1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
39 #include <asm/pgalloc.h>
43 #include <linux/hugetlb.h>
44 #include <linux/hugetlb_cgroup.h>
45 #include <linux/node.h>
46 #include <linux/page_owner.h>
48 #include "hugetlb_vmemmap.h"
50 int hugetlb_max_hstate __read_mostly;
51 unsigned int default_hstate_idx;
52 struct hstate hstates[HUGE_MAX_HSTATE];
55 static struct cma *hugetlb_cma[MAX_NUMNODES];
56 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
57 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
63 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
68 static unsigned long hugetlb_cma_size __initdata;
70 __initdata LIST_HEAD(huge_boot_pages);
72 /* for command line parsing */
73 static struct hstate * __initdata parsed_hstate;
74 static unsigned long __initdata default_hstate_max_huge_pages;
75 static bool __initdata parsed_valid_hugepagesz = true;
76 static bool __initdata parsed_default_hugepagesz;
77 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
80 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
81 * free_huge_pages, and surplus_huge_pages.
83 DEFINE_SPINLOCK(hugetlb_lock);
86 * Serializes faults on the same logical page. This is used to
87 * prevent spurious OOMs when the hugepage pool is fully utilized.
89 static int num_fault_mutexes;
90 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92 /* Forward declaration */
93 static int hugetlb_acct_memory(struct hstate *h, long delta);
94 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
95 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
96 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
97 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
98 unsigned long start, unsigned long end);
100 static inline bool subpool_is_free(struct hugepage_subpool *spool)
104 if (spool->max_hpages != -1)
105 return spool->used_hpages == 0;
106 if (spool->min_hpages != -1)
107 return spool->rsv_hpages == spool->min_hpages;
112 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
113 unsigned long irq_flags)
115 spin_unlock_irqrestore(&spool->lock, irq_flags);
117 /* If no pages are used, and no other handles to the subpool
118 * remain, give up any reservations based on minimum size and
119 * free the subpool */
120 if (subpool_is_free(spool)) {
121 if (spool->min_hpages != -1)
122 hugetlb_acct_memory(spool->hstate,
128 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 struct hugepage_subpool *spool;
133 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
137 spin_lock_init(&spool->lock);
139 spool->max_hpages = max_hpages;
141 spool->min_hpages = min_hpages;
143 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
147 spool->rsv_hpages = min_hpages;
152 void hugepage_put_subpool(struct hugepage_subpool *spool)
156 spin_lock_irqsave(&spool->lock, flags);
157 BUG_ON(!spool->count);
159 unlock_or_release_subpool(spool, flags);
163 * Subpool accounting for allocating and reserving pages.
164 * Return -ENOMEM if there are not enough resources to satisfy the
165 * request. Otherwise, return the number of pages by which the
166 * global pools must be adjusted (upward). The returned value may
167 * only be different than the passed value (delta) in the case where
168 * a subpool minimum size must be maintained.
170 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
178 spin_lock_irq(&spool->lock);
180 if (spool->max_hpages != -1) { /* maximum size accounting */
181 if ((spool->used_hpages + delta) <= spool->max_hpages)
182 spool->used_hpages += delta;
189 /* minimum size accounting */
190 if (spool->min_hpages != -1 && spool->rsv_hpages) {
191 if (delta > spool->rsv_hpages) {
193 * Asking for more reserves than those already taken on
194 * behalf of subpool. Return difference.
196 ret = delta - spool->rsv_hpages;
197 spool->rsv_hpages = 0;
199 ret = 0; /* reserves already accounted for */
200 spool->rsv_hpages -= delta;
205 spin_unlock_irq(&spool->lock);
210 * Subpool accounting for freeing and unreserving pages.
211 * Return the number of global page reservations that must be dropped.
212 * The return value may only be different than the passed value (delta)
213 * in the case where a subpool minimum size must be maintained.
215 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
224 spin_lock_irqsave(&spool->lock, flags);
226 if (spool->max_hpages != -1) /* maximum size accounting */
227 spool->used_hpages -= delta;
229 /* minimum size accounting */
230 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
231 if (spool->rsv_hpages + delta <= spool->min_hpages)
234 ret = spool->rsv_hpages + delta - spool->min_hpages;
236 spool->rsv_hpages += delta;
237 if (spool->rsv_hpages > spool->min_hpages)
238 spool->rsv_hpages = spool->min_hpages;
242 * If hugetlbfs_put_super couldn't free spool due to an outstanding
243 * quota reference, free it now.
245 unlock_or_release_subpool(spool, flags);
250 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
252 return HUGETLBFS_SB(inode->i_sb)->spool;
255 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
257 return subpool_inode(file_inode(vma->vm_file));
261 * hugetlb vma_lock helper routines
263 static bool __vma_shareable_lock(struct vm_area_struct *vma)
265 return vma->vm_flags & VM_MAYSHARE && vma->vm_private_data;
268 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
270 if (__vma_shareable_lock(vma)) {
271 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
273 down_read(&vma_lock->rw_sema);
277 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
279 if (__vma_shareable_lock(vma)) {
280 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
282 up_read(&vma_lock->rw_sema);
286 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
288 if (__vma_shareable_lock(vma)) {
289 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
291 down_write(&vma_lock->rw_sema);
295 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
297 if (__vma_shareable_lock(vma)) {
298 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
300 up_write(&vma_lock->rw_sema);
304 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
306 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
308 if (!__vma_shareable_lock(vma))
311 return down_write_trylock(&vma_lock->rw_sema);
314 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
316 if (__vma_shareable_lock(vma)) {
317 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
319 lockdep_assert_held(&vma_lock->rw_sema);
323 void hugetlb_vma_lock_release(struct kref *kref)
325 struct hugetlb_vma_lock *vma_lock = container_of(kref,
326 struct hugetlb_vma_lock, refs);
331 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
333 struct vm_area_struct *vma = vma_lock->vma;
336 * vma_lock structure may or not be released as a result of put,
337 * it certainly will no longer be attached to vma so clear pointer.
338 * Semaphore synchronizes access to vma_lock->vma field.
340 vma_lock->vma = NULL;
341 vma->vm_private_data = NULL;
342 up_write(&vma_lock->rw_sema);
343 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
346 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
348 if (__vma_shareable_lock(vma)) {
349 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
351 __hugetlb_vma_unlock_write_put(vma_lock);
355 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
358 * Only present in sharable vmas.
360 if (!vma || !__vma_shareable_lock(vma))
363 if (vma->vm_private_data) {
364 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
366 down_write(&vma_lock->rw_sema);
367 __hugetlb_vma_unlock_write_put(vma_lock);
371 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
373 struct hugetlb_vma_lock *vma_lock;
375 /* Only establish in (flags) sharable vmas */
376 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
379 /* Should never get here with non-NULL vm_private_data */
380 if (vma->vm_private_data)
383 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
386 * If we can not allocate structure, then vma can not
387 * participate in pmd sharing. This is only a possible
388 * performance enhancement and memory saving issue.
389 * However, the lock is also used to synchronize page
390 * faults with truncation. If the lock is not present,
391 * unlikely races could leave pages in a file past i_size
392 * until the file is removed. Warn in the unlikely case of
393 * allocation failure.
395 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
399 kref_init(&vma_lock->refs);
400 init_rwsem(&vma_lock->rw_sema);
402 vma->vm_private_data = vma_lock;
405 /* Helper that removes a struct file_region from the resv_map cache and returns
408 static struct file_region *
409 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
411 struct file_region *nrg;
413 VM_BUG_ON(resv->region_cache_count <= 0);
415 resv->region_cache_count--;
416 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
417 list_del(&nrg->link);
425 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
426 struct file_region *rg)
428 #ifdef CONFIG_CGROUP_HUGETLB
429 nrg->reservation_counter = rg->reservation_counter;
436 /* Helper that records hugetlb_cgroup uncharge info. */
437 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
439 struct resv_map *resv,
440 struct file_region *nrg)
442 #ifdef CONFIG_CGROUP_HUGETLB
444 nrg->reservation_counter =
445 &h_cg->rsvd_hugepage[hstate_index(h)];
446 nrg->css = &h_cg->css;
448 * The caller will hold exactly one h_cg->css reference for the
449 * whole contiguous reservation region. But this area might be
450 * scattered when there are already some file_regions reside in
451 * it. As a result, many file_regions may share only one css
452 * reference. In order to ensure that one file_region must hold
453 * exactly one h_cg->css reference, we should do css_get for
454 * each file_region and leave the reference held by caller
458 if (!resv->pages_per_hpage)
459 resv->pages_per_hpage = pages_per_huge_page(h);
460 /* pages_per_hpage should be the same for all entries in
463 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
465 nrg->reservation_counter = NULL;
471 static void put_uncharge_info(struct file_region *rg)
473 #ifdef CONFIG_CGROUP_HUGETLB
479 static bool has_same_uncharge_info(struct file_region *rg,
480 struct file_region *org)
482 #ifdef CONFIG_CGROUP_HUGETLB
483 return rg->reservation_counter == org->reservation_counter &&
491 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
493 struct file_region *nrg, *prg;
495 prg = list_prev_entry(rg, link);
496 if (&prg->link != &resv->regions && prg->to == rg->from &&
497 has_same_uncharge_info(prg, rg)) {
501 put_uncharge_info(rg);
507 nrg = list_next_entry(rg, link);
508 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
509 has_same_uncharge_info(nrg, rg)) {
510 nrg->from = rg->from;
513 put_uncharge_info(rg);
519 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
520 long to, struct hstate *h, struct hugetlb_cgroup *cg,
521 long *regions_needed)
523 struct file_region *nrg;
525 if (!regions_needed) {
526 nrg = get_file_region_entry_from_cache(map, from, to);
527 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
528 list_add(&nrg->link, rg);
529 coalesce_file_region(map, nrg);
531 *regions_needed += 1;
537 * Must be called with resv->lock held.
539 * Calling this with regions_needed != NULL will count the number of pages
540 * to be added but will not modify the linked list. And regions_needed will
541 * indicate the number of file_regions needed in the cache to carry out to add
542 * the regions for this range.
544 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
545 struct hugetlb_cgroup *h_cg,
546 struct hstate *h, long *regions_needed)
549 struct list_head *head = &resv->regions;
550 long last_accounted_offset = f;
551 struct file_region *iter, *trg = NULL;
552 struct list_head *rg = NULL;
557 /* In this loop, we essentially handle an entry for the range
558 * [last_accounted_offset, iter->from), at every iteration, with some
561 list_for_each_entry_safe(iter, trg, head, link) {
562 /* Skip irrelevant regions that start before our range. */
563 if (iter->from < f) {
564 /* If this region ends after the last accounted offset,
565 * then we need to update last_accounted_offset.
567 if (iter->to > last_accounted_offset)
568 last_accounted_offset = iter->to;
572 /* When we find a region that starts beyond our range, we've
575 if (iter->from >= t) {
576 rg = iter->link.prev;
580 /* Add an entry for last_accounted_offset -> iter->from, and
581 * update last_accounted_offset.
583 if (iter->from > last_accounted_offset)
584 add += hugetlb_resv_map_add(resv, iter->link.prev,
585 last_accounted_offset,
589 last_accounted_offset = iter->to;
592 /* Handle the case where our range extends beyond
593 * last_accounted_offset.
597 if (last_accounted_offset < t)
598 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
599 t, h, h_cg, regions_needed);
604 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
606 static int allocate_file_region_entries(struct resv_map *resv,
608 __must_hold(&resv->lock)
610 LIST_HEAD(allocated_regions);
611 int to_allocate = 0, i = 0;
612 struct file_region *trg = NULL, *rg = NULL;
614 VM_BUG_ON(regions_needed < 0);
617 * Check for sufficient descriptors in the cache to accommodate
618 * the number of in progress add operations plus regions_needed.
620 * This is a while loop because when we drop the lock, some other call
621 * to region_add or region_del may have consumed some region_entries,
622 * so we keep looping here until we finally have enough entries for
623 * (adds_in_progress + regions_needed).
625 while (resv->region_cache_count <
626 (resv->adds_in_progress + regions_needed)) {
627 to_allocate = resv->adds_in_progress + regions_needed -
628 resv->region_cache_count;
630 /* At this point, we should have enough entries in the cache
631 * for all the existing adds_in_progress. We should only be
632 * needing to allocate for regions_needed.
634 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
636 spin_unlock(&resv->lock);
637 for (i = 0; i < to_allocate; i++) {
638 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
641 list_add(&trg->link, &allocated_regions);
644 spin_lock(&resv->lock);
646 list_splice(&allocated_regions, &resv->region_cache);
647 resv->region_cache_count += to_allocate;
653 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
661 * Add the huge page range represented by [f, t) to the reserve
662 * map. Regions will be taken from the cache to fill in this range.
663 * Sufficient regions should exist in the cache due to the previous
664 * call to region_chg with the same range, but in some cases the cache will not
665 * have sufficient entries due to races with other code doing region_add or
666 * region_del. The extra needed entries will be allocated.
668 * regions_needed is the out value provided by a previous call to region_chg.
670 * Return the number of new huge pages added to the map. This number is greater
671 * than or equal to zero. If file_region entries needed to be allocated for
672 * this operation and we were not able to allocate, it returns -ENOMEM.
673 * region_add of regions of length 1 never allocate file_regions and cannot
674 * fail; region_chg will always allocate at least 1 entry and a region_add for
675 * 1 page will only require at most 1 entry.
677 static long region_add(struct resv_map *resv, long f, long t,
678 long in_regions_needed, struct hstate *h,
679 struct hugetlb_cgroup *h_cg)
681 long add = 0, actual_regions_needed = 0;
683 spin_lock(&resv->lock);
686 /* Count how many regions are actually needed to execute this add. */
687 add_reservation_in_range(resv, f, t, NULL, NULL,
688 &actual_regions_needed);
691 * Check for sufficient descriptors in the cache to accommodate
692 * this add operation. Note that actual_regions_needed may be greater
693 * than in_regions_needed, as the resv_map may have been modified since
694 * the region_chg call. In this case, we need to make sure that we
695 * allocate extra entries, such that we have enough for all the
696 * existing adds_in_progress, plus the excess needed for this
699 if (actual_regions_needed > in_regions_needed &&
700 resv->region_cache_count <
701 resv->adds_in_progress +
702 (actual_regions_needed - in_regions_needed)) {
703 /* region_add operation of range 1 should never need to
704 * allocate file_region entries.
706 VM_BUG_ON(t - f <= 1);
708 if (allocate_file_region_entries(
709 resv, actual_regions_needed - in_regions_needed)) {
716 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
718 resv->adds_in_progress -= in_regions_needed;
720 spin_unlock(&resv->lock);
725 * Examine the existing reserve map and determine how many
726 * huge pages in the specified range [f, t) are NOT currently
727 * represented. This routine is called before a subsequent
728 * call to region_add that will actually modify the reserve
729 * map to add the specified range [f, t). region_chg does
730 * not change the number of huge pages represented by the
731 * map. A number of new file_region structures is added to the cache as a
732 * placeholder, for the subsequent region_add call to use. At least 1
733 * file_region structure is added.
735 * out_regions_needed is the number of regions added to the
736 * resv->adds_in_progress. This value needs to be provided to a follow up call
737 * to region_add or region_abort for proper accounting.
739 * Returns the number of huge pages that need to be added to the existing
740 * reservation map for the range [f, t). This number is greater or equal to
741 * zero. -ENOMEM is returned if a new file_region structure or cache entry
742 * is needed and can not be allocated.
744 static long region_chg(struct resv_map *resv, long f, long t,
745 long *out_regions_needed)
749 spin_lock(&resv->lock);
751 /* Count how many hugepages in this range are NOT represented. */
752 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
755 if (*out_regions_needed == 0)
756 *out_regions_needed = 1;
758 if (allocate_file_region_entries(resv, *out_regions_needed))
761 resv->adds_in_progress += *out_regions_needed;
763 spin_unlock(&resv->lock);
768 * Abort the in progress add operation. The adds_in_progress field
769 * of the resv_map keeps track of the operations in progress between
770 * calls to region_chg and region_add. Operations are sometimes
771 * aborted after the call to region_chg. In such cases, region_abort
772 * is called to decrement the adds_in_progress counter. regions_needed
773 * is the value returned by the region_chg call, it is used to decrement
774 * the adds_in_progress counter.
776 * NOTE: The range arguments [f, t) are not needed or used in this
777 * routine. They are kept to make reading the calling code easier as
778 * arguments will match the associated region_chg call.
780 static void region_abort(struct resv_map *resv, long f, long t,
783 spin_lock(&resv->lock);
784 VM_BUG_ON(!resv->region_cache_count);
785 resv->adds_in_progress -= regions_needed;
786 spin_unlock(&resv->lock);
790 * Delete the specified range [f, t) from the reserve map. If the
791 * t parameter is LONG_MAX, this indicates that ALL regions after f
792 * should be deleted. Locate the regions which intersect [f, t)
793 * and either trim, delete or split the existing regions.
795 * Returns the number of huge pages deleted from the reserve map.
796 * In the normal case, the return value is zero or more. In the
797 * case where a region must be split, a new region descriptor must
798 * be allocated. If the allocation fails, -ENOMEM will be returned.
799 * NOTE: If the parameter t == LONG_MAX, then we will never split
800 * a region and possibly return -ENOMEM. Callers specifying
801 * t == LONG_MAX do not need to check for -ENOMEM error.
803 static long region_del(struct resv_map *resv, long f, long t)
805 struct list_head *head = &resv->regions;
806 struct file_region *rg, *trg;
807 struct file_region *nrg = NULL;
811 spin_lock(&resv->lock);
812 list_for_each_entry_safe(rg, trg, head, link) {
814 * Skip regions before the range to be deleted. file_region
815 * ranges are normally of the form [from, to). However, there
816 * may be a "placeholder" entry in the map which is of the form
817 * (from, to) with from == to. Check for placeholder entries
818 * at the beginning of the range to be deleted.
820 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
826 if (f > rg->from && t < rg->to) { /* Must split region */
828 * Check for an entry in the cache before dropping
829 * lock and attempting allocation.
832 resv->region_cache_count > resv->adds_in_progress) {
833 nrg = list_first_entry(&resv->region_cache,
836 list_del(&nrg->link);
837 resv->region_cache_count--;
841 spin_unlock(&resv->lock);
842 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
849 hugetlb_cgroup_uncharge_file_region(
850 resv, rg, t - f, false);
852 /* New entry for end of split region */
856 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
858 INIT_LIST_HEAD(&nrg->link);
860 /* Original entry is trimmed */
863 list_add(&nrg->link, &rg->link);
868 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
869 del += rg->to - rg->from;
870 hugetlb_cgroup_uncharge_file_region(resv, rg,
871 rg->to - rg->from, true);
877 if (f <= rg->from) { /* Trim beginning of region */
878 hugetlb_cgroup_uncharge_file_region(resv, rg,
879 t - rg->from, false);
883 } else { /* Trim end of region */
884 hugetlb_cgroup_uncharge_file_region(resv, rg,
892 spin_unlock(&resv->lock);
898 * A rare out of memory error was encountered which prevented removal of
899 * the reserve map region for a page. The huge page itself was free'ed
900 * and removed from the page cache. This routine will adjust the subpool
901 * usage count, and the global reserve count if needed. By incrementing
902 * these counts, the reserve map entry which could not be deleted will
903 * appear as a "reserved" entry instead of simply dangling with incorrect
906 void hugetlb_fix_reserve_counts(struct inode *inode)
908 struct hugepage_subpool *spool = subpool_inode(inode);
910 bool reserved = false;
912 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
913 if (rsv_adjust > 0) {
914 struct hstate *h = hstate_inode(inode);
916 if (!hugetlb_acct_memory(h, 1))
918 } else if (!rsv_adjust) {
923 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
927 * Count and return the number of huge pages in the reserve map
928 * that intersect with the range [f, t).
930 static long region_count(struct resv_map *resv, long f, long t)
932 struct list_head *head = &resv->regions;
933 struct file_region *rg;
936 spin_lock(&resv->lock);
937 /* Locate each segment we overlap with, and count that overlap. */
938 list_for_each_entry(rg, head, link) {
947 seg_from = max(rg->from, f);
948 seg_to = min(rg->to, t);
950 chg += seg_to - seg_from;
952 spin_unlock(&resv->lock);
958 * Convert the address within this vma to the page offset within
959 * the mapping, in pagecache page units; huge pages here.
961 static pgoff_t vma_hugecache_offset(struct hstate *h,
962 struct vm_area_struct *vma, unsigned long address)
964 return ((address - vma->vm_start) >> huge_page_shift(h)) +
965 (vma->vm_pgoff >> huge_page_order(h));
968 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
969 unsigned long address)
971 return vma_hugecache_offset(hstate_vma(vma), vma, address);
973 EXPORT_SYMBOL_GPL(linear_hugepage_index);
976 * Return the size of the pages allocated when backing a VMA. In the majority
977 * cases this will be same size as used by the page table entries.
979 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
981 if (vma->vm_ops && vma->vm_ops->pagesize)
982 return vma->vm_ops->pagesize(vma);
985 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
988 * Return the page size being used by the MMU to back a VMA. In the majority
989 * of cases, the page size used by the kernel matches the MMU size. On
990 * architectures where it differs, an architecture-specific 'strong'
991 * version of this symbol is required.
993 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
995 return vma_kernel_pagesize(vma);
999 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1000 * bits of the reservation map pointer, which are always clear due to
1003 #define HPAGE_RESV_OWNER (1UL << 0)
1004 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1005 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1008 * These helpers are used to track how many pages are reserved for
1009 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1010 * is guaranteed to have their future faults succeed.
1012 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1013 * the reserve counters are updated with the hugetlb_lock held. It is safe
1014 * to reset the VMA at fork() time as it is not in use yet and there is no
1015 * chance of the global counters getting corrupted as a result of the values.
1017 * The private mapping reservation is represented in a subtly different
1018 * manner to a shared mapping. A shared mapping has a region map associated
1019 * with the underlying file, this region map represents the backing file
1020 * pages which have ever had a reservation assigned which this persists even
1021 * after the page is instantiated. A private mapping has a region map
1022 * associated with the original mmap which is attached to all VMAs which
1023 * reference it, this region map represents those offsets which have consumed
1024 * reservation ie. where pages have been instantiated.
1026 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1028 return (unsigned long)vma->vm_private_data;
1031 static void set_vma_private_data(struct vm_area_struct *vma,
1032 unsigned long value)
1034 vma->vm_private_data = (void *)value;
1038 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1039 struct hugetlb_cgroup *h_cg,
1042 #ifdef CONFIG_CGROUP_HUGETLB
1044 resv_map->reservation_counter = NULL;
1045 resv_map->pages_per_hpage = 0;
1046 resv_map->css = NULL;
1048 resv_map->reservation_counter =
1049 &h_cg->rsvd_hugepage[hstate_index(h)];
1050 resv_map->pages_per_hpage = pages_per_huge_page(h);
1051 resv_map->css = &h_cg->css;
1056 struct resv_map *resv_map_alloc(void)
1058 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1059 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1061 if (!resv_map || !rg) {
1067 kref_init(&resv_map->refs);
1068 spin_lock_init(&resv_map->lock);
1069 INIT_LIST_HEAD(&resv_map->regions);
1071 resv_map->adds_in_progress = 0;
1073 * Initialize these to 0. On shared mappings, 0's here indicate these
1074 * fields don't do cgroup accounting. On private mappings, these will be
1075 * re-initialized to the proper values, to indicate that hugetlb cgroup
1076 * reservations are to be un-charged from here.
1078 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1080 INIT_LIST_HEAD(&resv_map->region_cache);
1081 list_add(&rg->link, &resv_map->region_cache);
1082 resv_map->region_cache_count = 1;
1087 void resv_map_release(struct kref *ref)
1089 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1090 struct list_head *head = &resv_map->region_cache;
1091 struct file_region *rg, *trg;
1093 /* Clear out any active regions before we release the map. */
1094 region_del(resv_map, 0, LONG_MAX);
1096 /* ... and any entries left in the cache */
1097 list_for_each_entry_safe(rg, trg, head, link) {
1098 list_del(&rg->link);
1102 VM_BUG_ON(resv_map->adds_in_progress);
1107 static inline struct resv_map *inode_resv_map(struct inode *inode)
1110 * At inode evict time, i_mapping may not point to the original
1111 * address space within the inode. This original address space
1112 * contains the pointer to the resv_map. So, always use the
1113 * address space embedded within the inode.
1114 * The VERY common case is inode->mapping == &inode->i_data but,
1115 * this may not be true for device special inodes.
1117 return (struct resv_map *)(&inode->i_data)->private_data;
1120 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1122 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1123 if (vma->vm_flags & VM_MAYSHARE) {
1124 struct address_space *mapping = vma->vm_file->f_mapping;
1125 struct inode *inode = mapping->host;
1127 return inode_resv_map(inode);
1130 return (struct resv_map *)(get_vma_private_data(vma) &
1135 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1137 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1138 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1140 set_vma_private_data(vma, (get_vma_private_data(vma) &
1141 HPAGE_RESV_MASK) | (unsigned long)map);
1144 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1146 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1147 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1149 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1152 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1154 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1156 return (get_vma_private_data(vma) & flag) != 0;
1159 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1161 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1163 * Clear vm_private_data
1164 * - For shared mappings this is a per-vma semaphore that may be
1165 * allocated in a subsequent call to hugetlb_vm_op_open.
1166 * Before clearing, make sure pointer is not associated with vma
1167 * as this will leak the structure. This is the case when called
1168 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1169 * been called to allocate a new structure.
1170 * - For MAP_PRIVATE mappings, this is the reserve map which does
1171 * not apply to children. Faults generated by the children are
1172 * not guaranteed to succeed, even if read-only.
1174 if (vma->vm_flags & VM_MAYSHARE) {
1175 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1177 if (vma_lock && vma_lock->vma != vma)
1178 vma->vm_private_data = NULL;
1180 vma->vm_private_data = NULL;
1184 * Reset and decrement one ref on hugepage private reservation.
1185 * Called with mm->mmap_lock writer semaphore held.
1186 * This function should be only used by move_vma() and operate on
1187 * same sized vma. It should never come here with last ref on the
1190 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1193 * Clear the old hugetlb private page reservation.
1194 * It has already been transferred to new_vma.
1196 * During a mremap() operation of a hugetlb vma we call move_vma()
1197 * which copies vma into new_vma and unmaps vma. After the copy
1198 * operation both new_vma and vma share a reference to the resv_map
1199 * struct, and at that point vma is about to be unmapped. We don't
1200 * want to return the reservation to the pool at unmap of vma because
1201 * the reservation still lives on in new_vma, so simply decrement the
1202 * ref here and remove the resv_map reference from this vma.
1204 struct resv_map *reservations = vma_resv_map(vma);
1206 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1207 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1208 kref_put(&reservations->refs, resv_map_release);
1211 hugetlb_dup_vma_private(vma);
1214 /* Returns true if the VMA has associated reserve pages */
1215 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1217 if (vma->vm_flags & VM_NORESERVE) {
1219 * This address is already reserved by other process(chg == 0),
1220 * so, we should decrement reserved count. Without decrementing,
1221 * reserve count remains after releasing inode, because this
1222 * allocated page will go into page cache and is regarded as
1223 * coming from reserved pool in releasing step. Currently, we
1224 * don't have any other solution to deal with this situation
1225 * properly, so add work-around here.
1227 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1233 /* Shared mappings always use reserves */
1234 if (vma->vm_flags & VM_MAYSHARE) {
1236 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1237 * be a region map for all pages. The only situation where
1238 * there is no region map is if a hole was punched via
1239 * fallocate. In this case, there really are no reserves to
1240 * use. This situation is indicated if chg != 0.
1249 * Only the process that called mmap() has reserves for
1252 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1254 * Like the shared case above, a hole punch or truncate
1255 * could have been performed on the private mapping.
1256 * Examine the value of chg to determine if reserves
1257 * actually exist or were previously consumed.
1258 * Very Subtle - The value of chg comes from a previous
1259 * call to vma_needs_reserves(). The reserve map for
1260 * private mappings has different (opposite) semantics
1261 * than that of shared mappings. vma_needs_reserves()
1262 * has already taken this difference in semantics into
1263 * account. Therefore, the meaning of chg is the same
1264 * as in the shared case above. Code could easily be
1265 * combined, but keeping it separate draws attention to
1266 * subtle differences.
1277 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1279 int nid = folio_nid(folio);
1281 lockdep_assert_held(&hugetlb_lock);
1282 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1284 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1285 h->free_huge_pages++;
1286 h->free_huge_pages_node[nid]++;
1287 folio_set_hugetlb_freed(folio);
1290 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1293 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1295 lockdep_assert_held(&hugetlb_lock);
1296 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1297 if (pin && !is_longterm_pinnable_page(page))
1300 if (PageHWPoison(page))
1303 list_move(&page->lru, &h->hugepage_activelist);
1304 set_page_refcounted(page);
1305 ClearHPageFreed(page);
1306 h->free_huge_pages--;
1307 h->free_huge_pages_node[nid]--;
1314 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1317 unsigned int cpuset_mems_cookie;
1318 struct zonelist *zonelist;
1321 int node = NUMA_NO_NODE;
1323 zonelist = node_zonelist(nid, gfp_mask);
1326 cpuset_mems_cookie = read_mems_allowed_begin();
1327 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1330 if (!cpuset_zone_allowed(zone, gfp_mask))
1333 * no need to ask again on the same node. Pool is node rather than
1336 if (zone_to_nid(zone) == node)
1338 node = zone_to_nid(zone);
1340 page = dequeue_huge_page_node_exact(h, node);
1344 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1350 static unsigned long available_huge_pages(struct hstate *h)
1352 return h->free_huge_pages - h->resv_huge_pages;
1355 static struct page *dequeue_huge_page_vma(struct hstate *h,
1356 struct vm_area_struct *vma,
1357 unsigned long address, int avoid_reserve,
1360 struct page *page = NULL;
1361 struct mempolicy *mpol;
1363 nodemask_t *nodemask;
1367 * A child process with MAP_PRIVATE mappings created by their parent
1368 * have no page reserves. This check ensures that reservations are
1369 * not "stolen". The child may still get SIGKILLed
1371 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1374 /* If reserves cannot be used, ensure enough pages are in the pool */
1375 if (avoid_reserve && !available_huge_pages(h))
1378 gfp_mask = htlb_alloc_mask(h);
1379 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1381 if (mpol_is_preferred_many(mpol)) {
1382 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1384 /* Fallback to all nodes if page==NULL */
1389 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1391 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1392 SetHPageRestoreReserve(page);
1393 h->resv_huge_pages--;
1396 mpol_cond_put(mpol);
1404 * common helper functions for hstate_next_node_to_{alloc|free}.
1405 * We may have allocated or freed a huge page based on a different
1406 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1407 * be outside of *nodes_allowed. Ensure that we use an allowed
1408 * node for alloc or free.
1410 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1412 nid = next_node_in(nid, *nodes_allowed);
1413 VM_BUG_ON(nid >= MAX_NUMNODES);
1418 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1420 if (!node_isset(nid, *nodes_allowed))
1421 nid = next_node_allowed(nid, nodes_allowed);
1426 * returns the previously saved node ["this node"] from which to
1427 * allocate a persistent huge page for the pool and advance the
1428 * next node from which to allocate, handling wrap at end of node
1431 static int hstate_next_node_to_alloc(struct hstate *h,
1432 nodemask_t *nodes_allowed)
1436 VM_BUG_ON(!nodes_allowed);
1438 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1439 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1445 * helper for remove_pool_huge_page() - return the previously saved
1446 * node ["this node"] from which to free a huge page. Advance the
1447 * next node id whether or not we find a free huge page to free so
1448 * that the next attempt to free addresses the next node.
1450 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1454 VM_BUG_ON(!nodes_allowed);
1456 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1457 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1462 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1463 for (nr_nodes = nodes_weight(*mask); \
1465 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1468 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1469 for (nr_nodes = nodes_weight(*mask); \
1471 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1474 /* used to demote non-gigantic_huge pages as well */
1475 static void __destroy_compound_gigantic_folio(struct folio *folio,
1476 unsigned int order, bool demote)
1479 int nr_pages = 1 << order;
1482 atomic_set(folio_mapcount_ptr(folio), 0);
1483 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
1484 atomic_set(folio_pincount_ptr(folio), 0);
1486 for (i = 1; i < nr_pages; i++) {
1487 p = folio_page(folio, i);
1489 clear_compound_head(p);
1491 set_page_refcounted(p);
1494 folio_set_order(folio, 0);
1495 __folio_clear_head(folio);
1498 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1501 __destroy_compound_gigantic_folio(folio, order, true);
1504 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1505 static void destroy_compound_gigantic_folio(struct folio *folio,
1508 __destroy_compound_gigantic_folio(folio, order, false);
1511 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1514 * If the page isn't allocated using the cma allocator,
1515 * cma_release() returns false.
1518 int nid = folio_nid(folio);
1520 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1524 free_contig_range(folio_pfn(folio), 1 << order);
1527 #ifdef CONFIG_CONTIG_ALLOC
1528 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1529 int nid, nodemask_t *nodemask)
1532 unsigned long nr_pages = pages_per_huge_page(h);
1533 if (nid == NUMA_NO_NODE)
1534 nid = numa_mem_id();
1540 if (hugetlb_cma[nid]) {
1541 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1542 huge_page_order(h), true);
1544 return page_folio(page);
1547 if (!(gfp_mask & __GFP_THISNODE)) {
1548 for_each_node_mask(node, *nodemask) {
1549 if (node == nid || !hugetlb_cma[node])
1552 page = cma_alloc(hugetlb_cma[node], nr_pages,
1553 huge_page_order(h), true);
1555 return page_folio(page);
1561 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1562 return page ? page_folio(page) : NULL;
1565 #else /* !CONFIG_CONTIG_ALLOC */
1566 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1567 int nid, nodemask_t *nodemask)
1571 #endif /* CONFIG_CONTIG_ALLOC */
1573 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1574 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1575 int nid, nodemask_t *nodemask)
1579 static inline void free_gigantic_folio(struct folio *folio,
1580 unsigned int order) { }
1581 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1582 unsigned int order) { }
1586 * Remove hugetlb folio from lists, and update dtor so that the folio appears
1587 * as just a compound page.
1589 * A reference is held on the folio, except in the case of demote.
1591 * Must be called with hugetlb lock held.
1593 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1594 bool adjust_surplus,
1597 int nid = folio_nid(folio);
1599 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1600 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1602 lockdep_assert_held(&hugetlb_lock);
1603 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1606 list_del(&folio->lru);
1608 if (folio_test_hugetlb_freed(folio)) {
1609 h->free_huge_pages--;
1610 h->free_huge_pages_node[nid]--;
1612 if (adjust_surplus) {
1613 h->surplus_huge_pages--;
1614 h->surplus_huge_pages_node[nid]--;
1620 * For non-gigantic pages set the destructor to the normal compound
1621 * page dtor. This is needed in case someone takes an additional
1622 * temporary ref to the page, and freeing is delayed until they drop
1625 * For gigantic pages set the destructor to the null dtor. This
1626 * destructor will never be called. Before freeing the gigantic
1627 * page destroy_compound_gigantic_folio will turn the folio into a
1628 * simple group of pages. After this the destructor does not
1631 * This handles the case where more than one ref is held when and
1632 * after update_and_free_hugetlb_folio is called.
1634 * In the case of demote we do not ref count the page as it will soon
1635 * be turned into a page of smaller size.
1638 folio_ref_unfreeze(folio, 1);
1639 if (hstate_is_gigantic(h))
1640 folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR);
1642 folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR);
1645 h->nr_huge_pages_node[nid]--;
1648 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1649 bool adjust_surplus)
1651 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1654 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1655 bool adjust_surplus)
1657 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1660 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1661 bool adjust_surplus)
1664 int nid = folio_nid(folio);
1666 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1668 lockdep_assert_held(&hugetlb_lock);
1670 INIT_LIST_HEAD(&folio->lru);
1672 h->nr_huge_pages_node[nid]++;
1674 if (adjust_surplus) {
1675 h->surplus_huge_pages++;
1676 h->surplus_huge_pages_node[nid]++;
1679 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1680 folio_change_private(folio, NULL);
1682 * We have to set hugetlb_vmemmap_optimized again as above
1683 * folio_change_private(folio, NULL) cleared it.
1685 folio_set_hugetlb_vmemmap_optimized(folio);
1688 * This folio is about to be managed by the hugetlb allocator and
1689 * should have no users. Drop our reference, and check for others
1692 zeroed = folio_put_testzero(folio);
1693 if (unlikely(!zeroed))
1695 * It is VERY unlikely soneone else has taken a ref on
1696 * the page. In this case, we simply return as the
1697 * hugetlb destructor (free_huge_page) will be called
1698 * when this other ref is dropped.
1702 arch_clear_hugepage_flags(&folio->page);
1703 enqueue_hugetlb_folio(h, folio);
1706 static void __update_and_free_page(struct hstate *h, struct page *page)
1709 struct folio *folio = page_folio(page);
1710 struct page *subpage;
1712 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1716 * If we don't know which subpages are hwpoisoned, we can't free
1717 * the hugepage, so it's leaked intentionally.
1719 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1722 if (hugetlb_vmemmap_restore(h, page)) {
1723 spin_lock_irq(&hugetlb_lock);
1725 * If we cannot allocate vmemmap pages, just refuse to free the
1726 * page and put the page back on the hugetlb free list and treat
1727 * as a surplus page.
1729 add_hugetlb_folio(h, folio, true);
1730 spin_unlock_irq(&hugetlb_lock);
1735 * Move PageHWPoison flag from head page to the raw error pages,
1736 * which makes any healthy subpages reusable.
1738 if (unlikely(folio_test_hwpoison(folio)))
1739 hugetlb_clear_page_hwpoison(&folio->page);
1741 for (i = 0; i < pages_per_huge_page(h); i++) {
1742 subpage = folio_page(folio, i);
1743 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1744 1 << PG_referenced | 1 << PG_dirty |
1745 1 << PG_active | 1 << PG_private |
1750 * Non-gigantic pages demoted from CMA allocated gigantic pages
1751 * need to be given back to CMA in free_gigantic_folio.
1753 if (hstate_is_gigantic(h) ||
1754 hugetlb_cma_folio(folio, huge_page_order(h))) {
1755 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1756 free_gigantic_folio(folio, huge_page_order(h));
1758 __free_pages(page, huge_page_order(h));
1763 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1764 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1765 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1766 * the vmemmap pages.
1768 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1769 * freed and frees them one-by-one. As the page->mapping pointer is going
1770 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1771 * structure of a lockless linked list of huge pages to be freed.
1773 static LLIST_HEAD(hpage_freelist);
1775 static void free_hpage_workfn(struct work_struct *work)
1777 struct llist_node *node;
1779 node = llist_del_all(&hpage_freelist);
1785 page = container_of((struct address_space **)node,
1786 struct page, mapping);
1788 page->mapping = NULL;
1790 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1791 * is going to trigger because a previous call to
1792 * remove_hugetlb_folio() will call folio_set_compound_dtor
1793 * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate()
1796 h = size_to_hstate(page_size(page));
1798 __update_and_free_page(h, page);
1803 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1805 static inline void flush_free_hpage_work(struct hstate *h)
1807 if (hugetlb_vmemmap_optimizable(h))
1808 flush_work(&free_hpage_work);
1811 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1814 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1815 __update_and_free_page(h, &folio->page);
1820 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1822 * Only call schedule_work() if hpage_freelist is previously
1823 * empty. Otherwise, schedule_work() had been called but the workfn
1824 * hasn't retrieved the list yet.
1826 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1827 schedule_work(&free_hpage_work);
1830 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1832 struct page *page, *t_page;
1833 struct folio *folio;
1835 list_for_each_entry_safe(page, t_page, list, lru) {
1836 folio = page_folio(page);
1837 update_and_free_hugetlb_folio(h, folio, false);
1842 struct hstate *size_to_hstate(unsigned long size)
1846 for_each_hstate(h) {
1847 if (huge_page_size(h) == size)
1853 void free_huge_page(struct page *page)
1856 * Can't pass hstate in here because it is called from the
1857 * compound page destructor.
1859 struct folio *folio = page_folio(page);
1860 struct hstate *h = folio_hstate(folio);
1861 int nid = folio_nid(folio);
1862 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1863 bool restore_reserve;
1864 unsigned long flags;
1866 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1867 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1869 hugetlb_set_folio_subpool(folio, NULL);
1870 if (folio_test_anon(folio))
1871 __ClearPageAnonExclusive(&folio->page);
1872 folio->mapping = NULL;
1873 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1874 folio_clear_hugetlb_restore_reserve(folio);
1877 * If HPageRestoreReserve was set on page, page allocation consumed a
1878 * reservation. If the page was associated with a subpool, there
1879 * would have been a page reserved in the subpool before allocation
1880 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1881 * reservation, do not call hugepage_subpool_put_pages() as this will
1882 * remove the reserved page from the subpool.
1884 if (!restore_reserve) {
1886 * A return code of zero implies that the subpool will be
1887 * under its minimum size if the reservation is not restored
1888 * after page is free. Therefore, force restore_reserve
1891 if (hugepage_subpool_put_pages(spool, 1) == 0)
1892 restore_reserve = true;
1895 spin_lock_irqsave(&hugetlb_lock, flags);
1896 folio_clear_hugetlb_migratable(folio);
1897 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1898 pages_per_huge_page(h), folio);
1899 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1900 pages_per_huge_page(h), folio);
1901 if (restore_reserve)
1902 h->resv_huge_pages++;
1904 if (folio_test_hugetlb_temporary(folio)) {
1905 remove_hugetlb_folio(h, folio, false);
1906 spin_unlock_irqrestore(&hugetlb_lock, flags);
1907 update_and_free_hugetlb_folio(h, folio, true);
1908 } else if (h->surplus_huge_pages_node[nid]) {
1909 /* remove the page from active list */
1910 remove_hugetlb_folio(h, folio, true);
1911 spin_unlock_irqrestore(&hugetlb_lock, flags);
1912 update_and_free_hugetlb_folio(h, folio, true);
1914 arch_clear_hugepage_flags(page);
1915 enqueue_hugetlb_folio(h, folio);
1916 spin_unlock_irqrestore(&hugetlb_lock, flags);
1921 * Must be called with the hugetlb lock held
1923 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1925 lockdep_assert_held(&hugetlb_lock);
1927 h->nr_huge_pages_node[nid]++;
1930 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1932 hugetlb_vmemmap_optimize(h, &folio->page);
1933 INIT_LIST_HEAD(&folio->lru);
1934 folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR);
1935 hugetlb_set_folio_subpool(folio, NULL);
1936 set_hugetlb_cgroup(folio, NULL);
1937 set_hugetlb_cgroup_rsvd(folio, NULL);
1940 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1942 __prep_new_hugetlb_folio(h, folio);
1943 spin_lock_irq(&hugetlb_lock);
1944 __prep_account_new_huge_page(h, nid);
1945 spin_unlock_irq(&hugetlb_lock);
1948 static bool __prep_compound_gigantic_folio(struct folio *folio,
1949 unsigned int order, bool demote)
1952 int nr_pages = 1 << order;
1955 __folio_clear_reserved(folio);
1956 __folio_set_head(folio);
1957 /* we rely on prep_new_hugetlb_folio to set the destructor */
1958 folio_set_order(folio, order);
1959 for (i = 0; i < nr_pages; i++) {
1960 p = folio_page(folio, i);
1963 * For gigantic hugepages allocated through bootmem at
1964 * boot, it's safer to be consistent with the not-gigantic
1965 * hugepages and clear the PG_reserved bit from all tail pages
1966 * too. Otherwise drivers using get_user_pages() to access tail
1967 * pages may get the reference counting wrong if they see
1968 * PG_reserved set on a tail page (despite the head page not
1969 * having PG_reserved set). Enforcing this consistency between
1970 * head and tail pages allows drivers to optimize away a check
1971 * on the head page when they need know if put_page() is needed
1972 * after get_user_pages().
1974 if (i != 0) /* head page cleared above */
1975 __ClearPageReserved(p);
1977 * Subtle and very unlikely
1979 * Gigantic 'page allocators' such as memblock or cma will
1980 * return a set of pages with each page ref counted. We need
1981 * to turn this set of pages into a compound page with tail
1982 * page ref counts set to zero. Code such as speculative page
1983 * cache adding could take a ref on a 'to be' tail page.
1984 * We need to respect any increased ref count, and only set
1985 * the ref count to zero if count is currently 1. If count
1986 * is not 1, we return an error. An error return indicates
1987 * the set of pages can not be converted to a gigantic page.
1988 * The caller who allocated the pages should then discard the
1989 * pages using the appropriate free interface.
1991 * In the case of demote, the ref count will be zero.
1994 if (!page_ref_freeze(p, 1)) {
1995 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1999 VM_BUG_ON_PAGE(page_count(p), p);
2002 set_compound_head(p, &folio->page);
2004 atomic_set(folio_mapcount_ptr(folio), -1);
2005 atomic_set(folio_subpages_mapcount_ptr(folio), 0);
2006 atomic_set(folio_pincount_ptr(folio), 0);
2010 /* undo page modifications made above */
2011 for (j = 0; j < i; j++) {
2012 p = folio_page(folio, j);
2014 clear_compound_head(p);
2015 set_page_refcounted(p);
2017 /* need to clear PG_reserved on remaining tail pages */
2018 for (; j < nr_pages; j++) {
2019 p = folio_page(folio, j);
2020 __ClearPageReserved(p);
2022 folio_set_order(folio, 0);
2023 __folio_clear_head(folio);
2027 static bool prep_compound_gigantic_folio(struct folio *folio,
2030 return __prep_compound_gigantic_folio(folio, order, false);
2033 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2036 return __prep_compound_gigantic_folio(folio, order, true);
2040 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2041 * transparent huge pages. See the PageTransHuge() documentation for more
2044 int PageHuge(struct page *page)
2046 if (!PageCompound(page))
2049 page = compound_head(page);
2050 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
2052 EXPORT_SYMBOL_GPL(PageHuge);
2055 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
2056 * normal or transparent huge pages.
2058 int PageHeadHuge(struct page *page_head)
2060 if (!PageHead(page_head))
2063 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
2065 EXPORT_SYMBOL_GPL(PageHeadHuge);
2068 * Find and lock address space (mapping) in write mode.
2070 * Upon entry, the page is locked which means that page_mapping() is
2071 * stable. Due to locking order, we can only trylock_write. If we can
2072 * not get the lock, simply return NULL to caller.
2074 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2076 struct address_space *mapping = page_mapping(hpage);
2081 if (i_mmap_trylock_write(mapping))
2087 pgoff_t hugetlb_basepage_index(struct page *page)
2089 struct page *page_head = compound_head(page);
2090 pgoff_t index = page_index(page_head);
2091 unsigned long compound_idx;
2093 if (compound_order(page_head) >= MAX_ORDER)
2094 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2096 compound_idx = page - page_head;
2098 return (index << compound_order(page_head)) + compound_idx;
2101 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2102 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2103 nodemask_t *node_alloc_noretry)
2105 int order = huge_page_order(h);
2107 bool alloc_try_hard = true;
2111 * By default we always try hard to allocate the page with
2112 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2113 * a loop (to adjust global huge page counts) and previous allocation
2114 * failed, do not continue to try hard on the same node. Use the
2115 * node_alloc_noretry bitmap to manage this state information.
2117 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2118 alloc_try_hard = false;
2119 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2121 gfp_mask |= __GFP_RETRY_MAYFAIL;
2122 if (nid == NUMA_NO_NODE)
2123 nid = numa_mem_id();
2125 page = __alloc_pages(gfp_mask, order, nid, nmask);
2127 /* Freeze head page */
2128 if (page && !page_ref_freeze(page, 1)) {
2129 __free_pages(page, order);
2130 if (retry) { /* retry once */
2134 /* WOW! twice in a row. */
2135 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2140 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2141 * indicates an overall state change. Clear bit so that we resume
2142 * normal 'try hard' allocations.
2144 if (node_alloc_noretry && page && !alloc_try_hard)
2145 node_clear(nid, *node_alloc_noretry);
2148 * If we tried hard to get a page but failed, set bit so that
2149 * subsequent attempts will not try as hard until there is an
2150 * overall state change.
2152 if (node_alloc_noretry && !page && alloc_try_hard)
2153 node_set(nid, *node_alloc_noretry);
2156 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2160 __count_vm_event(HTLB_BUDDY_PGALLOC);
2161 return page_folio(page);
2165 * Common helper to allocate a fresh hugetlb page. All specific allocators
2166 * should use this function to get new hugetlb pages
2168 * Note that returned page is 'frozen': ref count of head page and all tail
2171 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2172 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2173 nodemask_t *node_alloc_noretry)
2175 struct folio *folio;
2179 if (hstate_is_gigantic(h))
2180 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2182 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2183 nid, nmask, node_alloc_noretry);
2186 if (hstate_is_gigantic(h)) {
2187 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2189 * Rare failure to convert pages to compound page.
2190 * Free pages and try again - ONCE!
2192 free_gigantic_folio(folio, huge_page_order(h));
2200 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2206 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2209 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2210 nodemask_t *node_alloc_noretry)
2212 struct folio *folio;
2214 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2216 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2217 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2218 nodes_allowed, node_alloc_noretry);
2220 free_huge_page(&folio->page); /* free it into the hugepage allocator */
2229 * Remove huge page from pool from next node to free. Attempt to keep
2230 * persistent huge pages more or less balanced over allowed nodes.
2231 * This routine only 'removes' the hugetlb page. The caller must make
2232 * an additional call to free the page to low level allocators.
2233 * Called with hugetlb_lock locked.
2235 static struct page *remove_pool_huge_page(struct hstate *h,
2236 nodemask_t *nodes_allowed,
2240 struct page *page = NULL;
2241 struct folio *folio;
2243 lockdep_assert_held(&hugetlb_lock);
2244 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2246 * If we're returning unused surplus pages, only examine
2247 * nodes with surplus pages.
2249 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2250 !list_empty(&h->hugepage_freelists[node])) {
2251 page = list_entry(h->hugepage_freelists[node].next,
2253 folio = page_folio(page);
2254 remove_hugetlb_folio(h, folio, acct_surplus);
2263 * Dissolve a given free hugepage into free buddy pages. This function does
2264 * nothing for in-use hugepages and non-hugepages.
2265 * This function returns values like below:
2267 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2268 * when the system is under memory pressure and the feature of
2269 * freeing unused vmemmap pages associated with each hugetlb page
2271 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2272 * (allocated or reserved.)
2273 * 0: successfully dissolved free hugepages or the page is not a
2274 * hugepage (considered as already dissolved)
2276 int dissolve_free_huge_page(struct page *page)
2279 struct folio *folio = page_folio(page);
2282 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2283 if (!folio_test_hugetlb(folio))
2286 spin_lock_irq(&hugetlb_lock);
2287 if (!folio_test_hugetlb(folio)) {
2292 if (!folio_ref_count(folio)) {
2293 struct hstate *h = folio_hstate(folio);
2294 if (!available_huge_pages(h))
2298 * We should make sure that the page is already on the free list
2299 * when it is dissolved.
2301 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2302 spin_unlock_irq(&hugetlb_lock);
2306 * Theoretically, we should return -EBUSY when we
2307 * encounter this race. In fact, we have a chance
2308 * to successfully dissolve the page if we do a
2309 * retry. Because the race window is quite small.
2310 * If we seize this opportunity, it is an optimization
2311 * for increasing the success rate of dissolving page.
2316 remove_hugetlb_folio(h, folio, false);
2317 h->max_huge_pages--;
2318 spin_unlock_irq(&hugetlb_lock);
2321 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2322 * before freeing the page. update_and_free_hugtlb_folio will fail to
2323 * free the page if it can not allocate required vmemmap. We
2324 * need to adjust max_huge_pages if the page is not freed.
2325 * Attempt to allocate vmemmmap here so that we can take
2326 * appropriate action on failure.
2328 rc = hugetlb_vmemmap_restore(h, &folio->page);
2330 update_and_free_hugetlb_folio(h, folio, false);
2332 spin_lock_irq(&hugetlb_lock);
2333 add_hugetlb_folio(h, folio, false);
2334 h->max_huge_pages++;
2335 spin_unlock_irq(&hugetlb_lock);
2341 spin_unlock_irq(&hugetlb_lock);
2346 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2347 * make specified memory blocks removable from the system.
2348 * Note that this will dissolve a free gigantic hugepage completely, if any
2349 * part of it lies within the given range.
2350 * Also note that if dissolve_free_huge_page() returns with an error, all
2351 * free hugepages that were dissolved before that error are lost.
2353 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2361 if (!hugepages_supported())
2364 order = huge_page_order(&default_hstate);
2366 order = min(order, huge_page_order(h));
2368 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2369 page = pfn_to_page(pfn);
2370 rc = dissolve_free_huge_page(page);
2379 * Allocates a fresh surplus page from the page allocator.
2381 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2382 int nid, nodemask_t *nmask)
2384 struct folio *folio = NULL;
2386 if (hstate_is_gigantic(h))
2389 spin_lock_irq(&hugetlb_lock);
2390 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2392 spin_unlock_irq(&hugetlb_lock);
2394 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2398 spin_lock_irq(&hugetlb_lock);
2400 * We could have raced with the pool size change.
2401 * Double check that and simply deallocate the new page
2402 * if we would end up overcommiting the surpluses. Abuse
2403 * temporary page to workaround the nasty free_huge_page
2406 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2407 folio_set_hugetlb_temporary(folio);
2408 spin_unlock_irq(&hugetlb_lock);
2409 free_huge_page(&folio->page);
2413 h->surplus_huge_pages++;
2414 h->surplus_huge_pages_node[folio_nid(folio)]++;
2417 spin_unlock_irq(&hugetlb_lock);
2419 return &folio->page;
2422 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2423 int nid, nodemask_t *nmask)
2425 struct folio *folio;
2427 if (hstate_is_gigantic(h))
2430 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2434 /* fresh huge pages are frozen */
2435 folio_ref_unfreeze(folio, 1);
2437 * We do not account these pages as surplus because they are only
2438 * temporary and will be released properly on the last reference
2440 folio_set_hugetlb_temporary(folio);
2442 return &folio->page;
2446 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2449 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2450 struct vm_area_struct *vma, unsigned long addr)
2452 struct page *page = NULL;
2453 struct mempolicy *mpol;
2454 gfp_t gfp_mask = htlb_alloc_mask(h);
2456 nodemask_t *nodemask;
2458 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2459 if (mpol_is_preferred_many(mpol)) {
2460 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2462 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2463 page = alloc_surplus_huge_page(h, gfp, nid, nodemask);
2465 /* Fallback to all nodes if page==NULL */
2470 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2471 mpol_cond_put(mpol);
2475 /* page migration callback function */
2476 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2477 nodemask_t *nmask, gfp_t gfp_mask)
2479 spin_lock_irq(&hugetlb_lock);
2480 if (available_huge_pages(h)) {
2483 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2485 spin_unlock_irq(&hugetlb_lock);
2489 spin_unlock_irq(&hugetlb_lock);
2491 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2494 /* mempolicy aware migration callback */
2495 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2496 unsigned long address)
2498 struct mempolicy *mpol;
2499 nodemask_t *nodemask;
2504 gfp_mask = htlb_alloc_mask(h);
2505 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2506 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2507 mpol_cond_put(mpol);
2513 * Increase the hugetlb pool such that it can accommodate a reservation
2516 static int gather_surplus_pages(struct hstate *h, long delta)
2517 __must_hold(&hugetlb_lock)
2519 LIST_HEAD(surplus_list);
2520 struct page *page, *tmp;
2523 long needed, allocated;
2524 bool alloc_ok = true;
2526 lockdep_assert_held(&hugetlb_lock);
2527 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2529 h->resv_huge_pages += delta;
2537 spin_unlock_irq(&hugetlb_lock);
2538 for (i = 0; i < needed; i++) {
2539 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2540 NUMA_NO_NODE, NULL);
2545 list_add(&page->lru, &surplus_list);
2551 * After retaking hugetlb_lock, we need to recalculate 'needed'
2552 * because either resv_huge_pages or free_huge_pages may have changed.
2554 spin_lock_irq(&hugetlb_lock);
2555 needed = (h->resv_huge_pages + delta) -
2556 (h->free_huge_pages + allocated);
2561 * We were not able to allocate enough pages to
2562 * satisfy the entire reservation so we free what
2563 * we've allocated so far.
2568 * The surplus_list now contains _at_least_ the number of extra pages
2569 * needed to accommodate the reservation. Add the appropriate number
2570 * of pages to the hugetlb pool and free the extras back to the buddy
2571 * allocator. Commit the entire reservation here to prevent another
2572 * process from stealing the pages as they are added to the pool but
2573 * before they are reserved.
2575 needed += allocated;
2576 h->resv_huge_pages += delta;
2579 /* Free the needed pages to the hugetlb pool */
2580 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2583 /* Add the page to the hugetlb allocator */
2584 enqueue_hugetlb_folio(h, page_folio(page));
2587 spin_unlock_irq(&hugetlb_lock);
2590 * Free unnecessary surplus pages to the buddy allocator.
2591 * Pages have no ref count, call free_huge_page directly.
2593 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2594 free_huge_page(page);
2595 spin_lock_irq(&hugetlb_lock);
2601 * This routine has two main purposes:
2602 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2603 * in unused_resv_pages. This corresponds to the prior adjustments made
2604 * to the associated reservation map.
2605 * 2) Free any unused surplus pages that may have been allocated to satisfy
2606 * the reservation. As many as unused_resv_pages may be freed.
2608 static void return_unused_surplus_pages(struct hstate *h,
2609 unsigned long unused_resv_pages)
2611 unsigned long nr_pages;
2613 LIST_HEAD(page_list);
2615 lockdep_assert_held(&hugetlb_lock);
2616 /* Uncommit the reservation */
2617 h->resv_huge_pages -= unused_resv_pages;
2619 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2623 * Part (or even all) of the reservation could have been backed
2624 * by pre-allocated pages. Only free surplus pages.
2626 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2629 * We want to release as many surplus pages as possible, spread
2630 * evenly across all nodes with memory. Iterate across these nodes
2631 * until we can no longer free unreserved surplus pages. This occurs
2632 * when the nodes with surplus pages have no free pages.
2633 * remove_pool_huge_page() will balance the freed pages across the
2634 * on-line nodes with memory and will handle the hstate accounting.
2636 while (nr_pages--) {
2637 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2641 list_add(&page->lru, &page_list);
2645 spin_unlock_irq(&hugetlb_lock);
2646 update_and_free_pages_bulk(h, &page_list);
2647 spin_lock_irq(&hugetlb_lock);
2652 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2653 * are used by the huge page allocation routines to manage reservations.
2655 * vma_needs_reservation is called to determine if the huge page at addr
2656 * within the vma has an associated reservation. If a reservation is
2657 * needed, the value 1 is returned. The caller is then responsible for
2658 * managing the global reservation and subpool usage counts. After
2659 * the huge page has been allocated, vma_commit_reservation is called
2660 * to add the page to the reservation map. If the page allocation fails,
2661 * the reservation must be ended instead of committed. vma_end_reservation
2662 * is called in such cases.
2664 * In the normal case, vma_commit_reservation returns the same value
2665 * as the preceding vma_needs_reservation call. The only time this
2666 * is not the case is if a reserve map was changed between calls. It
2667 * is the responsibility of the caller to notice the difference and
2668 * take appropriate action.
2670 * vma_add_reservation is used in error paths where a reservation must
2671 * be restored when a newly allocated huge page must be freed. It is
2672 * to be called after calling vma_needs_reservation to determine if a
2673 * reservation exists.
2675 * vma_del_reservation is used in error paths where an entry in the reserve
2676 * map was created during huge page allocation and must be removed. It is to
2677 * be called after calling vma_needs_reservation to determine if a reservation
2680 enum vma_resv_mode {
2687 static long __vma_reservation_common(struct hstate *h,
2688 struct vm_area_struct *vma, unsigned long addr,
2689 enum vma_resv_mode mode)
2691 struct resv_map *resv;
2694 long dummy_out_regions_needed;
2696 resv = vma_resv_map(vma);
2700 idx = vma_hugecache_offset(h, vma, addr);
2702 case VMA_NEEDS_RESV:
2703 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2704 /* We assume that vma_reservation_* routines always operate on
2705 * 1 page, and that adding to resv map a 1 page entry can only
2706 * ever require 1 region.
2708 VM_BUG_ON(dummy_out_regions_needed != 1);
2710 case VMA_COMMIT_RESV:
2711 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2712 /* region_add calls of range 1 should never fail. */
2716 region_abort(resv, idx, idx + 1, 1);
2720 if (vma->vm_flags & VM_MAYSHARE) {
2721 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2722 /* region_add calls of range 1 should never fail. */
2725 region_abort(resv, idx, idx + 1, 1);
2726 ret = region_del(resv, idx, idx + 1);
2730 if (vma->vm_flags & VM_MAYSHARE) {
2731 region_abort(resv, idx, idx + 1, 1);
2732 ret = region_del(resv, idx, idx + 1);
2734 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2735 /* region_add calls of range 1 should never fail. */
2743 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2746 * We know private mapping must have HPAGE_RESV_OWNER set.
2748 * In most cases, reserves always exist for private mappings.
2749 * However, a file associated with mapping could have been
2750 * hole punched or truncated after reserves were consumed.
2751 * As subsequent fault on such a range will not use reserves.
2752 * Subtle - The reserve map for private mappings has the
2753 * opposite meaning than that of shared mappings. If NO
2754 * entry is in the reserve map, it means a reservation exists.
2755 * If an entry exists in the reserve map, it means the
2756 * reservation has already been consumed. As a result, the
2757 * return value of this routine is the opposite of the
2758 * value returned from reserve map manipulation routines above.
2767 static long vma_needs_reservation(struct hstate *h,
2768 struct vm_area_struct *vma, unsigned long addr)
2770 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2773 static long vma_commit_reservation(struct hstate *h,
2774 struct vm_area_struct *vma, unsigned long addr)
2776 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2779 static void vma_end_reservation(struct hstate *h,
2780 struct vm_area_struct *vma, unsigned long addr)
2782 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2785 static long vma_add_reservation(struct hstate *h,
2786 struct vm_area_struct *vma, unsigned long addr)
2788 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2791 static long vma_del_reservation(struct hstate *h,
2792 struct vm_area_struct *vma, unsigned long addr)
2794 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2798 * This routine is called to restore reservation information on error paths.
2799 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2800 * the hugetlb mutex should remain held when calling this routine.
2802 * It handles two specific cases:
2803 * 1) A reservation was in place and the page consumed the reservation.
2804 * HPageRestoreReserve is set in the page.
2805 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2806 * not set. However, alloc_huge_page always updates the reserve map.
2808 * In case 1, free_huge_page later in the error path will increment the
2809 * global reserve count. But, free_huge_page does not have enough context
2810 * to adjust the reservation map. This case deals primarily with private
2811 * mappings. Adjust the reserve map here to be consistent with global
2812 * reserve count adjustments to be made by free_huge_page. Make sure the
2813 * reserve map indicates there is a reservation present.
2815 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2817 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2818 unsigned long address, struct page *page)
2820 long rc = vma_needs_reservation(h, vma, address);
2822 if (HPageRestoreReserve(page)) {
2823 if (unlikely(rc < 0))
2825 * Rare out of memory condition in reserve map
2826 * manipulation. Clear HPageRestoreReserve so that
2827 * global reserve count will not be incremented
2828 * by free_huge_page. This will make it appear
2829 * as though the reservation for this page was
2830 * consumed. This may prevent the task from
2831 * faulting in the page at a later time. This
2832 * is better than inconsistent global huge page
2833 * accounting of reserve counts.
2835 ClearHPageRestoreReserve(page);
2837 (void)vma_add_reservation(h, vma, address);
2839 vma_end_reservation(h, vma, address);
2843 * This indicates there is an entry in the reserve map
2844 * not added by alloc_huge_page. We know it was added
2845 * before the alloc_huge_page call, otherwise
2846 * HPageRestoreReserve would be set on the page.
2847 * Remove the entry so that a subsequent allocation
2848 * does not consume a reservation.
2850 rc = vma_del_reservation(h, vma, address);
2853 * VERY rare out of memory condition. Since
2854 * we can not delete the entry, set
2855 * HPageRestoreReserve so that the reserve
2856 * count will be incremented when the page
2857 * is freed. This reserve will be consumed
2858 * on a subsequent allocation.
2860 SetHPageRestoreReserve(page);
2861 } else if (rc < 0) {
2863 * Rare out of memory condition from
2864 * vma_needs_reservation call. Memory allocation is
2865 * only attempted if a new entry is needed. Therefore,
2866 * this implies there is not an entry in the
2869 * For shared mappings, no entry in the map indicates
2870 * no reservation. We are done.
2872 if (!(vma->vm_flags & VM_MAYSHARE))
2874 * For private mappings, no entry indicates
2875 * a reservation is present. Since we can
2876 * not add an entry, set SetHPageRestoreReserve
2877 * on the page so reserve count will be
2878 * incremented when freed. This reserve will
2879 * be consumed on a subsequent allocation.
2881 SetHPageRestoreReserve(page);
2884 * No reservation present, do nothing
2886 vma_end_reservation(h, vma, address);
2891 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2893 * @h: struct hstate old page belongs to
2894 * @old_folio: Old folio to dissolve
2895 * @list: List to isolate the page in case we need to
2896 * Returns 0 on success, otherwise negated error.
2898 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2899 struct folio *old_folio, struct list_head *list)
2901 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2902 int nid = folio_nid(old_folio);
2903 struct folio *new_folio;
2907 * Before dissolving the folio, we need to allocate a new one for the
2908 * pool to remain stable. Here, we allocate the folio and 'prep' it
2909 * by doing everything but actually updating counters and adding to
2910 * the pool. This simplifies and let us do most of the processing
2913 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2916 __prep_new_hugetlb_folio(h, new_folio);
2919 spin_lock_irq(&hugetlb_lock);
2920 if (!folio_test_hugetlb(old_folio)) {
2922 * Freed from under us. Drop new_folio too.
2925 } else if (folio_ref_count(old_folio)) {
2927 * Someone has grabbed the folio, try to isolate it here.
2928 * Fail with -EBUSY if not possible.
2930 spin_unlock_irq(&hugetlb_lock);
2931 ret = isolate_hugetlb(&old_folio->page, list);
2932 spin_lock_irq(&hugetlb_lock);
2934 } else if (!folio_test_hugetlb_freed(old_folio)) {
2936 * Folio's refcount is 0 but it has not been enqueued in the
2937 * freelist yet. Race window is small, so we can succeed here if
2940 spin_unlock_irq(&hugetlb_lock);
2945 * Ok, old_folio is still a genuine free hugepage. Remove it from
2946 * the freelist and decrease the counters. These will be
2947 * incremented again when calling __prep_account_new_huge_page()
2948 * and enqueue_hugetlb_folio() for new_folio. The counters will
2949 * remain stable since this happens under the lock.
2951 remove_hugetlb_folio(h, old_folio, false);
2954 * Ref count on new_folio is already zero as it was dropped
2955 * earlier. It can be directly added to the pool free list.
2957 __prep_account_new_huge_page(h, nid);
2958 enqueue_hugetlb_folio(h, new_folio);
2961 * Folio has been replaced, we can safely free the old one.
2963 spin_unlock_irq(&hugetlb_lock);
2964 update_and_free_hugetlb_folio(h, old_folio, false);
2970 spin_unlock_irq(&hugetlb_lock);
2971 /* Folio has a zero ref count, but needs a ref to be freed */
2972 folio_ref_unfreeze(new_folio, 1);
2973 update_and_free_hugetlb_folio(h, new_folio, false);
2978 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2981 struct folio *folio = page_folio(page);
2985 * The page might have been dissolved from under our feet, so make sure
2986 * to carefully check the state under the lock.
2987 * Return success when racing as if we dissolved the page ourselves.
2989 spin_lock_irq(&hugetlb_lock);
2990 if (folio_test_hugetlb(folio)) {
2991 h = folio_hstate(folio);
2993 spin_unlock_irq(&hugetlb_lock);
2996 spin_unlock_irq(&hugetlb_lock);
2999 * Fence off gigantic pages as there is a cyclic dependency between
3000 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3001 * of bailing out right away without further retrying.
3003 if (hstate_is_gigantic(h))
3006 if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list))
3008 else if (!folio_ref_count(folio))
3009 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3014 struct page *alloc_huge_page(struct vm_area_struct *vma,
3015 unsigned long addr, int avoid_reserve)
3017 struct hugepage_subpool *spool = subpool_vma(vma);
3018 struct hstate *h = hstate_vma(vma);
3020 struct folio *folio;
3021 long map_chg, map_commit;
3024 struct hugetlb_cgroup *h_cg;
3025 bool deferred_reserve;
3027 idx = hstate_index(h);
3029 * Examine the region/reserve map to determine if the process
3030 * has a reservation for the page to be allocated. A return
3031 * code of zero indicates a reservation exists (no change).
3033 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3035 return ERR_PTR(-ENOMEM);
3038 * Processes that did not create the mapping will have no
3039 * reserves as indicated by the region/reserve map. Check
3040 * that the allocation will not exceed the subpool limit.
3041 * Allocations for MAP_NORESERVE mappings also need to be
3042 * checked against any subpool limit.
3044 if (map_chg || avoid_reserve) {
3045 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3047 vma_end_reservation(h, vma, addr);
3048 return ERR_PTR(-ENOSPC);
3052 * Even though there was no reservation in the region/reserve
3053 * map, there could be reservations associated with the
3054 * subpool that can be used. This would be indicated if the
3055 * return value of hugepage_subpool_get_pages() is zero.
3056 * However, if avoid_reserve is specified we still avoid even
3057 * the subpool reservations.
3063 /* If this allocation is not consuming a reservation, charge it now.
3065 deferred_reserve = map_chg || avoid_reserve;
3066 if (deferred_reserve) {
3067 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3068 idx, pages_per_huge_page(h), &h_cg);
3070 goto out_subpool_put;
3073 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3075 goto out_uncharge_cgroup_reservation;
3077 spin_lock_irq(&hugetlb_lock);
3079 * glb_chg is passed to indicate whether or not a page must be taken
3080 * from the global free pool (global change). gbl_chg == 0 indicates
3081 * a reservation exists for the allocation.
3083 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
3085 spin_unlock_irq(&hugetlb_lock);
3086 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
3088 goto out_uncharge_cgroup;
3089 spin_lock_irq(&hugetlb_lock);
3090 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3091 SetHPageRestoreReserve(page);
3092 h->resv_huge_pages--;
3094 list_add(&page->lru, &h->hugepage_activelist);
3095 set_page_refcounted(page);
3098 folio = page_folio(page);
3099 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
3100 /* If allocation is not consuming a reservation, also store the
3101 * hugetlb_cgroup pointer on the page.
3103 if (deferred_reserve) {
3104 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3108 spin_unlock_irq(&hugetlb_lock);
3110 hugetlb_set_page_subpool(page, spool);
3112 map_commit = vma_commit_reservation(h, vma, addr);
3113 if (unlikely(map_chg > map_commit)) {
3115 * The page was added to the reservation map between
3116 * vma_needs_reservation and vma_commit_reservation.
3117 * This indicates a race with hugetlb_reserve_pages.
3118 * Adjust for the subpool count incremented above AND
3119 * in hugetlb_reserve_pages for the same page. Also,
3120 * the reservation count added in hugetlb_reserve_pages
3121 * no longer applies.
3125 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3126 hugetlb_acct_memory(h, -rsv_adjust);
3127 if (deferred_reserve)
3128 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3129 pages_per_huge_page(h), folio);
3133 out_uncharge_cgroup:
3134 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3135 out_uncharge_cgroup_reservation:
3136 if (deferred_reserve)
3137 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3140 if (map_chg || avoid_reserve)
3141 hugepage_subpool_put_pages(spool, 1);
3142 vma_end_reservation(h, vma, addr);
3143 return ERR_PTR(-ENOSPC);
3146 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3147 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3148 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3150 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3153 /* do node specific alloc */
3154 if (nid != NUMA_NO_NODE) {
3155 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3156 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3161 /* allocate from next node when distributing huge pages */
3162 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3163 m = memblock_alloc_try_nid_raw(
3164 huge_page_size(h), huge_page_size(h),
3165 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3167 * Use the beginning of the huge page to store the
3168 * huge_bootmem_page struct (until gather_bootmem
3169 * puts them into the mem_map).
3177 /* Put them into a private list first because mem_map is not up yet */
3178 INIT_LIST_HEAD(&m->list);
3179 list_add(&m->list, &huge_boot_pages);
3185 * Put bootmem huge pages into the standard lists after mem_map is up.
3186 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3188 static void __init gather_bootmem_prealloc(void)
3190 struct huge_bootmem_page *m;
3192 list_for_each_entry(m, &huge_boot_pages, list) {
3193 struct page *page = virt_to_page(m);
3194 struct folio *folio = page_folio(page);
3195 struct hstate *h = m->hstate;
3197 VM_BUG_ON(!hstate_is_gigantic(h));
3198 WARN_ON(folio_ref_count(folio) != 1);
3199 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3200 WARN_ON(folio_test_reserved(folio));
3201 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3202 free_huge_page(page); /* add to the hugepage allocator */
3204 /* VERY unlikely inflated ref count on a tail page */
3205 free_gigantic_folio(folio, huge_page_order(h));
3209 * We need to restore the 'stolen' pages to totalram_pages
3210 * in order to fix confusing memory reports from free(1) and
3211 * other side-effects, like CommitLimit going negative.
3213 adjust_managed_page_count(page, pages_per_huge_page(h));
3217 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3222 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3223 if (hstate_is_gigantic(h)) {
3224 if (!alloc_bootmem_huge_page(h, nid))
3227 struct folio *folio;
3228 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3230 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3231 &node_states[N_MEMORY], NULL);
3234 free_huge_page(&folio->page); /* free it into the hugepage allocator */
3238 if (i == h->max_huge_pages_node[nid])
3241 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3242 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3243 h->max_huge_pages_node[nid], buf, nid, i);
3244 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3245 h->max_huge_pages_node[nid] = i;
3248 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3251 nodemask_t *node_alloc_noretry;
3252 bool node_specific_alloc = false;
3254 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3255 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3256 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3260 /* do node specific alloc */
3261 for_each_online_node(i) {
3262 if (h->max_huge_pages_node[i] > 0) {
3263 hugetlb_hstate_alloc_pages_onenode(h, i);
3264 node_specific_alloc = true;
3268 if (node_specific_alloc)
3271 /* below will do all node balanced alloc */
3272 if (!hstate_is_gigantic(h)) {
3274 * Bit mask controlling how hard we retry per-node allocations.
3275 * Ignore errors as lower level routines can deal with
3276 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3277 * time, we are likely in bigger trouble.
3279 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3282 /* allocations done at boot time */
3283 node_alloc_noretry = NULL;
3286 /* bit mask controlling how hard we retry per-node allocations */
3287 if (node_alloc_noretry)
3288 nodes_clear(*node_alloc_noretry);
3290 for (i = 0; i < h->max_huge_pages; ++i) {
3291 if (hstate_is_gigantic(h)) {
3292 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3294 } else if (!alloc_pool_huge_page(h,
3295 &node_states[N_MEMORY],
3296 node_alloc_noretry))
3300 if (i < h->max_huge_pages) {
3303 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3304 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3305 h->max_huge_pages, buf, i);
3306 h->max_huge_pages = i;
3308 kfree(node_alloc_noretry);
3311 static void __init hugetlb_init_hstates(void)
3313 struct hstate *h, *h2;
3315 for_each_hstate(h) {
3316 /* oversize hugepages were init'ed in early boot */
3317 if (!hstate_is_gigantic(h))
3318 hugetlb_hstate_alloc_pages(h);
3321 * Set demote order for each hstate. Note that
3322 * h->demote_order is initially 0.
3323 * - We can not demote gigantic pages if runtime freeing
3324 * is not supported, so skip this.
3325 * - If CMA allocation is possible, we can not demote
3326 * HUGETLB_PAGE_ORDER or smaller size pages.
3328 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3330 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3332 for_each_hstate(h2) {
3335 if (h2->order < h->order &&
3336 h2->order > h->demote_order)
3337 h->demote_order = h2->order;
3342 static void __init report_hugepages(void)
3346 for_each_hstate(h) {
3349 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3350 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3351 buf, h->free_huge_pages);
3352 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3353 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3357 #ifdef CONFIG_HIGHMEM
3358 static void try_to_free_low(struct hstate *h, unsigned long count,
3359 nodemask_t *nodes_allowed)
3362 LIST_HEAD(page_list);
3364 lockdep_assert_held(&hugetlb_lock);
3365 if (hstate_is_gigantic(h))
3369 * Collect pages to be freed on a list, and free after dropping lock
3371 for_each_node_mask(i, *nodes_allowed) {
3372 struct page *page, *next;
3373 struct list_head *freel = &h->hugepage_freelists[i];
3374 list_for_each_entry_safe(page, next, freel, lru) {
3375 if (count >= h->nr_huge_pages)
3377 if (PageHighMem(page))
3379 remove_hugetlb_folio(h, page_folio(page), false);
3380 list_add(&page->lru, &page_list);
3385 spin_unlock_irq(&hugetlb_lock);
3386 update_and_free_pages_bulk(h, &page_list);
3387 spin_lock_irq(&hugetlb_lock);
3390 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3391 nodemask_t *nodes_allowed)
3397 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3398 * balanced by operating on them in a round-robin fashion.
3399 * Returns 1 if an adjustment was made.
3401 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3406 lockdep_assert_held(&hugetlb_lock);
3407 VM_BUG_ON(delta != -1 && delta != 1);
3410 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3411 if (h->surplus_huge_pages_node[node])
3415 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3416 if (h->surplus_huge_pages_node[node] <
3417 h->nr_huge_pages_node[node])
3424 h->surplus_huge_pages += delta;
3425 h->surplus_huge_pages_node[node] += delta;
3429 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3430 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3431 nodemask_t *nodes_allowed)
3433 unsigned long min_count, ret;
3435 LIST_HEAD(page_list);
3436 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3439 * Bit mask controlling how hard we retry per-node allocations.
3440 * If we can not allocate the bit mask, do not attempt to allocate
3441 * the requested huge pages.
3443 if (node_alloc_noretry)
3444 nodes_clear(*node_alloc_noretry);
3449 * resize_lock mutex prevents concurrent adjustments to number of
3450 * pages in hstate via the proc/sysfs interfaces.
3452 mutex_lock(&h->resize_lock);
3453 flush_free_hpage_work(h);
3454 spin_lock_irq(&hugetlb_lock);
3457 * Check for a node specific request.
3458 * Changing node specific huge page count may require a corresponding
3459 * change to the global count. In any case, the passed node mask
3460 * (nodes_allowed) will restrict alloc/free to the specified node.
3462 if (nid != NUMA_NO_NODE) {
3463 unsigned long old_count = count;
3465 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3467 * User may have specified a large count value which caused the
3468 * above calculation to overflow. In this case, they wanted
3469 * to allocate as many huge pages as possible. Set count to
3470 * largest possible value to align with their intention.
3472 if (count < old_count)
3477 * Gigantic pages runtime allocation depend on the capability for large
3478 * page range allocation.
3479 * If the system does not provide this feature, return an error when
3480 * the user tries to allocate gigantic pages but let the user free the
3481 * boottime allocated gigantic pages.
3483 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3484 if (count > persistent_huge_pages(h)) {
3485 spin_unlock_irq(&hugetlb_lock);
3486 mutex_unlock(&h->resize_lock);
3487 NODEMASK_FREE(node_alloc_noretry);
3490 /* Fall through to decrease pool */
3494 * Increase the pool size
3495 * First take pages out of surplus state. Then make up the
3496 * remaining difference by allocating fresh huge pages.
3498 * We might race with alloc_surplus_huge_page() here and be unable
3499 * to convert a surplus huge page to a normal huge page. That is
3500 * not critical, though, it just means the overall size of the
3501 * pool might be one hugepage larger than it needs to be, but
3502 * within all the constraints specified by the sysctls.
3504 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3505 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3509 while (count > persistent_huge_pages(h)) {
3511 * If this allocation races such that we no longer need the
3512 * page, free_huge_page will handle it by freeing the page
3513 * and reducing the surplus.
3515 spin_unlock_irq(&hugetlb_lock);
3517 /* yield cpu to avoid soft lockup */
3520 ret = alloc_pool_huge_page(h, nodes_allowed,
3521 node_alloc_noretry);
3522 spin_lock_irq(&hugetlb_lock);
3526 /* Bail for signals. Probably ctrl-c from user */
3527 if (signal_pending(current))
3532 * Decrease the pool size
3533 * First return free pages to the buddy allocator (being careful
3534 * to keep enough around to satisfy reservations). Then place
3535 * pages into surplus state as needed so the pool will shrink
3536 * to the desired size as pages become free.
3538 * By placing pages into the surplus state independent of the
3539 * overcommit value, we are allowing the surplus pool size to
3540 * exceed overcommit. There are few sane options here. Since
3541 * alloc_surplus_huge_page() is checking the global counter,
3542 * though, we'll note that we're not allowed to exceed surplus
3543 * and won't grow the pool anywhere else. Not until one of the
3544 * sysctls are changed, or the surplus pages go out of use.
3546 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3547 min_count = max(count, min_count);
3548 try_to_free_low(h, min_count, nodes_allowed);
3551 * Collect pages to be removed on list without dropping lock
3553 while (min_count < persistent_huge_pages(h)) {
3554 page = remove_pool_huge_page(h, nodes_allowed, 0);
3558 list_add(&page->lru, &page_list);
3560 /* free the pages after dropping lock */
3561 spin_unlock_irq(&hugetlb_lock);
3562 update_and_free_pages_bulk(h, &page_list);
3563 flush_free_hpage_work(h);
3564 spin_lock_irq(&hugetlb_lock);
3566 while (count < persistent_huge_pages(h)) {
3567 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3571 h->max_huge_pages = persistent_huge_pages(h);
3572 spin_unlock_irq(&hugetlb_lock);
3573 mutex_unlock(&h->resize_lock);
3575 NODEMASK_FREE(node_alloc_noretry);
3580 static int demote_free_huge_page(struct hstate *h, struct page *page)
3582 int i, nid = page_to_nid(page);
3583 struct hstate *target_hstate;
3584 struct folio *folio = page_folio(page);
3585 struct page *subpage;
3588 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3590 remove_hugetlb_folio_for_demote(h, folio, false);
3591 spin_unlock_irq(&hugetlb_lock);
3593 rc = hugetlb_vmemmap_restore(h, page);
3595 /* Allocation of vmemmmap failed, we can not demote page */
3596 spin_lock_irq(&hugetlb_lock);
3597 set_page_refcounted(page);
3598 add_hugetlb_folio(h, page_folio(page), false);
3603 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3604 * sizes as it will not ref count pages.
3606 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3609 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3610 * Without the mutex, pages added to target hstate could be marked
3613 * Note that we already hold h->resize_lock. To prevent deadlock,
3614 * use the convention of always taking larger size hstate mutex first.
3616 mutex_lock(&target_hstate->resize_lock);
3617 for (i = 0; i < pages_per_huge_page(h);
3618 i += pages_per_huge_page(target_hstate)) {
3619 subpage = nth_page(page, i);
3620 folio = page_folio(subpage);
3621 if (hstate_is_gigantic(target_hstate))
3622 prep_compound_gigantic_folio_for_demote(folio,
3623 target_hstate->order);
3625 prep_compound_page(subpage, target_hstate->order);
3626 set_page_private(subpage, 0);
3627 prep_new_hugetlb_folio(target_hstate, folio, nid);
3628 free_huge_page(subpage);
3630 mutex_unlock(&target_hstate->resize_lock);
3632 spin_lock_irq(&hugetlb_lock);
3635 * Not absolutely necessary, but for consistency update max_huge_pages
3636 * based on pool changes for the demoted page.
3638 h->max_huge_pages--;
3639 target_hstate->max_huge_pages +=
3640 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3645 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3646 __must_hold(&hugetlb_lock)
3651 lockdep_assert_held(&hugetlb_lock);
3653 /* We should never get here if no demote order */
3654 if (!h->demote_order) {
3655 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3656 return -EINVAL; /* internal error */
3659 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3660 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3661 if (PageHWPoison(page))
3664 return demote_free_huge_page(h, page);
3669 * Only way to get here is if all pages on free lists are poisoned.
3670 * Return -EBUSY so that caller will not retry.
3675 #define HSTATE_ATTR_RO(_name) \
3676 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3678 #define HSTATE_ATTR_WO(_name) \
3679 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3681 #define HSTATE_ATTR(_name) \
3682 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3684 static struct kobject *hugepages_kobj;
3685 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3687 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3689 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3693 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3694 if (hstate_kobjs[i] == kobj) {
3696 *nidp = NUMA_NO_NODE;
3700 return kobj_to_node_hstate(kobj, nidp);
3703 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3704 struct kobj_attribute *attr, char *buf)
3707 unsigned long nr_huge_pages;
3710 h = kobj_to_hstate(kobj, &nid);
3711 if (nid == NUMA_NO_NODE)
3712 nr_huge_pages = h->nr_huge_pages;
3714 nr_huge_pages = h->nr_huge_pages_node[nid];
3716 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3719 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3720 struct hstate *h, int nid,
3721 unsigned long count, size_t len)
3724 nodemask_t nodes_allowed, *n_mask;
3726 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3729 if (nid == NUMA_NO_NODE) {
3731 * global hstate attribute
3733 if (!(obey_mempolicy &&
3734 init_nodemask_of_mempolicy(&nodes_allowed)))
3735 n_mask = &node_states[N_MEMORY];
3737 n_mask = &nodes_allowed;
3740 * Node specific request. count adjustment happens in
3741 * set_max_huge_pages() after acquiring hugetlb_lock.
3743 init_nodemask_of_node(&nodes_allowed, nid);
3744 n_mask = &nodes_allowed;
3747 err = set_max_huge_pages(h, count, nid, n_mask);
3749 return err ? err : len;
3752 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3753 struct kobject *kobj, const char *buf,
3757 unsigned long count;
3761 err = kstrtoul(buf, 10, &count);
3765 h = kobj_to_hstate(kobj, &nid);
3766 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3769 static ssize_t nr_hugepages_show(struct kobject *kobj,
3770 struct kobj_attribute *attr, char *buf)
3772 return nr_hugepages_show_common(kobj, attr, buf);
3775 static ssize_t nr_hugepages_store(struct kobject *kobj,
3776 struct kobj_attribute *attr, const char *buf, size_t len)
3778 return nr_hugepages_store_common(false, kobj, buf, len);
3780 HSTATE_ATTR(nr_hugepages);
3785 * hstate attribute for optionally mempolicy-based constraint on persistent
3786 * huge page alloc/free.
3788 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3789 struct kobj_attribute *attr,
3792 return nr_hugepages_show_common(kobj, attr, buf);
3795 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3796 struct kobj_attribute *attr, const char *buf, size_t len)
3798 return nr_hugepages_store_common(true, kobj, buf, len);
3800 HSTATE_ATTR(nr_hugepages_mempolicy);
3804 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3805 struct kobj_attribute *attr, char *buf)
3807 struct hstate *h = kobj_to_hstate(kobj, NULL);
3808 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3811 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3812 struct kobj_attribute *attr, const char *buf, size_t count)
3815 unsigned long input;
3816 struct hstate *h = kobj_to_hstate(kobj, NULL);
3818 if (hstate_is_gigantic(h))
3821 err = kstrtoul(buf, 10, &input);
3825 spin_lock_irq(&hugetlb_lock);
3826 h->nr_overcommit_huge_pages = input;
3827 spin_unlock_irq(&hugetlb_lock);
3831 HSTATE_ATTR(nr_overcommit_hugepages);
3833 static ssize_t free_hugepages_show(struct kobject *kobj,
3834 struct kobj_attribute *attr, char *buf)
3837 unsigned long free_huge_pages;
3840 h = kobj_to_hstate(kobj, &nid);
3841 if (nid == NUMA_NO_NODE)
3842 free_huge_pages = h->free_huge_pages;
3844 free_huge_pages = h->free_huge_pages_node[nid];
3846 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3848 HSTATE_ATTR_RO(free_hugepages);
3850 static ssize_t resv_hugepages_show(struct kobject *kobj,
3851 struct kobj_attribute *attr, char *buf)
3853 struct hstate *h = kobj_to_hstate(kobj, NULL);
3854 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3856 HSTATE_ATTR_RO(resv_hugepages);
3858 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3859 struct kobj_attribute *attr, char *buf)
3862 unsigned long surplus_huge_pages;
3865 h = kobj_to_hstate(kobj, &nid);
3866 if (nid == NUMA_NO_NODE)
3867 surplus_huge_pages = h->surplus_huge_pages;
3869 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3871 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3873 HSTATE_ATTR_RO(surplus_hugepages);
3875 static ssize_t demote_store(struct kobject *kobj,
3876 struct kobj_attribute *attr, const char *buf, size_t len)
3878 unsigned long nr_demote;
3879 unsigned long nr_available;
3880 nodemask_t nodes_allowed, *n_mask;
3885 err = kstrtoul(buf, 10, &nr_demote);
3888 h = kobj_to_hstate(kobj, &nid);
3890 if (nid != NUMA_NO_NODE) {
3891 init_nodemask_of_node(&nodes_allowed, nid);
3892 n_mask = &nodes_allowed;
3894 n_mask = &node_states[N_MEMORY];
3897 /* Synchronize with other sysfs operations modifying huge pages */
3898 mutex_lock(&h->resize_lock);
3899 spin_lock_irq(&hugetlb_lock);
3903 * Check for available pages to demote each time thorough the
3904 * loop as demote_pool_huge_page will drop hugetlb_lock.
3906 if (nid != NUMA_NO_NODE)
3907 nr_available = h->free_huge_pages_node[nid];
3909 nr_available = h->free_huge_pages;
3910 nr_available -= h->resv_huge_pages;
3914 err = demote_pool_huge_page(h, n_mask);
3921 spin_unlock_irq(&hugetlb_lock);
3922 mutex_unlock(&h->resize_lock);
3928 HSTATE_ATTR_WO(demote);
3930 static ssize_t demote_size_show(struct kobject *kobj,
3931 struct kobj_attribute *attr, char *buf)
3933 struct hstate *h = kobj_to_hstate(kobj, NULL);
3934 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3936 return sysfs_emit(buf, "%lukB\n", demote_size);
3939 static ssize_t demote_size_store(struct kobject *kobj,
3940 struct kobj_attribute *attr,
3941 const char *buf, size_t count)
3943 struct hstate *h, *demote_hstate;
3944 unsigned long demote_size;
3945 unsigned int demote_order;
3947 demote_size = (unsigned long)memparse(buf, NULL);
3949 demote_hstate = size_to_hstate(demote_size);
3952 demote_order = demote_hstate->order;
3953 if (demote_order < HUGETLB_PAGE_ORDER)
3956 /* demote order must be smaller than hstate order */
3957 h = kobj_to_hstate(kobj, NULL);
3958 if (demote_order >= h->order)
3961 /* resize_lock synchronizes access to demote size and writes */
3962 mutex_lock(&h->resize_lock);
3963 h->demote_order = demote_order;
3964 mutex_unlock(&h->resize_lock);
3968 HSTATE_ATTR(demote_size);
3970 static struct attribute *hstate_attrs[] = {
3971 &nr_hugepages_attr.attr,
3972 &nr_overcommit_hugepages_attr.attr,
3973 &free_hugepages_attr.attr,
3974 &resv_hugepages_attr.attr,
3975 &surplus_hugepages_attr.attr,
3977 &nr_hugepages_mempolicy_attr.attr,
3982 static const struct attribute_group hstate_attr_group = {
3983 .attrs = hstate_attrs,
3986 static struct attribute *hstate_demote_attrs[] = {
3987 &demote_size_attr.attr,
3992 static const struct attribute_group hstate_demote_attr_group = {
3993 .attrs = hstate_demote_attrs,
3996 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3997 struct kobject **hstate_kobjs,
3998 const struct attribute_group *hstate_attr_group)
4001 int hi = hstate_index(h);
4003 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4004 if (!hstate_kobjs[hi])
4007 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4009 kobject_put(hstate_kobjs[hi]);
4010 hstate_kobjs[hi] = NULL;
4014 if (h->demote_order) {
4015 retval = sysfs_create_group(hstate_kobjs[hi],
4016 &hstate_demote_attr_group);
4018 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4019 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4020 kobject_put(hstate_kobjs[hi]);
4021 hstate_kobjs[hi] = NULL;
4030 static bool hugetlb_sysfs_initialized __ro_after_init;
4033 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4034 * with node devices in node_devices[] using a parallel array. The array
4035 * index of a node device or _hstate == node id.
4036 * This is here to avoid any static dependency of the node device driver, in
4037 * the base kernel, on the hugetlb module.
4039 struct node_hstate {
4040 struct kobject *hugepages_kobj;
4041 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4043 static struct node_hstate node_hstates[MAX_NUMNODES];
4046 * A subset of global hstate attributes for node devices
4048 static struct attribute *per_node_hstate_attrs[] = {
4049 &nr_hugepages_attr.attr,
4050 &free_hugepages_attr.attr,
4051 &surplus_hugepages_attr.attr,
4055 static const struct attribute_group per_node_hstate_attr_group = {
4056 .attrs = per_node_hstate_attrs,
4060 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4061 * Returns node id via non-NULL nidp.
4063 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4067 for (nid = 0; nid < nr_node_ids; nid++) {
4068 struct node_hstate *nhs = &node_hstates[nid];
4070 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4071 if (nhs->hstate_kobjs[i] == kobj) {
4083 * Unregister hstate attributes from a single node device.
4084 * No-op if no hstate attributes attached.
4086 void hugetlb_unregister_node(struct node *node)
4089 struct node_hstate *nhs = &node_hstates[node->dev.id];
4091 if (!nhs->hugepages_kobj)
4092 return; /* no hstate attributes */
4094 for_each_hstate(h) {
4095 int idx = hstate_index(h);
4096 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4100 if (h->demote_order)
4101 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4102 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4103 kobject_put(hstate_kobj);
4104 nhs->hstate_kobjs[idx] = NULL;
4107 kobject_put(nhs->hugepages_kobj);
4108 nhs->hugepages_kobj = NULL;
4113 * Register hstate attributes for a single node device.
4114 * No-op if attributes already registered.
4116 void hugetlb_register_node(struct node *node)
4119 struct node_hstate *nhs = &node_hstates[node->dev.id];
4122 if (!hugetlb_sysfs_initialized)
4125 if (nhs->hugepages_kobj)
4126 return; /* already allocated */
4128 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4130 if (!nhs->hugepages_kobj)
4133 for_each_hstate(h) {
4134 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4136 &per_node_hstate_attr_group);
4138 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4139 h->name, node->dev.id);
4140 hugetlb_unregister_node(node);
4147 * hugetlb init time: register hstate attributes for all registered node
4148 * devices of nodes that have memory. All on-line nodes should have
4149 * registered their associated device by this time.
4151 static void __init hugetlb_register_all_nodes(void)
4155 for_each_online_node(nid)
4156 hugetlb_register_node(node_devices[nid]);
4158 #else /* !CONFIG_NUMA */
4160 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4168 static void hugetlb_register_all_nodes(void) { }
4173 static void __init hugetlb_cma_check(void);
4175 static inline __init void hugetlb_cma_check(void)
4180 static void __init hugetlb_sysfs_init(void)
4185 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4186 if (!hugepages_kobj)
4189 for_each_hstate(h) {
4190 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4191 hstate_kobjs, &hstate_attr_group);
4193 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4197 hugetlb_sysfs_initialized = true;
4199 hugetlb_register_all_nodes();
4202 static int __init hugetlb_init(void)
4206 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4209 if (!hugepages_supported()) {
4210 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4211 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4216 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4217 * architectures depend on setup being done here.
4219 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4220 if (!parsed_default_hugepagesz) {
4222 * If we did not parse a default huge page size, set
4223 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4224 * number of huge pages for this default size was implicitly
4225 * specified, set that here as well.
4226 * Note that the implicit setting will overwrite an explicit
4227 * setting. A warning will be printed in this case.
4229 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4230 if (default_hstate_max_huge_pages) {
4231 if (default_hstate.max_huge_pages) {
4234 string_get_size(huge_page_size(&default_hstate),
4235 1, STRING_UNITS_2, buf, 32);
4236 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4237 default_hstate.max_huge_pages, buf);
4238 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4239 default_hstate_max_huge_pages);
4241 default_hstate.max_huge_pages =
4242 default_hstate_max_huge_pages;
4244 for_each_online_node(i)
4245 default_hstate.max_huge_pages_node[i] =
4246 default_hugepages_in_node[i];
4250 hugetlb_cma_check();
4251 hugetlb_init_hstates();
4252 gather_bootmem_prealloc();
4255 hugetlb_sysfs_init();
4256 hugetlb_cgroup_file_init();
4259 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4261 num_fault_mutexes = 1;
4263 hugetlb_fault_mutex_table =
4264 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4266 BUG_ON(!hugetlb_fault_mutex_table);
4268 for (i = 0; i < num_fault_mutexes; i++)
4269 mutex_init(&hugetlb_fault_mutex_table[i]);
4272 subsys_initcall(hugetlb_init);
4274 /* Overwritten by architectures with more huge page sizes */
4275 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4277 return size == HPAGE_SIZE;
4280 void __init hugetlb_add_hstate(unsigned int order)
4285 if (size_to_hstate(PAGE_SIZE << order)) {
4288 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4290 h = &hstates[hugetlb_max_hstate++];
4291 mutex_init(&h->resize_lock);
4293 h->mask = ~(huge_page_size(h) - 1);
4294 for (i = 0; i < MAX_NUMNODES; ++i)
4295 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4296 INIT_LIST_HEAD(&h->hugepage_activelist);
4297 h->next_nid_to_alloc = first_memory_node;
4298 h->next_nid_to_free = first_memory_node;
4299 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4300 huge_page_size(h)/SZ_1K);
4305 bool __init __weak hugetlb_node_alloc_supported(void)
4310 static void __init hugepages_clear_pages_in_node(void)
4312 if (!hugetlb_max_hstate) {
4313 default_hstate_max_huge_pages = 0;
4314 memset(default_hugepages_in_node, 0,
4315 sizeof(default_hugepages_in_node));
4317 parsed_hstate->max_huge_pages = 0;
4318 memset(parsed_hstate->max_huge_pages_node, 0,
4319 sizeof(parsed_hstate->max_huge_pages_node));
4324 * hugepages command line processing
4325 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4326 * specification. If not, ignore the hugepages value. hugepages can also
4327 * be the first huge page command line option in which case it implicitly
4328 * specifies the number of huge pages for the default size.
4330 static int __init hugepages_setup(char *s)
4333 static unsigned long *last_mhp;
4334 int node = NUMA_NO_NODE;
4339 if (!parsed_valid_hugepagesz) {
4340 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4341 parsed_valid_hugepagesz = true;
4346 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4347 * yet, so this hugepages= parameter goes to the "default hstate".
4348 * Otherwise, it goes with the previously parsed hugepagesz or
4349 * default_hugepagesz.
4351 else if (!hugetlb_max_hstate)
4352 mhp = &default_hstate_max_huge_pages;
4354 mhp = &parsed_hstate->max_huge_pages;
4356 if (mhp == last_mhp) {
4357 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4363 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4365 /* Parameter is node format */
4366 if (p[count] == ':') {
4367 if (!hugetlb_node_alloc_supported()) {
4368 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4371 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4373 node = array_index_nospec(tmp, MAX_NUMNODES);
4375 /* Parse hugepages */
4376 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4378 if (!hugetlb_max_hstate)
4379 default_hugepages_in_node[node] = tmp;
4381 parsed_hstate->max_huge_pages_node[node] = tmp;
4383 /* Go to parse next node*/
4384 if (p[count] == ',')
4397 * Global state is always initialized later in hugetlb_init.
4398 * But we need to allocate gigantic hstates here early to still
4399 * use the bootmem allocator.
4401 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4402 hugetlb_hstate_alloc_pages(parsed_hstate);
4409 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4410 hugepages_clear_pages_in_node();
4413 __setup("hugepages=", hugepages_setup);
4416 * hugepagesz command line processing
4417 * A specific huge page size can only be specified once with hugepagesz.
4418 * hugepagesz is followed by hugepages on the command line. The global
4419 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4420 * hugepagesz argument was valid.
4422 static int __init hugepagesz_setup(char *s)
4427 parsed_valid_hugepagesz = false;
4428 size = (unsigned long)memparse(s, NULL);
4430 if (!arch_hugetlb_valid_size(size)) {
4431 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4435 h = size_to_hstate(size);
4438 * hstate for this size already exists. This is normally
4439 * an error, but is allowed if the existing hstate is the
4440 * default hstate. More specifically, it is only allowed if
4441 * the number of huge pages for the default hstate was not
4442 * previously specified.
4444 if (!parsed_default_hugepagesz || h != &default_hstate ||
4445 default_hstate.max_huge_pages) {
4446 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4451 * No need to call hugetlb_add_hstate() as hstate already
4452 * exists. But, do set parsed_hstate so that a following
4453 * hugepages= parameter will be applied to this hstate.
4456 parsed_valid_hugepagesz = true;
4460 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4461 parsed_valid_hugepagesz = true;
4464 __setup("hugepagesz=", hugepagesz_setup);
4467 * default_hugepagesz command line input
4468 * Only one instance of default_hugepagesz allowed on command line.
4470 static int __init default_hugepagesz_setup(char *s)
4475 parsed_valid_hugepagesz = false;
4476 if (parsed_default_hugepagesz) {
4477 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4481 size = (unsigned long)memparse(s, NULL);
4483 if (!arch_hugetlb_valid_size(size)) {
4484 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4488 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4489 parsed_valid_hugepagesz = true;
4490 parsed_default_hugepagesz = true;
4491 default_hstate_idx = hstate_index(size_to_hstate(size));
4494 * The number of default huge pages (for this size) could have been
4495 * specified as the first hugetlb parameter: hugepages=X. If so,
4496 * then default_hstate_max_huge_pages is set. If the default huge
4497 * page size is gigantic (>= MAX_ORDER), then the pages must be
4498 * allocated here from bootmem allocator.
4500 if (default_hstate_max_huge_pages) {
4501 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4502 for_each_online_node(i)
4503 default_hstate.max_huge_pages_node[i] =
4504 default_hugepages_in_node[i];
4505 if (hstate_is_gigantic(&default_hstate))
4506 hugetlb_hstate_alloc_pages(&default_hstate);
4507 default_hstate_max_huge_pages = 0;
4512 __setup("default_hugepagesz=", default_hugepagesz_setup);
4514 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4517 struct mempolicy *mpol = get_task_policy(current);
4520 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4521 * (from policy_nodemask) specifically for hugetlb case
4523 if (mpol->mode == MPOL_BIND &&
4524 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4525 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4526 return &mpol->nodes;
4531 static unsigned int allowed_mems_nr(struct hstate *h)
4534 unsigned int nr = 0;
4535 nodemask_t *mbind_nodemask;
4536 unsigned int *array = h->free_huge_pages_node;
4537 gfp_t gfp_mask = htlb_alloc_mask(h);
4539 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4540 for_each_node_mask(node, cpuset_current_mems_allowed) {
4541 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4548 #ifdef CONFIG_SYSCTL
4549 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4550 void *buffer, size_t *length,
4551 loff_t *ppos, unsigned long *out)
4553 struct ctl_table dup_table;
4556 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4557 * can duplicate the @table and alter the duplicate of it.
4560 dup_table.data = out;
4562 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4565 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4566 struct ctl_table *table, int write,
4567 void *buffer, size_t *length, loff_t *ppos)
4569 struct hstate *h = &default_hstate;
4570 unsigned long tmp = h->max_huge_pages;
4573 if (!hugepages_supported())
4576 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4582 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4583 NUMA_NO_NODE, tmp, *length);
4588 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4589 void *buffer, size_t *length, loff_t *ppos)
4592 return hugetlb_sysctl_handler_common(false, table, write,
4593 buffer, length, ppos);
4597 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4598 void *buffer, size_t *length, loff_t *ppos)
4600 return hugetlb_sysctl_handler_common(true, table, write,
4601 buffer, length, ppos);
4603 #endif /* CONFIG_NUMA */
4605 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4606 void *buffer, size_t *length, loff_t *ppos)
4608 struct hstate *h = &default_hstate;
4612 if (!hugepages_supported())
4615 tmp = h->nr_overcommit_huge_pages;
4617 if (write && hstate_is_gigantic(h))
4620 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4626 spin_lock_irq(&hugetlb_lock);
4627 h->nr_overcommit_huge_pages = tmp;
4628 spin_unlock_irq(&hugetlb_lock);
4634 #endif /* CONFIG_SYSCTL */
4636 void hugetlb_report_meminfo(struct seq_file *m)
4639 unsigned long total = 0;
4641 if (!hugepages_supported())
4644 for_each_hstate(h) {
4645 unsigned long count = h->nr_huge_pages;
4647 total += huge_page_size(h) * count;
4649 if (h == &default_hstate)
4651 "HugePages_Total: %5lu\n"
4652 "HugePages_Free: %5lu\n"
4653 "HugePages_Rsvd: %5lu\n"
4654 "HugePages_Surp: %5lu\n"
4655 "Hugepagesize: %8lu kB\n",
4659 h->surplus_huge_pages,
4660 huge_page_size(h) / SZ_1K);
4663 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4666 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4668 struct hstate *h = &default_hstate;
4670 if (!hugepages_supported())
4673 return sysfs_emit_at(buf, len,
4674 "Node %d HugePages_Total: %5u\n"
4675 "Node %d HugePages_Free: %5u\n"
4676 "Node %d HugePages_Surp: %5u\n",
4677 nid, h->nr_huge_pages_node[nid],
4678 nid, h->free_huge_pages_node[nid],
4679 nid, h->surplus_huge_pages_node[nid]);
4682 void hugetlb_show_meminfo_node(int nid)
4686 if (!hugepages_supported())
4690 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4692 h->nr_huge_pages_node[nid],
4693 h->free_huge_pages_node[nid],
4694 h->surplus_huge_pages_node[nid],
4695 huge_page_size(h) / SZ_1K);
4698 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4700 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4701 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4704 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4705 unsigned long hugetlb_total_pages(void)
4708 unsigned long nr_total_pages = 0;
4711 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4712 return nr_total_pages;
4715 static int hugetlb_acct_memory(struct hstate *h, long delta)
4722 spin_lock_irq(&hugetlb_lock);
4724 * When cpuset is configured, it breaks the strict hugetlb page
4725 * reservation as the accounting is done on a global variable. Such
4726 * reservation is completely rubbish in the presence of cpuset because
4727 * the reservation is not checked against page availability for the
4728 * current cpuset. Application can still potentially OOM'ed by kernel
4729 * with lack of free htlb page in cpuset that the task is in.
4730 * Attempt to enforce strict accounting with cpuset is almost
4731 * impossible (or too ugly) because cpuset is too fluid that
4732 * task or memory node can be dynamically moved between cpusets.
4734 * The change of semantics for shared hugetlb mapping with cpuset is
4735 * undesirable. However, in order to preserve some of the semantics,
4736 * we fall back to check against current free page availability as
4737 * a best attempt and hopefully to minimize the impact of changing
4738 * semantics that cpuset has.
4740 * Apart from cpuset, we also have memory policy mechanism that
4741 * also determines from which node the kernel will allocate memory
4742 * in a NUMA system. So similar to cpuset, we also should consider
4743 * the memory policy of the current task. Similar to the description
4747 if (gather_surplus_pages(h, delta) < 0)
4750 if (delta > allowed_mems_nr(h)) {
4751 return_unused_surplus_pages(h, delta);
4758 return_unused_surplus_pages(h, (unsigned long) -delta);
4761 spin_unlock_irq(&hugetlb_lock);
4765 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4767 struct resv_map *resv = vma_resv_map(vma);
4770 * HPAGE_RESV_OWNER indicates a private mapping.
4771 * This new VMA should share its siblings reservation map if present.
4772 * The VMA will only ever have a valid reservation map pointer where
4773 * it is being copied for another still existing VMA. As that VMA
4774 * has a reference to the reservation map it cannot disappear until
4775 * after this open call completes. It is therefore safe to take a
4776 * new reference here without additional locking.
4778 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4779 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4780 kref_get(&resv->refs);
4784 * vma_lock structure for sharable mappings is vma specific.
4785 * Clear old pointer (if copied via vm_area_dup) and allocate
4786 * new structure. Before clearing, make sure vma_lock is not
4789 if (vma->vm_flags & VM_MAYSHARE) {
4790 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4793 if (vma_lock->vma != vma) {
4794 vma->vm_private_data = NULL;
4795 hugetlb_vma_lock_alloc(vma);
4797 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4799 hugetlb_vma_lock_alloc(vma);
4803 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4805 struct hstate *h = hstate_vma(vma);
4806 struct resv_map *resv;
4807 struct hugepage_subpool *spool = subpool_vma(vma);
4808 unsigned long reserve, start, end;
4811 hugetlb_vma_lock_free(vma);
4813 resv = vma_resv_map(vma);
4814 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4817 start = vma_hugecache_offset(h, vma, vma->vm_start);
4818 end = vma_hugecache_offset(h, vma, vma->vm_end);
4820 reserve = (end - start) - region_count(resv, start, end);
4821 hugetlb_cgroup_uncharge_counter(resv, start, end);
4824 * Decrement reserve counts. The global reserve count may be
4825 * adjusted if the subpool has a minimum size.
4827 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4828 hugetlb_acct_memory(h, -gbl_reserve);
4831 kref_put(&resv->refs, resv_map_release);
4834 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4836 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4840 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4841 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4842 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4844 if (addr & ~PUD_MASK) {
4846 * hugetlb_vm_op_split is called right before we attempt to
4847 * split the VMA. We will need to unshare PMDs in the old and
4848 * new VMAs, so let's unshare before we split.
4850 unsigned long floor = addr & PUD_MASK;
4851 unsigned long ceil = floor + PUD_SIZE;
4853 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4854 hugetlb_unshare_pmds(vma, floor, ceil);
4860 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4862 return huge_page_size(hstate_vma(vma));
4866 * We cannot handle pagefaults against hugetlb pages at all. They cause
4867 * handle_mm_fault() to try to instantiate regular-sized pages in the
4868 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4871 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4878 * When a new function is introduced to vm_operations_struct and added
4879 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4880 * This is because under System V memory model, mappings created via
4881 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4882 * their original vm_ops are overwritten with shm_vm_ops.
4884 const struct vm_operations_struct hugetlb_vm_ops = {
4885 .fault = hugetlb_vm_op_fault,
4886 .open = hugetlb_vm_op_open,
4887 .close = hugetlb_vm_op_close,
4888 .may_split = hugetlb_vm_op_split,
4889 .pagesize = hugetlb_vm_op_pagesize,
4892 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4896 unsigned int shift = huge_page_shift(hstate_vma(vma));
4899 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4900 vma->vm_page_prot)));
4902 entry = huge_pte_wrprotect(mk_huge_pte(page,
4903 vma->vm_page_prot));
4905 entry = pte_mkyoung(entry);
4906 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4911 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4912 unsigned long address, pte_t *ptep)
4916 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4917 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4918 update_mmu_cache(vma, address, ptep);
4921 bool is_hugetlb_entry_migration(pte_t pte)
4925 if (huge_pte_none(pte) || pte_present(pte))
4927 swp = pte_to_swp_entry(pte);
4928 if (is_migration_entry(swp))
4934 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4938 if (huge_pte_none(pte) || pte_present(pte))
4940 swp = pte_to_swp_entry(pte);
4941 if (is_hwpoison_entry(swp))
4948 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4949 struct page *new_page)
4951 __SetPageUptodate(new_page);
4952 hugepage_add_new_anon_rmap(new_page, vma, addr);
4953 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4954 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4955 SetHPageMigratable(new_page);
4958 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4959 struct vm_area_struct *dst_vma,
4960 struct vm_area_struct *src_vma)
4962 pte_t *src_pte, *dst_pte, entry;
4963 struct page *ptepage;
4965 bool cow = is_cow_mapping(src_vma->vm_flags);
4966 struct hstate *h = hstate_vma(src_vma);
4967 unsigned long sz = huge_page_size(h);
4968 unsigned long npages = pages_per_huge_page(h);
4969 struct mmu_notifier_range range;
4970 unsigned long last_addr_mask;
4974 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4977 mmu_notifier_invalidate_range_start(&range);
4978 mmap_assert_write_locked(src);
4979 raw_write_seqcount_begin(&src->write_protect_seq);
4982 * For shared mappings the vma lock must be held before
4983 * calling huge_pte_offset in the src vma. Otherwise, the
4984 * returned ptep could go away if part of a shared pmd and
4985 * another thread calls huge_pmd_unshare.
4987 hugetlb_vma_lock_read(src_vma);
4990 last_addr_mask = hugetlb_mask_last_page(h);
4991 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4992 spinlock_t *src_ptl, *dst_ptl;
4993 src_pte = huge_pte_offset(src, addr, sz);
4995 addr |= last_addr_mask;
4998 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5005 * If the pagetables are shared don't copy or take references.
5007 * dst_pte == src_pte is the common case of src/dest sharing.
5008 * However, src could have 'unshared' and dst shares with
5009 * another vma. So page_count of ptep page is checked instead
5010 * to reliably determine whether pte is shared.
5012 if (page_count(virt_to_page(dst_pte)) > 1) {
5013 addr |= last_addr_mask;
5017 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5018 src_ptl = huge_pte_lockptr(h, src, src_pte);
5019 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5020 entry = huge_ptep_get(src_pte);
5022 if (huge_pte_none(entry)) {
5024 * Skip if src entry none.
5027 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5028 bool uffd_wp = huge_pte_uffd_wp(entry);
5030 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5031 entry = huge_pte_clear_uffd_wp(entry);
5032 set_huge_pte_at(dst, addr, dst_pte, entry);
5033 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5034 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5035 bool uffd_wp = huge_pte_uffd_wp(entry);
5037 if (!is_readable_migration_entry(swp_entry) && cow) {
5039 * COW mappings require pages in both
5040 * parent and child to be set to read.
5042 swp_entry = make_readable_migration_entry(
5043 swp_offset(swp_entry));
5044 entry = swp_entry_to_pte(swp_entry);
5045 if (userfaultfd_wp(src_vma) && uffd_wp)
5046 entry = huge_pte_mkuffd_wp(entry);
5047 set_huge_pte_at(src, addr, src_pte, entry);
5049 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5050 entry = huge_pte_clear_uffd_wp(entry);
5051 set_huge_pte_at(dst, addr, dst_pte, entry);
5052 } else if (unlikely(is_pte_marker(entry))) {
5053 /* No swap on hugetlb */
5055 is_swapin_error_entry(pte_to_swp_entry(entry)));
5057 * We copy the pte marker only if the dst vma has
5060 if (userfaultfd_wp(dst_vma))
5061 set_huge_pte_at(dst, addr, dst_pte, entry);
5063 entry = huge_ptep_get(src_pte);
5064 ptepage = pte_page(entry);
5068 * Failing to duplicate the anon rmap is a rare case
5069 * where we see pinned hugetlb pages while they're
5070 * prone to COW. We need to do the COW earlier during
5073 * When pre-allocating the page or copying data, we
5074 * need to be without the pgtable locks since we could
5075 * sleep during the process.
5077 if (!PageAnon(ptepage)) {
5078 page_dup_file_rmap(ptepage, true);
5079 } else if (page_try_dup_anon_rmap(ptepage, true,
5081 pte_t src_pte_old = entry;
5084 spin_unlock(src_ptl);
5085 spin_unlock(dst_ptl);
5086 /* Do not use reserve as it's private owned */
5087 new = alloc_huge_page(dst_vma, addr, 1);
5093 copy_user_huge_page(new, ptepage, addr, dst_vma,
5097 /* Install the new huge page if src pte stable */
5098 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5099 src_ptl = huge_pte_lockptr(h, src, src_pte);
5100 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5101 entry = huge_ptep_get(src_pte);
5102 if (!pte_same(src_pte_old, entry)) {
5103 restore_reserve_on_error(h, dst_vma, addr,
5106 /* huge_ptep of dst_pte won't change as in child */
5109 hugetlb_install_page(dst_vma, dst_pte, addr, new);
5110 spin_unlock(src_ptl);
5111 spin_unlock(dst_ptl);
5117 * No need to notify as we are downgrading page
5118 * table protection not changing it to point
5121 * See Documentation/mm/mmu_notifier.rst
5123 huge_ptep_set_wrprotect(src, addr, src_pte);
5124 entry = huge_pte_wrprotect(entry);
5127 set_huge_pte_at(dst, addr, dst_pte, entry);
5128 hugetlb_count_add(npages, dst);
5130 spin_unlock(src_ptl);
5131 spin_unlock(dst_ptl);
5135 raw_write_seqcount_end(&src->write_protect_seq);
5136 mmu_notifier_invalidate_range_end(&range);
5138 hugetlb_vma_unlock_read(src_vma);
5144 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5145 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5147 struct hstate *h = hstate_vma(vma);
5148 struct mm_struct *mm = vma->vm_mm;
5149 spinlock_t *src_ptl, *dst_ptl;
5152 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5153 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5156 * We don't have to worry about the ordering of src and dst ptlocks
5157 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5159 if (src_ptl != dst_ptl)
5160 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5162 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5163 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5165 if (src_ptl != dst_ptl)
5166 spin_unlock(src_ptl);
5167 spin_unlock(dst_ptl);
5170 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5171 struct vm_area_struct *new_vma,
5172 unsigned long old_addr, unsigned long new_addr,
5175 struct hstate *h = hstate_vma(vma);
5176 struct address_space *mapping = vma->vm_file->f_mapping;
5177 unsigned long sz = huge_page_size(h);
5178 struct mm_struct *mm = vma->vm_mm;
5179 unsigned long old_end = old_addr + len;
5180 unsigned long last_addr_mask;
5181 pte_t *src_pte, *dst_pte;
5182 struct mmu_notifier_range range;
5183 bool shared_pmd = false;
5185 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
5187 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5189 * In case of shared PMDs, we should cover the maximum possible
5192 flush_cache_range(vma, range.start, range.end);
5194 mmu_notifier_invalidate_range_start(&range);
5195 last_addr_mask = hugetlb_mask_last_page(h);
5196 /* Prevent race with file truncation */
5197 hugetlb_vma_lock_write(vma);
5198 i_mmap_lock_write(mapping);
5199 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5200 src_pte = huge_pte_offset(mm, old_addr, sz);
5202 old_addr |= last_addr_mask;
5203 new_addr |= last_addr_mask;
5206 if (huge_pte_none(huge_ptep_get(src_pte)))
5209 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5211 old_addr |= last_addr_mask;
5212 new_addr |= last_addr_mask;
5216 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5220 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5224 flush_tlb_range(vma, range.start, range.end);
5226 flush_tlb_range(vma, old_end - len, old_end);
5227 mmu_notifier_invalidate_range_end(&range);
5228 i_mmap_unlock_write(mapping);
5229 hugetlb_vma_unlock_write(vma);
5231 return len + old_addr - old_end;
5234 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5235 unsigned long start, unsigned long end,
5236 struct page *ref_page, zap_flags_t zap_flags)
5238 struct mm_struct *mm = vma->vm_mm;
5239 unsigned long address;
5244 struct hstate *h = hstate_vma(vma);
5245 unsigned long sz = huge_page_size(h);
5246 unsigned long last_addr_mask;
5247 bool force_flush = false;
5249 WARN_ON(!is_vm_hugetlb_page(vma));
5250 BUG_ON(start & ~huge_page_mask(h));
5251 BUG_ON(end & ~huge_page_mask(h));
5254 * This is a hugetlb vma, all the pte entries should point
5257 tlb_change_page_size(tlb, sz);
5258 tlb_start_vma(tlb, vma);
5260 last_addr_mask = hugetlb_mask_last_page(h);
5262 for (; address < end; address += sz) {
5263 ptep = huge_pte_offset(mm, address, sz);
5265 address |= last_addr_mask;
5269 ptl = huge_pte_lock(h, mm, ptep);
5270 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5272 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5274 address |= last_addr_mask;
5278 pte = huge_ptep_get(ptep);
5279 if (huge_pte_none(pte)) {
5285 * Migrating hugepage or HWPoisoned hugepage is already
5286 * unmapped and its refcount is dropped, so just clear pte here.
5288 if (unlikely(!pte_present(pte))) {
5290 * If the pte was wr-protected by uffd-wp in any of the
5291 * swap forms, meanwhile the caller does not want to
5292 * drop the uffd-wp bit in this zap, then replace the
5293 * pte with a marker.
5295 if (pte_swp_uffd_wp_any(pte) &&
5296 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5297 set_huge_pte_at(mm, address, ptep,
5298 make_pte_marker(PTE_MARKER_UFFD_WP));
5300 huge_pte_clear(mm, address, ptep, sz);
5305 page = pte_page(pte);
5307 * If a reference page is supplied, it is because a specific
5308 * page is being unmapped, not a range. Ensure the page we
5309 * are about to unmap is the actual page of interest.
5312 if (page != ref_page) {
5317 * Mark the VMA as having unmapped its page so that
5318 * future faults in this VMA will fail rather than
5319 * looking like data was lost
5321 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5324 pte = huge_ptep_get_and_clear(mm, address, ptep);
5325 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5326 if (huge_pte_dirty(pte))
5327 set_page_dirty(page);
5328 /* Leave a uffd-wp pte marker if needed */
5329 if (huge_pte_uffd_wp(pte) &&
5330 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5331 set_huge_pte_at(mm, address, ptep,
5332 make_pte_marker(PTE_MARKER_UFFD_WP));
5333 hugetlb_count_sub(pages_per_huge_page(h), mm);
5334 page_remove_rmap(page, vma, true);
5337 tlb_remove_page_size(tlb, page, huge_page_size(h));
5339 * Bail out after unmapping reference page if supplied
5344 tlb_end_vma(tlb, vma);
5347 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5348 * could defer the flush until now, since by holding i_mmap_rwsem we
5349 * guaranteed that the last refernece would not be dropped. But we must
5350 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5351 * dropped and the last reference to the shared PMDs page might be
5354 * In theory we could defer the freeing of the PMD pages as well, but
5355 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5356 * detect sharing, so we cannot defer the release of the page either.
5357 * Instead, do flush now.
5360 tlb_flush_mmu_tlbonly(tlb);
5363 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5364 struct vm_area_struct *vma, unsigned long start,
5365 unsigned long end, struct page *ref_page,
5366 zap_flags_t zap_flags)
5368 hugetlb_vma_lock_write(vma);
5369 i_mmap_lock_write(vma->vm_file->f_mapping);
5371 /* mmu notification performed in caller */
5372 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5374 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5376 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5377 * When the vma_lock is freed, this makes the vma ineligible
5378 * for pmd sharing. And, i_mmap_rwsem is required to set up
5379 * pmd sharing. This is important as page tables for this
5380 * unmapped range will be asynchrously deleted. If the page
5381 * tables are shared, there will be issues when accessed by
5384 __hugetlb_vma_unlock_write_free(vma);
5385 i_mmap_unlock_write(vma->vm_file->f_mapping);
5387 i_mmap_unlock_write(vma->vm_file->f_mapping);
5388 hugetlb_vma_unlock_write(vma);
5392 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5393 unsigned long end, struct page *ref_page,
5394 zap_flags_t zap_flags)
5396 struct mmu_notifier_range range;
5397 struct mmu_gather tlb;
5399 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
5401 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5402 mmu_notifier_invalidate_range_start(&range);
5403 tlb_gather_mmu(&tlb, vma->vm_mm);
5405 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5407 mmu_notifier_invalidate_range_end(&range);
5408 tlb_finish_mmu(&tlb);
5412 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5413 * mapping it owns the reserve page for. The intention is to unmap the page
5414 * from other VMAs and let the children be SIGKILLed if they are faulting the
5417 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5418 struct page *page, unsigned long address)
5420 struct hstate *h = hstate_vma(vma);
5421 struct vm_area_struct *iter_vma;
5422 struct address_space *mapping;
5426 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5427 * from page cache lookup which is in HPAGE_SIZE units.
5429 address = address & huge_page_mask(h);
5430 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5432 mapping = vma->vm_file->f_mapping;
5435 * Take the mapping lock for the duration of the table walk. As
5436 * this mapping should be shared between all the VMAs,
5437 * __unmap_hugepage_range() is called as the lock is already held
5439 i_mmap_lock_write(mapping);
5440 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5441 /* Do not unmap the current VMA */
5442 if (iter_vma == vma)
5446 * Shared VMAs have their own reserves and do not affect
5447 * MAP_PRIVATE accounting but it is possible that a shared
5448 * VMA is using the same page so check and skip such VMAs.
5450 if (iter_vma->vm_flags & VM_MAYSHARE)
5454 * Unmap the page from other VMAs without their own reserves.
5455 * They get marked to be SIGKILLed if they fault in these
5456 * areas. This is because a future no-page fault on this VMA
5457 * could insert a zeroed page instead of the data existing
5458 * from the time of fork. This would look like data corruption
5460 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5461 unmap_hugepage_range(iter_vma, address,
5462 address + huge_page_size(h), page, 0);
5464 i_mmap_unlock_write(mapping);
5468 * hugetlb_wp() should be called with page lock of the original hugepage held.
5469 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5470 * cannot race with other handlers or page migration.
5471 * Keep the pte_same checks anyway to make transition from the mutex easier.
5473 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5474 unsigned long address, pte_t *ptep, unsigned int flags,
5475 struct page *pagecache_page, spinlock_t *ptl)
5477 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5479 struct hstate *h = hstate_vma(vma);
5480 struct page *old_page, *new_page;
5481 int outside_reserve = 0;
5483 unsigned long haddr = address & huge_page_mask(h);
5484 struct mmu_notifier_range range;
5487 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5488 * PTE mapped R/O such as maybe_mkwrite() would do.
5490 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5491 return VM_FAULT_SIGSEGV;
5493 /* Let's take out MAP_SHARED mappings first. */
5494 if (vma->vm_flags & VM_MAYSHARE) {
5495 set_huge_ptep_writable(vma, haddr, ptep);
5499 pte = huge_ptep_get(ptep);
5500 old_page = pte_page(pte);
5502 delayacct_wpcopy_start();
5506 * If no-one else is actually using this page, we're the exclusive
5507 * owner and can reuse this page.
5509 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5510 if (!PageAnonExclusive(old_page))
5511 page_move_anon_rmap(old_page, vma);
5512 if (likely(!unshare))
5513 set_huge_ptep_writable(vma, haddr, ptep);
5515 delayacct_wpcopy_end();
5518 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5522 * If the process that created a MAP_PRIVATE mapping is about to
5523 * perform a COW due to a shared page count, attempt to satisfy
5524 * the allocation without using the existing reserves. The pagecache
5525 * page is used to determine if the reserve at this address was
5526 * consumed or not. If reserves were used, a partial faulted mapping
5527 * at the time of fork() could consume its reserves on COW instead
5528 * of the full address range.
5530 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5531 old_page != pagecache_page)
5532 outside_reserve = 1;
5537 * Drop page table lock as buddy allocator may be called. It will
5538 * be acquired again before returning to the caller, as expected.
5541 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5543 if (IS_ERR(new_page)) {
5545 * If a process owning a MAP_PRIVATE mapping fails to COW,
5546 * it is due to references held by a child and an insufficient
5547 * huge page pool. To guarantee the original mappers
5548 * reliability, unmap the page from child processes. The child
5549 * may get SIGKILLed if it later faults.
5551 if (outside_reserve) {
5552 struct address_space *mapping = vma->vm_file->f_mapping;
5558 * Drop hugetlb_fault_mutex and vma_lock before
5559 * unmapping. unmapping needs to hold vma_lock
5560 * in write mode. Dropping vma_lock in read mode
5561 * here is OK as COW mappings do not interact with
5564 * Reacquire both after unmap operation.
5566 idx = vma_hugecache_offset(h, vma, haddr);
5567 hash = hugetlb_fault_mutex_hash(mapping, idx);
5568 hugetlb_vma_unlock_read(vma);
5569 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5571 unmap_ref_private(mm, vma, old_page, haddr);
5573 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5574 hugetlb_vma_lock_read(vma);
5576 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5578 pte_same(huge_ptep_get(ptep), pte)))
5579 goto retry_avoidcopy;
5581 * race occurs while re-acquiring page table
5582 * lock, and our job is done.
5584 delayacct_wpcopy_end();
5588 ret = vmf_error(PTR_ERR(new_page));
5589 goto out_release_old;
5593 * When the original hugepage is shared one, it does not have
5594 * anon_vma prepared.
5596 if (unlikely(anon_vma_prepare(vma))) {
5598 goto out_release_all;
5601 copy_user_huge_page(new_page, old_page, address, vma,
5602 pages_per_huge_page(h));
5603 __SetPageUptodate(new_page);
5605 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5606 haddr + huge_page_size(h));
5607 mmu_notifier_invalidate_range_start(&range);
5610 * Retake the page table lock to check for racing updates
5611 * before the page tables are altered
5614 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5615 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5616 /* Break COW or unshare */
5617 huge_ptep_clear_flush(vma, haddr, ptep);
5618 mmu_notifier_invalidate_range(mm, range.start, range.end);
5619 page_remove_rmap(old_page, vma, true);
5620 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5621 set_huge_pte_at(mm, haddr, ptep,
5622 make_huge_pte(vma, new_page, !unshare));
5623 SetHPageMigratable(new_page);
5624 /* Make the old page be freed below */
5625 new_page = old_page;
5628 mmu_notifier_invalidate_range_end(&range);
5631 * No restore in case of successful pagetable update (Break COW or
5634 if (new_page != old_page)
5635 restore_reserve_on_error(h, vma, haddr, new_page);
5640 spin_lock(ptl); /* Caller expects lock to be held */
5642 delayacct_wpcopy_end();
5647 * Return whether there is a pagecache page to back given address within VMA.
5648 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5650 static bool hugetlbfs_pagecache_present(struct hstate *h,
5651 struct vm_area_struct *vma, unsigned long address)
5653 struct address_space *mapping;
5657 mapping = vma->vm_file->f_mapping;
5658 idx = vma_hugecache_offset(h, vma, address);
5660 page = find_get_page(mapping, idx);
5663 return page != NULL;
5666 int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping,
5669 struct folio *folio = page_folio(page);
5670 struct inode *inode = mapping->host;
5671 struct hstate *h = hstate_inode(inode);
5674 __folio_set_locked(folio);
5675 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5677 if (unlikely(err)) {
5678 __folio_clear_locked(folio);
5681 ClearHPageRestoreReserve(page);
5684 * mark folio dirty so that it will not be removed from cache/file
5685 * by non-hugetlbfs specific code paths.
5687 folio_mark_dirty(folio);
5689 spin_lock(&inode->i_lock);
5690 inode->i_blocks += blocks_per_huge_page(h);
5691 spin_unlock(&inode->i_lock);
5695 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5696 struct address_space *mapping,
5699 unsigned long haddr,
5701 unsigned long reason)
5704 struct vm_fault vmf = {
5707 .real_address = addr,
5711 * Hard to debug if it ends up being
5712 * used by a callee that assumes
5713 * something about the other
5714 * uninitialized fields... same as in
5720 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5721 * userfault. Also mmap_lock could be dropped due to handling
5722 * userfault, any vma operation should be careful from here.
5724 hugetlb_vma_unlock_read(vma);
5725 hash = hugetlb_fault_mutex_hash(mapping, idx);
5726 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5727 return handle_userfault(&vmf, reason);
5731 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5732 * false if pte changed or is changing.
5734 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5735 pte_t *ptep, pte_t old_pte)
5740 ptl = huge_pte_lock(h, mm, ptep);
5741 same = pte_same(huge_ptep_get(ptep), old_pte);
5747 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5748 struct vm_area_struct *vma,
5749 struct address_space *mapping, pgoff_t idx,
5750 unsigned long address, pte_t *ptep,
5751 pte_t old_pte, unsigned int flags)
5753 struct hstate *h = hstate_vma(vma);
5754 vm_fault_t ret = VM_FAULT_SIGBUS;
5760 unsigned long haddr = address & huge_page_mask(h);
5761 bool new_page, new_pagecache_page = false;
5762 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5765 * Currently, we are forced to kill the process in the event the
5766 * original mapper has unmapped pages from the child due to a failed
5767 * COW/unsharing. Warn that such a situation has occurred as it may not
5770 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5771 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5777 * Use page lock to guard against racing truncation
5778 * before we get page_table_lock.
5781 page = find_lock_page(mapping, idx);
5783 size = i_size_read(mapping->host) >> huge_page_shift(h);
5786 /* Check for page in userfault range */
5787 if (userfaultfd_missing(vma)) {
5789 * Since hugetlb_no_page() was examining pte
5790 * without pgtable lock, we need to re-test under
5791 * lock because the pte may not be stable and could
5792 * have changed from under us. Try to detect
5793 * either changed or during-changing ptes and retry
5794 * properly when needed.
5796 * Note that userfaultfd is actually fine with
5797 * false positives (e.g. caused by pte changed),
5798 * but not wrong logical events (e.g. caused by
5799 * reading a pte during changing). The latter can
5800 * confuse the userspace, so the strictness is very
5801 * much preferred. E.g., MISSING event should
5802 * never happen on the page after UFFDIO_COPY has
5803 * correctly installed the page and returned.
5805 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5810 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5815 page = alloc_huge_page(vma, haddr, 0);
5818 * Returning error will result in faulting task being
5819 * sent SIGBUS. The hugetlb fault mutex prevents two
5820 * tasks from racing to fault in the same page which
5821 * could result in false unable to allocate errors.
5822 * Page migration does not take the fault mutex, but
5823 * does a clear then write of pte's under page table
5824 * lock. Page fault code could race with migration,
5825 * notice the clear pte and try to allocate a page
5826 * here. Before returning error, get ptl and make
5827 * sure there really is no pte entry.
5829 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5830 ret = vmf_error(PTR_ERR(page));
5835 clear_huge_page(page, address, pages_per_huge_page(h));
5836 __SetPageUptodate(page);
5839 if (vma->vm_flags & VM_MAYSHARE) {
5840 int err = hugetlb_add_to_page_cache(page, mapping, idx);
5843 * err can't be -EEXIST which implies someone
5844 * else consumed the reservation since hugetlb
5845 * fault mutex is held when add a hugetlb page
5846 * to the page cache. So it's safe to call
5847 * restore_reserve_on_error() here.
5849 restore_reserve_on_error(h, vma, haddr, page);
5853 new_pagecache_page = true;
5856 if (unlikely(anon_vma_prepare(vma))) {
5858 goto backout_unlocked;
5864 * If memory error occurs between mmap() and fault, some process
5865 * don't have hwpoisoned swap entry for errored virtual address.
5866 * So we need to block hugepage fault by PG_hwpoison bit check.
5868 if (unlikely(PageHWPoison(page))) {
5869 ret = VM_FAULT_HWPOISON_LARGE |
5870 VM_FAULT_SET_HINDEX(hstate_index(h));
5871 goto backout_unlocked;
5874 /* Check for page in userfault range. */
5875 if (userfaultfd_minor(vma)) {
5878 /* See comment in userfaultfd_missing() block above */
5879 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5883 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5890 * If we are going to COW a private mapping later, we examine the
5891 * pending reservations for this page now. This will ensure that
5892 * any allocations necessary to record that reservation occur outside
5895 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5896 if (vma_needs_reservation(h, vma, haddr) < 0) {
5898 goto backout_unlocked;
5900 /* Just decrements count, does not deallocate */
5901 vma_end_reservation(h, vma, haddr);
5904 ptl = huge_pte_lock(h, mm, ptep);
5906 /* If pte changed from under us, retry */
5907 if (!pte_same(huge_ptep_get(ptep), old_pte))
5911 hugepage_add_new_anon_rmap(page, vma, haddr);
5913 page_dup_file_rmap(page, true);
5914 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5915 && (vma->vm_flags & VM_SHARED)));
5917 * If this pte was previously wr-protected, keep it wr-protected even
5920 if (unlikely(pte_marker_uffd_wp(old_pte)))
5921 new_pte = huge_pte_mkuffd_wp(new_pte);
5922 set_huge_pte_at(mm, haddr, ptep, new_pte);
5924 hugetlb_count_add(pages_per_huge_page(h), mm);
5925 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5926 /* Optimization, do the COW without a second fault */
5927 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5933 * Only set HPageMigratable in newly allocated pages. Existing pages
5934 * found in the pagecache may not have HPageMigratableset if they have
5935 * been isolated for migration.
5938 SetHPageMigratable(page);
5942 hugetlb_vma_unlock_read(vma);
5943 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5949 if (new_page && !new_pagecache_page)
5950 restore_reserve_on_error(h, vma, haddr, page);
5958 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5960 unsigned long key[2];
5963 key[0] = (unsigned long) mapping;
5966 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5968 return hash & (num_fault_mutexes - 1);
5972 * For uniprocessor systems we always use a single mutex, so just
5973 * return 0 and avoid the hashing overhead.
5975 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5981 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5982 unsigned long address, unsigned int flags)
5989 struct page *page = NULL;
5990 struct page *pagecache_page = NULL;
5991 struct hstate *h = hstate_vma(vma);
5992 struct address_space *mapping;
5993 int need_wait_lock = 0;
5994 unsigned long haddr = address & huge_page_mask(h);
5996 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5999 * Since we hold no locks, ptep could be stale. That is
6000 * OK as we are only making decisions based on content and
6001 * not actually modifying content here.
6003 entry = huge_ptep_get(ptep);
6004 if (unlikely(is_hugetlb_entry_migration(entry))) {
6005 migration_entry_wait_huge(vma, ptep);
6007 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6008 return VM_FAULT_HWPOISON_LARGE |
6009 VM_FAULT_SET_HINDEX(hstate_index(h));
6013 * Serialize hugepage allocation and instantiation, so that we don't
6014 * get spurious allocation failures if two CPUs race to instantiate
6015 * the same page in the page cache.
6017 mapping = vma->vm_file->f_mapping;
6018 idx = vma_hugecache_offset(h, vma, haddr);
6019 hash = hugetlb_fault_mutex_hash(mapping, idx);
6020 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6023 * Acquire vma lock before calling huge_pte_alloc and hold
6024 * until finished with ptep. This prevents huge_pmd_unshare from
6025 * being called elsewhere and making the ptep no longer valid.
6027 * ptep could have already be assigned via huge_pte_offset. That
6028 * is OK, as huge_pte_alloc will return the same value unless
6029 * something has changed.
6031 hugetlb_vma_lock_read(vma);
6032 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6034 hugetlb_vma_unlock_read(vma);
6035 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6036 return VM_FAULT_OOM;
6039 entry = huge_ptep_get(ptep);
6040 /* PTE markers should be handled the same way as none pte */
6041 if (huge_pte_none_mostly(entry))
6043 * hugetlb_no_page will drop vma lock and hugetlb fault
6044 * mutex internally, which make us return immediately.
6046 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6052 * entry could be a migration/hwpoison entry at this point, so this
6053 * check prevents the kernel from going below assuming that we have
6054 * an active hugepage in pagecache. This goto expects the 2nd page
6055 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6056 * properly handle it.
6058 if (!pte_present(entry))
6062 * If we are going to COW/unshare the mapping later, we examine the
6063 * pending reservations for this page now. This will ensure that any
6064 * allocations necessary to record that reservation occur outside the
6065 * spinlock. Also lookup the pagecache page now as it is used to
6066 * determine if a reservation has been consumed.
6068 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6069 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6070 if (vma_needs_reservation(h, vma, haddr) < 0) {
6074 /* Just decrements count, does not deallocate */
6075 vma_end_reservation(h, vma, haddr);
6077 pagecache_page = find_lock_page(mapping, idx);
6080 ptl = huge_pte_lock(h, mm, ptep);
6082 /* Check for a racing update before calling hugetlb_wp() */
6083 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6086 /* Handle userfault-wp first, before trying to lock more pages */
6087 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6088 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6089 struct vm_fault vmf = {
6092 .real_address = address,
6097 if (pagecache_page) {
6098 unlock_page(pagecache_page);
6099 put_page(pagecache_page);
6101 hugetlb_vma_unlock_read(vma);
6102 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6103 return handle_userfault(&vmf, VM_UFFD_WP);
6107 * hugetlb_wp() requires page locks of pte_page(entry) and
6108 * pagecache_page, so here we need take the former one
6109 * when page != pagecache_page or !pagecache_page.
6111 page = pte_page(entry);
6112 if (page != pagecache_page)
6113 if (!trylock_page(page)) {
6120 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6121 if (!huge_pte_write(entry)) {
6122 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6123 pagecache_page, ptl);
6125 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6126 entry = huge_pte_mkdirty(entry);
6129 entry = pte_mkyoung(entry);
6130 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6131 flags & FAULT_FLAG_WRITE))
6132 update_mmu_cache(vma, haddr, ptep);
6134 if (page != pagecache_page)
6140 if (pagecache_page) {
6141 unlock_page(pagecache_page);
6142 put_page(pagecache_page);
6145 hugetlb_vma_unlock_read(vma);
6146 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6148 * Generally it's safe to hold refcount during waiting page lock. But
6149 * here we just wait to defer the next page fault to avoid busy loop and
6150 * the page is not used after unlocked before returning from the current
6151 * page fault. So we are safe from accessing freed page, even if we wait
6152 * here without taking refcount.
6155 wait_on_page_locked(page);
6159 #ifdef CONFIG_USERFAULTFD
6161 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
6162 * modifications for huge pages.
6164 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
6166 struct vm_area_struct *dst_vma,
6167 unsigned long dst_addr,
6168 unsigned long src_addr,
6169 enum mcopy_atomic_mode mode,
6170 struct page **pagep,
6173 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
6174 struct hstate *h = hstate_vma(dst_vma);
6175 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6176 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6178 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6184 bool page_in_pagecache = false;
6188 page = find_lock_page(mapping, idx);
6191 page_in_pagecache = true;
6192 } else if (!*pagep) {
6193 /* If a page already exists, then it's UFFDIO_COPY for
6194 * a non-missing case. Return -EEXIST.
6197 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6202 page = alloc_huge_page(dst_vma, dst_addr, 0);
6208 ret = copy_huge_page_from_user(page,
6209 (const void __user *) src_addr,
6210 pages_per_huge_page(h), false);
6212 /* fallback to copy_from_user outside mmap_lock */
6213 if (unlikely(ret)) {
6215 /* Free the allocated page which may have
6216 * consumed a reservation.
6218 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6221 /* Allocate a temporary page to hold the copied
6224 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
6230 /* Set the outparam pagep and return to the caller to
6231 * copy the contents outside the lock. Don't free the
6238 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6245 page = alloc_huge_page(dst_vma, dst_addr, 0);
6252 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
6253 pages_per_huge_page(h));
6259 * The memory barrier inside __SetPageUptodate makes sure that
6260 * preceding stores to the page contents become visible before
6261 * the set_pte_at() write.
6263 __SetPageUptodate(page);
6265 /* Add shared, newly allocated pages to the page cache. */
6266 if (vm_shared && !is_continue) {
6267 size = i_size_read(mapping->host) >> huge_page_shift(h);
6270 goto out_release_nounlock;
6273 * Serialization between remove_inode_hugepages() and
6274 * hugetlb_add_to_page_cache() below happens through the
6275 * hugetlb_fault_mutex_table that here must be hold by
6278 ret = hugetlb_add_to_page_cache(page, mapping, idx);
6280 goto out_release_nounlock;
6281 page_in_pagecache = true;
6284 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6287 if (PageHWPoison(page))
6288 goto out_release_unlock;
6291 * We allow to overwrite a pte marker: consider when both MISSING|WP
6292 * registered, we firstly wr-protect a none pte which has no page cache
6293 * page backing it, then access the page.
6296 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6297 goto out_release_unlock;
6299 if (page_in_pagecache)
6300 page_dup_file_rmap(page, true);
6302 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6305 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6306 * with wp flag set, don't set pte write bit.
6308 if (wp_copy || (is_continue && !vm_shared))
6311 writable = dst_vma->vm_flags & VM_WRITE;
6313 _dst_pte = make_huge_pte(dst_vma, page, writable);
6315 * Always mark UFFDIO_COPY page dirty; note that this may not be
6316 * extremely important for hugetlbfs for now since swapping is not
6317 * supported, but we should still be clear in that this page cannot be
6318 * thrown away at will, even if write bit not set.
6320 _dst_pte = huge_pte_mkdirty(_dst_pte);
6321 _dst_pte = pte_mkyoung(_dst_pte);
6324 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6326 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6328 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6330 /* No need to invalidate - it was non-present before */
6331 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6335 SetHPageMigratable(page);
6336 if (vm_shared || is_continue)
6343 if (vm_shared || is_continue)
6345 out_release_nounlock:
6346 if (!page_in_pagecache)
6347 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6351 #endif /* CONFIG_USERFAULTFD */
6353 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6354 int refs, struct page **pages,
6355 struct vm_area_struct **vmas)
6359 for (nr = 0; nr < refs; nr++) {
6361 pages[nr] = nth_page(page, nr);
6367 static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma,
6368 unsigned int flags, pte_t *pte,
6371 pte_t pteval = huge_ptep_get(pte);
6374 if (is_swap_pte(pteval))
6376 if (huge_pte_write(pteval))
6378 if (flags & FOLL_WRITE)
6380 if (gup_must_unshare(vma, flags, pte_page(pteval))) {
6387 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6388 unsigned long address, unsigned int flags)
6390 struct hstate *h = hstate_vma(vma);
6391 struct mm_struct *mm = vma->vm_mm;
6392 unsigned long haddr = address & huge_page_mask(h);
6393 struct page *page = NULL;
6398 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
6399 * follow_hugetlb_page().
6401 if (WARN_ON_ONCE(flags & FOLL_PIN))
6405 pte = huge_pte_offset(mm, haddr, huge_page_size(h));
6409 ptl = huge_pte_lock(h, mm, pte);
6410 entry = huge_ptep_get(pte);
6411 if (pte_present(entry)) {
6412 page = pte_page(entry) +
6413 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6415 * Note that page may be a sub-page, and with vmemmap
6416 * optimizations the page struct may be read only.
6417 * try_grab_page() will increase the ref count on the
6418 * head page, so this will be OK.
6420 * try_grab_page() should always be able to get the page here,
6421 * because we hold the ptl lock and have verified pte_present().
6423 if (try_grab_page(page, flags)) {
6428 if (is_hugetlb_entry_migration(entry)) {
6430 __migration_entry_wait_huge(pte, ptl);
6434 * hwpoisoned entry is treated as no_page_table in
6435 * follow_page_mask().
6443 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6444 struct page **pages, struct vm_area_struct **vmas,
6445 unsigned long *position, unsigned long *nr_pages,
6446 long i, unsigned int flags, int *locked)
6448 unsigned long pfn_offset;
6449 unsigned long vaddr = *position;
6450 unsigned long remainder = *nr_pages;
6451 struct hstate *h = hstate_vma(vma);
6452 int err = -EFAULT, refs;
6454 while (vaddr < vma->vm_end && remainder) {
6456 spinlock_t *ptl = NULL;
6457 bool unshare = false;
6462 * If we have a pending SIGKILL, don't keep faulting pages and
6463 * potentially allocating memory.
6465 if (fatal_signal_pending(current)) {
6471 * Some archs (sparc64, sh*) have multiple pte_ts to
6472 * each hugepage. We have to make sure we get the
6473 * first, for the page indexing below to work.
6475 * Note that page table lock is not held when pte is null.
6477 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6480 ptl = huge_pte_lock(h, mm, pte);
6481 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6484 * When coredumping, it suits get_dump_page if we just return
6485 * an error where there's an empty slot with no huge pagecache
6486 * to back it. This way, we avoid allocating a hugepage, and
6487 * the sparse dumpfile avoids allocating disk blocks, but its
6488 * huge holes still show up with zeroes where they need to be.
6490 if (absent && (flags & FOLL_DUMP) &&
6491 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6499 * We need call hugetlb_fault for both hugepages under migration
6500 * (in which case hugetlb_fault waits for the migration,) and
6501 * hwpoisoned hugepages (in which case we need to prevent the
6502 * caller from accessing to them.) In order to do this, we use
6503 * here is_swap_pte instead of is_hugetlb_entry_migration and
6504 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6505 * both cases, and because we can't follow correct pages
6506 * directly from any kind of swap entries.
6509 __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) {
6511 unsigned int fault_flags = 0;
6515 if (flags & FOLL_WRITE)
6516 fault_flags |= FAULT_FLAG_WRITE;
6518 fault_flags |= FAULT_FLAG_UNSHARE;
6520 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6521 FAULT_FLAG_KILLABLE;
6522 if (flags & FOLL_INTERRUPTIBLE)
6523 fault_flags |= FAULT_FLAG_INTERRUPTIBLE;
6525 if (flags & FOLL_NOWAIT)
6526 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6527 FAULT_FLAG_RETRY_NOWAIT;
6528 if (flags & FOLL_TRIED) {
6530 * Note: FAULT_FLAG_ALLOW_RETRY and
6531 * FAULT_FLAG_TRIED can co-exist
6533 fault_flags |= FAULT_FLAG_TRIED;
6535 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6536 if (ret & VM_FAULT_ERROR) {
6537 err = vm_fault_to_errno(ret, flags);
6541 if (ret & VM_FAULT_RETRY) {
6543 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6547 * VM_FAULT_RETRY must not return an
6548 * error, it will return zero
6551 * No need to update "position" as the
6552 * caller will not check it after
6553 * *nr_pages is set to 0.
6560 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6561 page = pte_page(huge_ptep_get(pte));
6563 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6564 !PageAnonExclusive(page), page);
6567 * If subpage information not requested, update counters
6568 * and skip the same_page loop below.
6570 if (!pages && !vmas && !pfn_offset &&
6571 (vaddr + huge_page_size(h) < vma->vm_end) &&
6572 (remainder >= pages_per_huge_page(h))) {
6573 vaddr += huge_page_size(h);
6574 remainder -= pages_per_huge_page(h);
6575 i += pages_per_huge_page(h);
6580 /* vaddr may not be aligned to PAGE_SIZE */
6581 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6582 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6585 record_subpages_vmas(nth_page(page, pfn_offset),
6587 likely(pages) ? pages + i : NULL,
6588 vmas ? vmas + i : NULL);
6592 * try_grab_folio() should always succeed here,
6593 * because: a) we hold the ptl lock, and b) we've just
6594 * checked that the huge page is present in the page
6595 * tables. If the huge page is present, then the tail
6596 * pages must also be present. The ptl prevents the
6597 * head page and tail pages from being rearranged in
6598 * any way. As this is hugetlb, the pages will never
6599 * be p2pdma or not longterm pinable. So this page
6600 * must be available at this point, unless the page
6601 * refcount overflowed:
6603 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6612 vaddr += (refs << PAGE_SHIFT);
6618 *nr_pages = remainder;
6620 * setting position is actually required only if remainder is
6621 * not zero but it's faster not to add a "if (remainder)"
6629 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6630 unsigned long address, unsigned long end,
6631 pgprot_t newprot, unsigned long cp_flags)
6633 struct mm_struct *mm = vma->vm_mm;
6634 unsigned long start = address;
6637 struct hstate *h = hstate_vma(vma);
6638 unsigned long pages = 0, psize = huge_page_size(h);
6639 bool shared_pmd = false;
6640 struct mmu_notifier_range range;
6641 unsigned long last_addr_mask;
6642 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6643 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6646 * In the case of shared PMDs, the area to flush could be beyond
6647 * start/end. Set range.start/range.end to cover the maximum possible
6648 * range if PMD sharing is possible.
6650 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6651 0, vma, mm, start, end);
6652 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6654 BUG_ON(address >= end);
6655 flush_cache_range(vma, range.start, range.end);
6657 mmu_notifier_invalidate_range_start(&range);
6658 hugetlb_vma_lock_write(vma);
6659 i_mmap_lock_write(vma->vm_file->f_mapping);
6660 last_addr_mask = hugetlb_mask_last_page(h);
6661 for (; address < end; address += psize) {
6663 ptep = huge_pte_offset(mm, address, psize);
6666 address |= last_addr_mask;
6670 * Userfaultfd wr-protect requires pgtable
6671 * pre-allocations to install pte markers.
6673 ptep = huge_pte_alloc(mm, vma, address, psize);
6677 ptl = huge_pte_lock(h, mm, ptep);
6678 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6680 * When uffd-wp is enabled on the vma, unshare
6681 * shouldn't happen at all. Warn about it if it
6682 * happened due to some reason.
6684 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6688 address |= last_addr_mask;
6691 pte = huge_ptep_get(ptep);
6692 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6693 /* Nothing to do. */
6694 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6695 swp_entry_t entry = pte_to_swp_entry(pte);
6696 struct page *page = pfn_swap_entry_to_page(entry);
6699 if (is_writable_migration_entry(entry)) {
6701 entry = make_readable_exclusive_migration_entry(
6704 entry = make_readable_migration_entry(
6706 newpte = swp_entry_to_pte(entry);
6711 newpte = pte_swp_mkuffd_wp(newpte);
6712 else if (uffd_wp_resolve)
6713 newpte = pte_swp_clear_uffd_wp(newpte);
6714 if (!pte_same(pte, newpte))
6715 set_huge_pte_at(mm, address, ptep, newpte);
6716 } else if (unlikely(is_pte_marker(pte))) {
6717 /* No other markers apply for now. */
6718 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6719 if (uffd_wp_resolve)
6720 /* Safe to modify directly (non-present->none). */
6721 huge_pte_clear(mm, address, ptep, psize);
6722 } else if (!huge_pte_none(pte)) {
6724 unsigned int shift = huge_page_shift(hstate_vma(vma));
6726 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6727 pte = huge_pte_modify(old_pte, newprot);
6728 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6730 pte = huge_pte_mkuffd_wp(pte);
6731 else if (uffd_wp_resolve)
6732 pte = huge_pte_clear_uffd_wp(pte);
6733 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6737 if (unlikely(uffd_wp))
6738 /* Safe to modify directly (none->non-present). */
6739 set_huge_pte_at(mm, address, ptep,
6740 make_pte_marker(PTE_MARKER_UFFD_WP));
6745 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6746 * may have cleared our pud entry and done put_page on the page table:
6747 * once we release i_mmap_rwsem, another task can do the final put_page
6748 * and that page table be reused and filled with junk. If we actually
6749 * did unshare a page of pmds, flush the range corresponding to the pud.
6752 flush_hugetlb_tlb_range(vma, range.start, range.end);
6754 flush_hugetlb_tlb_range(vma, start, end);
6756 * No need to call mmu_notifier_invalidate_range() we are downgrading
6757 * page table protection not changing it to point to a new page.
6759 * See Documentation/mm/mmu_notifier.rst
6761 i_mmap_unlock_write(vma->vm_file->f_mapping);
6762 hugetlb_vma_unlock_write(vma);
6763 mmu_notifier_invalidate_range_end(&range);
6765 return pages << h->order;
6768 /* Return true if reservation was successful, false otherwise. */
6769 bool hugetlb_reserve_pages(struct inode *inode,
6771 struct vm_area_struct *vma,
6772 vm_flags_t vm_flags)
6775 struct hstate *h = hstate_inode(inode);
6776 struct hugepage_subpool *spool = subpool_inode(inode);
6777 struct resv_map *resv_map;
6778 struct hugetlb_cgroup *h_cg = NULL;
6779 long gbl_reserve, regions_needed = 0;
6781 /* This should never happen */
6783 VM_WARN(1, "%s called with a negative range\n", __func__);
6788 * vma specific semaphore used for pmd sharing and fault/truncation
6791 hugetlb_vma_lock_alloc(vma);
6794 * Only apply hugepage reservation if asked. At fault time, an
6795 * attempt will be made for VM_NORESERVE to allocate a page
6796 * without using reserves
6798 if (vm_flags & VM_NORESERVE)
6802 * Shared mappings base their reservation on the number of pages that
6803 * are already allocated on behalf of the file. Private mappings need
6804 * to reserve the full area even if read-only as mprotect() may be
6805 * called to make the mapping read-write. Assume !vma is a shm mapping
6807 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6809 * resv_map can not be NULL as hugetlb_reserve_pages is only
6810 * called for inodes for which resv_maps were created (see
6811 * hugetlbfs_get_inode).
6813 resv_map = inode_resv_map(inode);
6815 chg = region_chg(resv_map, from, to, ®ions_needed);
6817 /* Private mapping. */
6818 resv_map = resv_map_alloc();
6824 set_vma_resv_map(vma, resv_map);
6825 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6831 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6832 chg * pages_per_huge_page(h), &h_cg) < 0)
6835 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6836 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6839 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6843 * There must be enough pages in the subpool for the mapping. If
6844 * the subpool has a minimum size, there may be some global
6845 * reservations already in place (gbl_reserve).
6847 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6848 if (gbl_reserve < 0)
6849 goto out_uncharge_cgroup;
6852 * Check enough hugepages are available for the reservation.
6853 * Hand the pages back to the subpool if there are not
6855 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6859 * Account for the reservations made. Shared mappings record regions
6860 * that have reservations as they are shared by multiple VMAs.
6861 * When the last VMA disappears, the region map says how much
6862 * the reservation was and the page cache tells how much of
6863 * the reservation was consumed. Private mappings are per-VMA and
6864 * only the consumed reservations are tracked. When the VMA
6865 * disappears, the original reservation is the VMA size and the
6866 * consumed reservations are stored in the map. Hence, nothing
6867 * else has to be done for private mappings here
6869 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6870 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6872 if (unlikely(add < 0)) {
6873 hugetlb_acct_memory(h, -gbl_reserve);
6875 } else if (unlikely(chg > add)) {
6877 * pages in this range were added to the reserve
6878 * map between region_chg and region_add. This
6879 * indicates a race with alloc_huge_page. Adjust
6880 * the subpool and reserve counts modified above
6881 * based on the difference.
6886 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6887 * reference to h_cg->css. See comment below for detail.
6889 hugetlb_cgroup_uncharge_cgroup_rsvd(
6891 (chg - add) * pages_per_huge_page(h), h_cg);
6893 rsv_adjust = hugepage_subpool_put_pages(spool,
6895 hugetlb_acct_memory(h, -rsv_adjust);
6898 * The file_regions will hold their own reference to
6899 * h_cg->css. So we should release the reference held
6900 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6903 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6909 /* put back original number of pages, chg */
6910 (void)hugepage_subpool_put_pages(spool, chg);
6911 out_uncharge_cgroup:
6912 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6913 chg * pages_per_huge_page(h), h_cg);
6915 hugetlb_vma_lock_free(vma);
6916 if (!vma || vma->vm_flags & VM_MAYSHARE)
6917 /* Only call region_abort if the region_chg succeeded but the
6918 * region_add failed or didn't run.
6920 if (chg >= 0 && add < 0)
6921 region_abort(resv_map, from, to, regions_needed);
6922 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6923 kref_put(&resv_map->refs, resv_map_release);
6927 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6930 struct hstate *h = hstate_inode(inode);
6931 struct resv_map *resv_map = inode_resv_map(inode);
6933 struct hugepage_subpool *spool = subpool_inode(inode);
6937 * Since this routine can be called in the evict inode path for all
6938 * hugetlbfs inodes, resv_map could be NULL.
6941 chg = region_del(resv_map, start, end);
6943 * region_del() can fail in the rare case where a region
6944 * must be split and another region descriptor can not be
6945 * allocated. If end == LONG_MAX, it will not fail.
6951 spin_lock(&inode->i_lock);
6952 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6953 spin_unlock(&inode->i_lock);
6956 * If the subpool has a minimum size, the number of global
6957 * reservations to be released may be adjusted.
6959 * Note that !resv_map implies freed == 0. So (chg - freed)
6960 * won't go negative.
6962 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6963 hugetlb_acct_memory(h, -gbl_reserve);
6968 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6969 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6970 struct vm_area_struct *vma,
6971 unsigned long addr, pgoff_t idx)
6973 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6975 unsigned long sbase = saddr & PUD_MASK;
6976 unsigned long s_end = sbase + PUD_SIZE;
6978 /* Allow segments to share if only one is marked locked */
6979 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6980 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6983 * match the virtual addresses, permission and the alignment of the
6986 * Also, vma_lock (vm_private_data) is required for sharing.
6988 if (pmd_index(addr) != pmd_index(saddr) ||
6989 vm_flags != svm_flags ||
6990 !range_in_vma(svma, sbase, s_end) ||
6991 !svma->vm_private_data)
6997 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6999 unsigned long start = addr & PUD_MASK;
7000 unsigned long end = start + PUD_SIZE;
7002 #ifdef CONFIG_USERFAULTFD
7003 if (uffd_disable_huge_pmd_share(vma))
7007 * check on proper vm_flags and page table alignment
7009 if (!(vma->vm_flags & VM_MAYSHARE))
7011 if (!vma->vm_private_data) /* vma lock required for sharing */
7013 if (!range_in_vma(vma, start, end))
7019 * Determine if start,end range within vma could be mapped by shared pmd.
7020 * If yes, adjust start and end to cover range associated with possible
7021 * shared pmd mappings.
7023 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7024 unsigned long *start, unsigned long *end)
7026 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7027 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7030 * vma needs to span at least one aligned PUD size, and the range
7031 * must be at least partially within in.
7033 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7034 (*end <= v_start) || (*start >= v_end))
7037 /* Extend the range to be PUD aligned for a worst case scenario */
7038 if (*start > v_start)
7039 *start = ALIGN_DOWN(*start, PUD_SIZE);
7042 *end = ALIGN(*end, PUD_SIZE);
7046 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7047 * and returns the corresponding pte. While this is not necessary for the
7048 * !shared pmd case because we can allocate the pmd later as well, it makes the
7049 * code much cleaner. pmd allocation is essential for the shared case because
7050 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7051 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7052 * bad pmd for sharing.
7054 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7055 unsigned long addr, pud_t *pud)
7057 struct address_space *mapping = vma->vm_file->f_mapping;
7058 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7060 struct vm_area_struct *svma;
7061 unsigned long saddr;
7066 i_mmap_lock_read(mapping);
7067 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7071 saddr = page_table_shareable(svma, vma, addr, idx);
7073 spte = huge_pte_offset(svma->vm_mm, saddr,
7074 vma_mmu_pagesize(svma));
7076 get_page(virt_to_page(spte));
7085 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
7086 if (pud_none(*pud)) {
7087 pud_populate(mm, pud,
7088 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7091 put_page(virt_to_page(spte));
7095 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7096 i_mmap_unlock_read(mapping);
7101 * unmap huge page backed by shared pte.
7103 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7104 * indicated by page_count > 1, unmap is achieved by clearing pud and
7105 * decrementing the ref count. If count == 1, the pte page is not shared.
7107 * Called with page table lock held.
7109 * returns: 1 successfully unmapped a shared pte page
7110 * 0 the underlying pte page is not shared, or it is the last user
7112 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7113 unsigned long addr, pte_t *ptep)
7115 pgd_t *pgd = pgd_offset(mm, addr);
7116 p4d_t *p4d = p4d_offset(pgd, addr);
7117 pud_t *pud = pud_offset(p4d, addr);
7119 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7120 hugetlb_vma_assert_locked(vma);
7121 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7122 if (page_count(virt_to_page(ptep)) == 1)
7126 put_page(virt_to_page(ptep));
7131 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7133 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7134 unsigned long addr, pud_t *pud)
7139 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7140 unsigned long addr, pte_t *ptep)
7145 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7146 unsigned long *start, unsigned long *end)
7150 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7154 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7156 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7157 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7158 unsigned long addr, unsigned long sz)
7165 pgd = pgd_offset(mm, addr);
7166 p4d = p4d_alloc(mm, pgd, addr);
7169 pud = pud_alloc(mm, p4d, addr);
7171 if (sz == PUD_SIZE) {
7174 BUG_ON(sz != PMD_SIZE);
7175 if (want_pmd_share(vma, addr) && pud_none(*pud))
7176 pte = huge_pmd_share(mm, vma, addr, pud);
7178 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7181 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
7187 * huge_pte_offset() - Walk the page table to resolve the hugepage
7188 * entry at address @addr
7190 * Return: Pointer to page table entry (PUD or PMD) for
7191 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7192 * size @sz doesn't match the hugepage size at this level of the page
7195 pte_t *huge_pte_offset(struct mm_struct *mm,
7196 unsigned long addr, unsigned long sz)
7203 pgd = pgd_offset(mm, addr);
7204 if (!pgd_present(*pgd))
7206 p4d = p4d_offset(pgd, addr);
7207 if (!p4d_present(*p4d))
7210 pud = pud_offset(p4d, addr);
7212 /* must be pud huge, non-present or none */
7213 return (pte_t *)pud;
7214 if (!pud_present(*pud))
7216 /* must have a valid entry and size to go further */
7218 pmd = pmd_offset(pud, addr);
7219 /* must be pmd huge, non-present or none */
7220 return (pte_t *)pmd;
7224 * Return a mask that can be used to update an address to the last huge
7225 * page in a page table page mapping size. Used to skip non-present
7226 * page table entries when linearly scanning address ranges. Architectures
7227 * with unique huge page to page table relationships can define their own
7228 * version of this routine.
7230 unsigned long hugetlb_mask_last_page(struct hstate *h)
7232 unsigned long hp_size = huge_page_size(h);
7234 if (hp_size == PUD_SIZE)
7235 return P4D_SIZE - PUD_SIZE;
7236 else if (hp_size == PMD_SIZE)
7237 return PUD_SIZE - PMD_SIZE;
7244 /* See description above. Architectures can provide their own version. */
7245 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7247 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7248 if (huge_page_size(h) == PMD_SIZE)
7249 return PUD_SIZE - PMD_SIZE;
7254 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7257 * These functions are overwritable if your architecture needs its own
7260 int isolate_hugetlb(struct page *page, struct list_head *list)
7264 spin_lock_irq(&hugetlb_lock);
7265 if (!PageHeadHuge(page) ||
7266 !HPageMigratable(page) ||
7267 !get_page_unless_zero(page)) {
7271 ClearHPageMigratable(page);
7272 list_move_tail(&page->lru, list);
7274 spin_unlock_irq(&hugetlb_lock);
7278 int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison)
7283 spin_lock_irq(&hugetlb_lock);
7284 if (PageHeadHuge(page)) {
7286 if (HPageFreed(page))
7288 else if (HPageMigratable(page) || unpoison)
7289 ret = get_page_unless_zero(page);
7293 spin_unlock_irq(&hugetlb_lock);
7297 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7298 bool *migratable_cleared)
7302 spin_lock_irq(&hugetlb_lock);
7303 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7304 spin_unlock_irq(&hugetlb_lock);
7308 void putback_active_hugepage(struct page *page)
7310 spin_lock_irq(&hugetlb_lock);
7311 SetHPageMigratable(page);
7312 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7313 spin_unlock_irq(&hugetlb_lock);
7317 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7319 struct hstate *h = folio_hstate(old_folio);
7321 hugetlb_cgroup_migrate(old_folio, new_folio);
7322 set_page_owner_migrate_reason(&new_folio->page, reason);
7325 * transfer temporary state of the new hugetlb folio. This is
7326 * reverse to other transitions because the newpage is going to
7327 * be final while the old one will be freed so it takes over
7328 * the temporary status.
7330 * Also note that we have to transfer the per-node surplus state
7331 * here as well otherwise the global surplus count will not match
7334 if (folio_test_hugetlb_temporary(new_folio)) {
7335 int old_nid = folio_nid(old_folio);
7336 int new_nid = folio_nid(new_folio);
7338 folio_set_hugetlb_temporary(old_folio);
7339 folio_clear_hugetlb_temporary(new_folio);
7343 * There is no need to transfer the per-node surplus state
7344 * when we do not cross the node.
7346 if (new_nid == old_nid)
7348 spin_lock_irq(&hugetlb_lock);
7349 if (h->surplus_huge_pages_node[old_nid]) {
7350 h->surplus_huge_pages_node[old_nid]--;
7351 h->surplus_huge_pages_node[new_nid]++;
7353 spin_unlock_irq(&hugetlb_lock);
7357 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7358 unsigned long start,
7361 struct hstate *h = hstate_vma(vma);
7362 unsigned long sz = huge_page_size(h);
7363 struct mm_struct *mm = vma->vm_mm;
7364 struct mmu_notifier_range range;
7365 unsigned long address;
7369 if (!(vma->vm_flags & VM_MAYSHARE))
7375 flush_cache_range(vma, start, end);
7377 * No need to call adjust_range_if_pmd_sharing_possible(), because
7378 * we have already done the PUD_SIZE alignment.
7380 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7382 mmu_notifier_invalidate_range_start(&range);
7383 hugetlb_vma_lock_write(vma);
7384 i_mmap_lock_write(vma->vm_file->f_mapping);
7385 for (address = start; address < end; address += PUD_SIZE) {
7386 ptep = huge_pte_offset(mm, address, sz);
7389 ptl = huge_pte_lock(h, mm, ptep);
7390 huge_pmd_unshare(mm, vma, address, ptep);
7393 flush_hugetlb_tlb_range(vma, start, end);
7394 i_mmap_unlock_write(vma->vm_file->f_mapping);
7395 hugetlb_vma_unlock_write(vma);
7397 * No need to call mmu_notifier_invalidate_range(), see
7398 * Documentation/mm/mmu_notifier.rst.
7400 mmu_notifier_invalidate_range_end(&range);
7404 * This function will unconditionally remove all the shared pmd pgtable entries
7405 * within the specific vma for a hugetlbfs memory range.
7407 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7409 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7410 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7414 static bool cma_reserve_called __initdata;
7416 static int __init cmdline_parse_hugetlb_cma(char *p)
7423 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7426 if (s[count] == ':') {
7427 if (tmp >= MAX_NUMNODES)
7429 nid = array_index_nospec(tmp, MAX_NUMNODES);
7432 tmp = memparse(s, &s);
7433 hugetlb_cma_size_in_node[nid] = tmp;
7434 hugetlb_cma_size += tmp;
7437 * Skip the separator if have one, otherwise
7438 * break the parsing.
7445 hugetlb_cma_size = memparse(p, &p);
7453 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7455 void __init hugetlb_cma_reserve(int order)
7457 unsigned long size, reserved, per_node;
7458 bool node_specific_cma_alloc = false;
7461 cma_reserve_called = true;
7463 if (!hugetlb_cma_size)
7466 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7467 if (hugetlb_cma_size_in_node[nid] == 0)
7470 if (!node_online(nid)) {
7471 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7472 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7473 hugetlb_cma_size_in_node[nid] = 0;
7477 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7478 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7479 nid, (PAGE_SIZE << order) / SZ_1M);
7480 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7481 hugetlb_cma_size_in_node[nid] = 0;
7483 node_specific_cma_alloc = true;
7487 /* Validate the CMA size again in case some invalid nodes specified. */
7488 if (!hugetlb_cma_size)
7491 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7492 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7493 (PAGE_SIZE << order) / SZ_1M);
7494 hugetlb_cma_size = 0;
7498 if (!node_specific_cma_alloc) {
7500 * If 3 GB area is requested on a machine with 4 numa nodes,
7501 * let's allocate 1 GB on first three nodes and ignore the last one.
7503 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7504 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7505 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7509 for_each_online_node(nid) {
7511 char name[CMA_MAX_NAME];
7513 if (node_specific_cma_alloc) {
7514 if (hugetlb_cma_size_in_node[nid] == 0)
7517 size = hugetlb_cma_size_in_node[nid];
7519 size = min(per_node, hugetlb_cma_size - reserved);
7522 size = round_up(size, PAGE_SIZE << order);
7524 snprintf(name, sizeof(name), "hugetlb%d", nid);
7526 * Note that 'order per bit' is based on smallest size that
7527 * may be returned to CMA allocator in the case of
7528 * huge page demotion.
7530 res = cma_declare_contiguous_nid(0, size, 0,
7531 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7533 &hugetlb_cma[nid], nid);
7535 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7541 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7544 if (reserved >= hugetlb_cma_size)
7550 * hugetlb_cma_size is used to determine if allocations from
7551 * cma are possible. Set to zero if no cma regions are set up.
7553 hugetlb_cma_size = 0;
7556 static void __init hugetlb_cma_check(void)
7558 if (!hugetlb_cma_size || cma_reserve_called)
7561 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7564 #endif /* CONFIG_CMA */