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_page(struct page *page, unsigned int order)
59 return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page,
63 static bool hugetlb_cma_page(struct page *page, 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 | VM_SHARED) &&
266 vma->vm_private_data;
269 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
271 if (__vma_shareable_lock(vma)) {
272 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
274 down_read(&vma_lock->rw_sema);
278 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
280 if (__vma_shareable_lock(vma)) {
281 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
283 up_read(&vma_lock->rw_sema);
287 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
289 if (__vma_shareable_lock(vma)) {
290 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
292 down_write(&vma_lock->rw_sema);
296 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
298 if (__vma_shareable_lock(vma)) {
299 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
301 up_write(&vma_lock->rw_sema);
305 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
307 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
309 if (!__vma_shareable_lock(vma))
312 return down_write_trylock(&vma_lock->rw_sema);
315 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
317 if (__vma_shareable_lock(vma)) {
318 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
320 lockdep_assert_held(&vma_lock->rw_sema);
324 void hugetlb_vma_lock_release(struct kref *kref)
326 struct hugetlb_vma_lock *vma_lock = container_of(kref,
327 struct hugetlb_vma_lock, refs);
332 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
334 struct vm_area_struct *vma = vma_lock->vma;
337 * vma_lock structure may or not be released as a result of put,
338 * it certainly will no longer be attached to vma so clear pointer.
339 * Semaphore synchronizes access to vma_lock->vma field.
341 vma_lock->vma = NULL;
342 vma->vm_private_data = NULL;
343 up_write(&vma_lock->rw_sema);
344 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
347 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
349 if (__vma_shareable_lock(vma)) {
350 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
352 __hugetlb_vma_unlock_write_put(vma_lock);
356 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
359 * Only present in sharable vmas.
361 if (!vma || !__vma_shareable_lock(vma))
364 if (vma->vm_private_data) {
365 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
367 down_write(&vma_lock->rw_sema);
368 __hugetlb_vma_unlock_write_put(vma_lock);
372 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
374 struct hugetlb_vma_lock *vma_lock;
376 /* Only establish in (flags) sharable vmas */
377 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
380 /* Should never get here with non-NULL vm_private_data */
381 if (vma->vm_private_data)
384 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
387 * If we can not allocate structure, then vma can not
388 * participate in pmd sharing. This is only a possible
389 * performance enhancement and memory saving issue.
390 * However, the lock is also used to synchronize page
391 * faults with truncation. If the lock is not present,
392 * unlikely races could leave pages in a file past i_size
393 * until the file is removed. Warn in the unlikely case of
394 * allocation failure.
396 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
400 kref_init(&vma_lock->refs);
401 init_rwsem(&vma_lock->rw_sema);
403 vma->vm_private_data = vma_lock;
406 /* Helper that removes a struct file_region from the resv_map cache and returns
409 static struct file_region *
410 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
412 struct file_region *nrg;
414 VM_BUG_ON(resv->region_cache_count <= 0);
416 resv->region_cache_count--;
417 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
418 list_del(&nrg->link);
426 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
427 struct file_region *rg)
429 #ifdef CONFIG_CGROUP_HUGETLB
430 nrg->reservation_counter = rg->reservation_counter;
437 /* Helper that records hugetlb_cgroup uncharge info. */
438 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
440 struct resv_map *resv,
441 struct file_region *nrg)
443 #ifdef CONFIG_CGROUP_HUGETLB
445 nrg->reservation_counter =
446 &h_cg->rsvd_hugepage[hstate_index(h)];
447 nrg->css = &h_cg->css;
449 * The caller will hold exactly one h_cg->css reference for the
450 * whole contiguous reservation region. But this area might be
451 * scattered when there are already some file_regions reside in
452 * it. As a result, many file_regions may share only one css
453 * reference. In order to ensure that one file_region must hold
454 * exactly one h_cg->css reference, we should do css_get for
455 * each file_region and leave the reference held by caller
459 if (!resv->pages_per_hpage)
460 resv->pages_per_hpage = pages_per_huge_page(h);
461 /* pages_per_hpage should be the same for all entries in
464 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
466 nrg->reservation_counter = NULL;
472 static void put_uncharge_info(struct file_region *rg)
474 #ifdef CONFIG_CGROUP_HUGETLB
480 static bool has_same_uncharge_info(struct file_region *rg,
481 struct file_region *org)
483 #ifdef CONFIG_CGROUP_HUGETLB
484 return rg->reservation_counter == org->reservation_counter &&
492 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
494 struct file_region *nrg, *prg;
496 prg = list_prev_entry(rg, link);
497 if (&prg->link != &resv->regions && prg->to == rg->from &&
498 has_same_uncharge_info(prg, rg)) {
502 put_uncharge_info(rg);
508 nrg = list_next_entry(rg, link);
509 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
510 has_same_uncharge_info(nrg, rg)) {
511 nrg->from = rg->from;
514 put_uncharge_info(rg);
520 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
521 long to, struct hstate *h, struct hugetlb_cgroup *cg,
522 long *regions_needed)
524 struct file_region *nrg;
526 if (!regions_needed) {
527 nrg = get_file_region_entry_from_cache(map, from, to);
528 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
529 list_add(&nrg->link, rg);
530 coalesce_file_region(map, nrg);
532 *regions_needed += 1;
538 * Must be called with resv->lock held.
540 * Calling this with regions_needed != NULL will count the number of pages
541 * to be added but will not modify the linked list. And regions_needed will
542 * indicate the number of file_regions needed in the cache to carry out to add
543 * the regions for this range.
545 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
546 struct hugetlb_cgroup *h_cg,
547 struct hstate *h, long *regions_needed)
550 struct list_head *head = &resv->regions;
551 long last_accounted_offset = f;
552 struct file_region *iter, *trg = NULL;
553 struct list_head *rg = NULL;
558 /* In this loop, we essentially handle an entry for the range
559 * [last_accounted_offset, iter->from), at every iteration, with some
562 list_for_each_entry_safe(iter, trg, head, link) {
563 /* Skip irrelevant regions that start before our range. */
564 if (iter->from < f) {
565 /* If this region ends after the last accounted offset,
566 * then we need to update last_accounted_offset.
568 if (iter->to > last_accounted_offset)
569 last_accounted_offset = iter->to;
573 /* When we find a region that starts beyond our range, we've
576 if (iter->from >= t) {
577 rg = iter->link.prev;
581 /* Add an entry for last_accounted_offset -> iter->from, and
582 * update last_accounted_offset.
584 if (iter->from > last_accounted_offset)
585 add += hugetlb_resv_map_add(resv, iter->link.prev,
586 last_accounted_offset,
590 last_accounted_offset = iter->to;
593 /* Handle the case where our range extends beyond
594 * last_accounted_offset.
598 if (last_accounted_offset < t)
599 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
600 t, h, h_cg, regions_needed);
605 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
607 static int allocate_file_region_entries(struct resv_map *resv,
609 __must_hold(&resv->lock)
611 LIST_HEAD(allocated_regions);
612 int to_allocate = 0, i = 0;
613 struct file_region *trg = NULL, *rg = NULL;
615 VM_BUG_ON(regions_needed < 0);
618 * Check for sufficient descriptors in the cache to accommodate
619 * the number of in progress add operations plus regions_needed.
621 * This is a while loop because when we drop the lock, some other call
622 * to region_add or region_del may have consumed some region_entries,
623 * so we keep looping here until we finally have enough entries for
624 * (adds_in_progress + regions_needed).
626 while (resv->region_cache_count <
627 (resv->adds_in_progress + regions_needed)) {
628 to_allocate = resv->adds_in_progress + regions_needed -
629 resv->region_cache_count;
631 /* At this point, we should have enough entries in the cache
632 * for all the existing adds_in_progress. We should only be
633 * needing to allocate for regions_needed.
635 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
637 spin_unlock(&resv->lock);
638 for (i = 0; i < to_allocate; i++) {
639 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
642 list_add(&trg->link, &allocated_regions);
645 spin_lock(&resv->lock);
647 list_splice(&allocated_regions, &resv->region_cache);
648 resv->region_cache_count += to_allocate;
654 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
662 * Add the huge page range represented by [f, t) to the reserve
663 * map. Regions will be taken from the cache to fill in this range.
664 * Sufficient regions should exist in the cache due to the previous
665 * call to region_chg with the same range, but in some cases the cache will not
666 * have sufficient entries due to races with other code doing region_add or
667 * region_del. The extra needed entries will be allocated.
669 * regions_needed is the out value provided by a previous call to region_chg.
671 * Return the number of new huge pages added to the map. This number is greater
672 * than or equal to zero. If file_region entries needed to be allocated for
673 * this operation and we were not able to allocate, it returns -ENOMEM.
674 * region_add of regions of length 1 never allocate file_regions and cannot
675 * fail; region_chg will always allocate at least 1 entry and a region_add for
676 * 1 page will only require at most 1 entry.
678 static long region_add(struct resv_map *resv, long f, long t,
679 long in_regions_needed, struct hstate *h,
680 struct hugetlb_cgroup *h_cg)
682 long add = 0, actual_regions_needed = 0;
684 spin_lock(&resv->lock);
687 /* Count how many regions are actually needed to execute this add. */
688 add_reservation_in_range(resv, f, t, NULL, NULL,
689 &actual_regions_needed);
692 * Check for sufficient descriptors in the cache to accommodate
693 * this add operation. Note that actual_regions_needed may be greater
694 * than in_regions_needed, as the resv_map may have been modified since
695 * the region_chg call. In this case, we need to make sure that we
696 * allocate extra entries, such that we have enough for all the
697 * existing adds_in_progress, plus the excess needed for this
700 if (actual_regions_needed > in_regions_needed &&
701 resv->region_cache_count <
702 resv->adds_in_progress +
703 (actual_regions_needed - in_regions_needed)) {
704 /* region_add operation of range 1 should never need to
705 * allocate file_region entries.
707 VM_BUG_ON(t - f <= 1);
709 if (allocate_file_region_entries(
710 resv, actual_regions_needed - in_regions_needed)) {
717 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
719 resv->adds_in_progress -= in_regions_needed;
721 spin_unlock(&resv->lock);
726 * Examine the existing reserve map and determine how many
727 * huge pages in the specified range [f, t) are NOT currently
728 * represented. This routine is called before a subsequent
729 * call to region_add that will actually modify the reserve
730 * map to add the specified range [f, t). region_chg does
731 * not change the number of huge pages represented by the
732 * map. A number of new file_region structures is added to the cache as a
733 * placeholder, for the subsequent region_add call to use. At least 1
734 * file_region structure is added.
736 * out_regions_needed is the number of regions added to the
737 * resv->adds_in_progress. This value needs to be provided to a follow up call
738 * to region_add or region_abort for proper accounting.
740 * Returns the number of huge pages that need to be added to the existing
741 * reservation map for the range [f, t). This number is greater or equal to
742 * zero. -ENOMEM is returned if a new file_region structure or cache entry
743 * is needed and can not be allocated.
745 static long region_chg(struct resv_map *resv, long f, long t,
746 long *out_regions_needed)
750 spin_lock(&resv->lock);
752 /* Count how many hugepages in this range are NOT represented. */
753 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
756 if (*out_regions_needed == 0)
757 *out_regions_needed = 1;
759 if (allocate_file_region_entries(resv, *out_regions_needed))
762 resv->adds_in_progress += *out_regions_needed;
764 spin_unlock(&resv->lock);
769 * Abort the in progress add operation. The adds_in_progress field
770 * of the resv_map keeps track of the operations in progress between
771 * calls to region_chg and region_add. Operations are sometimes
772 * aborted after the call to region_chg. In such cases, region_abort
773 * is called to decrement the adds_in_progress counter. regions_needed
774 * is the value returned by the region_chg call, it is used to decrement
775 * the adds_in_progress counter.
777 * NOTE: The range arguments [f, t) are not needed or used in this
778 * routine. They are kept to make reading the calling code easier as
779 * arguments will match the associated region_chg call.
781 static void region_abort(struct resv_map *resv, long f, long t,
784 spin_lock(&resv->lock);
785 VM_BUG_ON(!resv->region_cache_count);
786 resv->adds_in_progress -= regions_needed;
787 spin_unlock(&resv->lock);
791 * Delete the specified range [f, t) from the reserve map. If the
792 * t parameter is LONG_MAX, this indicates that ALL regions after f
793 * should be deleted. Locate the regions which intersect [f, t)
794 * and either trim, delete or split the existing regions.
796 * Returns the number of huge pages deleted from the reserve map.
797 * In the normal case, the return value is zero or more. In the
798 * case where a region must be split, a new region descriptor must
799 * be allocated. If the allocation fails, -ENOMEM will be returned.
800 * NOTE: If the parameter t == LONG_MAX, then we will never split
801 * a region and possibly return -ENOMEM. Callers specifying
802 * t == LONG_MAX do not need to check for -ENOMEM error.
804 static long region_del(struct resv_map *resv, long f, long t)
806 struct list_head *head = &resv->regions;
807 struct file_region *rg, *trg;
808 struct file_region *nrg = NULL;
812 spin_lock(&resv->lock);
813 list_for_each_entry_safe(rg, trg, head, link) {
815 * Skip regions before the range to be deleted. file_region
816 * ranges are normally of the form [from, to). However, there
817 * may be a "placeholder" entry in the map which is of the form
818 * (from, to) with from == to. Check for placeholder entries
819 * at the beginning of the range to be deleted.
821 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
827 if (f > rg->from && t < rg->to) { /* Must split region */
829 * Check for an entry in the cache before dropping
830 * lock and attempting allocation.
833 resv->region_cache_count > resv->adds_in_progress) {
834 nrg = list_first_entry(&resv->region_cache,
837 list_del(&nrg->link);
838 resv->region_cache_count--;
842 spin_unlock(&resv->lock);
843 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
850 hugetlb_cgroup_uncharge_file_region(
851 resv, rg, t - f, false);
853 /* New entry for end of split region */
857 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
859 INIT_LIST_HEAD(&nrg->link);
861 /* Original entry is trimmed */
864 list_add(&nrg->link, &rg->link);
869 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
870 del += rg->to - rg->from;
871 hugetlb_cgroup_uncharge_file_region(resv, rg,
872 rg->to - rg->from, true);
878 if (f <= rg->from) { /* Trim beginning of region */
879 hugetlb_cgroup_uncharge_file_region(resv, rg,
880 t - rg->from, false);
884 } else { /* Trim end of region */
885 hugetlb_cgroup_uncharge_file_region(resv, rg,
893 spin_unlock(&resv->lock);
899 * A rare out of memory error was encountered which prevented removal of
900 * the reserve map region for a page. The huge page itself was free'ed
901 * and removed from the page cache. This routine will adjust the subpool
902 * usage count, and the global reserve count if needed. By incrementing
903 * these counts, the reserve map entry which could not be deleted will
904 * appear as a "reserved" entry instead of simply dangling with incorrect
907 void hugetlb_fix_reserve_counts(struct inode *inode)
909 struct hugepage_subpool *spool = subpool_inode(inode);
911 bool reserved = false;
913 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
914 if (rsv_adjust > 0) {
915 struct hstate *h = hstate_inode(inode);
917 if (!hugetlb_acct_memory(h, 1))
919 } else if (!rsv_adjust) {
924 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
928 * Count and return the number of huge pages in the reserve map
929 * that intersect with the range [f, t).
931 static long region_count(struct resv_map *resv, long f, long t)
933 struct list_head *head = &resv->regions;
934 struct file_region *rg;
937 spin_lock(&resv->lock);
938 /* Locate each segment we overlap with, and count that overlap. */
939 list_for_each_entry(rg, head, link) {
948 seg_from = max(rg->from, f);
949 seg_to = min(rg->to, t);
951 chg += seg_to - seg_from;
953 spin_unlock(&resv->lock);
959 * Convert the address within this vma to the page offset within
960 * the mapping, in pagecache page units; huge pages here.
962 static pgoff_t vma_hugecache_offset(struct hstate *h,
963 struct vm_area_struct *vma, unsigned long address)
965 return ((address - vma->vm_start) >> huge_page_shift(h)) +
966 (vma->vm_pgoff >> huge_page_order(h));
969 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
970 unsigned long address)
972 return vma_hugecache_offset(hstate_vma(vma), vma, address);
974 EXPORT_SYMBOL_GPL(linear_hugepage_index);
977 * Return the size of the pages allocated when backing a VMA. In the majority
978 * cases this will be same size as used by the page table entries.
980 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
982 if (vma->vm_ops && vma->vm_ops->pagesize)
983 return vma->vm_ops->pagesize(vma);
986 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
989 * Return the page size being used by the MMU to back a VMA. In the majority
990 * of cases, the page size used by the kernel matches the MMU size. On
991 * architectures where it differs, an architecture-specific 'strong'
992 * version of this symbol is required.
994 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
996 return vma_kernel_pagesize(vma);
1000 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1001 * bits of the reservation map pointer, which are always clear due to
1004 #define HPAGE_RESV_OWNER (1UL << 0)
1005 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1006 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1009 * These helpers are used to track how many pages are reserved for
1010 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1011 * is guaranteed to have their future faults succeed.
1013 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1014 * the reserve counters are updated with the hugetlb_lock held. It is safe
1015 * to reset the VMA at fork() time as it is not in use yet and there is no
1016 * chance of the global counters getting corrupted as a result of the values.
1018 * The private mapping reservation is represented in a subtly different
1019 * manner to a shared mapping. A shared mapping has a region map associated
1020 * with the underlying file, this region map represents the backing file
1021 * pages which have ever had a reservation assigned which this persists even
1022 * after the page is instantiated. A private mapping has a region map
1023 * associated with the original mmap which is attached to all VMAs which
1024 * reference it, this region map represents those offsets which have consumed
1025 * reservation ie. where pages have been instantiated.
1027 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1029 return (unsigned long)vma->vm_private_data;
1032 static void set_vma_private_data(struct vm_area_struct *vma,
1033 unsigned long value)
1035 vma->vm_private_data = (void *)value;
1039 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1040 struct hugetlb_cgroup *h_cg,
1043 #ifdef CONFIG_CGROUP_HUGETLB
1045 resv_map->reservation_counter = NULL;
1046 resv_map->pages_per_hpage = 0;
1047 resv_map->css = NULL;
1049 resv_map->reservation_counter =
1050 &h_cg->rsvd_hugepage[hstate_index(h)];
1051 resv_map->pages_per_hpage = pages_per_huge_page(h);
1052 resv_map->css = &h_cg->css;
1057 struct resv_map *resv_map_alloc(void)
1059 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1060 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1062 if (!resv_map || !rg) {
1068 kref_init(&resv_map->refs);
1069 spin_lock_init(&resv_map->lock);
1070 INIT_LIST_HEAD(&resv_map->regions);
1072 resv_map->adds_in_progress = 0;
1074 * Initialize these to 0. On shared mappings, 0's here indicate these
1075 * fields don't do cgroup accounting. On private mappings, these will be
1076 * re-initialized to the proper values, to indicate that hugetlb cgroup
1077 * reservations are to be un-charged from here.
1079 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1081 INIT_LIST_HEAD(&resv_map->region_cache);
1082 list_add(&rg->link, &resv_map->region_cache);
1083 resv_map->region_cache_count = 1;
1088 void resv_map_release(struct kref *ref)
1090 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1091 struct list_head *head = &resv_map->region_cache;
1092 struct file_region *rg, *trg;
1094 /* Clear out any active regions before we release the map. */
1095 region_del(resv_map, 0, LONG_MAX);
1097 /* ... and any entries left in the cache */
1098 list_for_each_entry_safe(rg, trg, head, link) {
1099 list_del(&rg->link);
1103 VM_BUG_ON(resv_map->adds_in_progress);
1108 static inline struct resv_map *inode_resv_map(struct inode *inode)
1111 * At inode evict time, i_mapping may not point to the original
1112 * address space within the inode. This original address space
1113 * contains the pointer to the resv_map. So, always use the
1114 * address space embedded within the inode.
1115 * The VERY common case is inode->mapping == &inode->i_data but,
1116 * this may not be true for device special inodes.
1118 return (struct resv_map *)(&inode->i_data)->private_data;
1121 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1123 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1124 if (vma->vm_flags & VM_MAYSHARE) {
1125 struct address_space *mapping = vma->vm_file->f_mapping;
1126 struct inode *inode = mapping->host;
1128 return inode_resv_map(inode);
1131 return (struct resv_map *)(get_vma_private_data(vma) &
1136 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1138 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1139 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1141 set_vma_private_data(vma, (get_vma_private_data(vma) &
1142 HPAGE_RESV_MASK) | (unsigned long)map);
1145 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1147 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1148 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1150 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1153 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1155 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1157 return (get_vma_private_data(vma) & flag) != 0;
1160 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1162 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1164 * Clear vm_private_data
1165 * - For shared mappings this is a per-vma semaphore that may be
1166 * allocated in a subsequent call to hugetlb_vm_op_open.
1167 * Before clearing, make sure pointer is not associated with vma
1168 * as this will leak the structure. This is the case when called
1169 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1170 * been called to allocate a new structure.
1171 * - For MAP_PRIVATE mappings, this is the reserve map which does
1172 * not apply to children. Faults generated by the children are
1173 * not guaranteed to succeed, even if read-only.
1175 if (vma->vm_flags & VM_MAYSHARE) {
1176 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1178 if (vma_lock && vma_lock->vma != vma)
1179 vma->vm_private_data = NULL;
1181 vma->vm_private_data = NULL;
1185 * Reset and decrement one ref on hugepage private reservation.
1186 * Called with mm->mmap_sem writer semaphore held.
1187 * This function should be only used by move_vma() and operate on
1188 * same sized vma. It should never come here with last ref on the
1191 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1194 * Clear the old hugetlb private page reservation.
1195 * It has already been transferred to new_vma.
1197 * During a mremap() operation of a hugetlb vma we call move_vma()
1198 * which copies vma into new_vma and unmaps vma. After the copy
1199 * operation both new_vma and vma share a reference to the resv_map
1200 * struct, and at that point vma is about to be unmapped. We don't
1201 * want to return the reservation to the pool at unmap of vma because
1202 * the reservation still lives on in new_vma, so simply decrement the
1203 * ref here and remove the resv_map reference from this vma.
1205 struct resv_map *reservations = vma_resv_map(vma);
1207 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1208 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1209 kref_put(&reservations->refs, resv_map_release);
1212 hugetlb_dup_vma_private(vma);
1215 /* Returns true if the VMA has associated reserve pages */
1216 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1218 if (vma->vm_flags & VM_NORESERVE) {
1220 * This address is already reserved by other process(chg == 0),
1221 * so, we should decrement reserved count. Without decrementing,
1222 * reserve count remains after releasing inode, because this
1223 * allocated page will go into page cache and is regarded as
1224 * coming from reserved pool in releasing step. Currently, we
1225 * don't have any other solution to deal with this situation
1226 * properly, so add work-around here.
1228 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1234 /* Shared mappings always use reserves */
1235 if (vma->vm_flags & VM_MAYSHARE) {
1237 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1238 * be a region map for all pages. The only situation where
1239 * there is no region map is if a hole was punched via
1240 * fallocate. In this case, there really are no reserves to
1241 * use. This situation is indicated if chg != 0.
1250 * Only the process that called mmap() has reserves for
1253 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1255 * Like the shared case above, a hole punch or truncate
1256 * could have been performed on the private mapping.
1257 * Examine the value of chg to determine if reserves
1258 * actually exist or were previously consumed.
1259 * Very Subtle - The value of chg comes from a previous
1260 * call to vma_needs_reserves(). The reserve map for
1261 * private mappings has different (opposite) semantics
1262 * than that of shared mappings. vma_needs_reserves()
1263 * has already taken this difference in semantics into
1264 * account. Therefore, the meaning of chg is the same
1265 * as in the shared case above. Code could easily be
1266 * combined, but keeping it separate draws attention to
1267 * subtle differences.
1278 static void enqueue_huge_page(struct hstate *h, struct page *page)
1280 int nid = page_to_nid(page);
1282 lockdep_assert_held(&hugetlb_lock);
1283 VM_BUG_ON_PAGE(page_count(page), page);
1285 list_move(&page->lru, &h->hugepage_freelists[nid]);
1286 h->free_huge_pages++;
1287 h->free_huge_pages_node[nid]++;
1288 SetHPageFreed(page);
1291 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1294 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1296 lockdep_assert_held(&hugetlb_lock);
1297 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1298 if (pin && !is_longterm_pinnable_page(page))
1301 if (PageHWPoison(page))
1304 list_move(&page->lru, &h->hugepage_activelist);
1305 set_page_refcounted(page);
1306 ClearHPageFreed(page);
1307 h->free_huge_pages--;
1308 h->free_huge_pages_node[nid]--;
1315 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1318 unsigned int cpuset_mems_cookie;
1319 struct zonelist *zonelist;
1322 int node = NUMA_NO_NODE;
1324 zonelist = node_zonelist(nid, gfp_mask);
1327 cpuset_mems_cookie = read_mems_allowed_begin();
1328 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1331 if (!cpuset_zone_allowed(zone, gfp_mask))
1334 * no need to ask again on the same node. Pool is node rather than
1337 if (zone_to_nid(zone) == node)
1339 node = zone_to_nid(zone);
1341 page = dequeue_huge_page_node_exact(h, node);
1345 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1351 static unsigned long available_huge_pages(struct hstate *h)
1353 return h->free_huge_pages - h->resv_huge_pages;
1356 static struct page *dequeue_huge_page_vma(struct hstate *h,
1357 struct vm_area_struct *vma,
1358 unsigned long address, int avoid_reserve,
1361 struct page *page = NULL;
1362 struct mempolicy *mpol;
1364 nodemask_t *nodemask;
1368 * A child process with MAP_PRIVATE mappings created by their parent
1369 * have no page reserves. This check ensures that reservations are
1370 * not "stolen". The child may still get SIGKILLed
1372 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1375 /* If reserves cannot be used, ensure enough pages are in the pool */
1376 if (avoid_reserve && !available_huge_pages(h))
1379 gfp_mask = htlb_alloc_mask(h);
1380 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1382 if (mpol_is_preferred_many(mpol)) {
1383 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1385 /* Fallback to all nodes if page==NULL */
1390 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1392 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1393 SetHPageRestoreReserve(page);
1394 h->resv_huge_pages--;
1397 mpol_cond_put(mpol);
1405 * common helper functions for hstate_next_node_to_{alloc|free}.
1406 * We may have allocated or freed a huge page based on a different
1407 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1408 * be outside of *nodes_allowed. Ensure that we use an allowed
1409 * node for alloc or free.
1411 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1413 nid = next_node_in(nid, *nodes_allowed);
1414 VM_BUG_ON(nid >= MAX_NUMNODES);
1419 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1421 if (!node_isset(nid, *nodes_allowed))
1422 nid = next_node_allowed(nid, nodes_allowed);
1427 * returns the previously saved node ["this node"] from which to
1428 * allocate a persistent huge page for the pool and advance the
1429 * next node from which to allocate, handling wrap at end of node
1432 static int hstate_next_node_to_alloc(struct hstate *h,
1433 nodemask_t *nodes_allowed)
1437 VM_BUG_ON(!nodes_allowed);
1439 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1440 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1446 * helper for remove_pool_huge_page() - return the previously saved
1447 * node ["this node"] from which to free a huge page. Advance the
1448 * next node id whether or not we find a free huge page to free so
1449 * that the next attempt to free addresses the next node.
1451 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1455 VM_BUG_ON(!nodes_allowed);
1457 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1458 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1463 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1464 for (nr_nodes = nodes_weight(*mask); \
1466 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1469 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1470 for (nr_nodes = nodes_weight(*mask); \
1472 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1475 /* used to demote non-gigantic_huge pages as well */
1476 static void __destroy_compound_gigantic_page(struct page *page,
1477 unsigned int order, bool demote)
1480 int nr_pages = 1 << order;
1483 atomic_set(compound_mapcount_ptr(page), 0);
1484 atomic_set(compound_pincount_ptr(page), 0);
1486 for (i = 1; i < nr_pages; i++) {
1487 p = nth_page(page, i);
1489 clear_compound_head(p);
1491 set_page_refcounted(p);
1494 set_compound_order(page, 0);
1496 page[1].compound_nr = 0;
1498 __ClearPageHead(page);
1501 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1504 __destroy_compound_gigantic_page(page, order, true);
1507 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1508 static void destroy_compound_gigantic_page(struct page *page,
1511 __destroy_compound_gigantic_page(page, order, false);
1514 static void free_gigantic_page(struct page *page, unsigned int order)
1517 * If the page isn't allocated using the cma allocator,
1518 * cma_release() returns false.
1521 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1525 free_contig_range(page_to_pfn(page), 1 << order);
1528 #ifdef CONFIG_CONTIG_ALLOC
1529 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1530 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();
1541 if (hugetlb_cma[nid]) {
1542 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1543 huge_page_order(h), true);
1548 if (!(gfp_mask & __GFP_THISNODE)) {
1549 for_each_node_mask(node, *nodemask) {
1550 if (node == nid || !hugetlb_cma[node])
1553 page = cma_alloc(hugetlb_cma[node], nr_pages,
1554 huge_page_order(h), true);
1562 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1565 #else /* !CONFIG_CONTIG_ALLOC */
1566 static struct page *alloc_gigantic_page(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 page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1575 int nid, nodemask_t *nodemask)
1579 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1580 static inline void destroy_compound_gigantic_page(struct page *page,
1581 unsigned int order) { }
1584 static inline void __clear_hugetlb_destructor(struct hstate *h,
1587 lockdep_assert_held(&hugetlb_lock);
1592 * For non-gigantic pages set the destructor to the normal compound
1593 * page dtor. This is needed in case someone takes an additional
1594 * temporary ref to the page, and freeing is delayed until they drop
1597 * For gigantic pages set the destructor to the null dtor. This
1598 * destructor will never be called. Before freeing the gigantic
1599 * page destroy_compound_gigantic_folio will turn the folio into a
1600 * simple group of pages. After this the destructor does not
1604 if (hstate_is_gigantic(h))
1605 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1607 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1611 * Remove hugetlb page from lists.
1612 * If vmemmap exists for the page, update dtor so that the page appears
1613 * as just a compound page. Otherwise, wait until after allocating vmemmap
1616 * A reference is held on the page, except in the case of demote.
1618 * Must be called with hugetlb lock held.
1620 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1621 bool adjust_surplus,
1624 int nid = page_to_nid(page);
1626 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1627 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1629 lockdep_assert_held(&hugetlb_lock);
1630 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1633 list_del(&page->lru);
1635 if (HPageFreed(page)) {
1636 h->free_huge_pages--;
1637 h->free_huge_pages_node[nid]--;
1639 if (adjust_surplus) {
1640 h->surplus_huge_pages--;
1641 h->surplus_huge_pages_node[nid]--;
1645 * We can only clear the hugetlb destructor after allocating vmemmap
1646 * pages. Otherwise, someone (memory error handling) may try to write
1647 * to tail struct pages.
1649 if (!HPageVmemmapOptimized(page))
1650 __clear_hugetlb_destructor(h, page);
1653 * In the case of demote we do not ref count the page as it will soon
1654 * be turned into a page of smaller size.
1657 set_page_refcounted(page);
1660 h->nr_huge_pages_node[nid]--;
1663 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1664 bool adjust_surplus)
1666 __remove_hugetlb_page(h, page, adjust_surplus, false);
1669 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1670 bool adjust_surplus)
1672 __remove_hugetlb_page(h, page, adjust_surplus, true);
1675 static void add_hugetlb_page(struct hstate *h, struct page *page,
1676 bool adjust_surplus)
1679 int nid = page_to_nid(page);
1681 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1683 lockdep_assert_held(&hugetlb_lock);
1685 INIT_LIST_HEAD(&page->lru);
1687 h->nr_huge_pages_node[nid]++;
1689 if (adjust_surplus) {
1690 h->surplus_huge_pages++;
1691 h->surplus_huge_pages_node[nid]++;
1694 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1695 set_page_private(page, 0);
1697 * We have to set HPageVmemmapOptimized again as above
1698 * set_page_private(page, 0) cleared it.
1700 SetHPageVmemmapOptimized(page);
1703 * This page is about to be managed by the hugetlb allocator and
1704 * should have no users. Drop our reference, and check for others
1707 zeroed = put_page_testzero(page);
1710 * It is VERY unlikely soneone else has taken a ref on
1711 * the page. In this case, we simply return as the
1712 * hugetlb destructor (free_huge_page) will be called
1713 * when this other ref is dropped.
1717 arch_clear_hugepage_flags(page);
1718 enqueue_huge_page(h, page);
1721 static void __update_and_free_page(struct hstate *h, struct page *page)
1724 struct page *subpage;
1725 bool clear_dtor = HPageVmemmapOptimized(page);
1727 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1731 * If we don't know which subpages are hwpoisoned, we can't free
1732 * the hugepage, so it's leaked intentionally.
1734 if (HPageRawHwpUnreliable(page))
1737 if (hugetlb_vmemmap_restore(h, page)) {
1738 spin_lock_irq(&hugetlb_lock);
1740 * If we cannot allocate vmemmap pages, just refuse to free the
1741 * page and put the page back on the hugetlb free list and treat
1742 * as a surplus page.
1744 add_hugetlb_page(h, page, true);
1745 spin_unlock_irq(&hugetlb_lock);
1750 * Move PageHWPoison flag from head page to the raw error pages,
1751 * which makes any healthy subpages reusable.
1753 if (unlikely(PageHWPoison(page)))
1754 hugetlb_clear_page_hwpoison(page);
1757 * If vmemmap pages were allocated above, then we need to clear the
1758 * hugetlb destructor under the hugetlb lock.
1761 spin_lock_irq(&hugetlb_lock);
1762 __clear_hugetlb_destructor(h, page);
1763 spin_unlock_irq(&hugetlb_lock);
1766 for (i = 0; i < pages_per_huge_page(h); i++) {
1767 subpage = nth_page(page, i);
1768 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1769 1 << PG_referenced | 1 << PG_dirty |
1770 1 << PG_active | 1 << PG_private |
1775 * Non-gigantic pages demoted from CMA allocated gigantic pages
1776 * need to be given back to CMA in free_gigantic_page.
1778 if (hstate_is_gigantic(h) ||
1779 hugetlb_cma_page(page, huge_page_order(h))) {
1780 destroy_compound_gigantic_page(page, huge_page_order(h));
1781 free_gigantic_page(page, huge_page_order(h));
1783 __free_pages(page, huge_page_order(h));
1788 * As update_and_free_page() can be called under any context, so we cannot
1789 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1790 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1791 * the vmemmap pages.
1793 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1794 * freed and frees them one-by-one. As the page->mapping pointer is going
1795 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1796 * structure of a lockless linked list of huge pages to be freed.
1798 static LLIST_HEAD(hpage_freelist);
1800 static void free_hpage_workfn(struct work_struct *work)
1802 struct llist_node *node;
1804 node = llist_del_all(&hpage_freelist);
1810 page = container_of((struct address_space **)node,
1811 struct page, mapping);
1813 page->mapping = NULL;
1815 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1816 * is going to trigger because a previous call to
1817 * remove_hugetlb_page() will set_compound_page_dtor(page,
1818 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1820 h = size_to_hstate(page_size(page));
1822 __update_and_free_page(h, page);
1827 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1829 static inline void flush_free_hpage_work(struct hstate *h)
1831 if (hugetlb_vmemmap_optimizable(h))
1832 flush_work(&free_hpage_work);
1835 static void update_and_free_page(struct hstate *h, struct page *page,
1838 if (!HPageVmemmapOptimized(page) || !atomic) {
1839 __update_and_free_page(h, page);
1844 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1846 * Only call schedule_work() if hpage_freelist is previously
1847 * empty. Otherwise, schedule_work() had been called but the workfn
1848 * hasn't retrieved the list yet.
1850 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1851 schedule_work(&free_hpage_work);
1854 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1856 struct page *page, *t_page;
1858 list_for_each_entry_safe(page, t_page, list, lru) {
1859 update_and_free_page(h, page, false);
1864 struct hstate *size_to_hstate(unsigned long size)
1868 for_each_hstate(h) {
1869 if (huge_page_size(h) == size)
1875 void free_huge_page(struct page *page)
1878 * Can't pass hstate in here because it is called from the
1879 * compound page destructor.
1881 struct hstate *h = page_hstate(page);
1882 int nid = page_to_nid(page);
1883 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1884 bool restore_reserve;
1885 unsigned long flags;
1887 VM_BUG_ON_PAGE(page_count(page), page);
1888 VM_BUG_ON_PAGE(page_mapcount(page), page);
1890 hugetlb_set_page_subpool(page, NULL);
1892 __ClearPageAnonExclusive(page);
1893 page->mapping = NULL;
1894 restore_reserve = HPageRestoreReserve(page);
1895 ClearHPageRestoreReserve(page);
1898 * If HPageRestoreReserve was set on page, page allocation consumed a
1899 * reservation. If the page was associated with a subpool, there
1900 * would have been a page reserved in the subpool before allocation
1901 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1902 * reservation, do not call hugepage_subpool_put_pages() as this will
1903 * remove the reserved page from the subpool.
1905 if (!restore_reserve) {
1907 * A return code of zero implies that the subpool will be
1908 * under its minimum size if the reservation is not restored
1909 * after page is free. Therefore, force restore_reserve
1912 if (hugepage_subpool_put_pages(spool, 1) == 0)
1913 restore_reserve = true;
1916 spin_lock_irqsave(&hugetlb_lock, flags);
1917 ClearHPageMigratable(page);
1918 hugetlb_cgroup_uncharge_page(hstate_index(h),
1919 pages_per_huge_page(h), page);
1920 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1921 pages_per_huge_page(h), page);
1922 if (restore_reserve)
1923 h->resv_huge_pages++;
1925 if (HPageTemporary(page)) {
1926 remove_hugetlb_page(h, page, false);
1927 spin_unlock_irqrestore(&hugetlb_lock, flags);
1928 update_and_free_page(h, page, true);
1929 } else if (h->surplus_huge_pages_node[nid]) {
1930 /* remove the page from active list */
1931 remove_hugetlb_page(h, page, true);
1932 spin_unlock_irqrestore(&hugetlb_lock, flags);
1933 update_and_free_page(h, page, true);
1935 arch_clear_hugepage_flags(page);
1936 enqueue_huge_page(h, page);
1937 spin_unlock_irqrestore(&hugetlb_lock, flags);
1942 * Must be called with the hugetlb lock held
1944 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1946 lockdep_assert_held(&hugetlb_lock);
1948 h->nr_huge_pages_node[nid]++;
1951 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1953 hugetlb_vmemmap_optimize(h, page);
1954 INIT_LIST_HEAD(&page->lru);
1955 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1956 hugetlb_set_page_subpool(page, NULL);
1957 set_hugetlb_cgroup(page, NULL);
1958 set_hugetlb_cgroup_rsvd(page, NULL);
1961 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1963 __prep_new_huge_page(h, page);
1964 spin_lock_irq(&hugetlb_lock);
1965 __prep_account_new_huge_page(h, nid);
1966 spin_unlock_irq(&hugetlb_lock);
1969 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1973 int nr_pages = 1 << order;
1976 /* we rely on prep_new_huge_page to set the destructor */
1977 set_compound_order(page, order);
1978 __ClearPageReserved(page);
1979 __SetPageHead(page);
1980 for (i = 0; i < nr_pages; i++) {
1981 p = nth_page(page, i);
1984 * For gigantic hugepages allocated through bootmem at
1985 * boot, it's safer to be consistent with the not-gigantic
1986 * hugepages and clear the PG_reserved bit from all tail pages
1987 * too. Otherwise drivers using get_user_pages() to access tail
1988 * pages may get the reference counting wrong if they see
1989 * PG_reserved set on a tail page (despite the head page not
1990 * having PG_reserved set). Enforcing this consistency between
1991 * head and tail pages allows drivers to optimize away a check
1992 * on the head page when they need know if put_page() is needed
1993 * after get_user_pages().
1995 if (i != 0) /* head page cleared above */
1996 __ClearPageReserved(p);
1998 * Subtle and very unlikely
2000 * Gigantic 'page allocators' such as memblock or cma will
2001 * return a set of pages with each page ref counted. We need
2002 * to turn this set of pages into a compound page with tail
2003 * page ref counts set to zero. Code such as speculative page
2004 * cache adding could take a ref on a 'to be' tail page.
2005 * We need to respect any increased ref count, and only set
2006 * the ref count to zero if count is currently 1. If count
2007 * is not 1, we return an error. An error return indicates
2008 * the set of pages can not be converted to a gigantic page.
2009 * The caller who allocated the pages should then discard the
2010 * pages using the appropriate free interface.
2012 * In the case of demote, the ref count will be zero.
2015 if (!page_ref_freeze(p, 1)) {
2016 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2020 VM_BUG_ON_PAGE(page_count(p), p);
2023 set_compound_head(p, page);
2025 atomic_set(compound_mapcount_ptr(page), -1);
2026 atomic_set(compound_pincount_ptr(page), 0);
2030 /* undo page modifications made above */
2031 for (j = 0; j < i; j++) {
2032 p = nth_page(page, j);
2034 clear_compound_head(p);
2035 set_page_refcounted(p);
2037 /* need to clear PG_reserved on remaining tail pages */
2038 for (; j < nr_pages; j++) {
2039 p = nth_page(page, j);
2040 __ClearPageReserved(p);
2042 set_compound_order(page, 0);
2044 page[1].compound_nr = 0;
2046 __ClearPageHead(page);
2050 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
2052 return __prep_compound_gigantic_page(page, order, false);
2055 static bool prep_compound_gigantic_page_for_demote(struct page *page,
2058 return __prep_compound_gigantic_page(page, order, true);
2062 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2063 * transparent huge pages. See the PageTransHuge() documentation for more
2066 int PageHuge(struct page *page)
2068 if (!PageCompound(page))
2071 page = compound_head(page);
2072 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
2074 EXPORT_SYMBOL_GPL(PageHuge);
2077 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
2078 * normal or transparent huge pages.
2080 int PageHeadHuge(struct page *page_head)
2082 if (!PageHead(page_head))
2085 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
2087 EXPORT_SYMBOL_GPL(PageHeadHuge);
2090 * Find and lock address space (mapping) in write mode.
2092 * Upon entry, the page is locked which means that page_mapping() is
2093 * stable. Due to locking order, we can only trylock_write. If we can
2094 * not get the lock, simply return NULL to caller.
2096 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2098 struct address_space *mapping = page_mapping(hpage);
2103 if (i_mmap_trylock_write(mapping))
2109 pgoff_t hugetlb_basepage_index(struct page *page)
2111 struct page *page_head = compound_head(page);
2112 pgoff_t index = page_index(page_head);
2113 unsigned long compound_idx;
2115 if (compound_order(page_head) >= MAX_ORDER)
2116 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2118 compound_idx = page - page_head;
2120 return (index << compound_order(page_head)) + compound_idx;
2123 static struct page *alloc_buddy_huge_page(struct hstate *h,
2124 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2125 nodemask_t *node_alloc_noretry)
2127 int order = huge_page_order(h);
2129 bool alloc_try_hard = true;
2133 * By default we always try hard to allocate the page with
2134 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2135 * a loop (to adjust global huge page counts) and previous allocation
2136 * failed, do not continue to try hard on the same node. Use the
2137 * node_alloc_noretry bitmap to manage this state information.
2139 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2140 alloc_try_hard = false;
2141 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2143 gfp_mask |= __GFP_RETRY_MAYFAIL;
2144 if (nid == NUMA_NO_NODE)
2145 nid = numa_mem_id();
2147 page = __alloc_pages(gfp_mask, order, nid, nmask);
2149 /* Freeze head page */
2150 if (page && !page_ref_freeze(page, 1)) {
2151 __free_pages(page, order);
2152 if (retry) { /* retry once */
2156 /* WOW! twice in a row. */
2157 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2162 __count_vm_event(HTLB_BUDDY_PGALLOC);
2164 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2167 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2168 * indicates an overall state change. Clear bit so that we resume
2169 * normal 'try hard' allocations.
2171 if (node_alloc_noretry && page && !alloc_try_hard)
2172 node_clear(nid, *node_alloc_noretry);
2175 * If we tried hard to get a page but failed, set bit so that
2176 * subsequent attempts will not try as hard until there is an
2177 * overall state change.
2179 if (node_alloc_noretry && !page && alloc_try_hard)
2180 node_set(nid, *node_alloc_noretry);
2186 * Common helper to allocate a fresh hugetlb page. All specific allocators
2187 * should use this function to get new hugetlb pages
2189 * Note that returned page is 'frozen': ref count of head page and all tail
2192 static struct page *alloc_fresh_huge_page(struct hstate *h,
2193 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2194 nodemask_t *node_alloc_noretry)
2200 if (hstate_is_gigantic(h))
2201 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
2203 page = alloc_buddy_huge_page(h, gfp_mask,
2204 nid, nmask, node_alloc_noretry);
2208 if (hstate_is_gigantic(h)) {
2209 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
2211 * Rare failure to convert pages to compound page.
2212 * Free pages and try again - ONCE!
2214 free_gigantic_page(page, huge_page_order(h));
2222 prep_new_huge_page(h, page, page_to_nid(page));
2228 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2231 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2232 nodemask_t *node_alloc_noretry)
2236 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2238 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2239 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
2240 node_alloc_noretry);
2248 free_huge_page(page); /* free it into the hugepage allocator */
2254 * Remove huge page from pool from next node to free. Attempt to keep
2255 * persistent huge pages more or less balanced over allowed nodes.
2256 * This routine only 'removes' the hugetlb page. The caller must make
2257 * an additional call to free the page to low level allocators.
2258 * Called with hugetlb_lock locked.
2260 static struct page *remove_pool_huge_page(struct hstate *h,
2261 nodemask_t *nodes_allowed,
2265 struct page *page = NULL;
2267 lockdep_assert_held(&hugetlb_lock);
2268 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2270 * If we're returning unused surplus pages, only examine
2271 * nodes with surplus pages.
2273 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2274 !list_empty(&h->hugepage_freelists[node])) {
2275 page = list_entry(h->hugepage_freelists[node].next,
2277 remove_hugetlb_page(h, page, acct_surplus);
2286 * Dissolve a given free hugepage into free buddy pages. This function does
2287 * nothing for in-use hugepages and non-hugepages.
2288 * This function returns values like below:
2290 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2291 * when the system is under memory pressure and the feature of
2292 * freeing unused vmemmap pages associated with each hugetlb page
2294 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2295 * (allocated or reserved.)
2296 * 0: successfully dissolved free hugepages or the page is not a
2297 * hugepage (considered as already dissolved)
2299 int dissolve_free_huge_page(struct page *page)
2304 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2305 if (!PageHuge(page))
2308 spin_lock_irq(&hugetlb_lock);
2309 if (!PageHuge(page)) {
2314 if (!page_count(page)) {
2315 struct page *head = compound_head(page);
2316 struct hstate *h = page_hstate(head);
2317 if (!available_huge_pages(h))
2321 * We should make sure that the page is already on the free list
2322 * when it is dissolved.
2324 if (unlikely(!HPageFreed(head))) {
2325 spin_unlock_irq(&hugetlb_lock);
2329 * Theoretically, we should return -EBUSY when we
2330 * encounter this race. In fact, we have a chance
2331 * to successfully dissolve the page if we do a
2332 * retry. Because the race window is quite small.
2333 * If we seize this opportunity, it is an optimization
2334 * for increasing the success rate of dissolving page.
2339 remove_hugetlb_page(h, head, false);
2340 h->max_huge_pages--;
2341 spin_unlock_irq(&hugetlb_lock);
2344 * Normally update_and_free_page will allocate required vmemmmap
2345 * before freeing the page. update_and_free_page will fail to
2346 * free the page if it can not allocate required vmemmap. We
2347 * need to adjust max_huge_pages if the page is not freed.
2348 * Attempt to allocate vmemmmap here so that we can take
2349 * appropriate action on failure.
2351 rc = hugetlb_vmemmap_restore(h, head);
2353 update_and_free_page(h, head, false);
2355 spin_lock_irq(&hugetlb_lock);
2356 add_hugetlb_page(h, head, false);
2357 h->max_huge_pages++;
2358 spin_unlock_irq(&hugetlb_lock);
2364 spin_unlock_irq(&hugetlb_lock);
2369 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2370 * make specified memory blocks removable from the system.
2371 * Note that this will dissolve a free gigantic hugepage completely, if any
2372 * part of it lies within the given range.
2373 * Also note that if dissolve_free_huge_page() returns with an error, all
2374 * free hugepages that were dissolved before that error are lost.
2376 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2384 if (!hugepages_supported())
2387 order = huge_page_order(&default_hstate);
2389 order = min(order, huge_page_order(h));
2391 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2392 page = pfn_to_page(pfn);
2393 rc = dissolve_free_huge_page(page);
2402 * Allocates a fresh surplus page from the page allocator.
2404 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2405 int nid, nodemask_t *nmask)
2407 struct page *page = NULL;
2409 if (hstate_is_gigantic(h))
2412 spin_lock_irq(&hugetlb_lock);
2413 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2415 spin_unlock_irq(&hugetlb_lock);
2417 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2421 spin_lock_irq(&hugetlb_lock);
2423 * We could have raced with the pool size change.
2424 * Double check that and simply deallocate the new page
2425 * if we would end up overcommiting the surpluses. Abuse
2426 * temporary page to workaround the nasty free_huge_page
2429 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2430 SetHPageTemporary(page);
2431 spin_unlock_irq(&hugetlb_lock);
2432 free_huge_page(page);
2436 h->surplus_huge_pages++;
2437 h->surplus_huge_pages_node[page_to_nid(page)]++;
2440 spin_unlock_irq(&hugetlb_lock);
2445 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2446 int nid, nodemask_t *nmask)
2450 if (hstate_is_gigantic(h))
2453 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2457 /* fresh huge pages are frozen */
2458 set_page_refcounted(page);
2461 * We do not account these pages as surplus because they are only
2462 * temporary and will be released properly on the last reference
2464 SetHPageTemporary(page);
2470 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2473 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2474 struct vm_area_struct *vma, unsigned long addr)
2476 struct page *page = NULL;
2477 struct mempolicy *mpol;
2478 gfp_t gfp_mask = htlb_alloc_mask(h);
2480 nodemask_t *nodemask;
2482 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2483 if (mpol_is_preferred_many(mpol)) {
2484 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2486 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2487 page = alloc_surplus_huge_page(h, gfp, nid, nodemask);
2489 /* Fallback to all nodes if page==NULL */
2494 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2495 mpol_cond_put(mpol);
2499 /* page migration callback function */
2500 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2501 nodemask_t *nmask, gfp_t gfp_mask)
2503 spin_lock_irq(&hugetlb_lock);
2504 if (available_huge_pages(h)) {
2507 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2509 spin_unlock_irq(&hugetlb_lock);
2513 spin_unlock_irq(&hugetlb_lock);
2515 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2518 /* mempolicy aware migration callback */
2519 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2520 unsigned long address)
2522 struct mempolicy *mpol;
2523 nodemask_t *nodemask;
2528 gfp_mask = htlb_alloc_mask(h);
2529 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2530 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2531 mpol_cond_put(mpol);
2537 * Increase the hugetlb pool such that it can accommodate a reservation
2540 static int gather_surplus_pages(struct hstate *h, long delta)
2541 __must_hold(&hugetlb_lock)
2543 LIST_HEAD(surplus_list);
2544 struct page *page, *tmp;
2547 long needed, allocated;
2548 bool alloc_ok = true;
2550 lockdep_assert_held(&hugetlb_lock);
2551 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2553 h->resv_huge_pages += delta;
2561 spin_unlock_irq(&hugetlb_lock);
2562 for (i = 0; i < needed; i++) {
2563 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2564 NUMA_NO_NODE, NULL);
2569 list_add(&page->lru, &surplus_list);
2575 * After retaking hugetlb_lock, we need to recalculate 'needed'
2576 * because either resv_huge_pages or free_huge_pages may have changed.
2578 spin_lock_irq(&hugetlb_lock);
2579 needed = (h->resv_huge_pages + delta) -
2580 (h->free_huge_pages + allocated);
2585 * We were not able to allocate enough pages to
2586 * satisfy the entire reservation so we free what
2587 * we've allocated so far.
2592 * The surplus_list now contains _at_least_ the number of extra pages
2593 * needed to accommodate the reservation. Add the appropriate number
2594 * of pages to the hugetlb pool and free the extras back to the buddy
2595 * allocator. Commit the entire reservation here to prevent another
2596 * process from stealing the pages as they are added to the pool but
2597 * before they are reserved.
2599 needed += allocated;
2600 h->resv_huge_pages += delta;
2603 /* Free the needed pages to the hugetlb pool */
2604 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2607 /* Add the page to the hugetlb allocator */
2608 enqueue_huge_page(h, page);
2611 spin_unlock_irq(&hugetlb_lock);
2614 * Free unnecessary surplus pages to the buddy allocator.
2615 * Pages have no ref count, call free_huge_page directly.
2617 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2618 free_huge_page(page);
2619 spin_lock_irq(&hugetlb_lock);
2625 * This routine has two main purposes:
2626 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2627 * in unused_resv_pages. This corresponds to the prior adjustments made
2628 * to the associated reservation map.
2629 * 2) Free any unused surplus pages that may have been allocated to satisfy
2630 * the reservation. As many as unused_resv_pages may be freed.
2632 static void return_unused_surplus_pages(struct hstate *h,
2633 unsigned long unused_resv_pages)
2635 unsigned long nr_pages;
2637 LIST_HEAD(page_list);
2639 lockdep_assert_held(&hugetlb_lock);
2640 /* Uncommit the reservation */
2641 h->resv_huge_pages -= unused_resv_pages;
2643 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2647 * Part (or even all) of the reservation could have been backed
2648 * by pre-allocated pages. Only free surplus pages.
2650 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2653 * We want to release as many surplus pages as possible, spread
2654 * evenly across all nodes with memory. Iterate across these nodes
2655 * until we can no longer free unreserved surplus pages. This occurs
2656 * when the nodes with surplus pages have no free pages.
2657 * remove_pool_huge_page() will balance the freed pages across the
2658 * on-line nodes with memory and will handle the hstate accounting.
2660 while (nr_pages--) {
2661 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2665 list_add(&page->lru, &page_list);
2669 spin_unlock_irq(&hugetlb_lock);
2670 update_and_free_pages_bulk(h, &page_list);
2671 spin_lock_irq(&hugetlb_lock);
2676 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2677 * are used by the huge page allocation routines to manage reservations.
2679 * vma_needs_reservation is called to determine if the huge page at addr
2680 * within the vma has an associated reservation. If a reservation is
2681 * needed, the value 1 is returned. The caller is then responsible for
2682 * managing the global reservation and subpool usage counts. After
2683 * the huge page has been allocated, vma_commit_reservation is called
2684 * to add the page to the reservation map. If the page allocation fails,
2685 * the reservation must be ended instead of committed. vma_end_reservation
2686 * is called in such cases.
2688 * In the normal case, vma_commit_reservation returns the same value
2689 * as the preceding vma_needs_reservation call. The only time this
2690 * is not the case is if a reserve map was changed between calls. It
2691 * is the responsibility of the caller to notice the difference and
2692 * take appropriate action.
2694 * vma_add_reservation is used in error paths where a reservation must
2695 * be restored when a newly allocated huge page must be freed. It is
2696 * to be called after calling vma_needs_reservation to determine if a
2697 * reservation exists.
2699 * vma_del_reservation is used in error paths where an entry in the reserve
2700 * map was created during huge page allocation and must be removed. It is to
2701 * be called after calling vma_needs_reservation to determine if a reservation
2704 enum vma_resv_mode {
2711 static long __vma_reservation_common(struct hstate *h,
2712 struct vm_area_struct *vma, unsigned long addr,
2713 enum vma_resv_mode mode)
2715 struct resv_map *resv;
2718 long dummy_out_regions_needed;
2720 resv = vma_resv_map(vma);
2724 idx = vma_hugecache_offset(h, vma, addr);
2726 case VMA_NEEDS_RESV:
2727 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2728 /* We assume that vma_reservation_* routines always operate on
2729 * 1 page, and that adding to resv map a 1 page entry can only
2730 * ever require 1 region.
2732 VM_BUG_ON(dummy_out_regions_needed != 1);
2734 case VMA_COMMIT_RESV:
2735 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2736 /* region_add calls of range 1 should never fail. */
2740 region_abort(resv, idx, idx + 1, 1);
2744 if (vma->vm_flags & VM_MAYSHARE) {
2745 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2746 /* region_add calls of range 1 should never fail. */
2749 region_abort(resv, idx, idx + 1, 1);
2750 ret = region_del(resv, idx, idx + 1);
2754 if (vma->vm_flags & VM_MAYSHARE) {
2755 region_abort(resv, idx, idx + 1, 1);
2756 ret = region_del(resv, idx, idx + 1);
2758 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2759 /* region_add calls of range 1 should never fail. */
2767 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2770 * We know private mapping must have HPAGE_RESV_OWNER set.
2772 * In most cases, reserves always exist for private mappings.
2773 * However, a file associated with mapping could have been
2774 * hole punched or truncated after reserves were consumed.
2775 * As subsequent fault on such a range will not use reserves.
2776 * Subtle - The reserve map for private mappings has the
2777 * opposite meaning than that of shared mappings. If NO
2778 * entry is in the reserve map, it means a reservation exists.
2779 * If an entry exists in the reserve map, it means the
2780 * reservation has already been consumed. As a result, the
2781 * return value of this routine is the opposite of the
2782 * value returned from reserve map manipulation routines above.
2791 static long vma_needs_reservation(struct hstate *h,
2792 struct vm_area_struct *vma, unsigned long addr)
2794 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2797 static long vma_commit_reservation(struct hstate *h,
2798 struct vm_area_struct *vma, unsigned long addr)
2800 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2803 static void vma_end_reservation(struct hstate *h,
2804 struct vm_area_struct *vma, unsigned long addr)
2806 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2809 static long vma_add_reservation(struct hstate *h,
2810 struct vm_area_struct *vma, unsigned long addr)
2812 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2815 static long vma_del_reservation(struct hstate *h,
2816 struct vm_area_struct *vma, unsigned long addr)
2818 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2822 * This routine is called to restore reservation information on error paths.
2823 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2824 * the hugetlb mutex should remain held when calling this routine.
2826 * It handles two specific cases:
2827 * 1) A reservation was in place and the page consumed the reservation.
2828 * HPageRestoreReserve is set in the page.
2829 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2830 * not set. However, alloc_huge_page always updates the reserve map.
2832 * In case 1, free_huge_page later in the error path will increment the
2833 * global reserve count. But, free_huge_page does not have enough context
2834 * to adjust the reservation map. This case deals primarily with private
2835 * mappings. Adjust the reserve map here to be consistent with global
2836 * reserve count adjustments to be made by free_huge_page. Make sure the
2837 * reserve map indicates there is a reservation present.
2839 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2841 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2842 unsigned long address, struct page *page)
2844 long rc = vma_needs_reservation(h, vma, address);
2846 if (HPageRestoreReserve(page)) {
2847 if (unlikely(rc < 0))
2849 * Rare out of memory condition in reserve map
2850 * manipulation. Clear HPageRestoreReserve so that
2851 * global reserve count will not be incremented
2852 * by free_huge_page. This will make it appear
2853 * as though the reservation for this page was
2854 * consumed. This may prevent the task from
2855 * faulting in the page at a later time. This
2856 * is better than inconsistent global huge page
2857 * accounting of reserve counts.
2859 ClearHPageRestoreReserve(page);
2861 (void)vma_add_reservation(h, vma, address);
2863 vma_end_reservation(h, vma, address);
2867 * This indicates there is an entry in the reserve map
2868 * not added by alloc_huge_page. We know it was added
2869 * before the alloc_huge_page call, otherwise
2870 * HPageRestoreReserve would be set on the page.
2871 * Remove the entry so that a subsequent allocation
2872 * does not consume a reservation.
2874 rc = vma_del_reservation(h, vma, address);
2877 * VERY rare out of memory condition. Since
2878 * we can not delete the entry, set
2879 * HPageRestoreReserve so that the reserve
2880 * count will be incremented when the page
2881 * is freed. This reserve will be consumed
2882 * on a subsequent allocation.
2884 SetHPageRestoreReserve(page);
2885 } else if (rc < 0) {
2887 * Rare out of memory condition from
2888 * vma_needs_reservation call. Memory allocation is
2889 * only attempted if a new entry is needed. Therefore,
2890 * this implies there is not an entry in the
2893 * For shared mappings, no entry in the map indicates
2894 * no reservation. We are done.
2896 if (!(vma->vm_flags & VM_MAYSHARE))
2898 * For private mappings, no entry indicates
2899 * a reservation is present. Since we can
2900 * not add an entry, set SetHPageRestoreReserve
2901 * on the page so reserve count will be
2902 * incremented when freed. This reserve will
2903 * be consumed on a subsequent allocation.
2905 SetHPageRestoreReserve(page);
2908 * No reservation present, do nothing
2910 vma_end_reservation(h, vma, address);
2915 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2916 * @h: struct hstate old page belongs to
2917 * @old_page: Old page to dissolve
2918 * @list: List to isolate the page in case we need to
2919 * Returns 0 on success, otherwise negated error.
2921 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2922 struct list_head *list)
2924 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2925 int nid = page_to_nid(old_page);
2926 struct page *new_page;
2930 * Before dissolving the page, we need to allocate a new one for the
2931 * pool to remain stable. Here, we allocate the page and 'prep' it
2932 * by doing everything but actually updating counters and adding to
2933 * the pool. This simplifies and let us do most of the processing
2936 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2939 __prep_new_huge_page(h, new_page);
2942 spin_lock_irq(&hugetlb_lock);
2943 if (!PageHuge(old_page)) {
2945 * Freed from under us. Drop new_page too.
2948 } else if (page_count(old_page)) {
2950 * Someone has grabbed the page, try to isolate it here.
2951 * Fail with -EBUSY if not possible.
2953 spin_unlock_irq(&hugetlb_lock);
2954 ret = isolate_hugetlb(old_page, list);
2955 spin_lock_irq(&hugetlb_lock);
2957 } else if (!HPageFreed(old_page)) {
2959 * Page's refcount is 0 but it has not been enqueued in the
2960 * freelist yet. Race window is small, so we can succeed here if
2963 spin_unlock_irq(&hugetlb_lock);
2968 * Ok, old_page is still a genuine free hugepage. Remove it from
2969 * the freelist and decrease the counters. These will be
2970 * incremented again when calling __prep_account_new_huge_page()
2971 * and enqueue_huge_page() for new_page. The counters will remain
2972 * stable since this happens under the lock.
2974 remove_hugetlb_page(h, old_page, false);
2977 * Ref count on new page is already zero as it was dropped
2978 * earlier. It can be directly added to the pool free list.
2980 __prep_account_new_huge_page(h, nid);
2981 enqueue_huge_page(h, new_page);
2984 * Pages have been replaced, we can safely free the old one.
2986 spin_unlock_irq(&hugetlb_lock);
2987 update_and_free_page(h, old_page, false);
2993 spin_unlock_irq(&hugetlb_lock);
2994 /* Page has a zero ref count, but needs a ref to be freed */
2995 set_page_refcounted(new_page);
2996 update_and_free_page(h, new_page, false);
3001 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3008 * The page might have been dissolved from under our feet, so make sure
3009 * to carefully check the state under the lock.
3010 * Return success when racing as if we dissolved the page ourselves.
3012 spin_lock_irq(&hugetlb_lock);
3013 if (PageHuge(page)) {
3014 head = compound_head(page);
3015 h = page_hstate(head);
3017 spin_unlock_irq(&hugetlb_lock);
3020 spin_unlock_irq(&hugetlb_lock);
3023 * Fence off gigantic pages as there is a cyclic dependency between
3024 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3025 * of bailing out right away without further retrying.
3027 if (hstate_is_gigantic(h))
3030 if (page_count(head) && !isolate_hugetlb(head, list))
3032 else if (!page_count(head))
3033 ret = alloc_and_dissolve_huge_page(h, head, list);
3038 struct page *alloc_huge_page(struct vm_area_struct *vma,
3039 unsigned long addr, int avoid_reserve)
3041 struct hugepage_subpool *spool = subpool_vma(vma);
3042 struct hstate *h = hstate_vma(vma);
3044 long map_chg, map_commit;
3047 struct hugetlb_cgroup *h_cg;
3048 bool deferred_reserve;
3050 idx = hstate_index(h);
3052 * Examine the region/reserve map to determine if the process
3053 * has a reservation for the page to be allocated. A return
3054 * code of zero indicates a reservation exists (no change).
3056 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3058 return ERR_PTR(-ENOMEM);
3061 * Processes that did not create the mapping will have no
3062 * reserves as indicated by the region/reserve map. Check
3063 * that the allocation will not exceed the subpool limit.
3064 * Allocations for MAP_NORESERVE mappings also need to be
3065 * checked against any subpool limit.
3067 if (map_chg || avoid_reserve) {
3068 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3070 vma_end_reservation(h, vma, addr);
3071 return ERR_PTR(-ENOSPC);
3075 * Even though there was no reservation in the region/reserve
3076 * map, there could be reservations associated with the
3077 * subpool that can be used. This would be indicated if the
3078 * return value of hugepage_subpool_get_pages() is zero.
3079 * However, if avoid_reserve is specified we still avoid even
3080 * the subpool reservations.
3086 /* If this allocation is not consuming a reservation, charge it now.
3088 deferred_reserve = map_chg || avoid_reserve;
3089 if (deferred_reserve) {
3090 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3091 idx, pages_per_huge_page(h), &h_cg);
3093 goto out_subpool_put;
3096 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3098 goto out_uncharge_cgroup_reservation;
3100 spin_lock_irq(&hugetlb_lock);
3102 * glb_chg is passed to indicate whether or not a page must be taken
3103 * from the global free pool (global change). gbl_chg == 0 indicates
3104 * a reservation exists for the allocation.
3106 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
3108 spin_unlock_irq(&hugetlb_lock);
3109 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
3111 goto out_uncharge_cgroup;
3112 spin_lock_irq(&hugetlb_lock);
3113 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3114 SetHPageRestoreReserve(page);
3115 h->resv_huge_pages--;
3117 list_add(&page->lru, &h->hugepage_activelist);
3118 set_page_refcounted(page);
3121 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
3122 /* If allocation is not consuming a reservation, also store the
3123 * hugetlb_cgroup pointer on the page.
3125 if (deferred_reserve) {
3126 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3130 spin_unlock_irq(&hugetlb_lock);
3132 hugetlb_set_page_subpool(page, spool);
3134 map_commit = vma_commit_reservation(h, vma, addr);
3135 if (unlikely(map_chg > map_commit)) {
3137 * The page was added to the reservation map between
3138 * vma_needs_reservation and vma_commit_reservation.
3139 * This indicates a race with hugetlb_reserve_pages.
3140 * Adjust for the subpool count incremented above AND
3141 * in hugetlb_reserve_pages for the same page. Also,
3142 * the reservation count added in hugetlb_reserve_pages
3143 * no longer applies.
3147 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3148 hugetlb_acct_memory(h, -rsv_adjust);
3149 if (deferred_reserve)
3150 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
3151 pages_per_huge_page(h), page);
3155 out_uncharge_cgroup:
3156 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3157 out_uncharge_cgroup_reservation:
3158 if (deferred_reserve)
3159 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3162 if (map_chg || avoid_reserve)
3163 hugepage_subpool_put_pages(spool, 1);
3164 vma_end_reservation(h, vma, addr);
3165 return ERR_PTR(-ENOSPC);
3168 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3169 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3170 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3172 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3175 /* do node specific alloc */
3176 if (nid != NUMA_NO_NODE) {
3177 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3178 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3183 /* allocate from next node when distributing huge pages */
3184 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3185 m = memblock_alloc_try_nid_raw(
3186 huge_page_size(h), huge_page_size(h),
3187 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3189 * Use the beginning of the huge page to store the
3190 * huge_bootmem_page struct (until gather_bootmem
3191 * puts them into the mem_map).
3199 /* Put them into a private list first because mem_map is not up yet */
3200 INIT_LIST_HEAD(&m->list);
3201 list_add(&m->list, &huge_boot_pages);
3207 * Put bootmem huge pages into the standard lists after mem_map is up.
3208 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3210 static void __init gather_bootmem_prealloc(void)
3212 struct huge_bootmem_page *m;
3214 list_for_each_entry(m, &huge_boot_pages, list) {
3215 struct page *page = virt_to_page(m);
3216 struct hstate *h = m->hstate;
3218 VM_BUG_ON(!hstate_is_gigantic(h));
3219 WARN_ON(page_count(page) != 1);
3220 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3221 WARN_ON(PageReserved(page));
3222 prep_new_huge_page(h, page, page_to_nid(page));
3223 free_huge_page(page); /* add to the hugepage allocator */
3225 /* VERY unlikely inflated ref count on a tail page */
3226 free_gigantic_page(page, huge_page_order(h));
3230 * We need to restore the 'stolen' pages to totalram_pages
3231 * in order to fix confusing memory reports from free(1) and
3232 * other side-effects, like CommitLimit going negative.
3234 adjust_managed_page_count(page, pages_per_huge_page(h));
3238 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3243 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3244 if (hstate_is_gigantic(h)) {
3245 if (!alloc_bootmem_huge_page(h, nid))
3249 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3251 page = alloc_fresh_huge_page(h, gfp_mask, nid,
3252 &node_states[N_MEMORY], NULL);
3255 free_huge_page(page); /* free it into the hugepage allocator */
3259 if (i == h->max_huge_pages_node[nid])
3262 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3263 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3264 h->max_huge_pages_node[nid], buf, nid, i);
3265 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3266 h->max_huge_pages_node[nid] = i;
3269 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3272 nodemask_t *node_alloc_noretry;
3273 bool node_specific_alloc = false;
3275 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3276 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3277 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3281 /* do node specific alloc */
3282 for_each_online_node(i) {
3283 if (h->max_huge_pages_node[i] > 0) {
3284 hugetlb_hstate_alloc_pages_onenode(h, i);
3285 node_specific_alloc = true;
3289 if (node_specific_alloc)
3292 /* below will do all node balanced alloc */
3293 if (!hstate_is_gigantic(h)) {
3295 * Bit mask controlling how hard we retry per-node allocations.
3296 * Ignore errors as lower level routines can deal with
3297 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3298 * time, we are likely in bigger trouble.
3300 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3303 /* allocations done at boot time */
3304 node_alloc_noretry = NULL;
3307 /* bit mask controlling how hard we retry per-node allocations */
3308 if (node_alloc_noretry)
3309 nodes_clear(*node_alloc_noretry);
3311 for (i = 0; i < h->max_huge_pages; ++i) {
3312 if (hstate_is_gigantic(h)) {
3313 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3315 } else if (!alloc_pool_huge_page(h,
3316 &node_states[N_MEMORY],
3317 node_alloc_noretry))
3321 if (i < h->max_huge_pages) {
3324 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3325 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3326 h->max_huge_pages, buf, i);
3327 h->max_huge_pages = i;
3329 kfree(node_alloc_noretry);
3332 static void __init hugetlb_init_hstates(void)
3334 struct hstate *h, *h2;
3336 for_each_hstate(h) {
3337 /* oversize hugepages were init'ed in early boot */
3338 if (!hstate_is_gigantic(h))
3339 hugetlb_hstate_alloc_pages(h);
3342 * Set demote order for each hstate. Note that
3343 * h->demote_order is initially 0.
3344 * - We can not demote gigantic pages if runtime freeing
3345 * is not supported, so skip this.
3346 * - If CMA allocation is possible, we can not demote
3347 * HUGETLB_PAGE_ORDER or smaller size pages.
3349 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3351 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3353 for_each_hstate(h2) {
3356 if (h2->order < h->order &&
3357 h2->order > h->demote_order)
3358 h->demote_order = h2->order;
3363 static void __init report_hugepages(void)
3367 for_each_hstate(h) {
3370 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3371 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3372 buf, h->free_huge_pages);
3373 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3374 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3378 #ifdef CONFIG_HIGHMEM
3379 static void try_to_free_low(struct hstate *h, unsigned long count,
3380 nodemask_t *nodes_allowed)
3383 LIST_HEAD(page_list);
3385 lockdep_assert_held(&hugetlb_lock);
3386 if (hstate_is_gigantic(h))
3390 * Collect pages to be freed on a list, and free after dropping lock
3392 for_each_node_mask(i, *nodes_allowed) {
3393 struct page *page, *next;
3394 struct list_head *freel = &h->hugepage_freelists[i];
3395 list_for_each_entry_safe(page, next, freel, lru) {
3396 if (count >= h->nr_huge_pages)
3398 if (PageHighMem(page))
3400 remove_hugetlb_page(h, page, false);
3401 list_add(&page->lru, &page_list);
3406 spin_unlock_irq(&hugetlb_lock);
3407 update_and_free_pages_bulk(h, &page_list);
3408 spin_lock_irq(&hugetlb_lock);
3411 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3412 nodemask_t *nodes_allowed)
3418 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3419 * balanced by operating on them in a round-robin fashion.
3420 * Returns 1 if an adjustment was made.
3422 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3427 lockdep_assert_held(&hugetlb_lock);
3428 VM_BUG_ON(delta != -1 && delta != 1);
3431 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3432 if (h->surplus_huge_pages_node[node])
3436 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3437 if (h->surplus_huge_pages_node[node] <
3438 h->nr_huge_pages_node[node])
3445 h->surplus_huge_pages += delta;
3446 h->surplus_huge_pages_node[node] += delta;
3450 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3451 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3452 nodemask_t *nodes_allowed)
3454 unsigned long min_count, ret;
3456 LIST_HEAD(page_list);
3457 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3460 * Bit mask controlling how hard we retry per-node allocations.
3461 * If we can not allocate the bit mask, do not attempt to allocate
3462 * the requested huge pages.
3464 if (node_alloc_noretry)
3465 nodes_clear(*node_alloc_noretry);
3470 * resize_lock mutex prevents concurrent adjustments to number of
3471 * pages in hstate via the proc/sysfs interfaces.
3473 mutex_lock(&h->resize_lock);
3474 flush_free_hpage_work(h);
3475 spin_lock_irq(&hugetlb_lock);
3478 * Check for a node specific request.
3479 * Changing node specific huge page count may require a corresponding
3480 * change to the global count. In any case, the passed node mask
3481 * (nodes_allowed) will restrict alloc/free to the specified node.
3483 if (nid != NUMA_NO_NODE) {
3484 unsigned long old_count = count;
3486 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3488 * User may have specified a large count value which caused the
3489 * above calculation to overflow. In this case, they wanted
3490 * to allocate as many huge pages as possible. Set count to
3491 * largest possible value to align with their intention.
3493 if (count < old_count)
3498 * Gigantic pages runtime allocation depend on the capability for large
3499 * page range allocation.
3500 * If the system does not provide this feature, return an error when
3501 * the user tries to allocate gigantic pages but let the user free the
3502 * boottime allocated gigantic pages.
3504 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3505 if (count > persistent_huge_pages(h)) {
3506 spin_unlock_irq(&hugetlb_lock);
3507 mutex_unlock(&h->resize_lock);
3508 NODEMASK_FREE(node_alloc_noretry);
3511 /* Fall through to decrease pool */
3515 * Increase the pool size
3516 * First take pages out of surplus state. Then make up the
3517 * remaining difference by allocating fresh huge pages.
3519 * We might race with alloc_surplus_huge_page() here and be unable
3520 * to convert a surplus huge page to a normal huge page. That is
3521 * not critical, though, it just means the overall size of the
3522 * pool might be one hugepage larger than it needs to be, but
3523 * within all the constraints specified by the sysctls.
3525 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3526 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3530 while (count > persistent_huge_pages(h)) {
3532 * If this allocation races such that we no longer need the
3533 * page, free_huge_page will handle it by freeing the page
3534 * and reducing the surplus.
3536 spin_unlock_irq(&hugetlb_lock);
3538 /* yield cpu to avoid soft lockup */
3541 ret = alloc_pool_huge_page(h, nodes_allowed,
3542 node_alloc_noretry);
3543 spin_lock_irq(&hugetlb_lock);
3547 /* Bail for signals. Probably ctrl-c from user */
3548 if (signal_pending(current))
3553 * Decrease the pool size
3554 * First return free pages to the buddy allocator (being careful
3555 * to keep enough around to satisfy reservations). Then place
3556 * pages into surplus state as needed so the pool will shrink
3557 * to the desired size as pages become free.
3559 * By placing pages into the surplus state independent of the
3560 * overcommit value, we are allowing the surplus pool size to
3561 * exceed overcommit. There are few sane options here. Since
3562 * alloc_surplus_huge_page() is checking the global counter,
3563 * though, we'll note that we're not allowed to exceed surplus
3564 * and won't grow the pool anywhere else. Not until one of the
3565 * sysctls are changed, or the surplus pages go out of use.
3567 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3568 min_count = max(count, min_count);
3569 try_to_free_low(h, min_count, nodes_allowed);
3572 * Collect pages to be removed on list without dropping lock
3574 while (min_count < persistent_huge_pages(h)) {
3575 page = remove_pool_huge_page(h, nodes_allowed, 0);
3579 list_add(&page->lru, &page_list);
3581 /* free the pages after dropping lock */
3582 spin_unlock_irq(&hugetlb_lock);
3583 update_and_free_pages_bulk(h, &page_list);
3584 flush_free_hpage_work(h);
3585 spin_lock_irq(&hugetlb_lock);
3587 while (count < persistent_huge_pages(h)) {
3588 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3592 h->max_huge_pages = persistent_huge_pages(h);
3593 spin_unlock_irq(&hugetlb_lock);
3594 mutex_unlock(&h->resize_lock);
3596 NODEMASK_FREE(node_alloc_noretry);
3601 static int demote_free_huge_page(struct hstate *h, struct page *page)
3603 int i, nid = page_to_nid(page);
3604 struct hstate *target_hstate;
3605 struct page *subpage;
3608 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3610 remove_hugetlb_page_for_demote(h, page, false);
3611 spin_unlock_irq(&hugetlb_lock);
3613 rc = hugetlb_vmemmap_restore(h, page);
3615 /* Allocation of vmemmmap failed, we can not demote page */
3616 spin_lock_irq(&hugetlb_lock);
3617 set_page_refcounted(page);
3618 add_hugetlb_page(h, page, false);
3623 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3624 * sizes as it will not ref count pages.
3626 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3629 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3630 * Without the mutex, pages added to target hstate could be marked
3633 * Note that we already hold h->resize_lock. To prevent deadlock,
3634 * use the convention of always taking larger size hstate mutex first.
3636 mutex_lock(&target_hstate->resize_lock);
3637 for (i = 0; i < pages_per_huge_page(h);
3638 i += pages_per_huge_page(target_hstate)) {
3639 subpage = nth_page(page, i);
3640 if (hstate_is_gigantic(target_hstate))
3641 prep_compound_gigantic_page_for_demote(subpage,
3642 target_hstate->order);
3644 prep_compound_page(subpage, target_hstate->order);
3645 set_page_private(subpage, 0);
3646 prep_new_huge_page(target_hstate, subpage, nid);
3647 free_huge_page(subpage);
3649 mutex_unlock(&target_hstate->resize_lock);
3651 spin_lock_irq(&hugetlb_lock);
3654 * Not absolutely necessary, but for consistency update max_huge_pages
3655 * based on pool changes for the demoted page.
3657 h->max_huge_pages--;
3658 target_hstate->max_huge_pages +=
3659 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3664 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3665 __must_hold(&hugetlb_lock)
3670 lockdep_assert_held(&hugetlb_lock);
3672 /* We should never get here if no demote order */
3673 if (!h->demote_order) {
3674 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3675 return -EINVAL; /* internal error */
3678 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3679 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3680 if (PageHWPoison(page))
3683 return demote_free_huge_page(h, page);
3688 * Only way to get here is if all pages on free lists are poisoned.
3689 * Return -EBUSY so that caller will not retry.
3694 #define HSTATE_ATTR_RO(_name) \
3695 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3697 #define HSTATE_ATTR_WO(_name) \
3698 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3700 #define HSTATE_ATTR(_name) \
3701 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3703 static struct kobject *hugepages_kobj;
3704 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3706 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3708 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3712 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3713 if (hstate_kobjs[i] == kobj) {
3715 *nidp = NUMA_NO_NODE;
3719 return kobj_to_node_hstate(kobj, nidp);
3722 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3723 struct kobj_attribute *attr, char *buf)
3726 unsigned long nr_huge_pages;
3729 h = kobj_to_hstate(kobj, &nid);
3730 if (nid == NUMA_NO_NODE)
3731 nr_huge_pages = h->nr_huge_pages;
3733 nr_huge_pages = h->nr_huge_pages_node[nid];
3735 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3738 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3739 struct hstate *h, int nid,
3740 unsigned long count, size_t len)
3743 nodemask_t nodes_allowed, *n_mask;
3745 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3748 if (nid == NUMA_NO_NODE) {
3750 * global hstate attribute
3752 if (!(obey_mempolicy &&
3753 init_nodemask_of_mempolicy(&nodes_allowed)))
3754 n_mask = &node_states[N_MEMORY];
3756 n_mask = &nodes_allowed;
3759 * Node specific request. count adjustment happens in
3760 * set_max_huge_pages() after acquiring hugetlb_lock.
3762 init_nodemask_of_node(&nodes_allowed, nid);
3763 n_mask = &nodes_allowed;
3766 err = set_max_huge_pages(h, count, nid, n_mask);
3768 return err ? err : len;
3771 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3772 struct kobject *kobj, const char *buf,
3776 unsigned long count;
3780 err = kstrtoul(buf, 10, &count);
3784 h = kobj_to_hstate(kobj, &nid);
3785 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3788 static ssize_t nr_hugepages_show(struct kobject *kobj,
3789 struct kobj_attribute *attr, char *buf)
3791 return nr_hugepages_show_common(kobj, attr, buf);
3794 static ssize_t nr_hugepages_store(struct kobject *kobj,
3795 struct kobj_attribute *attr, const char *buf, size_t len)
3797 return nr_hugepages_store_common(false, kobj, buf, len);
3799 HSTATE_ATTR(nr_hugepages);
3804 * hstate attribute for optionally mempolicy-based constraint on persistent
3805 * huge page alloc/free.
3807 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3808 struct kobj_attribute *attr,
3811 return nr_hugepages_show_common(kobj, attr, buf);
3814 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3815 struct kobj_attribute *attr, const char *buf, size_t len)
3817 return nr_hugepages_store_common(true, kobj, buf, len);
3819 HSTATE_ATTR(nr_hugepages_mempolicy);
3823 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3824 struct kobj_attribute *attr, char *buf)
3826 struct hstate *h = kobj_to_hstate(kobj, NULL);
3827 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3830 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3831 struct kobj_attribute *attr, const char *buf, size_t count)
3834 unsigned long input;
3835 struct hstate *h = kobj_to_hstate(kobj, NULL);
3837 if (hstate_is_gigantic(h))
3840 err = kstrtoul(buf, 10, &input);
3844 spin_lock_irq(&hugetlb_lock);
3845 h->nr_overcommit_huge_pages = input;
3846 spin_unlock_irq(&hugetlb_lock);
3850 HSTATE_ATTR(nr_overcommit_hugepages);
3852 static ssize_t free_hugepages_show(struct kobject *kobj,
3853 struct kobj_attribute *attr, char *buf)
3856 unsigned long free_huge_pages;
3859 h = kobj_to_hstate(kobj, &nid);
3860 if (nid == NUMA_NO_NODE)
3861 free_huge_pages = h->free_huge_pages;
3863 free_huge_pages = h->free_huge_pages_node[nid];
3865 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3867 HSTATE_ATTR_RO(free_hugepages);
3869 static ssize_t resv_hugepages_show(struct kobject *kobj,
3870 struct kobj_attribute *attr, char *buf)
3872 struct hstate *h = kobj_to_hstate(kobj, NULL);
3873 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3875 HSTATE_ATTR_RO(resv_hugepages);
3877 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3878 struct kobj_attribute *attr, char *buf)
3881 unsigned long surplus_huge_pages;
3884 h = kobj_to_hstate(kobj, &nid);
3885 if (nid == NUMA_NO_NODE)
3886 surplus_huge_pages = h->surplus_huge_pages;
3888 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3890 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3892 HSTATE_ATTR_RO(surplus_hugepages);
3894 static ssize_t demote_store(struct kobject *kobj,
3895 struct kobj_attribute *attr, const char *buf, size_t len)
3897 unsigned long nr_demote;
3898 unsigned long nr_available;
3899 nodemask_t nodes_allowed, *n_mask;
3904 err = kstrtoul(buf, 10, &nr_demote);
3907 h = kobj_to_hstate(kobj, &nid);
3909 if (nid != NUMA_NO_NODE) {
3910 init_nodemask_of_node(&nodes_allowed, nid);
3911 n_mask = &nodes_allowed;
3913 n_mask = &node_states[N_MEMORY];
3916 /* Synchronize with other sysfs operations modifying huge pages */
3917 mutex_lock(&h->resize_lock);
3918 spin_lock_irq(&hugetlb_lock);
3922 * Check for available pages to demote each time thorough the
3923 * loop as demote_pool_huge_page will drop hugetlb_lock.
3925 if (nid != NUMA_NO_NODE)
3926 nr_available = h->free_huge_pages_node[nid];
3928 nr_available = h->free_huge_pages;
3929 nr_available -= h->resv_huge_pages;
3933 err = demote_pool_huge_page(h, n_mask);
3940 spin_unlock_irq(&hugetlb_lock);
3941 mutex_unlock(&h->resize_lock);
3947 HSTATE_ATTR_WO(demote);
3949 static ssize_t demote_size_show(struct kobject *kobj,
3950 struct kobj_attribute *attr, char *buf)
3952 struct hstate *h = kobj_to_hstate(kobj, NULL);
3953 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3955 return sysfs_emit(buf, "%lukB\n", demote_size);
3958 static ssize_t demote_size_store(struct kobject *kobj,
3959 struct kobj_attribute *attr,
3960 const char *buf, size_t count)
3962 struct hstate *h, *demote_hstate;
3963 unsigned long demote_size;
3964 unsigned int demote_order;
3966 demote_size = (unsigned long)memparse(buf, NULL);
3968 demote_hstate = size_to_hstate(demote_size);
3971 demote_order = demote_hstate->order;
3972 if (demote_order < HUGETLB_PAGE_ORDER)
3975 /* demote order must be smaller than hstate order */
3976 h = kobj_to_hstate(kobj, NULL);
3977 if (demote_order >= h->order)
3980 /* resize_lock synchronizes access to demote size and writes */
3981 mutex_lock(&h->resize_lock);
3982 h->demote_order = demote_order;
3983 mutex_unlock(&h->resize_lock);
3987 HSTATE_ATTR(demote_size);
3989 static struct attribute *hstate_attrs[] = {
3990 &nr_hugepages_attr.attr,
3991 &nr_overcommit_hugepages_attr.attr,
3992 &free_hugepages_attr.attr,
3993 &resv_hugepages_attr.attr,
3994 &surplus_hugepages_attr.attr,
3996 &nr_hugepages_mempolicy_attr.attr,
4001 static const struct attribute_group hstate_attr_group = {
4002 .attrs = hstate_attrs,
4005 static struct attribute *hstate_demote_attrs[] = {
4006 &demote_size_attr.attr,
4011 static const struct attribute_group hstate_demote_attr_group = {
4012 .attrs = hstate_demote_attrs,
4015 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4016 struct kobject **hstate_kobjs,
4017 const struct attribute_group *hstate_attr_group)
4020 int hi = hstate_index(h);
4022 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4023 if (!hstate_kobjs[hi])
4026 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4028 kobject_put(hstate_kobjs[hi]);
4029 hstate_kobjs[hi] = NULL;
4033 if (h->demote_order) {
4034 retval = sysfs_create_group(hstate_kobjs[hi],
4035 &hstate_demote_attr_group);
4037 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4038 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4039 kobject_put(hstate_kobjs[hi]);
4040 hstate_kobjs[hi] = NULL;
4049 static bool hugetlb_sysfs_initialized __ro_after_init;
4052 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4053 * with node devices in node_devices[] using a parallel array. The array
4054 * index of a node device or _hstate == node id.
4055 * This is here to avoid any static dependency of the node device driver, in
4056 * the base kernel, on the hugetlb module.
4058 struct node_hstate {
4059 struct kobject *hugepages_kobj;
4060 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4062 static struct node_hstate node_hstates[MAX_NUMNODES];
4065 * A subset of global hstate attributes for node devices
4067 static struct attribute *per_node_hstate_attrs[] = {
4068 &nr_hugepages_attr.attr,
4069 &free_hugepages_attr.attr,
4070 &surplus_hugepages_attr.attr,
4074 static const struct attribute_group per_node_hstate_attr_group = {
4075 .attrs = per_node_hstate_attrs,
4079 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4080 * Returns node id via non-NULL nidp.
4082 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4086 for (nid = 0; nid < nr_node_ids; nid++) {
4087 struct node_hstate *nhs = &node_hstates[nid];
4089 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4090 if (nhs->hstate_kobjs[i] == kobj) {
4102 * Unregister hstate attributes from a single node device.
4103 * No-op if no hstate attributes attached.
4105 void hugetlb_unregister_node(struct node *node)
4108 struct node_hstate *nhs = &node_hstates[node->dev.id];
4110 if (!nhs->hugepages_kobj)
4111 return; /* no hstate attributes */
4113 for_each_hstate(h) {
4114 int idx = hstate_index(h);
4115 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4119 if (h->demote_order)
4120 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4121 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4122 kobject_put(hstate_kobj);
4123 nhs->hstate_kobjs[idx] = NULL;
4126 kobject_put(nhs->hugepages_kobj);
4127 nhs->hugepages_kobj = NULL;
4132 * Register hstate attributes for a single node device.
4133 * No-op if attributes already registered.
4135 void hugetlb_register_node(struct node *node)
4138 struct node_hstate *nhs = &node_hstates[node->dev.id];
4141 if (!hugetlb_sysfs_initialized)
4144 if (nhs->hugepages_kobj)
4145 return; /* already allocated */
4147 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4149 if (!nhs->hugepages_kobj)
4152 for_each_hstate(h) {
4153 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4155 &per_node_hstate_attr_group);
4157 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4158 h->name, node->dev.id);
4159 hugetlb_unregister_node(node);
4166 * hugetlb init time: register hstate attributes for all registered node
4167 * devices of nodes that have memory. All on-line nodes should have
4168 * registered their associated device by this time.
4170 static void __init hugetlb_register_all_nodes(void)
4174 for_each_online_node(nid)
4175 hugetlb_register_node(node_devices[nid]);
4177 #else /* !CONFIG_NUMA */
4179 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4187 static void hugetlb_register_all_nodes(void) { }
4192 static void __init hugetlb_cma_check(void);
4194 static inline __init void hugetlb_cma_check(void)
4199 static void __init hugetlb_sysfs_init(void)
4204 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4205 if (!hugepages_kobj)
4208 for_each_hstate(h) {
4209 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4210 hstate_kobjs, &hstate_attr_group);
4212 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4216 hugetlb_sysfs_initialized = true;
4218 hugetlb_register_all_nodes();
4221 static int __init hugetlb_init(void)
4225 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4228 if (!hugepages_supported()) {
4229 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4230 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4235 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4236 * architectures depend on setup being done here.
4238 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4239 if (!parsed_default_hugepagesz) {
4241 * If we did not parse a default huge page size, set
4242 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4243 * number of huge pages for this default size was implicitly
4244 * specified, set that here as well.
4245 * Note that the implicit setting will overwrite an explicit
4246 * setting. A warning will be printed in this case.
4248 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4249 if (default_hstate_max_huge_pages) {
4250 if (default_hstate.max_huge_pages) {
4253 string_get_size(huge_page_size(&default_hstate),
4254 1, STRING_UNITS_2, buf, 32);
4255 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4256 default_hstate.max_huge_pages, buf);
4257 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4258 default_hstate_max_huge_pages);
4260 default_hstate.max_huge_pages =
4261 default_hstate_max_huge_pages;
4263 for_each_online_node(i)
4264 default_hstate.max_huge_pages_node[i] =
4265 default_hugepages_in_node[i];
4269 hugetlb_cma_check();
4270 hugetlb_init_hstates();
4271 gather_bootmem_prealloc();
4274 hugetlb_sysfs_init();
4275 hugetlb_cgroup_file_init();
4278 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4280 num_fault_mutexes = 1;
4282 hugetlb_fault_mutex_table =
4283 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4285 BUG_ON(!hugetlb_fault_mutex_table);
4287 for (i = 0; i < num_fault_mutexes; i++)
4288 mutex_init(&hugetlb_fault_mutex_table[i]);
4291 subsys_initcall(hugetlb_init);
4293 /* Overwritten by architectures with more huge page sizes */
4294 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4296 return size == HPAGE_SIZE;
4299 void __init hugetlb_add_hstate(unsigned int order)
4304 if (size_to_hstate(PAGE_SIZE << order)) {
4307 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4309 h = &hstates[hugetlb_max_hstate++];
4310 mutex_init(&h->resize_lock);
4312 h->mask = ~(huge_page_size(h) - 1);
4313 for (i = 0; i < MAX_NUMNODES; ++i)
4314 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4315 INIT_LIST_HEAD(&h->hugepage_activelist);
4316 h->next_nid_to_alloc = first_memory_node;
4317 h->next_nid_to_free = first_memory_node;
4318 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4319 huge_page_size(h)/SZ_1K);
4324 bool __init __weak hugetlb_node_alloc_supported(void)
4329 static void __init hugepages_clear_pages_in_node(void)
4331 if (!hugetlb_max_hstate) {
4332 default_hstate_max_huge_pages = 0;
4333 memset(default_hugepages_in_node, 0,
4334 sizeof(default_hugepages_in_node));
4336 parsed_hstate->max_huge_pages = 0;
4337 memset(parsed_hstate->max_huge_pages_node, 0,
4338 sizeof(parsed_hstate->max_huge_pages_node));
4343 * hugepages command line processing
4344 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4345 * specification. If not, ignore the hugepages value. hugepages can also
4346 * be the first huge page command line option in which case it implicitly
4347 * specifies the number of huge pages for the default size.
4349 static int __init hugepages_setup(char *s)
4352 static unsigned long *last_mhp;
4353 int node = NUMA_NO_NODE;
4358 if (!parsed_valid_hugepagesz) {
4359 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4360 parsed_valid_hugepagesz = true;
4365 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4366 * yet, so this hugepages= parameter goes to the "default hstate".
4367 * Otherwise, it goes with the previously parsed hugepagesz or
4368 * default_hugepagesz.
4370 else if (!hugetlb_max_hstate)
4371 mhp = &default_hstate_max_huge_pages;
4373 mhp = &parsed_hstate->max_huge_pages;
4375 if (mhp == last_mhp) {
4376 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4382 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4384 /* Parameter is node format */
4385 if (p[count] == ':') {
4386 if (!hugetlb_node_alloc_supported()) {
4387 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4390 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4392 node = array_index_nospec(tmp, MAX_NUMNODES);
4394 /* Parse hugepages */
4395 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4397 if (!hugetlb_max_hstate)
4398 default_hugepages_in_node[node] = tmp;
4400 parsed_hstate->max_huge_pages_node[node] = tmp;
4402 /* Go to parse next node*/
4403 if (p[count] == ',')
4416 * Global state is always initialized later in hugetlb_init.
4417 * But we need to allocate gigantic hstates here early to still
4418 * use the bootmem allocator.
4420 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4421 hugetlb_hstate_alloc_pages(parsed_hstate);
4428 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4429 hugepages_clear_pages_in_node();
4432 __setup("hugepages=", hugepages_setup);
4435 * hugepagesz command line processing
4436 * A specific huge page size can only be specified once with hugepagesz.
4437 * hugepagesz is followed by hugepages on the command line. The global
4438 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4439 * hugepagesz argument was valid.
4441 static int __init hugepagesz_setup(char *s)
4446 parsed_valid_hugepagesz = false;
4447 size = (unsigned long)memparse(s, NULL);
4449 if (!arch_hugetlb_valid_size(size)) {
4450 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4454 h = size_to_hstate(size);
4457 * hstate for this size already exists. This is normally
4458 * an error, but is allowed if the existing hstate is the
4459 * default hstate. More specifically, it is only allowed if
4460 * the number of huge pages for the default hstate was not
4461 * previously specified.
4463 if (!parsed_default_hugepagesz || h != &default_hstate ||
4464 default_hstate.max_huge_pages) {
4465 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4470 * No need to call hugetlb_add_hstate() as hstate already
4471 * exists. But, do set parsed_hstate so that a following
4472 * hugepages= parameter will be applied to this hstate.
4475 parsed_valid_hugepagesz = true;
4479 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4480 parsed_valid_hugepagesz = true;
4483 __setup("hugepagesz=", hugepagesz_setup);
4486 * default_hugepagesz command line input
4487 * Only one instance of default_hugepagesz allowed on command line.
4489 static int __init default_hugepagesz_setup(char *s)
4494 parsed_valid_hugepagesz = false;
4495 if (parsed_default_hugepagesz) {
4496 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4500 size = (unsigned long)memparse(s, NULL);
4502 if (!arch_hugetlb_valid_size(size)) {
4503 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4507 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4508 parsed_valid_hugepagesz = true;
4509 parsed_default_hugepagesz = true;
4510 default_hstate_idx = hstate_index(size_to_hstate(size));
4513 * The number of default huge pages (for this size) could have been
4514 * specified as the first hugetlb parameter: hugepages=X. If so,
4515 * then default_hstate_max_huge_pages is set. If the default huge
4516 * page size is gigantic (>= MAX_ORDER), then the pages must be
4517 * allocated here from bootmem allocator.
4519 if (default_hstate_max_huge_pages) {
4520 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4521 for_each_online_node(i)
4522 default_hstate.max_huge_pages_node[i] =
4523 default_hugepages_in_node[i];
4524 if (hstate_is_gigantic(&default_hstate))
4525 hugetlb_hstate_alloc_pages(&default_hstate);
4526 default_hstate_max_huge_pages = 0;
4531 __setup("default_hugepagesz=", default_hugepagesz_setup);
4533 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4536 struct mempolicy *mpol = get_task_policy(current);
4539 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4540 * (from policy_nodemask) specifically for hugetlb case
4542 if (mpol->mode == MPOL_BIND &&
4543 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4544 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4545 return &mpol->nodes;
4550 static unsigned int allowed_mems_nr(struct hstate *h)
4553 unsigned int nr = 0;
4554 nodemask_t *mbind_nodemask;
4555 unsigned int *array = h->free_huge_pages_node;
4556 gfp_t gfp_mask = htlb_alloc_mask(h);
4558 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4559 for_each_node_mask(node, cpuset_current_mems_allowed) {
4560 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4567 #ifdef CONFIG_SYSCTL
4568 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4569 void *buffer, size_t *length,
4570 loff_t *ppos, unsigned long *out)
4572 struct ctl_table dup_table;
4575 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4576 * can duplicate the @table and alter the duplicate of it.
4579 dup_table.data = out;
4581 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4584 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4585 struct ctl_table *table, int write,
4586 void *buffer, size_t *length, loff_t *ppos)
4588 struct hstate *h = &default_hstate;
4589 unsigned long tmp = h->max_huge_pages;
4592 if (!hugepages_supported())
4595 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4601 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4602 NUMA_NO_NODE, tmp, *length);
4607 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4608 void *buffer, size_t *length, loff_t *ppos)
4611 return hugetlb_sysctl_handler_common(false, table, write,
4612 buffer, length, ppos);
4616 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4617 void *buffer, size_t *length, loff_t *ppos)
4619 return hugetlb_sysctl_handler_common(true, table, write,
4620 buffer, length, ppos);
4622 #endif /* CONFIG_NUMA */
4624 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4625 void *buffer, size_t *length, loff_t *ppos)
4627 struct hstate *h = &default_hstate;
4631 if (!hugepages_supported())
4634 tmp = h->nr_overcommit_huge_pages;
4636 if (write && hstate_is_gigantic(h))
4639 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4645 spin_lock_irq(&hugetlb_lock);
4646 h->nr_overcommit_huge_pages = tmp;
4647 spin_unlock_irq(&hugetlb_lock);
4653 #endif /* CONFIG_SYSCTL */
4655 void hugetlb_report_meminfo(struct seq_file *m)
4658 unsigned long total = 0;
4660 if (!hugepages_supported())
4663 for_each_hstate(h) {
4664 unsigned long count = h->nr_huge_pages;
4666 total += huge_page_size(h) * count;
4668 if (h == &default_hstate)
4670 "HugePages_Total: %5lu\n"
4671 "HugePages_Free: %5lu\n"
4672 "HugePages_Rsvd: %5lu\n"
4673 "HugePages_Surp: %5lu\n"
4674 "Hugepagesize: %8lu kB\n",
4678 h->surplus_huge_pages,
4679 huge_page_size(h) / SZ_1K);
4682 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4685 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4687 struct hstate *h = &default_hstate;
4689 if (!hugepages_supported())
4692 return sysfs_emit_at(buf, len,
4693 "Node %d HugePages_Total: %5u\n"
4694 "Node %d HugePages_Free: %5u\n"
4695 "Node %d HugePages_Surp: %5u\n",
4696 nid, h->nr_huge_pages_node[nid],
4697 nid, h->free_huge_pages_node[nid],
4698 nid, h->surplus_huge_pages_node[nid]);
4701 void hugetlb_show_meminfo_node(int nid)
4705 if (!hugepages_supported())
4709 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4711 h->nr_huge_pages_node[nid],
4712 h->free_huge_pages_node[nid],
4713 h->surplus_huge_pages_node[nid],
4714 huge_page_size(h) / SZ_1K);
4717 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4719 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4720 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4723 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4724 unsigned long hugetlb_total_pages(void)
4727 unsigned long nr_total_pages = 0;
4730 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4731 return nr_total_pages;
4734 static int hugetlb_acct_memory(struct hstate *h, long delta)
4741 spin_lock_irq(&hugetlb_lock);
4743 * When cpuset is configured, it breaks the strict hugetlb page
4744 * reservation as the accounting is done on a global variable. Such
4745 * reservation is completely rubbish in the presence of cpuset because
4746 * the reservation is not checked against page availability for the
4747 * current cpuset. Application can still potentially OOM'ed by kernel
4748 * with lack of free htlb page in cpuset that the task is in.
4749 * Attempt to enforce strict accounting with cpuset is almost
4750 * impossible (or too ugly) because cpuset is too fluid that
4751 * task or memory node can be dynamically moved between cpusets.
4753 * The change of semantics for shared hugetlb mapping with cpuset is
4754 * undesirable. However, in order to preserve some of the semantics,
4755 * we fall back to check against current free page availability as
4756 * a best attempt and hopefully to minimize the impact of changing
4757 * semantics that cpuset has.
4759 * Apart from cpuset, we also have memory policy mechanism that
4760 * also determines from which node the kernel will allocate memory
4761 * in a NUMA system. So similar to cpuset, we also should consider
4762 * the memory policy of the current task. Similar to the description
4766 if (gather_surplus_pages(h, delta) < 0)
4769 if (delta > allowed_mems_nr(h)) {
4770 return_unused_surplus_pages(h, delta);
4777 return_unused_surplus_pages(h, (unsigned long) -delta);
4780 spin_unlock_irq(&hugetlb_lock);
4784 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4786 struct resv_map *resv = vma_resv_map(vma);
4789 * HPAGE_RESV_OWNER indicates a private mapping.
4790 * This new VMA should share its siblings reservation map if present.
4791 * The VMA will only ever have a valid reservation map pointer where
4792 * it is being copied for another still existing VMA. As that VMA
4793 * has a reference to the reservation map it cannot disappear until
4794 * after this open call completes. It is therefore safe to take a
4795 * new reference here without additional locking.
4797 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4798 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4799 kref_get(&resv->refs);
4803 * vma_lock structure for sharable mappings is vma specific.
4804 * Clear old pointer (if copied via vm_area_dup) and allocate
4805 * new structure. Before clearing, make sure vma_lock is not
4808 if (vma->vm_flags & VM_MAYSHARE) {
4809 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4812 if (vma_lock->vma != vma) {
4813 vma->vm_private_data = NULL;
4814 hugetlb_vma_lock_alloc(vma);
4816 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4818 hugetlb_vma_lock_alloc(vma);
4822 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4824 struct hstate *h = hstate_vma(vma);
4825 struct resv_map *resv;
4826 struct hugepage_subpool *spool = subpool_vma(vma);
4827 unsigned long reserve, start, end;
4830 hugetlb_vma_lock_free(vma);
4832 resv = vma_resv_map(vma);
4833 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4836 start = vma_hugecache_offset(h, vma, vma->vm_start);
4837 end = vma_hugecache_offset(h, vma, vma->vm_end);
4839 reserve = (end - start) - region_count(resv, start, end);
4840 hugetlb_cgroup_uncharge_counter(resv, start, end);
4843 * Decrement reserve counts. The global reserve count may be
4844 * adjusted if the subpool has a minimum size.
4846 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4847 hugetlb_acct_memory(h, -gbl_reserve);
4850 kref_put(&resv->refs, resv_map_release);
4853 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4855 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4859 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4860 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4861 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4863 if (addr & ~PUD_MASK) {
4865 * hugetlb_vm_op_split is called right before we attempt to
4866 * split the VMA. We will need to unshare PMDs in the old and
4867 * new VMAs, so let's unshare before we split.
4869 unsigned long floor = addr & PUD_MASK;
4870 unsigned long ceil = floor + PUD_SIZE;
4872 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4873 hugetlb_unshare_pmds(vma, floor, ceil);
4879 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4881 return huge_page_size(hstate_vma(vma));
4885 * We cannot handle pagefaults against hugetlb pages at all. They cause
4886 * handle_mm_fault() to try to instantiate regular-sized pages in the
4887 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4890 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4897 * When a new function is introduced to vm_operations_struct and added
4898 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4899 * This is because under System V memory model, mappings created via
4900 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4901 * their original vm_ops are overwritten with shm_vm_ops.
4903 const struct vm_operations_struct hugetlb_vm_ops = {
4904 .fault = hugetlb_vm_op_fault,
4905 .open = hugetlb_vm_op_open,
4906 .close = hugetlb_vm_op_close,
4907 .may_split = hugetlb_vm_op_split,
4908 .pagesize = hugetlb_vm_op_pagesize,
4911 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4915 unsigned int shift = huge_page_shift(hstate_vma(vma));
4918 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4919 vma->vm_page_prot)));
4921 entry = huge_pte_wrprotect(mk_huge_pte(page,
4922 vma->vm_page_prot));
4924 entry = pte_mkyoung(entry);
4925 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4930 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4931 unsigned long address, pte_t *ptep)
4935 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4936 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4937 update_mmu_cache(vma, address, ptep);
4940 bool is_hugetlb_entry_migration(pte_t pte)
4944 if (huge_pte_none(pte) || pte_present(pte))
4946 swp = pte_to_swp_entry(pte);
4947 if (is_migration_entry(swp))
4953 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4957 if (huge_pte_none(pte) || pte_present(pte))
4959 swp = pte_to_swp_entry(pte);
4960 if (is_hwpoison_entry(swp))
4967 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4968 struct page *new_page)
4970 __SetPageUptodate(new_page);
4971 hugepage_add_new_anon_rmap(new_page, vma, addr);
4972 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4973 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4974 ClearHPageRestoreReserve(new_page);
4975 SetHPageMigratable(new_page);
4978 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4979 struct vm_area_struct *dst_vma,
4980 struct vm_area_struct *src_vma)
4982 pte_t *src_pte, *dst_pte, entry;
4983 struct page *ptepage;
4985 bool cow = is_cow_mapping(src_vma->vm_flags);
4986 struct hstate *h = hstate_vma(src_vma);
4987 unsigned long sz = huge_page_size(h);
4988 unsigned long npages = pages_per_huge_page(h);
4989 struct mmu_notifier_range range;
4990 unsigned long last_addr_mask;
4994 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4997 mmu_notifier_invalidate_range_start(&range);
4998 mmap_assert_write_locked(src);
4999 raw_write_seqcount_begin(&src->write_protect_seq);
5002 * For shared mappings the vma lock must be held before
5003 * calling huge_pte_offset in the src vma. Otherwise, the
5004 * returned ptep could go away if part of a shared pmd and
5005 * another thread calls huge_pmd_unshare.
5007 hugetlb_vma_lock_read(src_vma);
5010 last_addr_mask = hugetlb_mask_last_page(h);
5011 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5012 spinlock_t *src_ptl, *dst_ptl;
5013 src_pte = huge_pte_offset(src, addr, sz);
5015 addr |= last_addr_mask;
5018 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5025 * If the pagetables are shared don't copy or take references.
5027 * dst_pte == src_pte is the common case of src/dest sharing.
5028 * However, src could have 'unshared' and dst shares with
5029 * another vma. So page_count of ptep page is checked instead
5030 * to reliably determine whether pte is shared.
5032 if (page_count(virt_to_page(dst_pte)) > 1) {
5033 addr |= last_addr_mask;
5037 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5038 src_ptl = huge_pte_lockptr(h, src, src_pte);
5039 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5040 entry = huge_ptep_get(src_pte);
5042 if (huge_pte_none(entry)) {
5044 * Skip if src entry none.
5047 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5048 bool uffd_wp = huge_pte_uffd_wp(entry);
5050 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5051 entry = huge_pte_clear_uffd_wp(entry);
5052 set_huge_pte_at(dst, addr, dst_pte, entry);
5053 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5054 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5055 bool uffd_wp = huge_pte_uffd_wp(entry);
5057 if (!is_readable_migration_entry(swp_entry) && cow) {
5059 * COW mappings require pages in both
5060 * parent and child to be set to read.
5062 swp_entry = make_readable_migration_entry(
5063 swp_offset(swp_entry));
5064 entry = swp_entry_to_pte(swp_entry);
5065 if (userfaultfd_wp(src_vma) && uffd_wp)
5066 entry = huge_pte_mkuffd_wp(entry);
5067 set_huge_pte_at(src, addr, src_pte, entry);
5069 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5070 entry = huge_pte_clear_uffd_wp(entry);
5071 set_huge_pte_at(dst, addr, dst_pte, entry);
5072 } else if (unlikely(is_pte_marker(entry))) {
5074 * We copy the pte marker only if the dst vma has
5077 if (userfaultfd_wp(dst_vma))
5078 set_huge_pte_at(dst, addr, dst_pte, entry);
5080 entry = huge_ptep_get(src_pte);
5081 ptepage = pte_page(entry);
5085 * Failing to duplicate the anon rmap is a rare case
5086 * where we see pinned hugetlb pages while they're
5087 * prone to COW. We need to do the COW earlier during
5090 * When pre-allocating the page or copying data, we
5091 * need to be without the pgtable locks since we could
5092 * sleep during the process.
5094 if (!PageAnon(ptepage)) {
5095 page_dup_file_rmap(ptepage, true);
5096 } else if (page_try_dup_anon_rmap(ptepage, true,
5098 pte_t src_pte_old = entry;
5101 spin_unlock(src_ptl);
5102 spin_unlock(dst_ptl);
5103 /* Do not use reserve as it's private owned */
5104 new = alloc_huge_page(dst_vma, addr, 1);
5110 copy_user_huge_page(new, ptepage, addr, dst_vma,
5114 /* Install the new huge page if src pte stable */
5115 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5116 src_ptl = huge_pte_lockptr(h, src, src_pte);
5117 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5118 entry = huge_ptep_get(src_pte);
5119 if (!pte_same(src_pte_old, entry)) {
5120 restore_reserve_on_error(h, dst_vma, addr,
5123 /* huge_ptep of dst_pte won't change as in child */
5126 hugetlb_install_page(dst_vma, dst_pte, addr, new);
5127 spin_unlock(src_ptl);
5128 spin_unlock(dst_ptl);
5134 * No need to notify as we are downgrading page
5135 * table protection not changing it to point
5138 * See Documentation/mm/mmu_notifier.rst
5140 huge_ptep_set_wrprotect(src, addr, src_pte);
5141 entry = huge_pte_wrprotect(entry);
5144 set_huge_pte_at(dst, addr, dst_pte, entry);
5145 hugetlb_count_add(npages, dst);
5147 spin_unlock(src_ptl);
5148 spin_unlock(dst_ptl);
5152 raw_write_seqcount_end(&src->write_protect_seq);
5153 mmu_notifier_invalidate_range_end(&range);
5155 hugetlb_vma_unlock_read(src_vma);
5161 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5162 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5164 struct hstate *h = hstate_vma(vma);
5165 struct mm_struct *mm = vma->vm_mm;
5166 spinlock_t *src_ptl, *dst_ptl;
5169 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5170 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5173 * We don't have to worry about the ordering of src and dst ptlocks
5174 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
5176 if (src_ptl != dst_ptl)
5177 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5179 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5180 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5182 if (src_ptl != dst_ptl)
5183 spin_unlock(src_ptl);
5184 spin_unlock(dst_ptl);
5187 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5188 struct vm_area_struct *new_vma,
5189 unsigned long old_addr, unsigned long new_addr,
5192 struct hstate *h = hstate_vma(vma);
5193 struct address_space *mapping = vma->vm_file->f_mapping;
5194 unsigned long sz = huge_page_size(h);
5195 struct mm_struct *mm = vma->vm_mm;
5196 unsigned long old_end = old_addr + len;
5197 unsigned long last_addr_mask;
5198 pte_t *src_pte, *dst_pte;
5199 struct mmu_notifier_range range;
5200 bool shared_pmd = false;
5202 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
5204 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5206 * In case of shared PMDs, we should cover the maximum possible
5209 flush_cache_range(vma, range.start, range.end);
5211 mmu_notifier_invalidate_range_start(&range);
5212 last_addr_mask = hugetlb_mask_last_page(h);
5213 /* Prevent race with file truncation */
5214 hugetlb_vma_lock_write(vma);
5215 i_mmap_lock_write(mapping);
5216 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5217 src_pte = huge_pte_offset(mm, old_addr, sz);
5219 old_addr |= last_addr_mask;
5220 new_addr |= last_addr_mask;
5223 if (huge_pte_none(huge_ptep_get(src_pte)))
5226 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5228 old_addr |= last_addr_mask;
5229 new_addr |= last_addr_mask;
5233 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5237 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5241 flush_tlb_range(vma, range.start, range.end);
5243 flush_tlb_range(vma, old_end - len, old_end);
5244 mmu_notifier_invalidate_range_end(&range);
5245 i_mmap_unlock_write(mapping);
5246 hugetlb_vma_unlock_write(vma);
5248 return len + old_addr - old_end;
5251 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5252 unsigned long start, unsigned long end,
5253 struct page *ref_page, zap_flags_t zap_flags)
5255 struct mm_struct *mm = vma->vm_mm;
5256 unsigned long address;
5261 struct hstate *h = hstate_vma(vma);
5262 unsigned long sz = huge_page_size(h);
5263 struct mmu_notifier_range range;
5264 unsigned long last_addr_mask;
5265 bool force_flush = false;
5267 WARN_ON(!is_vm_hugetlb_page(vma));
5268 BUG_ON(start & ~huge_page_mask(h));
5269 BUG_ON(end & ~huge_page_mask(h));
5272 * This is a hugetlb vma, all the pte entries should point
5275 tlb_change_page_size(tlb, sz);
5276 tlb_start_vma(tlb, vma);
5279 * If sharing possible, alert mmu notifiers of worst case.
5281 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
5283 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5284 mmu_notifier_invalidate_range_start(&range);
5285 last_addr_mask = hugetlb_mask_last_page(h);
5287 for (; address < end; address += sz) {
5288 ptep = huge_pte_offset(mm, address, sz);
5290 address |= last_addr_mask;
5294 ptl = huge_pte_lock(h, mm, ptep);
5295 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5297 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5299 address |= last_addr_mask;
5303 pte = huge_ptep_get(ptep);
5304 if (huge_pte_none(pte)) {
5310 * Migrating hugepage or HWPoisoned hugepage is already
5311 * unmapped and its refcount is dropped, so just clear pte here.
5313 if (unlikely(!pte_present(pte))) {
5314 #ifdef CONFIG_PTE_MARKER_UFFD_WP
5316 * If the pte was wr-protected by uffd-wp in any of the
5317 * swap forms, meanwhile the caller does not want to
5318 * drop the uffd-wp bit in this zap, then replace the
5319 * pte with a marker.
5321 if (pte_swp_uffd_wp_any(pte) &&
5322 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5323 set_huge_pte_at(mm, address, ptep,
5324 make_pte_marker(PTE_MARKER_UFFD_WP));
5327 huge_pte_clear(mm, address, ptep, sz);
5332 page = pte_page(pte);
5334 * If a reference page is supplied, it is because a specific
5335 * page is being unmapped, not a range. Ensure the page we
5336 * are about to unmap is the actual page of interest.
5339 if (page != ref_page) {
5344 * Mark the VMA as having unmapped its page so that
5345 * future faults in this VMA will fail rather than
5346 * looking like data was lost
5348 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5351 pte = huge_ptep_get_and_clear(mm, address, ptep);
5352 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5353 if (huge_pte_dirty(pte))
5354 set_page_dirty(page);
5355 #ifdef CONFIG_PTE_MARKER_UFFD_WP
5356 /* Leave a uffd-wp pte marker if needed */
5357 if (huge_pte_uffd_wp(pte) &&
5358 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5359 set_huge_pte_at(mm, address, ptep,
5360 make_pte_marker(PTE_MARKER_UFFD_WP));
5362 hugetlb_count_sub(pages_per_huge_page(h), mm);
5363 page_remove_rmap(page, vma, true);
5366 tlb_remove_page_size(tlb, page, huge_page_size(h));
5368 * Bail out after unmapping reference page if supplied
5373 mmu_notifier_invalidate_range_end(&range);
5374 tlb_end_vma(tlb, vma);
5377 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5378 * could defer the flush until now, since by holding i_mmap_rwsem we
5379 * guaranteed that the last refernece would not be dropped. But we must
5380 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5381 * dropped and the last reference to the shared PMDs page might be
5384 * In theory we could defer the freeing of the PMD pages as well, but
5385 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5386 * detect sharing, so we cannot defer the release of the page either.
5387 * Instead, do flush now.
5390 tlb_flush_mmu_tlbonly(tlb);
5393 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5394 struct vm_area_struct *vma, unsigned long start,
5395 unsigned long end, struct page *ref_page,
5396 zap_flags_t zap_flags)
5398 hugetlb_vma_lock_write(vma);
5399 i_mmap_lock_write(vma->vm_file->f_mapping);
5401 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5403 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5405 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5406 * When the vma_lock is freed, this makes the vma ineligible
5407 * for pmd sharing. And, i_mmap_rwsem is required to set up
5408 * pmd sharing. This is important as page tables for this
5409 * unmapped range will be asynchrously deleted. If the page
5410 * tables are shared, there will be issues when accessed by
5413 __hugetlb_vma_unlock_write_free(vma);
5414 i_mmap_unlock_write(vma->vm_file->f_mapping);
5416 i_mmap_unlock_write(vma->vm_file->f_mapping);
5417 hugetlb_vma_unlock_write(vma);
5421 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5422 unsigned long end, struct page *ref_page,
5423 zap_flags_t zap_flags)
5425 struct mmu_gather tlb;
5427 tlb_gather_mmu(&tlb, vma->vm_mm);
5428 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5429 tlb_finish_mmu(&tlb);
5433 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5434 * mapping it owns the reserve page for. The intention is to unmap the page
5435 * from other VMAs and let the children be SIGKILLed if they are faulting the
5438 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5439 struct page *page, unsigned long address)
5441 struct hstate *h = hstate_vma(vma);
5442 struct vm_area_struct *iter_vma;
5443 struct address_space *mapping;
5447 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5448 * from page cache lookup which is in HPAGE_SIZE units.
5450 address = address & huge_page_mask(h);
5451 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5453 mapping = vma->vm_file->f_mapping;
5456 * Take the mapping lock for the duration of the table walk. As
5457 * this mapping should be shared between all the VMAs,
5458 * __unmap_hugepage_range() is called as the lock is already held
5460 i_mmap_lock_write(mapping);
5461 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5462 /* Do not unmap the current VMA */
5463 if (iter_vma == vma)
5467 * Shared VMAs have their own reserves and do not affect
5468 * MAP_PRIVATE accounting but it is possible that a shared
5469 * VMA is using the same page so check and skip such VMAs.
5471 if (iter_vma->vm_flags & VM_MAYSHARE)
5475 * Unmap the page from other VMAs without their own reserves.
5476 * They get marked to be SIGKILLed if they fault in these
5477 * areas. This is because a future no-page fault on this VMA
5478 * could insert a zeroed page instead of the data existing
5479 * from the time of fork. This would look like data corruption
5481 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5482 unmap_hugepage_range(iter_vma, address,
5483 address + huge_page_size(h), page, 0);
5485 i_mmap_unlock_write(mapping);
5489 * hugetlb_wp() should be called with page lock of the original hugepage held.
5490 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5491 * cannot race with other handlers or page migration.
5492 * Keep the pte_same checks anyway to make transition from the mutex easier.
5494 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5495 unsigned long address, pte_t *ptep, unsigned int flags,
5496 struct page *pagecache_page, spinlock_t *ptl)
5498 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5499 pte_t pte = huge_ptep_get(ptep);
5500 struct hstate *h = hstate_vma(vma);
5501 struct page *old_page, *new_page;
5502 int outside_reserve = 0;
5504 unsigned long haddr = address & huge_page_mask(h);
5505 struct mmu_notifier_range range;
5507 VM_BUG_ON(unshare && (flags & FOLL_WRITE));
5508 VM_BUG_ON(!unshare && !(flags & FOLL_WRITE));
5511 * Never handle CoW for uffd-wp protected pages. It should be only
5512 * handled when the uffd-wp protection is removed.
5514 * Note that only the CoW optimization path (in hugetlb_no_page())
5515 * can trigger this, because hugetlb_fault() will always resolve
5516 * uffd-wp bit first.
5518 if (!unshare && huge_pte_uffd_wp(pte))
5522 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5523 * PTE mapped R/O such as maybe_mkwrite() would do.
5525 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5526 return VM_FAULT_SIGSEGV;
5528 /* Let's take out MAP_SHARED mappings first. */
5529 if (vma->vm_flags & VM_MAYSHARE) {
5530 if (unlikely(unshare))
5532 set_huge_ptep_writable(vma, haddr, ptep);
5536 old_page = pte_page(pte);
5538 delayacct_wpcopy_start();
5542 * If no-one else is actually using this page, we're the exclusive
5543 * owner and can reuse this page.
5545 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5546 if (!PageAnonExclusive(old_page))
5547 page_move_anon_rmap(old_page, vma);
5548 if (likely(!unshare))
5549 set_huge_ptep_writable(vma, haddr, ptep);
5551 delayacct_wpcopy_end();
5554 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5558 * If the process that created a MAP_PRIVATE mapping is about to
5559 * perform a COW due to a shared page count, attempt to satisfy
5560 * the allocation without using the existing reserves. The pagecache
5561 * page is used to determine if the reserve at this address was
5562 * consumed or not. If reserves were used, a partial faulted mapping
5563 * at the time of fork() could consume its reserves on COW instead
5564 * of the full address range.
5566 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5567 old_page != pagecache_page)
5568 outside_reserve = 1;
5573 * Drop page table lock as buddy allocator may be called. It will
5574 * be acquired again before returning to the caller, as expected.
5577 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5579 if (IS_ERR(new_page)) {
5581 * If a process owning a MAP_PRIVATE mapping fails to COW,
5582 * it is due to references held by a child and an insufficient
5583 * huge page pool. To guarantee the original mappers
5584 * reliability, unmap the page from child processes. The child
5585 * may get SIGKILLed if it later faults.
5587 if (outside_reserve) {
5588 struct address_space *mapping = vma->vm_file->f_mapping;
5594 * Drop hugetlb_fault_mutex and vma_lock before
5595 * unmapping. unmapping needs to hold vma_lock
5596 * in write mode. Dropping vma_lock in read mode
5597 * here is OK as COW mappings do not interact with
5600 * Reacquire both after unmap operation.
5602 idx = vma_hugecache_offset(h, vma, haddr);
5603 hash = hugetlb_fault_mutex_hash(mapping, idx);
5604 hugetlb_vma_unlock_read(vma);
5605 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5607 unmap_ref_private(mm, vma, old_page, haddr);
5609 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5610 hugetlb_vma_lock_read(vma);
5612 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5614 pte_same(huge_ptep_get(ptep), pte)))
5615 goto retry_avoidcopy;
5617 * race occurs while re-acquiring page table
5618 * lock, and our job is done.
5620 delayacct_wpcopy_end();
5624 ret = vmf_error(PTR_ERR(new_page));
5625 goto out_release_old;
5629 * When the original hugepage is shared one, it does not have
5630 * anon_vma prepared.
5632 if (unlikely(anon_vma_prepare(vma))) {
5634 goto out_release_all;
5637 copy_user_huge_page(new_page, old_page, address, vma,
5638 pages_per_huge_page(h));
5639 __SetPageUptodate(new_page);
5641 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5642 haddr + huge_page_size(h));
5643 mmu_notifier_invalidate_range_start(&range);
5646 * Retake the page table lock to check for racing updates
5647 * before the page tables are altered
5650 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5651 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5652 ClearHPageRestoreReserve(new_page);
5654 /* Break COW or unshare */
5655 huge_ptep_clear_flush(vma, haddr, ptep);
5656 mmu_notifier_invalidate_range(mm, range.start, range.end);
5657 page_remove_rmap(old_page, vma, true);
5658 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5659 set_huge_pte_at(mm, haddr, ptep,
5660 make_huge_pte(vma, new_page, !unshare));
5661 SetHPageMigratable(new_page);
5662 /* Make the old page be freed below */
5663 new_page = old_page;
5666 mmu_notifier_invalidate_range_end(&range);
5669 * No restore in case of successful pagetable update (Break COW or
5672 if (new_page != old_page)
5673 restore_reserve_on_error(h, vma, haddr, new_page);
5678 spin_lock(ptl); /* Caller expects lock to be held */
5680 delayacct_wpcopy_end();
5685 * Return whether there is a pagecache page to back given address within VMA.
5686 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5688 static bool hugetlbfs_pagecache_present(struct hstate *h,
5689 struct vm_area_struct *vma, unsigned long address)
5691 struct address_space *mapping;
5695 mapping = vma->vm_file->f_mapping;
5696 idx = vma_hugecache_offset(h, vma, address);
5698 page = find_get_page(mapping, idx);
5701 return page != NULL;
5704 int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping,
5707 struct folio *folio = page_folio(page);
5708 struct inode *inode = mapping->host;
5709 struct hstate *h = hstate_inode(inode);
5712 __folio_set_locked(folio);
5713 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5715 if (unlikely(err)) {
5716 __folio_clear_locked(folio);
5719 ClearHPageRestoreReserve(page);
5722 * mark folio dirty so that it will not be removed from cache/file
5723 * by non-hugetlbfs specific code paths.
5725 folio_mark_dirty(folio);
5727 spin_lock(&inode->i_lock);
5728 inode->i_blocks += blocks_per_huge_page(h);
5729 spin_unlock(&inode->i_lock);
5733 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5734 struct address_space *mapping,
5737 unsigned long haddr,
5739 unsigned long reason)
5742 struct vm_fault vmf = {
5745 .real_address = addr,
5749 * Hard to debug if it ends up being
5750 * used by a callee that assumes
5751 * something about the other
5752 * uninitialized fields... same as in
5758 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5759 * userfault. Also mmap_lock could be dropped due to handling
5760 * userfault, any vma operation should be careful from here.
5762 hugetlb_vma_unlock_read(vma);
5763 hash = hugetlb_fault_mutex_hash(mapping, idx);
5764 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5765 return handle_userfault(&vmf, reason);
5769 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5770 * false if pte changed or is changing.
5772 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5773 pte_t *ptep, pte_t old_pte)
5778 ptl = huge_pte_lock(h, mm, ptep);
5779 same = pte_same(huge_ptep_get(ptep), old_pte);
5785 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5786 struct vm_area_struct *vma,
5787 struct address_space *mapping, pgoff_t idx,
5788 unsigned long address, pte_t *ptep,
5789 pte_t old_pte, unsigned int flags)
5791 struct hstate *h = hstate_vma(vma);
5792 vm_fault_t ret = VM_FAULT_SIGBUS;
5798 unsigned long haddr = address & huge_page_mask(h);
5799 bool new_page, new_pagecache_page = false;
5800 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5803 * Currently, we are forced to kill the process in the event the
5804 * original mapper has unmapped pages from the child due to a failed
5805 * COW/unsharing. Warn that such a situation has occurred as it may not
5808 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5809 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5815 * Use page lock to guard against racing truncation
5816 * before we get page_table_lock.
5819 page = find_lock_page(mapping, idx);
5821 size = i_size_read(mapping->host) >> huge_page_shift(h);
5824 /* Check for page in userfault range */
5825 if (userfaultfd_missing(vma)) {
5827 * Since hugetlb_no_page() was examining pte
5828 * without pgtable lock, we need to re-test under
5829 * lock because the pte may not be stable and could
5830 * have changed from under us. Try to detect
5831 * either changed or during-changing ptes and retry
5832 * properly when needed.
5834 * Note that userfaultfd is actually fine with
5835 * false positives (e.g. caused by pte changed),
5836 * but not wrong logical events (e.g. caused by
5837 * reading a pte during changing). The latter can
5838 * confuse the userspace, so the strictness is very
5839 * much preferred. E.g., MISSING event should
5840 * never happen on the page after UFFDIO_COPY has
5841 * correctly installed the page and returned.
5843 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5848 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5853 page = alloc_huge_page(vma, haddr, 0);
5856 * Returning error will result in faulting task being
5857 * sent SIGBUS. The hugetlb fault mutex prevents two
5858 * tasks from racing to fault in the same page which
5859 * could result in false unable to allocate errors.
5860 * Page migration does not take the fault mutex, but
5861 * does a clear then write of pte's under page table
5862 * lock. Page fault code could race with migration,
5863 * notice the clear pte and try to allocate a page
5864 * here. Before returning error, get ptl and make
5865 * sure there really is no pte entry.
5867 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5868 ret = vmf_error(PTR_ERR(page));
5873 clear_huge_page(page, address, pages_per_huge_page(h));
5874 __SetPageUptodate(page);
5877 if (vma->vm_flags & VM_MAYSHARE) {
5878 int err = hugetlb_add_to_page_cache(page, mapping, idx);
5881 * err can't be -EEXIST which implies someone
5882 * else consumed the reservation since hugetlb
5883 * fault mutex is held when add a hugetlb page
5884 * to the page cache. So it's safe to call
5885 * restore_reserve_on_error() here.
5887 restore_reserve_on_error(h, vma, haddr, page);
5891 new_pagecache_page = true;
5894 if (unlikely(anon_vma_prepare(vma))) {
5896 goto backout_unlocked;
5902 * If memory error occurs between mmap() and fault, some process
5903 * don't have hwpoisoned swap entry for errored virtual address.
5904 * So we need to block hugepage fault by PG_hwpoison bit check.
5906 if (unlikely(PageHWPoison(page))) {
5907 ret = VM_FAULT_HWPOISON_LARGE |
5908 VM_FAULT_SET_HINDEX(hstate_index(h));
5909 goto backout_unlocked;
5912 /* Check for page in userfault range. */
5913 if (userfaultfd_minor(vma)) {
5916 /* See comment in userfaultfd_missing() block above */
5917 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5921 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5928 * If we are going to COW a private mapping later, we examine the
5929 * pending reservations for this page now. This will ensure that
5930 * any allocations necessary to record that reservation occur outside
5933 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5934 if (vma_needs_reservation(h, vma, haddr) < 0) {
5936 goto backout_unlocked;
5938 /* Just decrements count, does not deallocate */
5939 vma_end_reservation(h, vma, haddr);
5942 ptl = huge_pte_lock(h, mm, ptep);
5944 /* If pte changed from under us, retry */
5945 if (!pte_same(huge_ptep_get(ptep), old_pte))
5949 ClearHPageRestoreReserve(page);
5950 hugepage_add_new_anon_rmap(page, vma, haddr);
5952 page_dup_file_rmap(page, true);
5953 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5954 && (vma->vm_flags & VM_SHARED)));
5956 * If this pte was previously wr-protected, keep it wr-protected even
5959 if (unlikely(pte_marker_uffd_wp(old_pte)))
5960 new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte));
5961 set_huge_pte_at(mm, haddr, ptep, new_pte);
5963 hugetlb_count_add(pages_per_huge_page(h), mm);
5964 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5965 /* Optimization, do the COW without a second fault */
5966 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5972 * Only set HPageMigratable in newly allocated pages. Existing pages
5973 * found in the pagecache may not have HPageMigratableset if they have
5974 * been isolated for migration.
5977 SetHPageMigratable(page);
5981 hugetlb_vma_unlock_read(vma);
5982 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5988 if (new_page && !new_pagecache_page)
5989 restore_reserve_on_error(h, vma, haddr, page);
5997 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5999 unsigned long key[2];
6002 key[0] = (unsigned long) mapping;
6005 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6007 return hash & (num_fault_mutexes - 1);
6011 * For uniprocessor systems we always use a single mutex, so just
6012 * return 0 and avoid the hashing overhead.
6014 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6020 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6021 unsigned long address, unsigned int flags)
6028 struct page *page = NULL;
6029 struct page *pagecache_page = NULL;
6030 struct hstate *h = hstate_vma(vma);
6031 struct address_space *mapping;
6032 int need_wait_lock = 0;
6033 unsigned long haddr = address & huge_page_mask(h);
6035 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
6038 * Since we hold no locks, ptep could be stale. That is
6039 * OK as we are only making decisions based on content and
6040 * not actually modifying content here.
6042 entry = huge_ptep_get(ptep);
6043 if (unlikely(is_hugetlb_entry_migration(entry))) {
6044 migration_entry_wait_huge(vma, ptep);
6046 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6047 return VM_FAULT_HWPOISON_LARGE |
6048 VM_FAULT_SET_HINDEX(hstate_index(h));
6052 * Serialize hugepage allocation and instantiation, so that we don't
6053 * get spurious allocation failures if two CPUs race to instantiate
6054 * the same page in the page cache.
6056 mapping = vma->vm_file->f_mapping;
6057 idx = vma_hugecache_offset(h, vma, haddr);
6058 hash = hugetlb_fault_mutex_hash(mapping, idx);
6059 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6062 * Acquire vma lock before calling huge_pte_alloc and hold
6063 * until finished with ptep. This prevents huge_pmd_unshare from
6064 * being called elsewhere and making the ptep no longer valid.
6066 * ptep could have already be assigned via huge_pte_offset. That
6067 * is OK, as huge_pte_alloc will return the same value unless
6068 * something has changed.
6070 hugetlb_vma_lock_read(vma);
6071 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6073 hugetlb_vma_unlock_read(vma);
6074 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6075 return VM_FAULT_OOM;
6078 entry = huge_ptep_get(ptep);
6079 /* PTE markers should be handled the same way as none pte */
6080 if (huge_pte_none_mostly(entry))
6082 * hugetlb_no_page will drop vma lock and hugetlb fault
6083 * mutex internally, which make us return immediately.
6085 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6091 * entry could be a migration/hwpoison entry at this point, so this
6092 * check prevents the kernel from going below assuming that we have
6093 * an active hugepage in pagecache. This goto expects the 2nd page
6094 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6095 * properly handle it.
6097 if (!pte_present(entry))
6101 * If we are going to COW/unshare the mapping later, we examine the
6102 * pending reservations for this page now. This will ensure that any
6103 * allocations necessary to record that reservation occur outside the
6104 * spinlock. Also lookup the pagecache page now as it is used to
6105 * determine if a reservation has been consumed.
6107 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6108 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6109 if (vma_needs_reservation(h, vma, haddr) < 0) {
6113 /* Just decrements count, does not deallocate */
6114 vma_end_reservation(h, vma, haddr);
6116 pagecache_page = find_lock_page(mapping, idx);
6119 ptl = huge_pte_lock(h, mm, ptep);
6121 /* Check for a racing update before calling hugetlb_wp() */
6122 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6125 /* Handle userfault-wp first, before trying to lock more pages */
6126 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6127 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6128 struct vm_fault vmf = {
6131 .real_address = address,
6136 if (pagecache_page) {
6137 unlock_page(pagecache_page);
6138 put_page(pagecache_page);
6140 hugetlb_vma_unlock_read(vma);
6141 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6142 return handle_userfault(&vmf, VM_UFFD_WP);
6146 * hugetlb_wp() requires page locks of pte_page(entry) and
6147 * pagecache_page, so here we need take the former one
6148 * when page != pagecache_page or !pagecache_page.
6150 page = pte_page(entry);
6151 if (page != pagecache_page)
6152 if (!trylock_page(page)) {
6159 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6160 if (!huge_pte_write(entry)) {
6161 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6162 pagecache_page, ptl);
6164 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6165 entry = huge_pte_mkdirty(entry);
6168 entry = pte_mkyoung(entry);
6169 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6170 flags & FAULT_FLAG_WRITE))
6171 update_mmu_cache(vma, haddr, ptep);
6173 if (page != pagecache_page)
6179 if (pagecache_page) {
6180 unlock_page(pagecache_page);
6181 put_page(pagecache_page);
6184 hugetlb_vma_unlock_read(vma);
6185 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6187 * Generally it's safe to hold refcount during waiting page lock. But
6188 * here we just wait to defer the next page fault to avoid busy loop and
6189 * the page is not used after unlocked before returning from the current
6190 * page fault. So we are safe from accessing freed page, even if we wait
6191 * here without taking refcount.
6194 wait_on_page_locked(page);
6198 #ifdef CONFIG_USERFAULTFD
6200 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
6201 * modifications for huge pages.
6203 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
6205 struct vm_area_struct *dst_vma,
6206 unsigned long dst_addr,
6207 unsigned long src_addr,
6208 enum mcopy_atomic_mode mode,
6209 struct page **pagep,
6212 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
6213 struct hstate *h = hstate_vma(dst_vma);
6214 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6215 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6217 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6223 bool page_in_pagecache = false;
6227 page = find_lock_page(mapping, idx);
6230 page_in_pagecache = true;
6231 } else if (!*pagep) {
6232 /* If a page already exists, then it's UFFDIO_COPY for
6233 * a non-missing case. Return -EEXIST.
6236 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6241 page = alloc_huge_page(dst_vma, dst_addr, 0);
6247 ret = copy_huge_page_from_user(page,
6248 (const void __user *) src_addr,
6249 pages_per_huge_page(h), false);
6251 /* fallback to copy_from_user outside mmap_lock */
6252 if (unlikely(ret)) {
6254 /* Free the allocated page which may have
6255 * consumed a reservation.
6257 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6260 /* Allocate a temporary page to hold the copied
6263 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
6269 /* Set the outparam pagep and return to the caller to
6270 * copy the contents outside the lock. Don't free the
6277 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6284 page = alloc_huge_page(dst_vma, dst_addr, 0);
6291 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
6292 pages_per_huge_page(h));
6298 * The memory barrier inside __SetPageUptodate makes sure that
6299 * preceding stores to the page contents become visible before
6300 * the set_pte_at() write.
6302 __SetPageUptodate(page);
6304 /* Add shared, newly allocated pages to the page cache. */
6305 if (vm_shared && !is_continue) {
6306 size = i_size_read(mapping->host) >> huge_page_shift(h);
6309 goto out_release_nounlock;
6312 * Serialization between remove_inode_hugepages() and
6313 * hugetlb_add_to_page_cache() below happens through the
6314 * hugetlb_fault_mutex_table that here must be hold by
6317 ret = hugetlb_add_to_page_cache(page, mapping, idx);
6319 goto out_release_nounlock;
6320 page_in_pagecache = true;
6323 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6326 if (PageHWPoison(page))
6327 goto out_release_unlock;
6330 * We allow to overwrite a pte marker: consider when both MISSING|WP
6331 * registered, we firstly wr-protect a none pte which has no page cache
6332 * page backing it, then access the page.
6335 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6336 goto out_release_unlock;
6338 if (page_in_pagecache) {
6339 page_dup_file_rmap(page, true);
6341 ClearHPageRestoreReserve(page);
6342 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6346 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6347 * with wp flag set, don't set pte write bit.
6349 if (wp_copy || (is_continue && !vm_shared))
6352 writable = dst_vma->vm_flags & VM_WRITE;
6354 _dst_pte = make_huge_pte(dst_vma, page, writable);
6356 * Always mark UFFDIO_COPY page dirty; note that this may not be
6357 * extremely important for hugetlbfs for now since swapping is not
6358 * supported, but we should still be clear in that this page cannot be
6359 * thrown away at will, even if write bit not set.
6361 _dst_pte = huge_pte_mkdirty(_dst_pte);
6362 _dst_pte = pte_mkyoung(_dst_pte);
6365 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6367 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6369 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6371 /* No need to invalidate - it was non-present before */
6372 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6376 SetHPageMigratable(page);
6377 if (vm_shared || is_continue)
6384 if (vm_shared || is_continue)
6386 out_release_nounlock:
6387 if (!page_in_pagecache)
6388 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6392 #endif /* CONFIG_USERFAULTFD */
6394 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6395 int refs, struct page **pages,
6396 struct vm_area_struct **vmas)
6400 for (nr = 0; nr < refs; nr++) {
6402 pages[nr] = nth_page(page, nr);
6408 static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte,
6411 pte_t pteval = huge_ptep_get(pte);
6414 if (is_swap_pte(pteval))
6416 if (huge_pte_write(pteval))
6418 if (flags & FOLL_WRITE)
6420 if (gup_must_unshare(flags, pte_page(pteval))) {
6427 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6428 struct page **pages, struct vm_area_struct **vmas,
6429 unsigned long *position, unsigned long *nr_pages,
6430 long i, unsigned int flags, int *locked)
6432 unsigned long pfn_offset;
6433 unsigned long vaddr = *position;
6434 unsigned long remainder = *nr_pages;
6435 struct hstate *h = hstate_vma(vma);
6436 int err = -EFAULT, refs;
6438 while (vaddr < vma->vm_end && remainder) {
6440 spinlock_t *ptl = NULL;
6441 bool unshare = false;
6446 * If we have a pending SIGKILL, don't keep faulting pages and
6447 * potentially allocating memory.
6449 if (fatal_signal_pending(current)) {
6455 * Some archs (sparc64, sh*) have multiple pte_ts to
6456 * each hugepage. We have to make sure we get the
6457 * first, for the page indexing below to work.
6459 * Note that page table lock is not held when pte is null.
6461 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6464 ptl = huge_pte_lock(h, mm, pte);
6465 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6468 * When coredumping, it suits get_dump_page if we just return
6469 * an error where there's an empty slot with no huge pagecache
6470 * to back it. This way, we avoid allocating a hugepage, and
6471 * the sparse dumpfile avoids allocating disk blocks, but its
6472 * huge holes still show up with zeroes where they need to be.
6474 if (absent && (flags & FOLL_DUMP) &&
6475 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6483 * We need call hugetlb_fault for both hugepages under migration
6484 * (in which case hugetlb_fault waits for the migration,) and
6485 * hwpoisoned hugepages (in which case we need to prevent the
6486 * caller from accessing to them.) In order to do this, we use
6487 * here is_swap_pte instead of is_hugetlb_entry_migration and
6488 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6489 * both cases, and because we can't follow correct pages
6490 * directly from any kind of swap entries.
6493 __follow_hugetlb_must_fault(flags, pte, &unshare)) {
6495 unsigned int fault_flags = 0;
6499 if (flags & FOLL_WRITE)
6500 fault_flags |= FAULT_FLAG_WRITE;
6502 fault_flags |= FAULT_FLAG_UNSHARE;
6504 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6505 FAULT_FLAG_KILLABLE;
6506 if (flags & FOLL_NOWAIT)
6507 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6508 FAULT_FLAG_RETRY_NOWAIT;
6509 if (flags & FOLL_TRIED) {
6511 * Note: FAULT_FLAG_ALLOW_RETRY and
6512 * FAULT_FLAG_TRIED can co-exist
6514 fault_flags |= FAULT_FLAG_TRIED;
6516 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6517 if (ret & VM_FAULT_ERROR) {
6518 err = vm_fault_to_errno(ret, flags);
6522 if (ret & VM_FAULT_RETRY) {
6524 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6528 * VM_FAULT_RETRY must not return an
6529 * error, it will return zero
6532 * No need to update "position" as the
6533 * caller will not check it after
6534 * *nr_pages is set to 0.
6541 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6542 page = pte_page(huge_ptep_get(pte));
6544 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6545 !PageAnonExclusive(page), page);
6548 * If subpage information not requested, update counters
6549 * and skip the same_page loop below.
6551 if (!pages && !vmas && !pfn_offset &&
6552 (vaddr + huge_page_size(h) < vma->vm_end) &&
6553 (remainder >= pages_per_huge_page(h))) {
6554 vaddr += huge_page_size(h);
6555 remainder -= pages_per_huge_page(h);
6556 i += pages_per_huge_page(h);
6561 /* vaddr may not be aligned to PAGE_SIZE */
6562 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6563 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6566 record_subpages_vmas(nth_page(page, pfn_offset),
6568 likely(pages) ? pages + i : NULL,
6569 vmas ? vmas + i : NULL);
6573 * try_grab_folio() should always succeed here,
6574 * because: a) we hold the ptl lock, and b) we've just
6575 * checked that the huge page is present in the page
6576 * tables. If the huge page is present, then the tail
6577 * pages must also be present. The ptl prevents the
6578 * head page and tail pages from being rearranged in
6579 * any way. So this page must be available at this
6580 * point, unless the page refcount overflowed:
6582 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6591 vaddr += (refs << PAGE_SHIFT);
6597 *nr_pages = remainder;
6599 * setting position is actually required only if remainder is
6600 * not zero but it's faster not to add a "if (remainder)"
6608 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6609 unsigned long address, unsigned long end,
6610 pgprot_t newprot, unsigned long cp_flags)
6612 struct mm_struct *mm = vma->vm_mm;
6613 unsigned long start = address;
6616 struct hstate *h = hstate_vma(vma);
6617 unsigned long pages = 0, psize = huge_page_size(h);
6618 bool shared_pmd = false;
6619 struct mmu_notifier_range range;
6620 unsigned long last_addr_mask;
6621 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6622 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6625 * In the case of shared PMDs, the area to flush could be beyond
6626 * start/end. Set range.start/range.end to cover the maximum possible
6627 * range if PMD sharing is possible.
6629 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6630 0, vma, mm, start, end);
6631 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6633 BUG_ON(address >= end);
6634 flush_cache_range(vma, range.start, range.end);
6636 mmu_notifier_invalidate_range_start(&range);
6637 hugetlb_vma_lock_write(vma);
6638 i_mmap_lock_write(vma->vm_file->f_mapping);
6639 last_addr_mask = hugetlb_mask_last_page(h);
6640 for (; address < end; address += psize) {
6642 ptep = huge_pte_offset(mm, address, psize);
6645 address |= last_addr_mask;
6649 * Userfaultfd wr-protect requires pgtable
6650 * pre-allocations to install pte markers.
6652 ptep = huge_pte_alloc(mm, vma, address, psize);
6656 ptl = huge_pte_lock(h, mm, ptep);
6657 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6659 * When uffd-wp is enabled on the vma, unshare
6660 * shouldn't happen at all. Warn about it if it
6661 * happened due to some reason.
6663 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6667 address |= last_addr_mask;
6670 pte = huge_ptep_get(ptep);
6671 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6672 /* Nothing to do. */
6673 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6674 swp_entry_t entry = pte_to_swp_entry(pte);
6675 struct page *page = pfn_swap_entry_to_page(entry);
6678 if (is_writable_migration_entry(entry)) {
6680 entry = make_readable_exclusive_migration_entry(
6683 entry = make_readable_migration_entry(
6685 newpte = swp_entry_to_pte(entry);
6690 newpte = pte_swp_mkuffd_wp(newpte);
6691 else if (uffd_wp_resolve)
6692 newpte = pte_swp_clear_uffd_wp(newpte);
6693 if (!pte_same(pte, newpte))
6694 set_huge_pte_at(mm, address, ptep, newpte);
6695 } else if (unlikely(is_pte_marker(pte))) {
6696 /* No other markers apply for now. */
6697 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6698 if (uffd_wp_resolve)
6699 /* Safe to modify directly (non-present->none). */
6700 huge_pte_clear(mm, address, ptep, psize);
6701 } else if (!huge_pte_none(pte)) {
6703 unsigned int shift = huge_page_shift(hstate_vma(vma));
6705 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6706 pte = huge_pte_modify(old_pte, newprot);
6707 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6709 pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte));
6710 else if (uffd_wp_resolve)
6711 pte = huge_pte_clear_uffd_wp(pte);
6712 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6716 if (unlikely(uffd_wp))
6717 /* Safe to modify directly (none->non-present). */
6718 set_huge_pte_at(mm, address, ptep,
6719 make_pte_marker(PTE_MARKER_UFFD_WP));
6724 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6725 * may have cleared our pud entry and done put_page on the page table:
6726 * once we release i_mmap_rwsem, another task can do the final put_page
6727 * and that page table be reused and filled with junk. If we actually
6728 * did unshare a page of pmds, flush the range corresponding to the pud.
6731 flush_hugetlb_tlb_range(vma, range.start, range.end);
6733 flush_hugetlb_tlb_range(vma, start, end);
6735 * No need to call mmu_notifier_invalidate_range() we are downgrading
6736 * page table protection not changing it to point to a new page.
6738 * See Documentation/mm/mmu_notifier.rst
6740 i_mmap_unlock_write(vma->vm_file->f_mapping);
6741 hugetlb_vma_unlock_write(vma);
6742 mmu_notifier_invalidate_range_end(&range);
6744 return pages << h->order;
6747 /* Return true if reservation was successful, false otherwise. */
6748 bool hugetlb_reserve_pages(struct inode *inode,
6750 struct vm_area_struct *vma,
6751 vm_flags_t vm_flags)
6754 struct hstate *h = hstate_inode(inode);
6755 struct hugepage_subpool *spool = subpool_inode(inode);
6756 struct resv_map *resv_map;
6757 struct hugetlb_cgroup *h_cg = NULL;
6758 long gbl_reserve, regions_needed = 0;
6760 /* This should never happen */
6762 VM_WARN(1, "%s called with a negative range\n", __func__);
6767 * vma specific semaphore used for pmd sharing and fault/truncation
6770 hugetlb_vma_lock_alloc(vma);
6773 * Only apply hugepage reservation if asked. At fault time, an
6774 * attempt will be made for VM_NORESERVE to allocate a page
6775 * without using reserves
6777 if (vm_flags & VM_NORESERVE)
6781 * Shared mappings base their reservation on the number of pages that
6782 * are already allocated on behalf of the file. Private mappings need
6783 * to reserve the full area even if read-only as mprotect() may be
6784 * called to make the mapping read-write. Assume !vma is a shm mapping
6786 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6788 * resv_map can not be NULL as hugetlb_reserve_pages is only
6789 * called for inodes for which resv_maps were created (see
6790 * hugetlbfs_get_inode).
6792 resv_map = inode_resv_map(inode);
6794 chg = region_chg(resv_map, from, to, ®ions_needed);
6796 /* Private mapping. */
6797 resv_map = resv_map_alloc();
6803 set_vma_resv_map(vma, resv_map);
6804 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6810 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6811 chg * pages_per_huge_page(h), &h_cg) < 0)
6814 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6815 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6818 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6822 * There must be enough pages in the subpool for the mapping. If
6823 * the subpool has a minimum size, there may be some global
6824 * reservations already in place (gbl_reserve).
6826 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6827 if (gbl_reserve < 0)
6828 goto out_uncharge_cgroup;
6831 * Check enough hugepages are available for the reservation.
6832 * Hand the pages back to the subpool if there are not
6834 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6838 * Account for the reservations made. Shared mappings record regions
6839 * that have reservations as they are shared by multiple VMAs.
6840 * When the last VMA disappears, the region map says how much
6841 * the reservation was and the page cache tells how much of
6842 * the reservation was consumed. Private mappings are per-VMA and
6843 * only the consumed reservations are tracked. When the VMA
6844 * disappears, the original reservation is the VMA size and the
6845 * consumed reservations are stored in the map. Hence, nothing
6846 * else has to be done for private mappings here
6848 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6849 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6851 if (unlikely(add < 0)) {
6852 hugetlb_acct_memory(h, -gbl_reserve);
6854 } else if (unlikely(chg > add)) {
6856 * pages in this range were added to the reserve
6857 * map between region_chg and region_add. This
6858 * indicates a race with alloc_huge_page. Adjust
6859 * the subpool and reserve counts modified above
6860 * based on the difference.
6865 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6866 * reference to h_cg->css. See comment below for detail.
6868 hugetlb_cgroup_uncharge_cgroup_rsvd(
6870 (chg - add) * pages_per_huge_page(h), h_cg);
6872 rsv_adjust = hugepage_subpool_put_pages(spool,
6874 hugetlb_acct_memory(h, -rsv_adjust);
6877 * The file_regions will hold their own reference to
6878 * h_cg->css. So we should release the reference held
6879 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6882 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6888 /* put back original number of pages, chg */
6889 (void)hugepage_subpool_put_pages(spool, chg);
6890 out_uncharge_cgroup:
6891 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6892 chg * pages_per_huge_page(h), h_cg);
6894 hugetlb_vma_lock_free(vma);
6895 if (!vma || vma->vm_flags & VM_MAYSHARE)
6896 /* Only call region_abort if the region_chg succeeded but the
6897 * region_add failed or didn't run.
6899 if (chg >= 0 && add < 0)
6900 region_abort(resv_map, from, to, regions_needed);
6901 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6902 kref_put(&resv_map->refs, resv_map_release);
6906 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6909 struct hstate *h = hstate_inode(inode);
6910 struct resv_map *resv_map = inode_resv_map(inode);
6912 struct hugepage_subpool *spool = subpool_inode(inode);
6916 * Since this routine can be called in the evict inode path for all
6917 * hugetlbfs inodes, resv_map could be NULL.
6920 chg = region_del(resv_map, start, end);
6922 * region_del() can fail in the rare case where a region
6923 * must be split and another region descriptor can not be
6924 * allocated. If end == LONG_MAX, it will not fail.
6930 spin_lock(&inode->i_lock);
6931 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6932 spin_unlock(&inode->i_lock);
6935 * If the subpool has a minimum size, the number of global
6936 * reservations to be released may be adjusted.
6938 * Note that !resv_map implies freed == 0. So (chg - freed)
6939 * won't go negative.
6941 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6942 hugetlb_acct_memory(h, -gbl_reserve);
6947 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6948 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6949 struct vm_area_struct *vma,
6950 unsigned long addr, pgoff_t idx)
6952 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6954 unsigned long sbase = saddr & PUD_MASK;
6955 unsigned long s_end = sbase + PUD_SIZE;
6957 /* Allow segments to share if only one is marked locked */
6958 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6959 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6962 * match the virtual addresses, permission and the alignment of the
6965 * Also, vma_lock (vm_private_data) is required for sharing.
6967 if (pmd_index(addr) != pmd_index(saddr) ||
6968 vm_flags != svm_flags ||
6969 !range_in_vma(svma, sbase, s_end) ||
6970 !svma->vm_private_data)
6976 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6978 unsigned long start = addr & PUD_MASK;
6979 unsigned long end = start + PUD_SIZE;
6981 #ifdef CONFIG_USERFAULTFD
6982 if (uffd_disable_huge_pmd_share(vma))
6986 * check on proper vm_flags and page table alignment
6988 if (!(vma->vm_flags & VM_MAYSHARE))
6990 if (!vma->vm_private_data) /* vma lock required for sharing */
6992 if (!range_in_vma(vma, start, end))
6998 * Determine if start,end range within vma could be mapped by shared pmd.
6999 * If yes, adjust start and end to cover range associated with possible
7000 * shared pmd mappings.
7002 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7003 unsigned long *start, unsigned long *end)
7005 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7006 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7009 * vma needs to span at least one aligned PUD size, and the range
7010 * must be at least partially within in.
7012 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7013 (*end <= v_start) || (*start >= v_end))
7016 /* Extend the range to be PUD aligned for a worst case scenario */
7017 if (*start > v_start)
7018 *start = ALIGN_DOWN(*start, PUD_SIZE);
7021 *end = ALIGN(*end, PUD_SIZE);
7025 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7026 * and returns the corresponding pte. While this is not necessary for the
7027 * !shared pmd case because we can allocate the pmd later as well, it makes the
7028 * code much cleaner. pmd allocation is essential for the shared case because
7029 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7030 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7031 * bad pmd for sharing.
7033 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7034 unsigned long addr, pud_t *pud)
7036 struct address_space *mapping = vma->vm_file->f_mapping;
7037 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7039 struct vm_area_struct *svma;
7040 unsigned long saddr;
7045 i_mmap_lock_read(mapping);
7046 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7050 saddr = page_table_shareable(svma, vma, addr, idx);
7052 spte = huge_pte_offset(svma->vm_mm, saddr,
7053 vma_mmu_pagesize(svma));
7055 get_page(virt_to_page(spte));
7064 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
7065 if (pud_none(*pud)) {
7066 pud_populate(mm, pud,
7067 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7070 put_page(virt_to_page(spte));
7074 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7075 i_mmap_unlock_read(mapping);
7080 * unmap huge page backed by shared pte.
7082 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7083 * indicated by page_count > 1, unmap is achieved by clearing pud and
7084 * decrementing the ref count. If count == 1, the pte page is not shared.
7086 * Called with page table lock held.
7088 * returns: 1 successfully unmapped a shared pte page
7089 * 0 the underlying pte page is not shared, or it is the last user
7091 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7092 unsigned long addr, pte_t *ptep)
7094 pgd_t *pgd = pgd_offset(mm, addr);
7095 p4d_t *p4d = p4d_offset(pgd, addr);
7096 pud_t *pud = pud_offset(p4d, addr);
7098 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7099 hugetlb_vma_assert_locked(vma);
7100 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7101 if (page_count(virt_to_page(ptep)) == 1)
7105 put_page(virt_to_page(ptep));
7110 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7112 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7113 unsigned long addr, pud_t *pud)
7118 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7119 unsigned long addr, pte_t *ptep)
7124 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7125 unsigned long *start, unsigned long *end)
7129 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7133 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7135 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7136 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7137 unsigned long addr, unsigned long sz)
7144 pgd = pgd_offset(mm, addr);
7145 p4d = p4d_alloc(mm, pgd, addr);
7148 pud = pud_alloc(mm, p4d, addr);
7150 if (sz == PUD_SIZE) {
7153 BUG_ON(sz != PMD_SIZE);
7154 if (want_pmd_share(vma, addr) && pud_none(*pud))
7155 pte = huge_pmd_share(mm, vma, addr, pud);
7157 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7160 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
7166 * huge_pte_offset() - Walk the page table to resolve the hugepage
7167 * entry at address @addr
7169 * Return: Pointer to page table entry (PUD or PMD) for
7170 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7171 * size @sz doesn't match the hugepage size at this level of the page
7174 pte_t *huge_pte_offset(struct mm_struct *mm,
7175 unsigned long addr, unsigned long sz)
7182 pgd = pgd_offset(mm, addr);
7183 if (!pgd_present(*pgd))
7185 p4d = p4d_offset(pgd, addr);
7186 if (!p4d_present(*p4d))
7189 pud = pud_offset(p4d, addr);
7191 /* must be pud huge, non-present or none */
7192 return (pte_t *)pud;
7193 if (!pud_present(*pud))
7195 /* must have a valid entry and size to go further */
7197 pmd = pmd_offset(pud, addr);
7198 /* must be pmd huge, non-present or none */
7199 return (pte_t *)pmd;
7203 * Return a mask that can be used to update an address to the last huge
7204 * page in a page table page mapping size. Used to skip non-present
7205 * page table entries when linearly scanning address ranges. Architectures
7206 * with unique huge page to page table relationships can define their own
7207 * version of this routine.
7209 unsigned long hugetlb_mask_last_page(struct hstate *h)
7211 unsigned long hp_size = huge_page_size(h);
7213 if (hp_size == PUD_SIZE)
7214 return P4D_SIZE - PUD_SIZE;
7215 else if (hp_size == PMD_SIZE)
7216 return PUD_SIZE - PMD_SIZE;
7223 /* See description above. Architectures can provide their own version. */
7224 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7226 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7227 if (huge_page_size(h) == PMD_SIZE)
7228 return PUD_SIZE - PMD_SIZE;
7233 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7236 * These functions are overwritable if your architecture needs its own
7239 struct page * __weak
7240 follow_huge_addr(struct mm_struct *mm, unsigned long address,
7243 return ERR_PTR(-EINVAL);
7246 struct page * __weak
7247 follow_huge_pd(struct vm_area_struct *vma,
7248 unsigned long address, hugepd_t hpd, int flags, int pdshift)
7250 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
7254 struct page * __weak
7255 follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags)
7257 struct hstate *h = hstate_vma(vma);
7258 struct mm_struct *mm = vma->vm_mm;
7259 struct page *page = NULL;
7264 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
7265 * follow_hugetlb_page().
7267 if (WARN_ON_ONCE(flags & FOLL_PIN))
7271 ptep = huge_pte_offset(mm, address, huge_page_size(h));
7275 ptl = huge_pte_lock(h, mm, ptep);
7276 pte = huge_ptep_get(ptep);
7277 if (pte_present(pte)) {
7278 page = pte_page(pte) +
7279 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
7281 * try_grab_page() should always succeed here, because: a) we
7282 * hold the pmd (ptl) lock, and b) we've just checked that the
7283 * huge pmd (head) page is present in the page tables. The ptl
7284 * prevents the head page and tail pages from being rearranged
7285 * in any way. So this page must be available at this point,
7286 * unless the page refcount overflowed:
7288 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
7293 if (is_hugetlb_entry_migration(pte)) {
7295 __migration_entry_wait_huge(ptep, ptl);
7299 * hwpoisoned entry is treated as no_page_table in
7300 * follow_page_mask().
7308 struct page * __weak
7309 follow_huge_pud(struct mm_struct *mm, unsigned long address,
7310 pud_t *pud, int flags)
7312 struct page *page = NULL;
7316 if (WARN_ON_ONCE(flags & FOLL_PIN))
7320 ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud);
7321 if (!pud_huge(*pud))
7323 pte = huge_ptep_get((pte_t *)pud);
7324 if (pte_present(pte)) {
7325 page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
7326 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
7331 if (is_hugetlb_entry_migration(pte)) {
7333 __migration_entry_wait(mm, (pte_t *)pud, ptl);
7337 * hwpoisoned entry is treated as no_page_table in
7338 * follow_page_mask().
7346 struct page * __weak
7347 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
7349 if (flags & (FOLL_GET | FOLL_PIN))
7352 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
7355 int isolate_hugetlb(struct page *page, struct list_head *list)
7359 spin_lock_irq(&hugetlb_lock);
7360 if (!PageHeadHuge(page) ||
7361 !HPageMigratable(page) ||
7362 !get_page_unless_zero(page)) {
7366 ClearHPageMigratable(page);
7367 list_move_tail(&page->lru, list);
7369 spin_unlock_irq(&hugetlb_lock);
7373 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
7378 spin_lock_irq(&hugetlb_lock);
7379 if (PageHeadHuge(page)) {
7381 if (HPageFreed(page))
7383 else if (HPageMigratable(page))
7384 ret = get_page_unless_zero(page);
7388 spin_unlock_irq(&hugetlb_lock);
7392 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
7396 spin_lock_irq(&hugetlb_lock);
7397 ret = __get_huge_page_for_hwpoison(pfn, flags);
7398 spin_unlock_irq(&hugetlb_lock);
7402 void putback_active_hugepage(struct page *page)
7404 spin_lock_irq(&hugetlb_lock);
7405 SetHPageMigratable(page);
7406 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7407 spin_unlock_irq(&hugetlb_lock);
7411 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
7413 struct hstate *h = page_hstate(oldpage);
7415 hugetlb_cgroup_migrate(oldpage, newpage);
7416 set_page_owner_migrate_reason(newpage, reason);
7419 * transfer temporary state of the new huge page. This is
7420 * reverse to other transitions because the newpage is going to
7421 * be final while the old one will be freed so it takes over
7422 * the temporary status.
7424 * Also note that we have to transfer the per-node surplus state
7425 * here as well otherwise the global surplus count will not match
7428 if (HPageTemporary(newpage)) {
7429 int old_nid = page_to_nid(oldpage);
7430 int new_nid = page_to_nid(newpage);
7432 SetHPageTemporary(oldpage);
7433 ClearHPageTemporary(newpage);
7436 * There is no need to transfer the per-node surplus state
7437 * when we do not cross the node.
7439 if (new_nid == old_nid)
7441 spin_lock_irq(&hugetlb_lock);
7442 if (h->surplus_huge_pages_node[old_nid]) {
7443 h->surplus_huge_pages_node[old_nid]--;
7444 h->surplus_huge_pages_node[new_nid]++;
7446 spin_unlock_irq(&hugetlb_lock);
7450 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7451 unsigned long start,
7454 struct hstate *h = hstate_vma(vma);
7455 unsigned long sz = huge_page_size(h);
7456 struct mm_struct *mm = vma->vm_mm;
7457 struct mmu_notifier_range range;
7458 unsigned long address;
7462 if (!(vma->vm_flags & VM_MAYSHARE))
7468 flush_cache_range(vma, start, end);
7470 * No need to call adjust_range_if_pmd_sharing_possible(), because
7471 * we have already done the PUD_SIZE alignment.
7473 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7475 mmu_notifier_invalidate_range_start(&range);
7476 hugetlb_vma_lock_write(vma);
7477 i_mmap_lock_write(vma->vm_file->f_mapping);
7478 for (address = start; address < end; address += PUD_SIZE) {
7479 ptep = huge_pte_offset(mm, address, sz);
7482 ptl = huge_pte_lock(h, mm, ptep);
7483 huge_pmd_unshare(mm, vma, address, ptep);
7486 flush_hugetlb_tlb_range(vma, start, end);
7487 i_mmap_unlock_write(vma->vm_file->f_mapping);
7488 hugetlb_vma_unlock_write(vma);
7490 * No need to call mmu_notifier_invalidate_range(), see
7491 * Documentation/mm/mmu_notifier.rst.
7493 mmu_notifier_invalidate_range_end(&range);
7497 * This function will unconditionally remove all the shared pmd pgtable entries
7498 * within the specific vma for a hugetlbfs memory range.
7500 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7502 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7503 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7507 static bool cma_reserve_called __initdata;
7509 static int __init cmdline_parse_hugetlb_cma(char *p)
7516 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7519 if (s[count] == ':') {
7520 if (tmp >= MAX_NUMNODES)
7522 nid = array_index_nospec(tmp, MAX_NUMNODES);
7525 tmp = memparse(s, &s);
7526 hugetlb_cma_size_in_node[nid] = tmp;
7527 hugetlb_cma_size += tmp;
7530 * Skip the separator if have one, otherwise
7531 * break the parsing.
7538 hugetlb_cma_size = memparse(p, &p);
7546 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7548 void __init hugetlb_cma_reserve(int order)
7550 unsigned long size, reserved, per_node;
7551 bool node_specific_cma_alloc = false;
7554 cma_reserve_called = true;
7556 if (!hugetlb_cma_size)
7559 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7560 if (hugetlb_cma_size_in_node[nid] == 0)
7563 if (!node_online(nid)) {
7564 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7565 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7566 hugetlb_cma_size_in_node[nid] = 0;
7570 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7571 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7572 nid, (PAGE_SIZE << order) / SZ_1M);
7573 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7574 hugetlb_cma_size_in_node[nid] = 0;
7576 node_specific_cma_alloc = true;
7580 /* Validate the CMA size again in case some invalid nodes specified. */
7581 if (!hugetlb_cma_size)
7584 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7585 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7586 (PAGE_SIZE << order) / SZ_1M);
7587 hugetlb_cma_size = 0;
7591 if (!node_specific_cma_alloc) {
7593 * If 3 GB area is requested on a machine with 4 numa nodes,
7594 * let's allocate 1 GB on first three nodes and ignore the last one.
7596 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7597 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7598 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7602 for_each_online_node(nid) {
7604 char name[CMA_MAX_NAME];
7606 if (node_specific_cma_alloc) {
7607 if (hugetlb_cma_size_in_node[nid] == 0)
7610 size = hugetlb_cma_size_in_node[nid];
7612 size = min(per_node, hugetlb_cma_size - reserved);
7615 size = round_up(size, PAGE_SIZE << order);
7617 snprintf(name, sizeof(name), "hugetlb%d", nid);
7619 * Note that 'order per bit' is based on smallest size that
7620 * may be returned to CMA allocator in the case of
7621 * huge page demotion.
7623 res = cma_declare_contiguous_nid(0, size, 0,
7624 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7626 &hugetlb_cma[nid], nid);
7628 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7634 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7637 if (reserved >= hugetlb_cma_size)
7643 * hugetlb_cma_size is used to determine if allocations from
7644 * cma are possible. Set to zero if no cma regions are set up.
7646 hugetlb_cma_size = 0;
7649 static void __init hugetlb_cma_check(void)
7651 if (!hugetlb_cma_size || cma_reserve_called)
7654 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7657 #endif /* CONFIG_CMA */