
7264 /* Extend the range to be PUD aligned for a worst case scenario */
7265 if (*start > v_start)
7266 *start = ALIGN_DOWN(*start, PUD_SIZE);
7269 *end = ALIGN(*end, PUD_SIZE);
7273 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7274 * and returns the corresponding pte. While this is not necessary for the
7275 * !shared pmd case because we can allocate the pmd later as well, it makes the
7276 * code much cleaner. pmd allocation is essential for the shared case because
7277 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7278 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7279 * bad pmd for sharing.
7281 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7282 unsigned long addr, pud_t *pud)
7284 struct address_space *mapping = vma->vm_file->f_mapping;
7285 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7287 struct vm_area_struct *svma;
7288 unsigned long saddr;
7292 i_mmap_lock_read(mapping);
7293 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7297 saddr = page_table_shareable(svma, vma, addr, idx);
7299 spte = hugetlb_walk(svma, saddr,
7300 vma_mmu_pagesize(svma));
7302 get_page(virt_to_page(spte));
7311 spin_lock(&mm->page_table_lock);
7312 if (pud_none(*pud)) {
7313 pud_populate(mm, pud,
7314 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7317 put_page(virt_to_page(spte));
7319 spin_unlock(&mm->page_table_lock);
7321 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7322 i_mmap_unlock_read(mapping);
7327 * unmap huge page backed by shared pte.
7329 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7330 * indicated by page_count > 1, unmap is achieved by clearing pud and
7331 * decrementing the ref count. If count == 1, the pte page is not shared.
7333 * Called with page table lock held.
7335 * returns: 1 successfully unmapped a shared pte page
7336 * 0 the underlying pte page is not shared, or it is the last user
7338 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7339 unsigned long addr, pte_t *ptep)
7341 pgd_t *pgd = pgd_offset(mm, addr);
7342 p4d_t *p4d = p4d_offset(pgd, addr);
7343 pud_t *pud = pud_offset(p4d, addr);
7345 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7346 hugetlb_vma_assert_locked(vma);
7347 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7348 if (page_count(virt_to_page(ptep)) == 1)
7352 put_page(virt_to_page(ptep));
7357 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7359 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7360 unsigned long addr, pud_t *pud)
7365 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7366 unsigned long addr, pte_t *ptep)
7371 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7372 unsigned long *start, unsigned long *end)
7376 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7380 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7382 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7383 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7384 unsigned long addr, unsigned long sz)
7391 pgd = pgd_offset(mm, addr);
7392 p4d = p4d_alloc(mm, pgd, addr);
7395 pud = pud_alloc(mm, p4d, addr);
7397 if (sz == PUD_SIZE) {
7400 BUG_ON(sz != PMD_SIZE);
7401 if (want_pmd_share(vma, addr) && pud_none(*pud))
7402 pte = huge_pmd_share(mm, vma, addr, pud);
7404 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7409 pte_t pteval = ptep_get_lockless(pte);
7411 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7418 * huge_pte_offset() - Walk the page table to resolve the hugepage
7419 * entry at address @addr
7421 * Return: Pointer to page table entry (PUD or PMD) for
7422 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7423 * size @sz doesn't match the hugepage size at this level of the page
7426 pte_t *huge_pte_offset(struct mm_struct *mm,
7427 unsigned long addr, unsigned long sz)
7434 pgd = pgd_offset(mm, addr);
7435 if (!pgd_present(*pgd))
7437 p4d = p4d_offset(pgd, addr);
7438 if (!p4d_present(*p4d))
7441 pud = pud_offset(p4d, addr);
7443 /* must be pud huge, non-present or none */
7444 return (pte_t *)pud;
7445 if (!pud_present(*pud))
7447 /* must have a valid entry and size to go further */
7449 pmd = pmd_offset(pud, addr);
7450 /* must be pmd huge, non-present or none */
7451 return (pte_t *)pmd;
7455 * Return a mask that can be used to update an address to the last huge
7456 * page in a page table page mapping size. Used to skip non-present
7457 * page table entries when linearly scanning address ranges. Architectures
7458 * with unique huge page to page table relationships can define their own
7459 * version of this routine.
7461 unsigned long hugetlb_mask_last_page(struct hstate *h)
7463 unsigned long hp_size = huge_page_size(h);
7465 if (hp_size == PUD_SIZE)
7466 return P4D_SIZE - PUD_SIZE;
7467 else if (hp_size == PMD_SIZE)
7468 return PUD_SIZE - PMD_SIZE;
7475 /* See description above. Architectures can provide their own version. */
7476 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7478 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7479 if (huge_page_size(h) == PMD_SIZE)
7480 return PUD_SIZE - PMD_SIZE;
7485 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7488 * These functions are overwritable if your architecture needs its own
7491 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7495 spin_lock_irq(&hugetlb_lock);
7496 if (!folio_test_hugetlb(folio) ||
7497 !folio_test_hugetlb_migratable(folio) ||
7498 !folio_try_get(folio)) {
7502 folio_clear_hugetlb_migratable(folio);
7503 list_move_tail(&folio->lru, list);
7505 spin_unlock_irq(&hugetlb_lock);
7509 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7514 spin_lock_irq(&hugetlb_lock);
7515 if (folio_test_hugetlb(folio)) {
7517 if (folio_test_hugetlb_freed(folio))
7519 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7520 ret = folio_try_get(folio);
7524 spin_unlock_irq(&hugetlb_lock);
7528 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7529 bool *migratable_cleared)
7533 spin_lock_irq(&hugetlb_lock);
7534 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7535 spin_unlock_irq(&hugetlb_lock);
7539 void folio_putback_active_hugetlb(struct folio *folio)
7541 spin_lock_irq(&hugetlb_lock);
7542 folio_set_hugetlb_migratable(folio);
7543 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7544 spin_unlock_irq(&hugetlb_lock);
7548 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7550 struct hstate *h = folio_hstate(old_folio);
7552 hugetlb_cgroup_migrate(old_folio, new_folio);
7553 set_page_owner_migrate_reason(&new_folio->page, reason);
7556 * transfer temporary state of the new hugetlb folio. This is
7557 * reverse to other transitions because the newpage is going to
7558 * be final while the old one will be freed so it takes over
7559 * the temporary status.
7561 * Also note that we have to transfer the per-node surplus state
7562 * here as well otherwise the global surplus count will not match
7565 if (folio_test_hugetlb_temporary(new_folio)) {
7566 int old_nid = folio_nid(old_folio);
7567 int new_nid = folio_nid(new_folio);
7569 folio_set_hugetlb_temporary(old_folio);
7570 folio_clear_hugetlb_temporary(new_folio);
7574 * There is no need to transfer the per-node surplus state
7575 * when we do not cross the node.
7577 if (new_nid == old_nid)
7579 spin_lock_irq(&hugetlb_lock);
7580 if (h->surplus_huge_pages_node[old_nid]) {
7581 h->surplus_huge_pages_node[old_nid]--;
7582 h->surplus_huge_pages_node[new_nid]++;
7584 spin_unlock_irq(&hugetlb_lock);
7588 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7589 unsigned long start,
7592 struct hstate *h = hstate_vma(vma);
7593 unsigned long sz = huge_page_size(h);
7594 struct mm_struct *mm = vma->vm_mm;
7595 struct mmu_notifier_range range;
7596 unsigned long address;
7600 if (!(vma->vm_flags & VM_MAYSHARE))
7606 flush_cache_range(vma, start, end);
7608 * No need to call adjust_range_if_pmd_sharing_possible(), because
7609 * we have already done the PUD_SIZE alignment.
7611 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7613 mmu_notifier_invalidate_range_start(&range);
7614 hugetlb_vma_lock_write(vma);
7615 i_mmap_lock_write(vma->vm_file->f_mapping);
7616 for (address = start; address < end; address += PUD_SIZE) {
7617 ptep = hugetlb_walk(vma, address, sz);
7620 ptl = huge_pte_lock(h, mm, ptep);
7621 huge_pmd_unshare(mm, vma, address, ptep);
7624 flush_hugetlb_tlb_range(vma, start, end);
7625 i_mmap_unlock_write(vma->vm_file->f_mapping);
7626 hugetlb_vma_unlock_write(vma);
7628 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7629 * Documentation/mm/mmu_notifier.rst.
7631 mmu_notifier_invalidate_range_end(&range);
7635 * This function will unconditionally remove all the shared pmd pgtable entries
7636 * within the specific vma for a hugetlbfs memory range.
7638 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7640 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7641 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7645 static bool cma_reserve_called __initdata;
7647 static int __init cmdline_parse_hugetlb_cma(char *p)
7654 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7657 if (s[count] == ':') {
7658 if (tmp >= MAX_NUMNODES)
7660 nid = array_index_nospec(tmp, MAX_NUMNODES);
7663 tmp = memparse(s, &s);
7664 hugetlb_cma_size_in_node[nid] = tmp;
7665 hugetlb_cma_size += tmp;
7668 * Skip the separator if have one, otherwise
7669 * break the parsing.
7676 hugetlb_cma_size = memparse(p, &p);
7684 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7686 void __init hugetlb_cma_reserve(int order)
7688 unsigned long size, reserved, per_node;
7689 bool node_specific_cma_alloc = false;
7692 cma_reserve_called = true;
7694 if (!hugetlb_cma_size)
7697 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7698 if (hugetlb_cma_size_in_node[nid] == 0)
7701 if (!node_online(nid)) {
7702 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7703 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7704 hugetlb_cma_size_in_node[nid] = 0;
7708 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7709 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7710 nid, (PAGE_SIZE << order) / SZ_1M);
7711 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7712 hugetlb_cma_size_in_node[nid] = 0;
7714 node_specific_cma_alloc = true;
7718 /* Validate the CMA size again in case some invalid nodes specified. */
7719 if (!hugetlb_cma_size)
7722 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7723 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7724 (PAGE_SIZE << order) / SZ_1M);
7725 hugetlb_cma_size = 0;
7729 if (!node_specific_cma_alloc) {
7731 * If 3 GB area is requested on a machine with 4 numa nodes,
7732 * let's allocate 1 GB on first three nodes and ignore the last one.
7734 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7735 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7736 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7740 for_each_online_node(nid) {
7742 char name[CMA_MAX_NAME];
7744 if (node_specific_cma_alloc) {
7745 if (hugetlb_cma_size_in_node[nid] == 0)
7748 size = hugetlb_cma_size_in_node[nid];
7750 size = min(per_node, hugetlb_cma_size - reserved);
7753 size = round_up(size, PAGE_SIZE << order);
7755 snprintf(name, sizeof(name), "hugetlb%d", nid);
7757 * Note that 'order per bit' is based on smallest size that
7758 * may be returned to CMA allocator in the case of
7759 * huge page demotion.
7761 res = cma_declare_contiguous_nid(0, size, 0,
7762 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7764 &hugetlb_cma[nid], nid);
7766 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7772 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7775 if (reserved >= hugetlb_cma_size)
7781 * hugetlb_cma_size is used to determine if allocations from
7782 * cma are possible. Set to zero if no cma regions are set up.
7784 hugetlb_cma_size = 0;
7787 static void __init hugetlb_cma_check(void)
7789 if (!hugetlb_cma_size || cma_reserve_called)
7792 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7795 #endif /* CONFIG_CMA */