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) { }
1585 * Remove hugetlb page from lists, and update dtor so that page appears
1586 * as just a compound page.
1588 * A reference is held on the page, except in the case of demote.
1590 * Must be called with hugetlb lock held.
1592 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1593 bool adjust_surplus,
1596 int nid = page_to_nid(page);
1598 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1599 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1601 lockdep_assert_held(&hugetlb_lock);
1602 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1605 list_del(&page->lru);
1607 if (HPageFreed(page)) {
1608 h->free_huge_pages--;
1609 h->free_huge_pages_node[nid]--;
1611 if (adjust_surplus) {
1612 h->surplus_huge_pages--;
1613 h->surplus_huge_pages_node[nid]--;
1619 * For non-gigantic pages set the destructor to the normal compound
1620 * page dtor. This is needed in case someone takes an additional
1621 * temporary ref to the page, and freeing is delayed until they drop
1624 * For gigantic pages set the destructor to the null dtor. This
1625 * destructor will never be called. Before freeing the gigantic
1626 * page destroy_compound_gigantic_page will turn the compound page
1627 * into a simple group of pages. After this the destructor does not
1630 * This handles the case where more than one ref is held when and
1631 * after update_and_free_page is called.
1633 * In the case of demote we do not ref count the page as it will soon
1634 * be turned into a page of smaller size.
1637 set_page_refcounted(page);
1638 if (hstate_is_gigantic(h))
1639 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1641 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1644 h->nr_huge_pages_node[nid]--;
1647 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1648 bool adjust_surplus)
1650 __remove_hugetlb_page(h, page, adjust_surplus, false);
1653 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1654 bool adjust_surplus)
1656 __remove_hugetlb_page(h, page, adjust_surplus, true);
1659 static void add_hugetlb_page(struct hstate *h, struct page *page,
1660 bool adjust_surplus)
1663 int nid = page_to_nid(page);
1665 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1667 lockdep_assert_held(&hugetlb_lock);
1669 INIT_LIST_HEAD(&page->lru);
1671 h->nr_huge_pages_node[nid]++;
1673 if (adjust_surplus) {
1674 h->surplus_huge_pages++;
1675 h->surplus_huge_pages_node[nid]++;
1678 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1679 set_page_private(page, 0);
1681 * We have to set HPageVmemmapOptimized again as above
1682 * set_page_private(page, 0) cleared it.
1684 SetHPageVmemmapOptimized(page);
1687 * This page is about to be managed by the hugetlb allocator and
1688 * should have no users. Drop our reference, and check for others
1691 zeroed = put_page_testzero(page);
1694 * It is VERY unlikely soneone else has taken a ref on
1695 * the page. In this case, we simply return as the
1696 * hugetlb destructor (free_huge_page) will be called
1697 * when this other ref is dropped.
1701 arch_clear_hugepage_flags(page);
1702 enqueue_huge_page(h, page);
1705 static void __update_and_free_page(struct hstate *h, struct page *page)
1708 struct page *subpage;
1710 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1714 * If we don't know which subpages are hwpoisoned, we can't free
1715 * the hugepage, so it's leaked intentionally.
1717 if (HPageRawHwpUnreliable(page))
1720 if (hugetlb_vmemmap_restore(h, page)) {
1721 spin_lock_irq(&hugetlb_lock);
1723 * If we cannot allocate vmemmap pages, just refuse to free the
1724 * page and put the page back on the hugetlb free list and treat
1725 * as a surplus page.
1727 add_hugetlb_page(h, page, true);
1728 spin_unlock_irq(&hugetlb_lock);
1733 * Move PageHWPoison flag from head page to the raw error pages,
1734 * which makes any healthy subpages reusable.
1736 if (unlikely(PageHWPoison(page)))
1737 hugetlb_clear_page_hwpoison(page);
1739 for (i = 0; i < pages_per_huge_page(h); i++) {
1740 subpage = nth_page(page, i);
1741 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1742 1 << PG_referenced | 1 << PG_dirty |
1743 1 << PG_active | 1 << PG_private |
1748 * Non-gigantic pages demoted from CMA allocated gigantic pages
1749 * need to be given back to CMA in free_gigantic_page.
1751 if (hstate_is_gigantic(h) ||
1752 hugetlb_cma_page(page, huge_page_order(h))) {
1753 destroy_compound_gigantic_page(page, huge_page_order(h));
1754 free_gigantic_page(page, huge_page_order(h));
1756 __free_pages(page, huge_page_order(h));
1761 * As update_and_free_page() can be called under any context, so we cannot
1762 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1763 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1764 * the vmemmap pages.
1766 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1767 * freed and frees them one-by-one. As the page->mapping pointer is going
1768 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1769 * structure of a lockless linked list of huge pages to be freed.
1771 static LLIST_HEAD(hpage_freelist);
1773 static void free_hpage_workfn(struct work_struct *work)
1775 struct llist_node *node;
1777 node = llist_del_all(&hpage_freelist);
1783 page = container_of((struct address_space **)node,
1784 struct page, mapping);
1786 page->mapping = NULL;
1788 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1789 * is going to trigger because a previous call to
1790 * remove_hugetlb_page() will set_compound_page_dtor(page,
1791 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1793 h = size_to_hstate(page_size(page));
1795 __update_and_free_page(h, page);
1800 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1802 static inline void flush_free_hpage_work(struct hstate *h)
1804 if (hugetlb_vmemmap_optimizable(h))
1805 flush_work(&free_hpage_work);
1808 static void update_and_free_page(struct hstate *h, struct page *page,
1811 if (!HPageVmemmapOptimized(page) || !atomic) {
1812 __update_and_free_page(h, page);
1817 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1819 * Only call schedule_work() if hpage_freelist is previously
1820 * empty. Otherwise, schedule_work() had been called but the workfn
1821 * hasn't retrieved the list yet.
1823 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1824 schedule_work(&free_hpage_work);
1827 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1829 struct page *page, *t_page;
1831 list_for_each_entry_safe(page, t_page, list, lru) {
1832 update_and_free_page(h, page, false);
1837 struct hstate *size_to_hstate(unsigned long size)
1841 for_each_hstate(h) {
1842 if (huge_page_size(h) == size)
1848 void free_huge_page(struct page *page)
1851 * Can't pass hstate in here because it is called from the
1852 * compound page destructor.
1854 struct hstate *h = page_hstate(page);
1855 int nid = page_to_nid(page);
1856 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1857 bool restore_reserve;
1858 unsigned long flags;
1860 VM_BUG_ON_PAGE(page_count(page), page);
1861 VM_BUG_ON_PAGE(page_mapcount(page), page);
1863 hugetlb_set_page_subpool(page, NULL);
1865 __ClearPageAnonExclusive(page);
1866 page->mapping = NULL;
1867 restore_reserve = HPageRestoreReserve(page);
1868 ClearHPageRestoreReserve(page);
1871 * If HPageRestoreReserve was set on page, page allocation consumed a
1872 * reservation. If the page was associated with a subpool, there
1873 * would have been a page reserved in the subpool before allocation
1874 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1875 * reservation, do not call hugepage_subpool_put_pages() as this will
1876 * remove the reserved page from the subpool.
1878 if (!restore_reserve) {
1880 * A return code of zero implies that the subpool will be
1881 * under its minimum size if the reservation is not restored
1882 * after page is free. Therefore, force restore_reserve
1885 if (hugepage_subpool_put_pages(spool, 1) == 0)
1886 restore_reserve = true;
1889 spin_lock_irqsave(&hugetlb_lock, flags);
1890 ClearHPageMigratable(page);
1891 hugetlb_cgroup_uncharge_page(hstate_index(h),
1892 pages_per_huge_page(h), page);
1893 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1894 pages_per_huge_page(h), page);
1895 if (restore_reserve)
1896 h->resv_huge_pages++;
1898 if (HPageTemporary(page)) {
1899 remove_hugetlb_page(h, page, false);
1900 spin_unlock_irqrestore(&hugetlb_lock, flags);
1901 update_and_free_page(h, page, true);
1902 } else if (h->surplus_huge_pages_node[nid]) {
1903 /* remove the page from active list */
1904 remove_hugetlb_page(h, page, true);
1905 spin_unlock_irqrestore(&hugetlb_lock, flags);
1906 update_and_free_page(h, page, true);
1908 arch_clear_hugepage_flags(page);
1909 enqueue_huge_page(h, page);
1910 spin_unlock_irqrestore(&hugetlb_lock, flags);
1915 * Must be called with the hugetlb lock held
1917 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1919 lockdep_assert_held(&hugetlb_lock);
1921 h->nr_huge_pages_node[nid]++;
1924 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1926 hugetlb_vmemmap_optimize(h, page);
1927 INIT_LIST_HEAD(&page->lru);
1928 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1929 hugetlb_set_page_subpool(page, NULL);
1930 set_hugetlb_cgroup(page, NULL);
1931 set_hugetlb_cgroup_rsvd(page, NULL);
1934 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1936 __prep_new_huge_page(h, page);
1937 spin_lock_irq(&hugetlb_lock);
1938 __prep_account_new_huge_page(h, nid);
1939 spin_unlock_irq(&hugetlb_lock);
1942 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1946 int nr_pages = 1 << order;
1949 /* we rely on prep_new_huge_page to set the destructor */
1950 set_compound_order(page, order);
1951 __ClearPageReserved(page);
1952 __SetPageHead(page);
1953 for (i = 0; i < nr_pages; i++) {
1954 p = nth_page(page, i);
1957 * For gigantic hugepages allocated through bootmem at
1958 * boot, it's safer to be consistent with the not-gigantic
1959 * hugepages and clear the PG_reserved bit from all tail pages
1960 * too. Otherwise drivers using get_user_pages() to access tail
1961 * pages may get the reference counting wrong if they see
1962 * PG_reserved set on a tail page (despite the head page not
1963 * having PG_reserved set). Enforcing this consistency between
1964 * head and tail pages allows drivers to optimize away a check
1965 * on the head page when they need know if put_page() is needed
1966 * after get_user_pages().
1968 if (i != 0) /* head page cleared above */
1969 __ClearPageReserved(p);
1971 * Subtle and very unlikely
1973 * Gigantic 'page allocators' such as memblock or cma will
1974 * return a set of pages with each page ref counted. We need
1975 * to turn this set of pages into a compound page with tail
1976 * page ref counts set to zero. Code such as speculative page
1977 * cache adding could take a ref on a 'to be' tail page.
1978 * We need to respect any increased ref count, and only set
1979 * the ref count to zero if count is currently 1. If count
1980 * is not 1, we return an error. An error return indicates
1981 * the set of pages can not be converted to a gigantic page.
1982 * The caller who allocated the pages should then discard the
1983 * pages using the appropriate free interface.
1985 * In the case of demote, the ref count will be zero.
1988 if (!page_ref_freeze(p, 1)) {
1989 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1993 VM_BUG_ON_PAGE(page_count(p), p);
1996 set_compound_head(p, page);
1998 atomic_set(compound_mapcount_ptr(page), -1);
1999 atomic_set(compound_pincount_ptr(page), 0);
2003 /* undo page modifications made above */
2004 for (j = 0; j < i; j++) {
2005 p = nth_page(page, j);
2007 clear_compound_head(p);
2008 set_page_refcounted(p);
2010 /* need to clear PG_reserved on remaining tail pages */
2011 for (; j < nr_pages; j++) {
2012 p = nth_page(page, j);
2013 __ClearPageReserved(p);
2015 set_compound_order(page, 0);
2017 page[1].compound_nr = 0;
2019 __ClearPageHead(page);
2023 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
2025 return __prep_compound_gigantic_page(page, order, false);
2028 static bool prep_compound_gigantic_page_for_demote(struct page *page,
2031 return __prep_compound_gigantic_page(page, order, true);
2035 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2036 * transparent huge pages. See the PageTransHuge() documentation for more
2039 int PageHuge(struct page *page)
2041 if (!PageCompound(page))
2044 page = compound_head(page);
2045 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
2047 EXPORT_SYMBOL_GPL(PageHuge);
2050 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
2051 * normal or transparent huge pages.
2053 int PageHeadHuge(struct page *page_head)
2055 if (!PageHead(page_head))
2058 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
2060 EXPORT_SYMBOL_GPL(PageHeadHuge);
2063 * Find and lock address space (mapping) in write mode.
2065 * Upon entry, the page is locked which means that page_mapping() is
2066 * stable. Due to locking order, we can only trylock_write. If we can
2067 * not get the lock, simply return NULL to caller.
2069 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2071 struct address_space *mapping = page_mapping(hpage);
2076 if (i_mmap_trylock_write(mapping))
2082 pgoff_t hugetlb_basepage_index(struct page *page)
2084 struct page *page_head = compound_head(page);
2085 pgoff_t index = page_index(page_head);
2086 unsigned long compound_idx;
2088 if (compound_order(page_head) >= MAX_ORDER)
2089 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2091 compound_idx = page - page_head;
2093 return (index << compound_order(page_head)) + compound_idx;
2096 static struct page *alloc_buddy_huge_page(struct hstate *h,
2097 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2098 nodemask_t *node_alloc_noretry)
2100 int order = huge_page_order(h);
2102 bool alloc_try_hard = true;
2106 * By default we always try hard to allocate the page with
2107 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2108 * a loop (to adjust global huge page counts) and previous allocation
2109 * failed, do not continue to try hard on the same node. Use the
2110 * node_alloc_noretry bitmap to manage this state information.
2112 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2113 alloc_try_hard = false;
2114 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2116 gfp_mask |= __GFP_RETRY_MAYFAIL;
2117 if (nid == NUMA_NO_NODE)
2118 nid = numa_mem_id();
2120 page = __alloc_pages(gfp_mask, order, nid, nmask);
2122 /* Freeze head page */
2123 if (page && !page_ref_freeze(page, 1)) {
2124 __free_pages(page, order);
2125 if (retry) { /* retry once */
2129 /* WOW! twice in a row. */
2130 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2135 __count_vm_event(HTLB_BUDDY_PGALLOC);
2137 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2140 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2141 * indicates an overall state change. Clear bit so that we resume
2142 * normal 'try hard' allocations.
2144 if (node_alloc_noretry && page && !alloc_try_hard)
2145 node_clear(nid, *node_alloc_noretry);
2148 * If we tried hard to get a page but failed, set bit so that
2149 * subsequent attempts will not try as hard until there is an
2150 * overall state change.
2152 if (node_alloc_noretry && !page && alloc_try_hard)
2153 node_set(nid, *node_alloc_noretry);
2159 * Common helper to allocate a fresh hugetlb page. All specific allocators
2160 * should use this function to get new hugetlb pages
2162 * Note that returned page is 'frozen': ref count of head page and all tail
2165 static struct page *alloc_fresh_huge_page(struct hstate *h,
2166 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2167 nodemask_t *node_alloc_noretry)
2173 if (hstate_is_gigantic(h))
2174 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
2176 page = alloc_buddy_huge_page(h, gfp_mask,
2177 nid, nmask, node_alloc_noretry);
2181 if (hstate_is_gigantic(h)) {
2182 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
2184 * Rare failure to convert pages to compound page.
2185 * Free pages and try again - ONCE!
2187 free_gigantic_page(page, huge_page_order(h));
2195 prep_new_huge_page(h, page, page_to_nid(page));
2201 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2204 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2205 nodemask_t *node_alloc_noretry)
2209 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2211 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2212 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
2213 node_alloc_noretry);
2221 free_huge_page(page); /* free it into the hugepage allocator */
2227 * Remove huge page from pool from next node to free. Attempt to keep
2228 * persistent huge pages more or less balanced over allowed nodes.
2229 * This routine only 'removes' the hugetlb page. The caller must make
2230 * an additional call to free the page to low level allocators.
2231 * Called with hugetlb_lock locked.
2233 static struct page *remove_pool_huge_page(struct hstate *h,
2234 nodemask_t *nodes_allowed,
2238 struct page *page = NULL;
2240 lockdep_assert_held(&hugetlb_lock);
2241 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2243 * If we're returning unused surplus pages, only examine
2244 * nodes with surplus pages.
2246 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2247 !list_empty(&h->hugepage_freelists[node])) {
2248 page = list_entry(h->hugepage_freelists[node].next,
2250 remove_hugetlb_page(h, page, acct_surplus);
2259 * Dissolve a given free hugepage into free buddy pages. This function does
2260 * nothing for in-use hugepages and non-hugepages.
2261 * This function returns values like below:
2263 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2264 * when the system is under memory pressure and the feature of
2265 * freeing unused vmemmap pages associated with each hugetlb page
2267 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2268 * (allocated or reserved.)
2269 * 0: successfully dissolved free hugepages or the page is not a
2270 * hugepage (considered as already dissolved)
2272 int dissolve_free_huge_page(struct page *page)
2277 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2278 if (!PageHuge(page))
2281 spin_lock_irq(&hugetlb_lock);
2282 if (!PageHuge(page)) {
2287 if (!page_count(page)) {
2288 struct page *head = compound_head(page);
2289 struct hstate *h = page_hstate(head);
2290 if (!available_huge_pages(h))
2294 * We should make sure that the page is already on the free list
2295 * when it is dissolved.
2297 if (unlikely(!HPageFreed(head))) {
2298 spin_unlock_irq(&hugetlb_lock);
2302 * Theoretically, we should return -EBUSY when we
2303 * encounter this race. In fact, we have a chance
2304 * to successfully dissolve the page if we do a
2305 * retry. Because the race window is quite small.
2306 * If we seize this opportunity, it is an optimization
2307 * for increasing the success rate of dissolving page.
2312 remove_hugetlb_page(h, head, false);
2313 h->max_huge_pages--;
2314 spin_unlock_irq(&hugetlb_lock);
2317 * Normally update_and_free_page will allocate required vmemmmap
2318 * before freeing the page. update_and_free_page will fail to
2319 * free the page if it can not allocate required vmemmap. We
2320 * need to adjust max_huge_pages if the page is not freed.
2321 * Attempt to allocate vmemmmap here so that we can take
2322 * appropriate action on failure.
2324 rc = hugetlb_vmemmap_restore(h, head);
2326 update_and_free_page(h, head, false);
2328 spin_lock_irq(&hugetlb_lock);
2329 add_hugetlb_page(h, head, false);
2330 h->max_huge_pages++;
2331 spin_unlock_irq(&hugetlb_lock);
2337 spin_unlock_irq(&hugetlb_lock);
2342 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2343 * make specified memory blocks removable from the system.
2344 * Note that this will dissolve a free gigantic hugepage completely, if any
2345 * part of it lies within the given range.
2346 * Also note that if dissolve_free_huge_page() returns with an error, all
2347 * free hugepages that were dissolved before that error are lost.
2349 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2357 if (!hugepages_supported())
2360 order = huge_page_order(&default_hstate);
2362 order = min(order, huge_page_order(h));
2364 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2365 page = pfn_to_page(pfn);
2366 rc = dissolve_free_huge_page(page);
2375 * Allocates a fresh surplus page from the page allocator.
2377 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2378 int nid, nodemask_t *nmask)
2380 struct page *page = NULL;
2382 if (hstate_is_gigantic(h))
2385 spin_lock_irq(&hugetlb_lock);
2386 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2388 spin_unlock_irq(&hugetlb_lock);
2390 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2394 spin_lock_irq(&hugetlb_lock);
2396 * We could have raced with the pool size change.
2397 * Double check that and simply deallocate the new page
2398 * if we would end up overcommiting the surpluses. Abuse
2399 * temporary page to workaround the nasty free_huge_page
2402 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2403 SetHPageTemporary(page);
2404 spin_unlock_irq(&hugetlb_lock);
2405 free_huge_page(page);
2409 h->surplus_huge_pages++;
2410 h->surplus_huge_pages_node[page_to_nid(page)]++;
2413 spin_unlock_irq(&hugetlb_lock);
2418 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2419 int nid, nodemask_t *nmask)
2423 if (hstate_is_gigantic(h))
2426 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2430 /* fresh huge pages are frozen */
2431 set_page_refcounted(page);
2434 * We do not account these pages as surplus because they are only
2435 * temporary and will be released properly on the last reference
2437 SetHPageTemporary(page);
2443 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2446 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2447 struct vm_area_struct *vma, unsigned long addr)
2449 struct page *page = NULL;
2450 struct mempolicy *mpol;
2451 gfp_t gfp_mask = htlb_alloc_mask(h);
2453 nodemask_t *nodemask;
2455 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2456 if (mpol_is_preferred_many(mpol)) {
2457 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2459 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2460 page = alloc_surplus_huge_page(h, gfp, nid, nodemask);
2462 /* Fallback to all nodes if page==NULL */
2467 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2468 mpol_cond_put(mpol);
2472 /* page migration callback function */
2473 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2474 nodemask_t *nmask, gfp_t gfp_mask)
2476 spin_lock_irq(&hugetlb_lock);
2477 if (available_huge_pages(h)) {
2480 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2482 spin_unlock_irq(&hugetlb_lock);
2486 spin_unlock_irq(&hugetlb_lock);
2488 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2491 /* mempolicy aware migration callback */
2492 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2493 unsigned long address)
2495 struct mempolicy *mpol;
2496 nodemask_t *nodemask;
2501 gfp_mask = htlb_alloc_mask(h);
2502 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2503 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2504 mpol_cond_put(mpol);
2510 * Increase the hugetlb pool such that it can accommodate a reservation
2513 static int gather_surplus_pages(struct hstate *h, long delta)
2514 __must_hold(&hugetlb_lock)
2516 LIST_HEAD(surplus_list);
2517 struct page *page, *tmp;
2520 long needed, allocated;
2521 bool alloc_ok = true;
2523 lockdep_assert_held(&hugetlb_lock);
2524 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2526 h->resv_huge_pages += delta;
2534 spin_unlock_irq(&hugetlb_lock);
2535 for (i = 0; i < needed; i++) {
2536 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2537 NUMA_NO_NODE, NULL);
2542 list_add(&page->lru, &surplus_list);
2548 * After retaking hugetlb_lock, we need to recalculate 'needed'
2549 * because either resv_huge_pages or free_huge_pages may have changed.
2551 spin_lock_irq(&hugetlb_lock);
2552 needed = (h->resv_huge_pages + delta) -
2553 (h->free_huge_pages + allocated);
2558 * We were not able to allocate enough pages to
2559 * satisfy the entire reservation so we free what
2560 * we've allocated so far.
2565 * The surplus_list now contains _at_least_ the number of extra pages
2566 * needed to accommodate the reservation. Add the appropriate number
2567 * of pages to the hugetlb pool and free the extras back to the buddy
2568 * allocator. Commit the entire reservation here to prevent another
2569 * process from stealing the pages as they are added to the pool but
2570 * before they are reserved.
2572 needed += allocated;
2573 h->resv_huge_pages += delta;
2576 /* Free the needed pages to the hugetlb pool */
2577 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2580 /* Add the page to the hugetlb allocator */
2581 enqueue_huge_page(h, page);
2584 spin_unlock_irq(&hugetlb_lock);
2587 * Free unnecessary surplus pages to the buddy allocator.
2588 * Pages have no ref count, call free_huge_page directly.
2590 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2591 free_huge_page(page);
2592 spin_lock_irq(&hugetlb_lock);
2598 * This routine has two main purposes:
2599 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2600 * in unused_resv_pages. This corresponds to the prior adjustments made
2601 * to the associated reservation map.
2602 * 2) Free any unused surplus pages that may have been allocated to satisfy
2603 * the reservation. As many as unused_resv_pages may be freed.
2605 static void return_unused_surplus_pages(struct hstate *h,
2606 unsigned long unused_resv_pages)
2608 unsigned long nr_pages;
2610 LIST_HEAD(page_list);
2612 lockdep_assert_held(&hugetlb_lock);
2613 /* Uncommit the reservation */
2614 h->resv_huge_pages -= unused_resv_pages;
2616 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2620 * Part (or even all) of the reservation could have been backed
2621 * by pre-allocated pages. Only free surplus pages.
2623 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2626 * We want to release as many surplus pages as possible, spread
2627 * evenly across all nodes with memory. Iterate across these nodes
2628 * until we can no longer free unreserved surplus pages. This occurs
2629 * when the nodes with surplus pages have no free pages.
2630 * remove_pool_huge_page() will balance the freed pages across the
2631 * on-line nodes with memory and will handle the hstate accounting.
2633 while (nr_pages--) {
2634 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2638 list_add(&page->lru, &page_list);
2642 spin_unlock_irq(&hugetlb_lock);
2643 update_and_free_pages_bulk(h, &page_list);
2644 spin_lock_irq(&hugetlb_lock);
2649 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2650 * are used by the huge page allocation routines to manage reservations.
2652 * vma_needs_reservation is called to determine if the huge page at addr
2653 * within the vma has an associated reservation. If a reservation is
2654 * needed, the value 1 is returned. The caller is then responsible for
2655 * managing the global reservation and subpool usage counts. After
2656 * the huge page has been allocated, vma_commit_reservation is called
2657 * to add the page to the reservation map. If the page allocation fails,
2658 * the reservation must be ended instead of committed. vma_end_reservation
2659 * is called in such cases.
2661 * In the normal case, vma_commit_reservation returns the same value
2662 * as the preceding vma_needs_reservation call. The only time this
2663 * is not the case is if a reserve map was changed between calls. It
2664 * is the responsibility of the caller to notice the difference and
2665 * take appropriate action.
2667 * vma_add_reservation is used in error paths where a reservation must
2668 * be restored when a newly allocated huge page must be freed. It is
2669 * to be called after calling vma_needs_reservation to determine if a
2670 * reservation exists.
2672 * vma_del_reservation is used in error paths where an entry in the reserve
2673 * map was created during huge page allocation and must be removed. It is to
2674 * be called after calling vma_needs_reservation to determine if a reservation
2677 enum vma_resv_mode {
2684 static long __vma_reservation_common(struct hstate *h,
2685 struct vm_area_struct *vma, unsigned long addr,
2686 enum vma_resv_mode mode)
2688 struct resv_map *resv;
2691 long dummy_out_regions_needed;
2693 resv = vma_resv_map(vma);
2697 idx = vma_hugecache_offset(h, vma, addr);
2699 case VMA_NEEDS_RESV:
2700 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2701 /* We assume that vma_reservation_* routines always operate on
2702 * 1 page, and that adding to resv map a 1 page entry can only
2703 * ever require 1 region.
2705 VM_BUG_ON(dummy_out_regions_needed != 1);
2707 case VMA_COMMIT_RESV:
2708 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2709 /* region_add calls of range 1 should never fail. */
2713 region_abort(resv, idx, idx + 1, 1);
2717 if (vma->vm_flags & VM_MAYSHARE) {
2718 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2719 /* region_add calls of range 1 should never fail. */
2722 region_abort(resv, idx, idx + 1, 1);
2723 ret = region_del(resv, idx, idx + 1);
2727 if (vma->vm_flags & VM_MAYSHARE) {
2728 region_abort(resv, idx, idx + 1, 1);
2729 ret = region_del(resv, idx, idx + 1);
2731 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2732 /* region_add calls of range 1 should never fail. */
2740 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2743 * We know private mapping must have HPAGE_RESV_OWNER set.
2745 * In most cases, reserves always exist for private mappings.
2746 * However, a file associated with mapping could have been
2747 * hole punched or truncated after reserves were consumed.
2748 * As subsequent fault on such a range will not use reserves.
2749 * Subtle - The reserve map for private mappings has the
2750 * opposite meaning than that of shared mappings. If NO
2751 * entry is in the reserve map, it means a reservation exists.
2752 * If an entry exists in the reserve map, it means the
2753 * reservation has already been consumed. As a result, the
2754 * return value of this routine is the opposite of the
2755 * value returned from reserve map manipulation routines above.
2764 static long vma_needs_reservation(struct hstate *h,
2765 struct vm_area_struct *vma, unsigned long addr)
2767 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2770 static long vma_commit_reservation(struct hstate *h,
2771 struct vm_area_struct *vma, unsigned long addr)
2773 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2776 static void vma_end_reservation(struct hstate *h,
2777 struct vm_area_struct *vma, unsigned long addr)
2779 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2782 static long vma_add_reservation(struct hstate *h,
2783 struct vm_area_struct *vma, unsigned long addr)
2785 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2788 static long vma_del_reservation(struct hstate *h,
2789 struct vm_area_struct *vma, unsigned long addr)
2791 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2795 * This routine is called to restore reservation information on error paths.
2796 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2797 * the hugetlb mutex should remain held when calling this routine.
2799 * It handles two specific cases:
2800 * 1) A reservation was in place and the page consumed the reservation.
2801 * HPageRestoreReserve is set in the page.
2802 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2803 * not set. However, alloc_huge_page always updates the reserve map.
2805 * In case 1, free_huge_page later in the error path will increment the
2806 * global reserve count. But, free_huge_page does not have enough context
2807 * to adjust the reservation map. This case deals primarily with private
2808 * mappings. Adjust the reserve map here to be consistent with global
2809 * reserve count adjustments to be made by free_huge_page. Make sure the
2810 * reserve map indicates there is a reservation present.
2812 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2814 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2815 unsigned long address, struct page *page)
2817 long rc = vma_needs_reservation(h, vma, address);
2819 if (HPageRestoreReserve(page)) {
2820 if (unlikely(rc < 0))
2822 * Rare out of memory condition in reserve map
2823 * manipulation. Clear HPageRestoreReserve so that
2824 * global reserve count will not be incremented
2825 * by free_huge_page. This will make it appear
2826 * as though the reservation for this page was
2827 * consumed. This may prevent the task from
2828 * faulting in the page at a later time. This
2829 * is better than inconsistent global huge page
2830 * accounting of reserve counts.
2832 ClearHPageRestoreReserve(page);
2834 (void)vma_add_reservation(h, vma, address);
2836 vma_end_reservation(h, vma, address);
2840 * This indicates there is an entry in the reserve map
2841 * not added by alloc_huge_page. We know it was added
2842 * before the alloc_huge_page call, otherwise
2843 * HPageRestoreReserve would be set on the page.
2844 * Remove the entry so that a subsequent allocation
2845 * does not consume a reservation.
2847 rc = vma_del_reservation(h, vma, address);
2850 * VERY rare out of memory condition. Since
2851 * we can not delete the entry, set
2852 * HPageRestoreReserve so that the reserve
2853 * count will be incremented when the page
2854 * is freed. This reserve will be consumed
2855 * on a subsequent allocation.
2857 SetHPageRestoreReserve(page);
2858 } else if (rc < 0) {
2860 * Rare out of memory condition from
2861 * vma_needs_reservation call. Memory allocation is
2862 * only attempted if a new entry is needed. Therefore,
2863 * this implies there is not an entry in the
2866 * For shared mappings, no entry in the map indicates
2867 * no reservation. We are done.
2869 if (!(vma->vm_flags & VM_MAYSHARE))
2871 * For private mappings, no entry indicates
2872 * a reservation is present. Since we can
2873 * not add an entry, set SetHPageRestoreReserve
2874 * on the page so reserve count will be
2875 * incremented when freed. This reserve will
2876 * be consumed on a subsequent allocation.
2878 SetHPageRestoreReserve(page);
2881 * No reservation present, do nothing
2883 vma_end_reservation(h, vma, address);
2888 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2889 * @h: struct hstate old page belongs to
2890 * @old_page: Old page to dissolve
2891 * @list: List to isolate the page in case we need to
2892 * Returns 0 on success, otherwise negated error.
2894 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2895 struct list_head *list)
2897 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2898 int nid = page_to_nid(old_page);
2899 struct page *new_page;
2903 * Before dissolving the page, we need to allocate a new one for the
2904 * pool to remain stable. Here, we allocate the page and 'prep' it
2905 * by doing everything but actually updating counters and adding to
2906 * the pool. This simplifies and let us do most of the processing
2909 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2912 __prep_new_huge_page(h, new_page);
2915 spin_lock_irq(&hugetlb_lock);
2916 if (!PageHuge(old_page)) {
2918 * Freed from under us. Drop new_page too.
2921 } else if (page_count(old_page)) {
2923 * Someone has grabbed the page, try to isolate it here.
2924 * Fail with -EBUSY if not possible.
2926 spin_unlock_irq(&hugetlb_lock);
2927 ret = isolate_hugetlb(old_page, list);
2928 spin_lock_irq(&hugetlb_lock);
2930 } else if (!HPageFreed(old_page)) {
2932 * Page's refcount is 0 but it has not been enqueued in the
2933 * freelist yet. Race window is small, so we can succeed here if
2936 spin_unlock_irq(&hugetlb_lock);
2941 * Ok, old_page is still a genuine free hugepage. Remove it from
2942 * the freelist and decrease the counters. These will be
2943 * incremented again when calling __prep_account_new_huge_page()
2944 * and enqueue_huge_page() for new_page. The counters will remain
2945 * stable since this happens under the lock.
2947 remove_hugetlb_page(h, old_page, false);
2950 * Ref count on new page is already zero as it was dropped
2951 * earlier. It can be directly added to the pool free list.
2953 __prep_account_new_huge_page(h, nid);
2954 enqueue_huge_page(h, new_page);
2957 * Pages have been replaced, we can safely free the old one.
2959 spin_unlock_irq(&hugetlb_lock);
2960 update_and_free_page(h, old_page, false);
2966 spin_unlock_irq(&hugetlb_lock);
2967 /* Page has a zero ref count, but needs a ref to be freed */
2968 set_page_refcounted(new_page);
2969 update_and_free_page(h, new_page, false);
2974 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2981 * The page might have been dissolved from under our feet, so make sure
2982 * to carefully check the state under the lock.
2983 * Return success when racing as if we dissolved the page ourselves.
2985 spin_lock_irq(&hugetlb_lock);
2986 if (PageHuge(page)) {
2987 head = compound_head(page);
2988 h = page_hstate(head);
2990 spin_unlock_irq(&hugetlb_lock);
2993 spin_unlock_irq(&hugetlb_lock);
2996 * Fence off gigantic pages as there is a cyclic dependency between
2997 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2998 * of bailing out right away without further retrying.
3000 if (hstate_is_gigantic(h))
3003 if (page_count(head) && !isolate_hugetlb(head, list))
3005 else if (!page_count(head))
3006 ret = alloc_and_dissolve_huge_page(h, head, list);
3011 struct page *alloc_huge_page(struct vm_area_struct *vma,
3012 unsigned long addr, int avoid_reserve)
3014 struct hugepage_subpool *spool = subpool_vma(vma);
3015 struct hstate *h = hstate_vma(vma);
3017 long map_chg, map_commit;
3020 struct hugetlb_cgroup *h_cg;
3021 bool deferred_reserve;
3023 idx = hstate_index(h);
3025 * Examine the region/reserve map to determine if the process
3026 * has a reservation for the page to be allocated. A return
3027 * code of zero indicates a reservation exists (no change).
3029 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3031 return ERR_PTR(-ENOMEM);
3034 * Processes that did not create the mapping will have no
3035 * reserves as indicated by the region/reserve map. Check
3036 * that the allocation will not exceed the subpool limit.
3037 * Allocations for MAP_NORESERVE mappings also need to be
3038 * checked against any subpool limit.
3040 if (map_chg || avoid_reserve) {
3041 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3043 vma_end_reservation(h, vma, addr);
3044 return ERR_PTR(-ENOSPC);
3048 * Even though there was no reservation in the region/reserve
3049 * map, there could be reservations associated with the
3050 * subpool that can be used. This would be indicated if the
3051 * return value of hugepage_subpool_get_pages() is zero.
3052 * However, if avoid_reserve is specified we still avoid even
3053 * the subpool reservations.
3059 /* If this allocation is not consuming a reservation, charge it now.
3061 deferred_reserve = map_chg || avoid_reserve;
3062 if (deferred_reserve) {
3063 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3064 idx, pages_per_huge_page(h), &h_cg);
3066 goto out_subpool_put;
3069 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3071 goto out_uncharge_cgroup_reservation;
3073 spin_lock_irq(&hugetlb_lock);
3075 * glb_chg is passed to indicate whether or not a page must be taken
3076 * from the global free pool (global change). gbl_chg == 0 indicates
3077 * a reservation exists for the allocation.
3079 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
3081 spin_unlock_irq(&hugetlb_lock);
3082 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
3084 goto out_uncharge_cgroup;
3085 spin_lock_irq(&hugetlb_lock);
3086 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3087 SetHPageRestoreReserve(page);
3088 h->resv_huge_pages--;
3090 list_add(&page->lru, &h->hugepage_activelist);
3091 set_page_refcounted(page);
3094 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
3095 /* If allocation is not consuming a reservation, also store the
3096 * hugetlb_cgroup pointer on the page.
3098 if (deferred_reserve) {
3099 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3103 spin_unlock_irq(&hugetlb_lock);
3105 hugetlb_set_page_subpool(page, spool);
3107 map_commit = vma_commit_reservation(h, vma, addr);
3108 if (unlikely(map_chg > map_commit)) {
3110 * The page was added to the reservation map between
3111 * vma_needs_reservation and vma_commit_reservation.
3112 * This indicates a race with hugetlb_reserve_pages.
3113 * Adjust for the subpool count incremented above AND
3114 * in hugetlb_reserve_pages for the same page. Also,
3115 * the reservation count added in hugetlb_reserve_pages
3116 * no longer applies.
3120 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3121 hugetlb_acct_memory(h, -rsv_adjust);
3122 if (deferred_reserve)
3123 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
3124 pages_per_huge_page(h), page);
3128 out_uncharge_cgroup:
3129 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3130 out_uncharge_cgroup_reservation:
3131 if (deferred_reserve)
3132 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3135 if (map_chg || avoid_reserve)
3136 hugepage_subpool_put_pages(spool, 1);
3137 vma_end_reservation(h, vma, addr);
3138 return ERR_PTR(-ENOSPC);
3141 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3142 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3143 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3145 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3148 /* do node specific alloc */
3149 if (nid != NUMA_NO_NODE) {
3150 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3151 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3156 /* allocate from next node when distributing huge pages */
3157 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3158 m = memblock_alloc_try_nid_raw(
3159 huge_page_size(h), huge_page_size(h),
3160 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3162 * Use the beginning of the huge page to store the
3163 * huge_bootmem_page struct (until gather_bootmem
3164 * puts them into the mem_map).
3172 /* Put them into a private list first because mem_map is not up yet */
3173 INIT_LIST_HEAD(&m->list);
3174 list_add(&m->list, &huge_boot_pages);
3180 * Put bootmem huge pages into the standard lists after mem_map is up.
3181 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3183 static void __init gather_bootmem_prealloc(void)
3185 struct huge_bootmem_page *m;
3187 list_for_each_entry(m, &huge_boot_pages, list) {
3188 struct page *page = virt_to_page(m);
3189 struct hstate *h = m->hstate;
3191 VM_BUG_ON(!hstate_is_gigantic(h));
3192 WARN_ON(page_count(page) != 1);
3193 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3194 WARN_ON(PageReserved(page));
3195 prep_new_huge_page(h, page, page_to_nid(page));
3196 free_huge_page(page); /* add to the hugepage allocator */
3198 /* VERY unlikely inflated ref count on a tail page */
3199 free_gigantic_page(page, huge_page_order(h));
3203 * We need to restore the 'stolen' pages to totalram_pages
3204 * in order to fix confusing memory reports from free(1) and
3205 * other side-effects, like CommitLimit going negative.
3207 adjust_managed_page_count(page, pages_per_huge_page(h));
3211 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3216 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3217 if (hstate_is_gigantic(h)) {
3218 if (!alloc_bootmem_huge_page(h, nid))
3222 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3224 page = alloc_fresh_huge_page(h, gfp_mask, nid,
3225 &node_states[N_MEMORY], NULL);
3228 free_huge_page(page); /* free it into the hugepage allocator */
3232 if (i == h->max_huge_pages_node[nid])
3235 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3236 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3237 h->max_huge_pages_node[nid], buf, nid, i);
3238 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3239 h->max_huge_pages_node[nid] = i;
3242 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3245 nodemask_t *node_alloc_noretry;
3246 bool node_specific_alloc = false;
3248 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3249 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3250 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3254 /* do node specific alloc */
3255 for_each_online_node(i) {
3256 if (h->max_huge_pages_node[i] > 0) {
3257 hugetlb_hstate_alloc_pages_onenode(h, i);
3258 node_specific_alloc = true;
3262 if (node_specific_alloc)
3265 /* below will do all node balanced alloc */
3266 if (!hstate_is_gigantic(h)) {
3268 * Bit mask controlling how hard we retry per-node allocations.
3269 * Ignore errors as lower level routines can deal with
3270 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3271 * time, we are likely in bigger trouble.
3273 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3276 /* allocations done at boot time */
3277 node_alloc_noretry = NULL;
3280 /* bit mask controlling how hard we retry per-node allocations */
3281 if (node_alloc_noretry)
3282 nodes_clear(*node_alloc_noretry);
3284 for (i = 0; i < h->max_huge_pages; ++i) {
3285 if (hstate_is_gigantic(h)) {
3286 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3288 } else if (!alloc_pool_huge_page(h,
3289 &node_states[N_MEMORY],
3290 node_alloc_noretry))
3294 if (i < h->max_huge_pages) {
3297 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3298 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3299 h->max_huge_pages, buf, i);
3300 h->max_huge_pages = i;
3302 kfree(node_alloc_noretry);
3305 static void __init hugetlb_init_hstates(void)
3307 struct hstate *h, *h2;
3309 for_each_hstate(h) {
3310 /* oversize hugepages were init'ed in early boot */
3311 if (!hstate_is_gigantic(h))
3312 hugetlb_hstate_alloc_pages(h);
3315 * Set demote order for each hstate. Note that
3316 * h->demote_order is initially 0.
3317 * - We can not demote gigantic pages if runtime freeing
3318 * is not supported, so skip this.
3319 * - If CMA allocation is possible, we can not demote
3320 * HUGETLB_PAGE_ORDER or smaller size pages.
3322 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3324 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3326 for_each_hstate(h2) {
3329 if (h2->order < h->order &&
3330 h2->order > h->demote_order)
3331 h->demote_order = h2->order;
3336 static void __init report_hugepages(void)
3340 for_each_hstate(h) {
3343 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3344 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3345 buf, h->free_huge_pages);
3346 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3347 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3351 #ifdef CONFIG_HIGHMEM
3352 static void try_to_free_low(struct hstate *h, unsigned long count,
3353 nodemask_t *nodes_allowed)
3356 LIST_HEAD(page_list);
3358 lockdep_assert_held(&hugetlb_lock);
3359 if (hstate_is_gigantic(h))
3363 * Collect pages to be freed on a list, and free after dropping lock
3365 for_each_node_mask(i, *nodes_allowed) {
3366 struct page *page, *next;
3367 struct list_head *freel = &h->hugepage_freelists[i];
3368 list_for_each_entry_safe(page, next, freel, lru) {
3369 if (count >= h->nr_huge_pages)
3371 if (PageHighMem(page))
3373 remove_hugetlb_page(h, page, false);
3374 list_add(&page->lru, &page_list);
3379 spin_unlock_irq(&hugetlb_lock);
3380 update_and_free_pages_bulk(h, &page_list);
3381 spin_lock_irq(&hugetlb_lock);
3384 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3385 nodemask_t *nodes_allowed)
3391 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3392 * balanced by operating on them in a round-robin fashion.
3393 * Returns 1 if an adjustment was made.
3395 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3400 lockdep_assert_held(&hugetlb_lock);
3401 VM_BUG_ON(delta != -1 && delta != 1);
3404 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3405 if (h->surplus_huge_pages_node[node])
3409 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3410 if (h->surplus_huge_pages_node[node] <
3411 h->nr_huge_pages_node[node])
3418 h->surplus_huge_pages += delta;
3419 h->surplus_huge_pages_node[node] += delta;
3423 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3424 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3425 nodemask_t *nodes_allowed)
3427 unsigned long min_count, ret;
3429 LIST_HEAD(page_list);
3430 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3433 * Bit mask controlling how hard we retry per-node allocations.
3434 * If we can not allocate the bit mask, do not attempt to allocate
3435 * the requested huge pages.
3437 if (node_alloc_noretry)
3438 nodes_clear(*node_alloc_noretry);
3443 * resize_lock mutex prevents concurrent adjustments to number of
3444 * pages in hstate via the proc/sysfs interfaces.
3446 mutex_lock(&h->resize_lock);
3447 flush_free_hpage_work(h);
3448 spin_lock_irq(&hugetlb_lock);
3451 * Check for a node specific request.
3452 * Changing node specific huge page count may require a corresponding
3453 * change to the global count. In any case, the passed node mask
3454 * (nodes_allowed) will restrict alloc/free to the specified node.
3456 if (nid != NUMA_NO_NODE) {
3457 unsigned long old_count = count;
3459 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3461 * User may have specified a large count value which caused the
3462 * above calculation to overflow. In this case, they wanted
3463 * to allocate as many huge pages as possible. Set count to
3464 * largest possible value to align with their intention.
3466 if (count < old_count)
3471 * Gigantic pages runtime allocation depend on the capability for large
3472 * page range allocation.
3473 * If the system does not provide this feature, return an error when
3474 * the user tries to allocate gigantic pages but let the user free the
3475 * boottime allocated gigantic pages.
3477 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3478 if (count > persistent_huge_pages(h)) {
3479 spin_unlock_irq(&hugetlb_lock);
3480 mutex_unlock(&h->resize_lock);
3481 NODEMASK_FREE(node_alloc_noretry);
3484 /* Fall through to decrease pool */
3488 * Increase the pool size
3489 * First take pages out of surplus state. Then make up the
3490 * remaining difference by allocating fresh huge pages.
3492 * We might race with alloc_surplus_huge_page() here and be unable
3493 * to convert a surplus huge page to a normal huge page. That is
3494 * not critical, though, it just means the overall size of the
3495 * pool might be one hugepage larger than it needs to be, but
3496 * within all the constraints specified by the sysctls.
3498 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3499 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3503 while (count > persistent_huge_pages(h)) {
3505 * If this allocation races such that we no longer need the
3506 * page, free_huge_page will handle it by freeing the page
3507 * and reducing the surplus.
3509 spin_unlock_irq(&hugetlb_lock);
3511 /* yield cpu to avoid soft lockup */
3514 ret = alloc_pool_huge_page(h, nodes_allowed,
3515 node_alloc_noretry);
3516 spin_lock_irq(&hugetlb_lock);
3520 /* Bail for signals. Probably ctrl-c from user */
3521 if (signal_pending(current))
3526 * Decrease the pool size
3527 * First return free pages to the buddy allocator (being careful
3528 * to keep enough around to satisfy reservations). Then place
3529 * pages into surplus state as needed so the pool will shrink
3530 * to the desired size as pages become free.
3532 * By placing pages into the surplus state independent of the
3533 * overcommit value, we are allowing the surplus pool size to
3534 * exceed overcommit. There are few sane options here. Since
3535 * alloc_surplus_huge_page() is checking the global counter,
3536 * though, we'll note that we're not allowed to exceed surplus
3537 * and won't grow the pool anywhere else. Not until one of the
3538 * sysctls are changed, or the surplus pages go out of use.
3540 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3541 min_count = max(count, min_count);
3542 try_to_free_low(h, min_count, nodes_allowed);
3545 * Collect pages to be removed on list without dropping lock
3547 while (min_count < persistent_huge_pages(h)) {
3548 page = remove_pool_huge_page(h, nodes_allowed, 0);
3552 list_add(&page->lru, &page_list);
3554 /* free the pages after dropping lock */
3555 spin_unlock_irq(&hugetlb_lock);
3556 update_and_free_pages_bulk(h, &page_list);
3557 flush_free_hpage_work(h);
3558 spin_lock_irq(&hugetlb_lock);
3560 while (count < persistent_huge_pages(h)) {
3561 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3565 h->max_huge_pages = persistent_huge_pages(h);
3566 spin_unlock_irq(&hugetlb_lock);
3567 mutex_unlock(&h->resize_lock);
3569 NODEMASK_FREE(node_alloc_noretry);
3574 static int demote_free_huge_page(struct hstate *h, struct page *page)
3576 int i, nid = page_to_nid(page);
3577 struct hstate *target_hstate;
3578 struct page *subpage;
3581 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3583 remove_hugetlb_page_for_demote(h, page, false);
3584 spin_unlock_irq(&hugetlb_lock);
3586 rc = hugetlb_vmemmap_restore(h, page);
3588 /* Allocation of vmemmmap failed, we can not demote page */
3589 spin_lock_irq(&hugetlb_lock);
3590 set_page_refcounted(page);
3591 add_hugetlb_page(h, page, false);
3596 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3597 * sizes as it will not ref count pages.
3599 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3602 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3603 * Without the mutex, pages added to target hstate could be marked
3606 * Note that we already hold h->resize_lock. To prevent deadlock,
3607 * use the convention of always taking larger size hstate mutex first.
3609 mutex_lock(&target_hstate->resize_lock);
3610 for (i = 0; i < pages_per_huge_page(h);
3611 i += pages_per_huge_page(target_hstate)) {
3612 subpage = nth_page(page, i);
3613 if (hstate_is_gigantic(target_hstate))
3614 prep_compound_gigantic_page_for_demote(subpage,
3615 target_hstate->order);
3617 prep_compound_page(subpage, target_hstate->order);
3618 set_page_private(subpage, 0);
3619 prep_new_huge_page(target_hstate, subpage, nid);
3620 free_huge_page(subpage);
3622 mutex_unlock(&target_hstate->resize_lock);
3624 spin_lock_irq(&hugetlb_lock);
3627 * Not absolutely necessary, but for consistency update max_huge_pages
3628 * based on pool changes for the demoted page.
3630 h->max_huge_pages--;
3631 target_hstate->max_huge_pages +=
3632 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3637 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3638 __must_hold(&hugetlb_lock)
3643 lockdep_assert_held(&hugetlb_lock);
3645 /* We should never get here if no demote order */
3646 if (!h->demote_order) {
3647 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3648 return -EINVAL; /* internal error */
3651 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3652 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3653 if (PageHWPoison(page))
3656 return demote_free_huge_page(h, page);
3661 * Only way to get here is if all pages on free lists are poisoned.
3662 * Return -EBUSY so that caller will not retry.
3667 #define HSTATE_ATTR_RO(_name) \
3668 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3670 #define HSTATE_ATTR_WO(_name) \
3671 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3673 #define HSTATE_ATTR(_name) \
3674 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3676 static struct kobject *hugepages_kobj;
3677 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3679 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3681 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3685 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3686 if (hstate_kobjs[i] == kobj) {
3688 *nidp = NUMA_NO_NODE;
3692 return kobj_to_node_hstate(kobj, nidp);
3695 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3696 struct kobj_attribute *attr, char *buf)
3699 unsigned long nr_huge_pages;
3702 h = kobj_to_hstate(kobj, &nid);
3703 if (nid == NUMA_NO_NODE)
3704 nr_huge_pages = h->nr_huge_pages;
3706 nr_huge_pages = h->nr_huge_pages_node[nid];
3708 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3711 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3712 struct hstate *h, int nid,
3713 unsigned long count, size_t len)
3716 nodemask_t nodes_allowed, *n_mask;
3718 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3721 if (nid == NUMA_NO_NODE) {
3723 * global hstate attribute
3725 if (!(obey_mempolicy &&
3726 init_nodemask_of_mempolicy(&nodes_allowed)))
3727 n_mask = &node_states[N_MEMORY];
3729 n_mask = &nodes_allowed;
3732 * Node specific request. count adjustment happens in
3733 * set_max_huge_pages() after acquiring hugetlb_lock.
3735 init_nodemask_of_node(&nodes_allowed, nid);
3736 n_mask = &nodes_allowed;
3739 err = set_max_huge_pages(h, count, nid, n_mask);
3741 return err ? err : len;
3744 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3745 struct kobject *kobj, const char *buf,
3749 unsigned long count;
3753 err = kstrtoul(buf, 10, &count);
3757 h = kobj_to_hstate(kobj, &nid);
3758 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3761 static ssize_t nr_hugepages_show(struct kobject *kobj,
3762 struct kobj_attribute *attr, char *buf)
3764 return nr_hugepages_show_common(kobj, attr, buf);
3767 static ssize_t nr_hugepages_store(struct kobject *kobj,
3768 struct kobj_attribute *attr, const char *buf, size_t len)
3770 return nr_hugepages_store_common(false, kobj, buf, len);
3772 HSTATE_ATTR(nr_hugepages);
3777 * hstate attribute for optionally mempolicy-based constraint on persistent
3778 * huge page alloc/free.
3780 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3781 struct kobj_attribute *attr,
3784 return nr_hugepages_show_common(kobj, attr, buf);
3787 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3788 struct kobj_attribute *attr, const char *buf, size_t len)
3790 return nr_hugepages_store_common(true, kobj, buf, len);
3792 HSTATE_ATTR(nr_hugepages_mempolicy);
3796 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3797 struct kobj_attribute *attr, char *buf)
3799 struct hstate *h = kobj_to_hstate(kobj, NULL);
3800 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3803 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3804 struct kobj_attribute *attr, const char *buf, size_t count)
3807 unsigned long input;
3808 struct hstate *h = kobj_to_hstate(kobj, NULL);
3810 if (hstate_is_gigantic(h))
3813 err = kstrtoul(buf, 10, &input);
3817 spin_lock_irq(&hugetlb_lock);
3818 h->nr_overcommit_huge_pages = input;
3819 spin_unlock_irq(&hugetlb_lock);
3823 HSTATE_ATTR(nr_overcommit_hugepages);
3825 static ssize_t free_hugepages_show(struct kobject *kobj,
3826 struct kobj_attribute *attr, char *buf)
3829 unsigned long free_huge_pages;
3832 h = kobj_to_hstate(kobj, &nid);
3833 if (nid == NUMA_NO_NODE)
3834 free_huge_pages = h->free_huge_pages;
3836 free_huge_pages = h->free_huge_pages_node[nid];
3838 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3840 HSTATE_ATTR_RO(free_hugepages);
3842 static ssize_t resv_hugepages_show(struct kobject *kobj,
3843 struct kobj_attribute *attr, char *buf)
3845 struct hstate *h = kobj_to_hstate(kobj, NULL);
3846 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3848 HSTATE_ATTR_RO(resv_hugepages);
3850 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3851 struct kobj_attribute *attr, char *buf)
3854 unsigned long surplus_huge_pages;
3857 h = kobj_to_hstate(kobj, &nid);
3858 if (nid == NUMA_NO_NODE)
3859 surplus_huge_pages = h->surplus_huge_pages;
3861 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3863 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3865 HSTATE_ATTR_RO(surplus_hugepages);
3867 static ssize_t demote_store(struct kobject *kobj,
3868 struct kobj_attribute *attr, const char *buf, size_t len)
3870 unsigned long nr_demote;
3871 unsigned long nr_available;
3872 nodemask_t nodes_allowed, *n_mask;
3877 err = kstrtoul(buf, 10, &nr_demote);
3880 h = kobj_to_hstate(kobj, &nid);
3882 if (nid != NUMA_NO_NODE) {
3883 init_nodemask_of_node(&nodes_allowed, nid);
3884 n_mask = &nodes_allowed;
3886 n_mask = &node_states[N_MEMORY];
3889 /* Synchronize with other sysfs operations modifying huge pages */
3890 mutex_lock(&h->resize_lock);
3891 spin_lock_irq(&hugetlb_lock);
3895 * Check for available pages to demote each time thorough the
3896 * loop as demote_pool_huge_page will drop hugetlb_lock.
3898 if (nid != NUMA_NO_NODE)
3899 nr_available = h->free_huge_pages_node[nid];
3901 nr_available = h->free_huge_pages;
3902 nr_available -= h->resv_huge_pages;
3906 err = demote_pool_huge_page(h, n_mask);
3913 spin_unlock_irq(&hugetlb_lock);
3914 mutex_unlock(&h->resize_lock);
3920 HSTATE_ATTR_WO(demote);
3922 static ssize_t demote_size_show(struct kobject *kobj,
3923 struct kobj_attribute *attr, char *buf)
3925 struct hstate *h = kobj_to_hstate(kobj, NULL);
3926 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3928 return sysfs_emit(buf, "%lukB\n", demote_size);
3931 static ssize_t demote_size_store(struct kobject *kobj,
3932 struct kobj_attribute *attr,
3933 const char *buf, size_t count)
3935 struct hstate *h, *demote_hstate;
3936 unsigned long demote_size;
3937 unsigned int demote_order;
3939 demote_size = (unsigned long)memparse(buf, NULL);
3941 demote_hstate = size_to_hstate(demote_size);
3944 demote_order = demote_hstate->order;
3945 if (demote_order < HUGETLB_PAGE_ORDER)
3948 /* demote order must be smaller than hstate order */
3949 h = kobj_to_hstate(kobj, NULL);
3950 if (demote_order >= h->order)
3953 /* resize_lock synchronizes access to demote size and writes */
3954 mutex_lock(&h->resize_lock);
3955 h->demote_order = demote_order;
3956 mutex_unlock(&h->resize_lock);
3960 HSTATE_ATTR(demote_size);
3962 static struct attribute *hstate_attrs[] = {
3963 &nr_hugepages_attr.attr,
3964 &nr_overcommit_hugepages_attr.attr,
3965 &free_hugepages_attr.attr,
3966 &resv_hugepages_attr.attr,
3967 &surplus_hugepages_attr.attr,
3969 &nr_hugepages_mempolicy_attr.attr,
3974 static const struct attribute_group hstate_attr_group = {
3975 .attrs = hstate_attrs,
3978 static struct attribute *hstate_demote_attrs[] = {
3979 &demote_size_attr.attr,
3984 static const struct attribute_group hstate_demote_attr_group = {
3985 .attrs = hstate_demote_attrs,
3988 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3989 struct kobject **hstate_kobjs,
3990 const struct attribute_group *hstate_attr_group)
3993 int hi = hstate_index(h);
3995 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3996 if (!hstate_kobjs[hi])
3999 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4001 kobject_put(hstate_kobjs[hi]);
4002 hstate_kobjs[hi] = NULL;
4006 if (h->demote_order) {
4007 retval = sysfs_create_group(hstate_kobjs[hi],
4008 &hstate_demote_attr_group);
4010 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4011 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4012 kobject_put(hstate_kobjs[hi]);
4013 hstate_kobjs[hi] = NULL;
4022 static bool hugetlb_sysfs_initialized __ro_after_init;
4025 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4026 * with node devices in node_devices[] using a parallel array. The array
4027 * index of a node device or _hstate == node id.
4028 * This is here to avoid any static dependency of the node device driver, in
4029 * the base kernel, on the hugetlb module.
4031 struct node_hstate {
4032 struct kobject *hugepages_kobj;
4033 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4035 static struct node_hstate node_hstates[MAX_NUMNODES];
4038 * A subset of global hstate attributes for node devices
4040 static struct attribute *per_node_hstate_attrs[] = {
4041 &nr_hugepages_attr.attr,
4042 &free_hugepages_attr.attr,
4043 &surplus_hugepages_attr.attr,
4047 static const struct attribute_group per_node_hstate_attr_group = {
4048 .attrs = per_node_hstate_attrs,
4052 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4053 * Returns node id via non-NULL nidp.
4055 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4059 for (nid = 0; nid < nr_node_ids; nid++) {
4060 struct node_hstate *nhs = &node_hstates[nid];
4062 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4063 if (nhs->hstate_kobjs[i] == kobj) {
4075 * Unregister hstate attributes from a single node device.
4076 * No-op if no hstate attributes attached.
4078 void hugetlb_unregister_node(struct node *node)
4081 struct node_hstate *nhs = &node_hstates[node->dev.id];
4083 if (!nhs->hugepages_kobj)
4084 return; /* no hstate attributes */
4086 for_each_hstate(h) {
4087 int idx = hstate_index(h);
4088 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4092 if (h->demote_order)
4093 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4094 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4095 kobject_put(hstate_kobj);
4096 nhs->hstate_kobjs[idx] = NULL;
4099 kobject_put(nhs->hugepages_kobj);
4100 nhs->hugepages_kobj = NULL;
4105 * Register hstate attributes for a single node device.
4106 * No-op if attributes already registered.
4108 void hugetlb_register_node(struct node *node)
4111 struct node_hstate *nhs = &node_hstates[node->dev.id];
4114 if (!hugetlb_sysfs_initialized)
4117 if (nhs->hugepages_kobj)
4118 return; /* already allocated */
4120 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4122 if (!nhs->hugepages_kobj)
4125 for_each_hstate(h) {
4126 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4128 &per_node_hstate_attr_group);
4130 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4131 h->name, node->dev.id);
4132 hugetlb_unregister_node(node);
4139 * hugetlb init time: register hstate attributes for all registered node
4140 * devices of nodes that have memory. All on-line nodes should have
4141 * registered their associated device by this time.
4143 static void __init hugetlb_register_all_nodes(void)
4147 for_each_online_node(nid)
4148 hugetlb_register_node(node_devices[nid]);
4150 #else /* !CONFIG_NUMA */
4152 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4160 static void hugetlb_register_all_nodes(void) { }
4165 static void __init hugetlb_cma_check(void);
4167 static inline __init void hugetlb_cma_check(void)
4172 static void __init hugetlb_sysfs_init(void)
4177 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4178 if (!hugepages_kobj)
4181 for_each_hstate(h) {
4182 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4183 hstate_kobjs, &hstate_attr_group);
4185 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4189 hugetlb_sysfs_initialized = true;
4191 hugetlb_register_all_nodes();
4194 static int __init hugetlb_init(void)
4198 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4201 if (!hugepages_supported()) {
4202 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4203 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4208 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4209 * architectures depend on setup being done here.
4211 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4212 if (!parsed_default_hugepagesz) {
4214 * If we did not parse a default huge page size, set
4215 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4216 * number of huge pages for this default size was implicitly
4217 * specified, set that here as well.
4218 * Note that the implicit setting will overwrite an explicit
4219 * setting. A warning will be printed in this case.
4221 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4222 if (default_hstate_max_huge_pages) {
4223 if (default_hstate.max_huge_pages) {
4226 string_get_size(huge_page_size(&default_hstate),
4227 1, STRING_UNITS_2, buf, 32);
4228 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4229 default_hstate.max_huge_pages, buf);
4230 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4231 default_hstate_max_huge_pages);
4233 default_hstate.max_huge_pages =
4234 default_hstate_max_huge_pages;
4236 for_each_online_node(i)
4237 default_hstate.max_huge_pages_node[i] =
4238 default_hugepages_in_node[i];
4242 hugetlb_cma_check();
4243 hugetlb_init_hstates();
4244 gather_bootmem_prealloc();
4247 hugetlb_sysfs_init();
4248 hugetlb_cgroup_file_init();
4251 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4253 num_fault_mutexes = 1;
4255 hugetlb_fault_mutex_table =
4256 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4258 BUG_ON(!hugetlb_fault_mutex_table);
4260 for (i = 0; i < num_fault_mutexes; i++)
4261 mutex_init(&hugetlb_fault_mutex_table[i]);
4264 subsys_initcall(hugetlb_init);
4266 /* Overwritten by architectures with more huge page sizes */
4267 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4269 return size == HPAGE_SIZE;
4272 void __init hugetlb_add_hstate(unsigned int order)
4277 if (size_to_hstate(PAGE_SIZE << order)) {
4280 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4282 h = &hstates[hugetlb_max_hstate++];
4283 mutex_init(&h->resize_lock);
4285 h->mask = ~(huge_page_size(h) - 1);
4286 for (i = 0; i < MAX_NUMNODES; ++i)
4287 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4288 INIT_LIST_HEAD(&h->hugepage_activelist);
4289 h->next_nid_to_alloc = first_memory_node;
4290 h->next_nid_to_free = first_memory_node;
4291 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4292 huge_page_size(h)/SZ_1K);
4297 bool __init __weak hugetlb_node_alloc_supported(void)
4302 static void __init hugepages_clear_pages_in_node(void)
4304 if (!hugetlb_max_hstate) {
4305 default_hstate_max_huge_pages = 0;
4306 memset(default_hugepages_in_node, 0,
4307 sizeof(default_hugepages_in_node));
4309 parsed_hstate->max_huge_pages = 0;
4310 memset(parsed_hstate->max_huge_pages_node, 0,
4311 sizeof(parsed_hstate->max_huge_pages_node));
4316 * hugepages command line processing
4317 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4318 * specification. If not, ignore the hugepages value. hugepages can also
4319 * be the first huge page command line option in which case it implicitly
4320 * specifies the number of huge pages for the default size.
4322 static int __init hugepages_setup(char *s)
4325 static unsigned long *last_mhp;
4326 int node = NUMA_NO_NODE;
4331 if (!parsed_valid_hugepagesz) {
4332 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4333 parsed_valid_hugepagesz = true;
4338 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4339 * yet, so this hugepages= parameter goes to the "default hstate".
4340 * Otherwise, it goes with the previously parsed hugepagesz or
4341 * default_hugepagesz.
4343 else if (!hugetlb_max_hstate)
4344 mhp = &default_hstate_max_huge_pages;
4346 mhp = &parsed_hstate->max_huge_pages;
4348 if (mhp == last_mhp) {
4349 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4355 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4357 /* Parameter is node format */
4358 if (p[count] == ':') {
4359 if (!hugetlb_node_alloc_supported()) {
4360 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4363 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4365 node = array_index_nospec(tmp, MAX_NUMNODES);
4367 /* Parse hugepages */
4368 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4370 if (!hugetlb_max_hstate)
4371 default_hugepages_in_node[node] = tmp;
4373 parsed_hstate->max_huge_pages_node[node] = tmp;
4375 /* Go to parse next node*/
4376 if (p[count] == ',')
4389 * Global state is always initialized later in hugetlb_init.
4390 * But we need to allocate gigantic hstates here early to still
4391 * use the bootmem allocator.
4393 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4394 hugetlb_hstate_alloc_pages(parsed_hstate);
4401 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4402 hugepages_clear_pages_in_node();
4405 __setup("hugepages=", hugepages_setup);
4408 * hugepagesz command line processing
4409 * A specific huge page size can only be specified once with hugepagesz.
4410 * hugepagesz is followed by hugepages on the command line. The global
4411 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4412 * hugepagesz argument was valid.
4414 static int __init hugepagesz_setup(char *s)
4419 parsed_valid_hugepagesz = false;
4420 size = (unsigned long)memparse(s, NULL);
4422 if (!arch_hugetlb_valid_size(size)) {
4423 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4427 h = size_to_hstate(size);
4430 * hstate for this size already exists. This is normally
4431 * an error, but is allowed if the existing hstate is the
4432 * default hstate. More specifically, it is only allowed if
4433 * the number of huge pages for the default hstate was not
4434 * previously specified.
4436 if (!parsed_default_hugepagesz || h != &default_hstate ||
4437 default_hstate.max_huge_pages) {
4438 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4443 * No need to call hugetlb_add_hstate() as hstate already
4444 * exists. But, do set parsed_hstate so that a following
4445 * hugepages= parameter will be applied to this hstate.
4448 parsed_valid_hugepagesz = true;
4452 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4453 parsed_valid_hugepagesz = true;
4456 __setup("hugepagesz=", hugepagesz_setup);
4459 * default_hugepagesz command line input
4460 * Only one instance of default_hugepagesz allowed on command line.
4462 static int __init default_hugepagesz_setup(char *s)
4467 parsed_valid_hugepagesz = false;
4468 if (parsed_default_hugepagesz) {
4469 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4473 size = (unsigned long)memparse(s, NULL);
4475 if (!arch_hugetlb_valid_size(size)) {
4476 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4480 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4481 parsed_valid_hugepagesz = true;
4482 parsed_default_hugepagesz = true;
4483 default_hstate_idx = hstate_index(size_to_hstate(size));
4486 * The number of default huge pages (for this size) could have been
4487 * specified as the first hugetlb parameter: hugepages=X. If so,
4488 * then default_hstate_max_huge_pages is set. If the default huge
4489 * page size is gigantic (>= MAX_ORDER), then the pages must be
4490 * allocated here from bootmem allocator.
4492 if (default_hstate_max_huge_pages) {
4493 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4494 for_each_online_node(i)
4495 default_hstate.max_huge_pages_node[i] =
4496 default_hugepages_in_node[i];
4497 if (hstate_is_gigantic(&default_hstate))
4498 hugetlb_hstate_alloc_pages(&default_hstate);
4499 default_hstate_max_huge_pages = 0;
4504 __setup("default_hugepagesz=", default_hugepagesz_setup);
4506 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4509 struct mempolicy *mpol = get_task_policy(current);
4512 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4513 * (from policy_nodemask) specifically for hugetlb case
4515 if (mpol->mode == MPOL_BIND &&
4516 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4517 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4518 return &mpol->nodes;
4523 static unsigned int allowed_mems_nr(struct hstate *h)
4526 unsigned int nr = 0;
4527 nodemask_t *mbind_nodemask;
4528 unsigned int *array = h->free_huge_pages_node;
4529 gfp_t gfp_mask = htlb_alloc_mask(h);
4531 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4532 for_each_node_mask(node, cpuset_current_mems_allowed) {
4533 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4540 #ifdef CONFIG_SYSCTL
4541 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4542 void *buffer, size_t *length,
4543 loff_t *ppos, unsigned long *out)
4545 struct ctl_table dup_table;
4548 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4549 * can duplicate the @table and alter the duplicate of it.
4552 dup_table.data = out;
4554 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4557 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4558 struct ctl_table *table, int write,
4559 void *buffer, size_t *length, loff_t *ppos)
4561 struct hstate *h = &default_hstate;
4562 unsigned long tmp = h->max_huge_pages;
4565 if (!hugepages_supported())
4568 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4574 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4575 NUMA_NO_NODE, tmp, *length);
4580 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4581 void *buffer, size_t *length, loff_t *ppos)
4584 return hugetlb_sysctl_handler_common(false, table, write,
4585 buffer, length, ppos);
4589 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4590 void *buffer, size_t *length, loff_t *ppos)
4592 return hugetlb_sysctl_handler_common(true, table, write,
4593 buffer, length, ppos);
4595 #endif /* CONFIG_NUMA */
4597 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4598 void *buffer, size_t *length, loff_t *ppos)
4600 struct hstate *h = &default_hstate;
4604 if (!hugepages_supported())
4607 tmp = h->nr_overcommit_huge_pages;
4609 if (write && hstate_is_gigantic(h))
4612 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4618 spin_lock_irq(&hugetlb_lock);
4619 h->nr_overcommit_huge_pages = tmp;
4620 spin_unlock_irq(&hugetlb_lock);
4626 #endif /* CONFIG_SYSCTL */
4628 void hugetlb_report_meminfo(struct seq_file *m)
4631 unsigned long total = 0;
4633 if (!hugepages_supported())
4636 for_each_hstate(h) {
4637 unsigned long count = h->nr_huge_pages;
4639 total += huge_page_size(h) * count;
4641 if (h == &default_hstate)
4643 "HugePages_Total: %5lu\n"
4644 "HugePages_Free: %5lu\n"
4645 "HugePages_Rsvd: %5lu\n"
4646 "HugePages_Surp: %5lu\n"
4647 "Hugepagesize: %8lu kB\n",
4651 h->surplus_huge_pages,
4652 huge_page_size(h) / SZ_1K);
4655 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4658 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4660 struct hstate *h = &default_hstate;
4662 if (!hugepages_supported())
4665 return sysfs_emit_at(buf, len,
4666 "Node %d HugePages_Total: %5u\n"
4667 "Node %d HugePages_Free: %5u\n"
4668 "Node %d HugePages_Surp: %5u\n",
4669 nid, h->nr_huge_pages_node[nid],
4670 nid, h->free_huge_pages_node[nid],
4671 nid, h->surplus_huge_pages_node[nid]);
4674 void hugetlb_show_meminfo_node(int nid)
4678 if (!hugepages_supported())
4682 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4684 h->nr_huge_pages_node[nid],
4685 h->free_huge_pages_node[nid],
4686 h->surplus_huge_pages_node[nid],
4687 huge_page_size(h) / SZ_1K);
4690 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4692 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4693 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4696 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4697 unsigned long hugetlb_total_pages(void)
4700 unsigned long nr_total_pages = 0;
4703 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4704 return nr_total_pages;
4707 static int hugetlb_acct_memory(struct hstate *h, long delta)
4714 spin_lock_irq(&hugetlb_lock);
4716 * When cpuset is configured, it breaks the strict hugetlb page
4717 * reservation as the accounting is done on a global variable. Such
4718 * reservation is completely rubbish in the presence of cpuset because
4719 * the reservation is not checked against page availability for the
4720 * current cpuset. Application can still potentially OOM'ed by kernel
4721 * with lack of free htlb page in cpuset that the task is in.
4722 * Attempt to enforce strict accounting with cpuset is almost
4723 * impossible (or too ugly) because cpuset is too fluid that
4724 * task or memory node can be dynamically moved between cpusets.
4726 * The change of semantics for shared hugetlb mapping with cpuset is
4727 * undesirable. However, in order to preserve some of the semantics,
4728 * we fall back to check against current free page availability as
4729 * a best attempt and hopefully to minimize the impact of changing
4730 * semantics that cpuset has.
4732 * Apart from cpuset, we also have memory policy mechanism that
4733 * also determines from which node the kernel will allocate memory
4734 * in a NUMA system. So similar to cpuset, we also should consider
4735 * the memory policy of the current task. Similar to the description
4739 if (gather_surplus_pages(h, delta) < 0)
4742 if (delta > allowed_mems_nr(h)) {
4743 return_unused_surplus_pages(h, delta);
4750 return_unused_surplus_pages(h, (unsigned long) -delta);
4753 spin_unlock_irq(&hugetlb_lock);
4757 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4759 struct resv_map *resv = vma_resv_map(vma);
4762 * HPAGE_RESV_OWNER indicates a private mapping.
4763 * This new VMA should share its siblings reservation map if present.
4764 * The VMA will only ever have a valid reservation map pointer where
4765 * it is being copied for another still existing VMA. As that VMA
4766 * has a reference to the reservation map it cannot disappear until
4767 * after this open call completes. It is therefore safe to take a
4768 * new reference here without additional locking.
4770 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4771 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4772 kref_get(&resv->refs);
4776 * vma_lock structure for sharable mappings is vma specific.
4777 * Clear old pointer (if copied via vm_area_dup) and allocate
4778 * new structure. Before clearing, make sure vma_lock is not
4781 if (vma->vm_flags & VM_MAYSHARE) {
4782 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4785 if (vma_lock->vma != vma) {
4786 vma->vm_private_data = NULL;
4787 hugetlb_vma_lock_alloc(vma);
4789 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4791 hugetlb_vma_lock_alloc(vma);
4795 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4797 struct hstate *h = hstate_vma(vma);
4798 struct resv_map *resv;
4799 struct hugepage_subpool *spool = subpool_vma(vma);
4800 unsigned long reserve, start, end;
4803 hugetlb_vma_lock_free(vma);
4805 resv = vma_resv_map(vma);
4806 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4809 start = vma_hugecache_offset(h, vma, vma->vm_start);
4810 end = vma_hugecache_offset(h, vma, vma->vm_end);
4812 reserve = (end - start) - region_count(resv, start, end);
4813 hugetlb_cgroup_uncharge_counter(resv, start, end);
4816 * Decrement reserve counts. The global reserve count may be
4817 * adjusted if the subpool has a minimum size.
4819 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4820 hugetlb_acct_memory(h, -gbl_reserve);
4823 kref_put(&resv->refs, resv_map_release);
4826 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4828 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4832 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4833 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4834 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4836 if (addr & ~PUD_MASK) {
4838 * hugetlb_vm_op_split is called right before we attempt to
4839 * split the VMA. We will need to unshare PMDs in the old and
4840 * new VMAs, so let's unshare before we split.
4842 unsigned long floor = addr & PUD_MASK;
4843 unsigned long ceil = floor + PUD_SIZE;
4845 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4846 hugetlb_unshare_pmds(vma, floor, ceil);
4852 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4854 return huge_page_size(hstate_vma(vma));
4858 * We cannot handle pagefaults against hugetlb pages at all. They cause
4859 * handle_mm_fault() to try to instantiate regular-sized pages in the
4860 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4863 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4870 * When a new function is introduced to vm_operations_struct and added
4871 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4872 * This is because under System V memory model, mappings created via
4873 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4874 * their original vm_ops are overwritten with shm_vm_ops.
4876 const struct vm_operations_struct hugetlb_vm_ops = {
4877 .fault = hugetlb_vm_op_fault,
4878 .open = hugetlb_vm_op_open,
4879 .close = hugetlb_vm_op_close,
4880 .may_split = hugetlb_vm_op_split,
4881 .pagesize = hugetlb_vm_op_pagesize,
4884 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4888 unsigned int shift = huge_page_shift(hstate_vma(vma));
4891 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4892 vma->vm_page_prot)));
4894 entry = huge_pte_wrprotect(mk_huge_pte(page,
4895 vma->vm_page_prot));
4897 entry = pte_mkyoung(entry);
4898 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4903 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4904 unsigned long address, pte_t *ptep)
4908 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4909 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4910 update_mmu_cache(vma, address, ptep);
4913 bool is_hugetlb_entry_migration(pte_t pte)
4917 if (huge_pte_none(pte) || pte_present(pte))
4919 swp = pte_to_swp_entry(pte);
4920 if (is_migration_entry(swp))
4926 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4930 if (huge_pte_none(pte) || pte_present(pte))
4932 swp = pte_to_swp_entry(pte);
4933 if (is_hwpoison_entry(swp))
4940 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4941 struct page *new_page)
4943 __SetPageUptodate(new_page);
4944 hugepage_add_new_anon_rmap(new_page, vma, addr);
4945 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4946 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4947 ClearHPageRestoreReserve(new_page);
4948 SetHPageMigratable(new_page);
4951 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4952 struct vm_area_struct *dst_vma,
4953 struct vm_area_struct *src_vma)
4955 pte_t *src_pte, *dst_pte, entry;
4956 struct page *ptepage;
4958 bool cow = is_cow_mapping(src_vma->vm_flags);
4959 struct hstate *h = hstate_vma(src_vma);
4960 unsigned long sz = huge_page_size(h);
4961 unsigned long npages = pages_per_huge_page(h);
4962 struct mmu_notifier_range range;
4963 unsigned long last_addr_mask;
4967 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4970 mmu_notifier_invalidate_range_start(&range);
4971 mmap_assert_write_locked(src);
4972 raw_write_seqcount_begin(&src->write_protect_seq);
4975 * For shared mappings the vma lock must be held before
4976 * calling huge_pte_offset in the src vma. Otherwise, the
4977 * returned ptep could go away if part of a shared pmd and
4978 * another thread calls huge_pmd_unshare.
4980 hugetlb_vma_lock_read(src_vma);
4983 last_addr_mask = hugetlb_mask_last_page(h);
4984 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4985 spinlock_t *src_ptl, *dst_ptl;
4986 src_pte = huge_pte_offset(src, addr, sz);
4988 addr |= last_addr_mask;
4991 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
4998 * If the pagetables are shared don't copy or take references.
5000 * dst_pte == src_pte is the common case of src/dest sharing.
5001 * However, src could have 'unshared' and dst shares with
5002 * another vma. So page_count of ptep page is checked instead
5003 * to reliably determine whether pte is shared.
5005 if (page_count(virt_to_page(dst_pte)) > 1) {
5006 addr |= last_addr_mask;
5010 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5011 src_ptl = huge_pte_lockptr(h, src, src_pte);
5012 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5013 entry = huge_ptep_get(src_pte);
5015 if (huge_pte_none(entry)) {
5017 * Skip if src entry none.
5020 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5021 bool uffd_wp = huge_pte_uffd_wp(entry);
5023 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5024 entry = huge_pte_clear_uffd_wp(entry);
5025 set_huge_pte_at(dst, addr, dst_pte, entry);
5026 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5027 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5028 bool uffd_wp = huge_pte_uffd_wp(entry);
5030 if (!is_readable_migration_entry(swp_entry) && cow) {
5032 * COW mappings require pages in both
5033 * parent and child to be set to read.
5035 swp_entry = make_readable_migration_entry(
5036 swp_offset(swp_entry));
5037 entry = swp_entry_to_pte(swp_entry);
5038 if (userfaultfd_wp(src_vma) && uffd_wp)
5039 entry = huge_pte_mkuffd_wp(entry);
5040 set_huge_pte_at(src, addr, src_pte, entry);
5042 if (!userfaultfd_wp(dst_vma) && uffd_wp)
5043 entry = huge_pte_clear_uffd_wp(entry);
5044 set_huge_pte_at(dst, addr, dst_pte, entry);
5045 } else if (unlikely(is_pte_marker(entry))) {
5047 * We copy the pte marker only if the dst vma has
5050 if (userfaultfd_wp(dst_vma))
5051 set_huge_pte_at(dst, addr, dst_pte, entry);
5053 entry = huge_ptep_get(src_pte);
5054 ptepage = pte_page(entry);
5058 * Failing to duplicate the anon rmap is a rare case
5059 * where we see pinned hugetlb pages while they're
5060 * prone to COW. We need to do the COW earlier during
5063 * When pre-allocating the page or copying data, we
5064 * need to be without the pgtable locks since we could
5065 * sleep during the process.
5067 if (!PageAnon(ptepage)) {
5068 page_dup_file_rmap(ptepage, true);
5069 } else if (page_try_dup_anon_rmap(ptepage, true,
5071 pte_t src_pte_old = entry;
5074 spin_unlock(src_ptl);
5075 spin_unlock(dst_ptl);
5076 /* Do not use reserve as it's private owned */
5077 new = alloc_huge_page(dst_vma, addr, 1);
5083 copy_user_huge_page(new, ptepage, addr, dst_vma,
5087 /* Install the new huge page if src pte stable */
5088 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5089 src_ptl = huge_pte_lockptr(h, src, src_pte);
5090 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5091 entry = huge_ptep_get(src_pte);
5092 if (!pte_same(src_pte_old, entry)) {
5093 restore_reserve_on_error(h, dst_vma, addr,
5096 /* huge_ptep of dst_pte won't change as in child */
5099 hugetlb_install_page(dst_vma, dst_pte, addr, new);
5100 spin_unlock(src_ptl);
5101 spin_unlock(dst_ptl);
5107 * No need to notify as we are downgrading page
5108 * table protection not changing it to point
5111 * See Documentation/mm/mmu_notifier.rst
5113 huge_ptep_set_wrprotect(src, addr, src_pte);
5114 entry = huge_pte_wrprotect(entry);
5117 set_huge_pte_at(dst, addr, dst_pte, entry);
5118 hugetlb_count_add(npages, dst);
5120 spin_unlock(src_ptl);
5121 spin_unlock(dst_ptl);
5125 raw_write_seqcount_end(&src->write_protect_seq);
5126 mmu_notifier_invalidate_range_end(&range);
5128 hugetlb_vma_unlock_read(src_vma);
5134 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5135 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
5137 struct hstate *h = hstate_vma(vma);
5138 struct mm_struct *mm = vma->vm_mm;
5139 spinlock_t *src_ptl, *dst_ptl;
5142 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5143 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5146 * We don't have to worry about the ordering of src and dst ptlocks
5147 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
5149 if (src_ptl != dst_ptl)
5150 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5152 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5153 set_huge_pte_at(mm, new_addr, dst_pte, pte);
5155 if (src_ptl != dst_ptl)
5156 spin_unlock(src_ptl);
5157 spin_unlock(dst_ptl);
5160 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5161 struct vm_area_struct *new_vma,
5162 unsigned long old_addr, unsigned long new_addr,
5165 struct hstate *h = hstate_vma(vma);
5166 struct address_space *mapping = vma->vm_file->f_mapping;
5167 unsigned long sz = huge_page_size(h);
5168 struct mm_struct *mm = vma->vm_mm;
5169 unsigned long old_end = old_addr + len;
5170 unsigned long last_addr_mask;
5171 pte_t *src_pte, *dst_pte;
5172 struct mmu_notifier_range range;
5173 bool shared_pmd = false;
5175 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
5177 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5179 * In case of shared PMDs, we should cover the maximum possible
5182 flush_cache_range(vma, range.start, range.end);
5184 mmu_notifier_invalidate_range_start(&range);
5185 last_addr_mask = hugetlb_mask_last_page(h);
5186 /* Prevent race with file truncation */
5187 hugetlb_vma_lock_write(vma);
5188 i_mmap_lock_write(mapping);
5189 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5190 src_pte = huge_pte_offset(mm, old_addr, sz);
5192 old_addr |= last_addr_mask;
5193 new_addr |= last_addr_mask;
5196 if (huge_pte_none(huge_ptep_get(src_pte)))
5199 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5201 old_addr |= last_addr_mask;
5202 new_addr |= last_addr_mask;
5206 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5210 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
5214 flush_tlb_range(vma, range.start, range.end);
5216 flush_tlb_range(vma, old_end - len, old_end);
5217 mmu_notifier_invalidate_range_end(&range);
5218 i_mmap_unlock_write(mapping);
5219 hugetlb_vma_unlock_write(vma);
5221 return len + old_addr - old_end;
5224 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5225 unsigned long start, unsigned long end,
5226 struct page *ref_page, zap_flags_t zap_flags)
5228 struct mm_struct *mm = vma->vm_mm;
5229 unsigned long address;
5234 struct hstate *h = hstate_vma(vma);
5235 unsigned long sz = huge_page_size(h);
5236 struct mmu_notifier_range range;
5237 unsigned long last_addr_mask;
5238 bool force_flush = false;
5240 WARN_ON(!is_vm_hugetlb_page(vma));
5241 BUG_ON(start & ~huge_page_mask(h));
5242 BUG_ON(end & ~huge_page_mask(h));
5245 * This is a hugetlb vma, all the pte entries should point
5248 tlb_change_page_size(tlb, sz);
5249 tlb_start_vma(tlb, vma);
5252 * If sharing possible, alert mmu notifiers of worst case.
5254 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
5256 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5257 mmu_notifier_invalidate_range_start(&range);
5258 last_addr_mask = hugetlb_mask_last_page(h);
5260 for (; address < end; address += sz) {
5261 ptep = huge_pte_offset(mm, address, sz);
5263 address |= last_addr_mask;
5267 ptl = huge_pte_lock(h, mm, ptep);
5268 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5270 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5272 address |= last_addr_mask;
5276 pte = huge_ptep_get(ptep);
5277 if (huge_pte_none(pte)) {
5283 * Migrating hugepage or HWPoisoned hugepage is already
5284 * unmapped and its refcount is dropped, so just clear pte here.
5286 if (unlikely(!pte_present(pte))) {
5287 #ifdef CONFIG_PTE_MARKER_UFFD_WP
5289 * If the pte was wr-protected by uffd-wp in any of the
5290 * swap forms, meanwhile the caller does not want to
5291 * drop the uffd-wp bit in this zap, then replace the
5292 * pte with a marker.
5294 if (pte_swp_uffd_wp_any(pte) &&
5295 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5296 set_huge_pte_at(mm, address, ptep,
5297 make_pte_marker(PTE_MARKER_UFFD_WP));
5300 huge_pte_clear(mm, address, ptep, sz);
5305 page = pte_page(pte);
5307 * If a reference page is supplied, it is because a specific
5308 * page is being unmapped, not a range. Ensure the page we
5309 * are about to unmap is the actual page of interest.
5312 if (page != ref_page) {
5317 * Mark the VMA as having unmapped its page so that
5318 * future faults in this VMA will fail rather than
5319 * looking like data was lost
5321 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5324 pte = huge_ptep_get_and_clear(mm, address, ptep);
5325 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5326 if (huge_pte_dirty(pte))
5327 set_page_dirty(page);
5328 #ifdef CONFIG_PTE_MARKER_UFFD_WP
5329 /* Leave a uffd-wp pte marker if needed */
5330 if (huge_pte_uffd_wp(pte) &&
5331 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5332 set_huge_pte_at(mm, address, ptep,
5333 make_pte_marker(PTE_MARKER_UFFD_WP));
5335 hugetlb_count_sub(pages_per_huge_page(h), mm);
5336 page_remove_rmap(page, vma, true);
5339 tlb_remove_page_size(tlb, page, huge_page_size(h));
5341 * Bail out after unmapping reference page if supplied
5346 mmu_notifier_invalidate_range_end(&range);
5347 tlb_end_vma(tlb, vma);
5350 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5351 * could defer the flush until now, since by holding i_mmap_rwsem we
5352 * guaranteed that the last refernece would not be dropped. But we must
5353 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5354 * dropped and the last reference to the shared PMDs page might be
5357 * In theory we could defer the freeing of the PMD pages as well, but
5358 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5359 * detect sharing, so we cannot defer the release of the page either.
5360 * Instead, do flush now.
5363 tlb_flush_mmu_tlbonly(tlb);
5366 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5367 struct vm_area_struct *vma, unsigned long start,
5368 unsigned long end, struct page *ref_page,
5369 zap_flags_t zap_flags)
5371 hugetlb_vma_lock_write(vma);
5372 i_mmap_lock_write(vma->vm_file->f_mapping);
5374 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5376 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5378 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5379 * When the vma_lock is freed, this makes the vma ineligible
5380 * for pmd sharing. And, i_mmap_rwsem is required to set up
5381 * pmd sharing. This is important as page tables for this
5382 * unmapped range will be asynchrously deleted. If the page
5383 * tables are shared, there will be issues when accessed by
5386 __hugetlb_vma_unlock_write_free(vma);
5387 i_mmap_unlock_write(vma->vm_file->f_mapping);
5389 i_mmap_unlock_write(vma->vm_file->f_mapping);
5390 hugetlb_vma_unlock_write(vma);
5394 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5395 unsigned long end, struct page *ref_page,
5396 zap_flags_t zap_flags)
5398 struct mmu_gather tlb;
5400 tlb_gather_mmu(&tlb, vma->vm_mm);
5401 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5402 tlb_finish_mmu(&tlb);
5406 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5407 * mapping it owns the reserve page for. The intention is to unmap the page
5408 * from other VMAs and let the children be SIGKILLed if they are faulting the
5411 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5412 struct page *page, unsigned long address)
5414 struct hstate *h = hstate_vma(vma);
5415 struct vm_area_struct *iter_vma;
5416 struct address_space *mapping;
5420 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5421 * from page cache lookup which is in HPAGE_SIZE units.
5423 address = address & huge_page_mask(h);
5424 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5426 mapping = vma->vm_file->f_mapping;
5429 * Take the mapping lock for the duration of the table walk. As
5430 * this mapping should be shared between all the VMAs,
5431 * __unmap_hugepage_range() is called as the lock is already held
5433 i_mmap_lock_write(mapping);
5434 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5435 /* Do not unmap the current VMA */
5436 if (iter_vma == vma)
5440 * Shared VMAs have their own reserves and do not affect
5441 * MAP_PRIVATE accounting but it is possible that a shared
5442 * VMA is using the same page so check and skip such VMAs.
5444 if (iter_vma->vm_flags & VM_MAYSHARE)
5448 * Unmap the page from other VMAs without their own reserves.
5449 * They get marked to be SIGKILLed if they fault in these
5450 * areas. This is because a future no-page fault on this VMA
5451 * could insert a zeroed page instead of the data existing
5452 * from the time of fork. This would look like data corruption
5454 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5455 unmap_hugepage_range(iter_vma, address,
5456 address + huge_page_size(h), page, 0);
5458 i_mmap_unlock_write(mapping);
5462 * hugetlb_wp() should be called with page lock of the original hugepage held.
5463 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5464 * cannot race with other handlers or page migration.
5465 * Keep the pte_same checks anyway to make transition from the mutex easier.
5467 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5468 unsigned long address, pte_t *ptep, unsigned int flags,
5469 struct page *pagecache_page, spinlock_t *ptl)
5471 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5472 pte_t pte = huge_ptep_get(ptep);
5473 struct hstate *h = hstate_vma(vma);
5474 struct page *old_page, *new_page;
5475 int outside_reserve = 0;
5477 unsigned long haddr = address & huge_page_mask(h);
5478 struct mmu_notifier_range range;
5480 VM_BUG_ON(unshare && (flags & FOLL_WRITE));
5481 VM_BUG_ON(!unshare && !(flags & FOLL_WRITE));
5484 * Never handle CoW for uffd-wp protected pages. It should be only
5485 * handled when the uffd-wp protection is removed.
5487 * Note that only the CoW optimization path (in hugetlb_no_page())
5488 * can trigger this, because hugetlb_fault() will always resolve
5489 * uffd-wp bit first.
5491 if (!unshare && huge_pte_uffd_wp(pte))
5495 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5496 * PTE mapped R/O such as maybe_mkwrite() would do.
5498 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5499 return VM_FAULT_SIGSEGV;
5501 /* Let's take out MAP_SHARED mappings first. */
5502 if (vma->vm_flags & VM_MAYSHARE) {
5503 if (unlikely(unshare))
5505 set_huge_ptep_writable(vma, haddr, ptep);
5509 old_page = pte_page(pte);
5511 delayacct_wpcopy_start();
5515 * If no-one else is actually using this page, we're the exclusive
5516 * owner and can reuse this page.
5518 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5519 if (!PageAnonExclusive(old_page))
5520 page_move_anon_rmap(old_page, vma);
5521 if (likely(!unshare))
5522 set_huge_ptep_writable(vma, haddr, ptep);
5524 delayacct_wpcopy_end();
5527 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5531 * If the process that created a MAP_PRIVATE mapping is about to
5532 * perform a COW due to a shared page count, attempt to satisfy
5533 * the allocation without using the existing reserves. The pagecache
5534 * page is used to determine if the reserve at this address was
5535 * consumed or not. If reserves were used, a partial faulted mapping
5536 * at the time of fork() could consume its reserves on COW instead
5537 * of the full address range.
5539 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5540 old_page != pagecache_page)
5541 outside_reserve = 1;
5546 * Drop page table lock as buddy allocator may be called. It will
5547 * be acquired again before returning to the caller, as expected.
5550 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5552 if (IS_ERR(new_page)) {
5554 * If a process owning a MAP_PRIVATE mapping fails to COW,
5555 * it is due to references held by a child and an insufficient
5556 * huge page pool. To guarantee the original mappers
5557 * reliability, unmap the page from child processes. The child
5558 * may get SIGKILLed if it later faults.
5560 if (outside_reserve) {
5561 struct address_space *mapping = vma->vm_file->f_mapping;
5567 * Drop hugetlb_fault_mutex and vma_lock before
5568 * unmapping. unmapping needs to hold vma_lock
5569 * in write mode. Dropping vma_lock in read mode
5570 * here is OK as COW mappings do not interact with
5573 * Reacquire both after unmap operation.
5575 idx = vma_hugecache_offset(h, vma, haddr);
5576 hash = hugetlb_fault_mutex_hash(mapping, idx);
5577 hugetlb_vma_unlock_read(vma);
5578 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5580 unmap_ref_private(mm, vma, old_page, haddr);
5582 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5583 hugetlb_vma_lock_read(vma);
5585 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5587 pte_same(huge_ptep_get(ptep), pte)))
5588 goto retry_avoidcopy;
5590 * race occurs while re-acquiring page table
5591 * lock, and our job is done.
5593 delayacct_wpcopy_end();
5597 ret = vmf_error(PTR_ERR(new_page));
5598 goto out_release_old;
5602 * When the original hugepage is shared one, it does not have
5603 * anon_vma prepared.
5605 if (unlikely(anon_vma_prepare(vma))) {
5607 goto out_release_all;
5610 copy_user_huge_page(new_page, old_page, address, vma,
5611 pages_per_huge_page(h));
5612 __SetPageUptodate(new_page);
5614 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5615 haddr + huge_page_size(h));
5616 mmu_notifier_invalidate_range_start(&range);
5619 * Retake the page table lock to check for racing updates
5620 * before the page tables are altered
5623 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5624 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5625 ClearHPageRestoreReserve(new_page);
5627 /* Break COW or unshare */
5628 huge_ptep_clear_flush(vma, haddr, ptep);
5629 mmu_notifier_invalidate_range(mm, range.start, range.end);
5630 page_remove_rmap(old_page, vma, true);
5631 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5632 set_huge_pte_at(mm, haddr, ptep,
5633 make_huge_pte(vma, new_page, !unshare));
5634 SetHPageMigratable(new_page);
5635 /* Make the old page be freed below */
5636 new_page = old_page;
5639 mmu_notifier_invalidate_range_end(&range);
5642 * No restore in case of successful pagetable update (Break COW or
5645 if (new_page != old_page)
5646 restore_reserve_on_error(h, vma, haddr, new_page);
5651 spin_lock(ptl); /* Caller expects lock to be held */
5653 delayacct_wpcopy_end();
5658 * Return whether there is a pagecache page to back given address within VMA.
5659 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5661 static bool hugetlbfs_pagecache_present(struct hstate *h,
5662 struct vm_area_struct *vma, unsigned long address)
5664 struct address_space *mapping;
5668 mapping = vma->vm_file->f_mapping;
5669 idx = vma_hugecache_offset(h, vma, address);
5671 page = find_get_page(mapping, idx);
5674 return page != NULL;
5677 int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping,
5680 struct folio *folio = page_folio(page);
5681 struct inode *inode = mapping->host;
5682 struct hstate *h = hstate_inode(inode);
5685 __folio_set_locked(folio);
5686 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5688 if (unlikely(err)) {
5689 __folio_clear_locked(folio);
5692 ClearHPageRestoreReserve(page);
5695 * mark folio dirty so that it will not be removed from cache/file
5696 * by non-hugetlbfs specific code paths.
5698 folio_mark_dirty(folio);
5700 spin_lock(&inode->i_lock);
5701 inode->i_blocks += blocks_per_huge_page(h);
5702 spin_unlock(&inode->i_lock);
5706 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5707 struct address_space *mapping,
5710 unsigned long haddr,
5712 unsigned long reason)
5715 struct vm_fault vmf = {
5718 .real_address = addr,
5722 * Hard to debug if it ends up being
5723 * used by a callee that assumes
5724 * something about the other
5725 * uninitialized fields... same as in
5731 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5732 * userfault. Also mmap_lock could be dropped due to handling
5733 * userfault, any vma operation should be careful from here.
5735 hugetlb_vma_unlock_read(vma);
5736 hash = hugetlb_fault_mutex_hash(mapping, idx);
5737 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5738 return handle_userfault(&vmf, reason);
5742 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5743 * false if pte changed or is changing.
5745 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5746 pte_t *ptep, pte_t old_pte)
5751 ptl = huge_pte_lock(h, mm, ptep);
5752 same = pte_same(huge_ptep_get(ptep), old_pte);
5758 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5759 struct vm_area_struct *vma,
5760 struct address_space *mapping, pgoff_t idx,
5761 unsigned long address, pte_t *ptep,
5762 pte_t old_pte, unsigned int flags)
5764 struct hstate *h = hstate_vma(vma);
5765 vm_fault_t ret = VM_FAULT_SIGBUS;
5771 unsigned long haddr = address & huge_page_mask(h);
5772 bool new_page, new_pagecache_page = false;
5773 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5776 * Currently, we are forced to kill the process in the event the
5777 * original mapper has unmapped pages from the child due to a failed
5778 * COW/unsharing. Warn that such a situation has occurred as it may not
5781 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5782 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5788 * Use page lock to guard against racing truncation
5789 * before we get page_table_lock.
5792 page = find_lock_page(mapping, idx);
5794 size = i_size_read(mapping->host) >> huge_page_shift(h);
5797 /* Check for page in userfault range */
5798 if (userfaultfd_missing(vma)) {
5800 * Since hugetlb_no_page() was examining pte
5801 * without pgtable lock, we need to re-test under
5802 * lock because the pte may not be stable and could
5803 * have changed from under us. Try to detect
5804 * either changed or during-changing ptes and retry
5805 * properly when needed.
5807 * Note that userfaultfd is actually fine with
5808 * false positives (e.g. caused by pte changed),
5809 * but not wrong logical events (e.g. caused by
5810 * reading a pte during changing). The latter can
5811 * confuse the userspace, so the strictness is very
5812 * much preferred. E.g., MISSING event should
5813 * never happen on the page after UFFDIO_COPY has
5814 * correctly installed the page and returned.
5816 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5821 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5826 page = alloc_huge_page(vma, haddr, 0);
5829 * Returning error will result in faulting task being
5830 * sent SIGBUS. The hugetlb fault mutex prevents two
5831 * tasks from racing to fault in the same page which
5832 * could result in false unable to allocate errors.
5833 * Page migration does not take the fault mutex, but
5834 * does a clear then write of pte's under page table
5835 * lock. Page fault code could race with migration,
5836 * notice the clear pte and try to allocate a page
5837 * here. Before returning error, get ptl and make
5838 * sure there really is no pte entry.
5840 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5841 ret = vmf_error(PTR_ERR(page));
5846 clear_huge_page(page, address, pages_per_huge_page(h));
5847 __SetPageUptodate(page);
5850 if (vma->vm_flags & VM_MAYSHARE) {
5851 int err = hugetlb_add_to_page_cache(page, mapping, idx);
5854 * err can't be -EEXIST which implies someone
5855 * else consumed the reservation since hugetlb
5856 * fault mutex is held when add a hugetlb page
5857 * to the page cache. So it's safe to call
5858 * restore_reserve_on_error() here.
5860 restore_reserve_on_error(h, vma, haddr, page);
5864 new_pagecache_page = true;
5867 if (unlikely(anon_vma_prepare(vma))) {
5869 goto backout_unlocked;
5875 * If memory error occurs between mmap() and fault, some process
5876 * don't have hwpoisoned swap entry for errored virtual address.
5877 * So we need to block hugepage fault by PG_hwpoison bit check.
5879 if (unlikely(PageHWPoison(page))) {
5880 ret = VM_FAULT_HWPOISON_LARGE |
5881 VM_FAULT_SET_HINDEX(hstate_index(h));
5882 goto backout_unlocked;
5885 /* Check for page in userfault range. */
5886 if (userfaultfd_minor(vma)) {
5889 /* See comment in userfaultfd_missing() block above */
5890 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5894 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5901 * If we are going to COW a private mapping later, we examine the
5902 * pending reservations for this page now. This will ensure that
5903 * any allocations necessary to record that reservation occur outside
5906 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5907 if (vma_needs_reservation(h, vma, haddr) < 0) {
5909 goto backout_unlocked;
5911 /* Just decrements count, does not deallocate */
5912 vma_end_reservation(h, vma, haddr);
5915 ptl = huge_pte_lock(h, mm, ptep);
5917 /* If pte changed from under us, retry */
5918 if (!pte_same(huge_ptep_get(ptep), old_pte))
5922 ClearHPageRestoreReserve(page);
5923 hugepage_add_new_anon_rmap(page, vma, haddr);
5925 page_dup_file_rmap(page, true);
5926 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5927 && (vma->vm_flags & VM_SHARED)));
5929 * If this pte was previously wr-protected, keep it wr-protected even
5932 if (unlikely(pte_marker_uffd_wp(old_pte)))
5933 new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte));
5934 set_huge_pte_at(mm, haddr, ptep, new_pte);
5936 hugetlb_count_add(pages_per_huge_page(h), mm);
5937 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5938 /* Optimization, do the COW without a second fault */
5939 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5945 * Only set HPageMigratable in newly allocated pages. Existing pages
5946 * found in the pagecache may not have HPageMigratableset if they have
5947 * been isolated for migration.
5950 SetHPageMigratable(page);
5954 hugetlb_vma_unlock_read(vma);
5955 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5961 if (new_page && !new_pagecache_page)
5962 restore_reserve_on_error(h, vma, haddr, page);
5970 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5972 unsigned long key[2];
5975 key[0] = (unsigned long) mapping;
5978 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5980 return hash & (num_fault_mutexes - 1);
5984 * For uniprocessor systems we always use a single mutex, so just
5985 * return 0 and avoid the hashing overhead.
5987 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5993 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5994 unsigned long address, unsigned int flags)
6001 struct page *page = NULL;
6002 struct page *pagecache_page = NULL;
6003 struct hstate *h = hstate_vma(vma);
6004 struct address_space *mapping;
6005 int need_wait_lock = 0;
6006 unsigned long haddr = address & huge_page_mask(h);
6008 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
6011 * Since we hold no locks, ptep could be stale. That is
6012 * OK as we are only making decisions based on content and
6013 * not actually modifying content here.
6015 entry = huge_ptep_get(ptep);
6016 if (unlikely(is_hugetlb_entry_migration(entry))) {
6017 migration_entry_wait_huge(vma, ptep);
6019 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6020 return VM_FAULT_HWPOISON_LARGE |
6021 VM_FAULT_SET_HINDEX(hstate_index(h));
6025 * Serialize hugepage allocation and instantiation, so that we don't
6026 * get spurious allocation failures if two CPUs race to instantiate
6027 * the same page in the page cache.
6029 mapping = vma->vm_file->f_mapping;
6030 idx = vma_hugecache_offset(h, vma, haddr);
6031 hash = hugetlb_fault_mutex_hash(mapping, idx);
6032 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6035 * Acquire vma lock before calling huge_pte_alloc and hold
6036 * until finished with ptep. This prevents huge_pmd_unshare from
6037 * being called elsewhere and making the ptep no longer valid.
6039 * ptep could have already be assigned via huge_pte_offset. That
6040 * is OK, as huge_pte_alloc will return the same value unless
6041 * something has changed.
6043 hugetlb_vma_lock_read(vma);
6044 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6046 hugetlb_vma_unlock_read(vma);
6047 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6048 return VM_FAULT_OOM;
6051 entry = huge_ptep_get(ptep);
6052 /* PTE markers should be handled the same way as none pte */
6053 if (huge_pte_none_mostly(entry))
6055 * hugetlb_no_page will drop vma lock and hugetlb fault
6056 * mutex internally, which make us return immediately.
6058 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6064 * entry could be a migration/hwpoison entry at this point, so this
6065 * check prevents the kernel from going below assuming that we have
6066 * an active hugepage in pagecache. This goto expects the 2nd page
6067 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6068 * properly handle it.
6070 if (!pte_present(entry))
6074 * If we are going to COW/unshare the mapping later, we examine the
6075 * pending reservations for this page now. This will ensure that any
6076 * allocations necessary to record that reservation occur outside the
6077 * spinlock. Also lookup the pagecache page now as it is used to
6078 * determine if a reservation has been consumed.
6080 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6081 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6082 if (vma_needs_reservation(h, vma, haddr) < 0) {
6086 /* Just decrements count, does not deallocate */
6087 vma_end_reservation(h, vma, haddr);
6089 pagecache_page = find_lock_page(mapping, idx);
6092 ptl = huge_pte_lock(h, mm, ptep);
6094 /* Check for a racing update before calling hugetlb_wp() */
6095 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6098 /* Handle userfault-wp first, before trying to lock more pages */
6099 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6100 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6101 struct vm_fault vmf = {
6104 .real_address = address,
6109 if (pagecache_page) {
6110 unlock_page(pagecache_page);
6111 put_page(pagecache_page);
6113 hugetlb_vma_unlock_read(vma);
6114 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6115 return handle_userfault(&vmf, VM_UFFD_WP);
6119 * hugetlb_wp() requires page locks of pte_page(entry) and
6120 * pagecache_page, so here we need take the former one
6121 * when page != pagecache_page or !pagecache_page.
6123 page = pte_page(entry);
6124 if (page != pagecache_page)
6125 if (!trylock_page(page)) {
6132 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6133 if (!huge_pte_write(entry)) {
6134 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6135 pagecache_page, ptl);
6137 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6138 entry = huge_pte_mkdirty(entry);
6141 entry = pte_mkyoung(entry);
6142 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6143 flags & FAULT_FLAG_WRITE))
6144 update_mmu_cache(vma, haddr, ptep);
6146 if (page != pagecache_page)
6152 if (pagecache_page) {
6153 unlock_page(pagecache_page);
6154 put_page(pagecache_page);
6157 hugetlb_vma_unlock_read(vma);
6158 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6160 * Generally it's safe to hold refcount during waiting page lock. But
6161 * here we just wait to defer the next page fault to avoid busy loop and
6162 * the page is not used after unlocked before returning from the current
6163 * page fault. So we are safe from accessing freed page, even if we wait
6164 * here without taking refcount.
6167 wait_on_page_locked(page);
6171 #ifdef CONFIG_USERFAULTFD
6173 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
6174 * modifications for huge pages.
6176 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
6178 struct vm_area_struct *dst_vma,
6179 unsigned long dst_addr,
6180 unsigned long src_addr,
6181 enum mcopy_atomic_mode mode,
6182 struct page **pagep,
6185 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
6186 struct hstate *h = hstate_vma(dst_vma);
6187 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6188 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6190 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6196 bool page_in_pagecache = false;
6200 page = find_lock_page(mapping, idx);
6203 page_in_pagecache = true;
6204 } else if (!*pagep) {
6205 /* If a page already exists, then it's UFFDIO_COPY for
6206 * a non-missing case. Return -EEXIST.
6209 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6214 page = alloc_huge_page(dst_vma, dst_addr, 0);
6220 ret = copy_huge_page_from_user(page,
6221 (const void __user *) src_addr,
6222 pages_per_huge_page(h), false);
6224 /* fallback to copy_from_user outside mmap_lock */
6225 if (unlikely(ret)) {
6227 /* Free the allocated page which may have
6228 * consumed a reservation.
6230 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6233 /* Allocate a temporary page to hold the copied
6236 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
6242 /* Set the outparam pagep and return to the caller to
6243 * copy the contents outside the lock. Don't free the
6250 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6257 page = alloc_huge_page(dst_vma, dst_addr, 0);
6264 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
6265 pages_per_huge_page(h));
6271 * The memory barrier inside __SetPageUptodate makes sure that
6272 * preceding stores to the page contents become visible before
6273 * the set_pte_at() write.
6275 __SetPageUptodate(page);
6277 /* Add shared, newly allocated pages to the page cache. */
6278 if (vm_shared && !is_continue) {
6279 size = i_size_read(mapping->host) >> huge_page_shift(h);
6282 goto out_release_nounlock;
6285 * Serialization between remove_inode_hugepages() and
6286 * hugetlb_add_to_page_cache() below happens through the
6287 * hugetlb_fault_mutex_table that here must be hold by
6290 ret = hugetlb_add_to_page_cache(page, mapping, idx);
6292 goto out_release_nounlock;
6293 page_in_pagecache = true;
6296 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6299 if (PageHWPoison(page))
6300 goto out_release_unlock;
6303 * We allow to overwrite a pte marker: consider when both MISSING|WP
6304 * registered, we firstly wr-protect a none pte which has no page cache
6305 * page backing it, then access the page.
6308 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6309 goto out_release_unlock;
6311 if (page_in_pagecache) {
6312 page_dup_file_rmap(page, true);
6314 ClearHPageRestoreReserve(page);
6315 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6319 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6320 * with wp flag set, don't set pte write bit.
6322 if (wp_copy || (is_continue && !vm_shared))
6325 writable = dst_vma->vm_flags & VM_WRITE;
6327 _dst_pte = make_huge_pte(dst_vma, page, writable);
6329 * Always mark UFFDIO_COPY page dirty; note that this may not be
6330 * extremely important for hugetlbfs for now since swapping is not
6331 * supported, but we should still be clear in that this page cannot be
6332 * thrown away at will, even if write bit not set.
6334 _dst_pte = huge_pte_mkdirty(_dst_pte);
6335 _dst_pte = pte_mkyoung(_dst_pte);
6338 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6340 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6342 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6344 /* No need to invalidate - it was non-present before */
6345 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6349 SetHPageMigratable(page);
6350 if (vm_shared || is_continue)
6357 if (vm_shared || is_continue)
6359 out_release_nounlock:
6360 if (!page_in_pagecache)
6361 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6365 #endif /* CONFIG_USERFAULTFD */
6367 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6368 int refs, struct page **pages,
6369 struct vm_area_struct **vmas)
6373 for (nr = 0; nr < refs; nr++) {
6375 pages[nr] = nth_page(page, nr);
6381 static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte,
6384 pte_t pteval = huge_ptep_get(pte);
6387 if (is_swap_pte(pteval))
6389 if (huge_pte_write(pteval))
6391 if (flags & FOLL_WRITE)
6393 if (gup_must_unshare(flags, pte_page(pteval))) {
6400 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6401 struct page **pages, struct vm_area_struct **vmas,
6402 unsigned long *position, unsigned long *nr_pages,
6403 long i, unsigned int flags, int *locked)
6405 unsigned long pfn_offset;
6406 unsigned long vaddr = *position;
6407 unsigned long remainder = *nr_pages;
6408 struct hstate *h = hstate_vma(vma);
6409 int err = -EFAULT, refs;
6411 while (vaddr < vma->vm_end && remainder) {
6413 spinlock_t *ptl = NULL;
6414 bool unshare = false;
6419 * If we have a pending SIGKILL, don't keep faulting pages and
6420 * potentially allocating memory.
6422 if (fatal_signal_pending(current)) {
6428 * Some archs (sparc64, sh*) have multiple pte_ts to
6429 * each hugepage. We have to make sure we get the
6430 * first, for the page indexing below to work.
6432 * Note that page table lock is not held when pte is null.
6434 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6437 ptl = huge_pte_lock(h, mm, pte);
6438 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6441 * When coredumping, it suits get_dump_page if we just return
6442 * an error where there's an empty slot with no huge pagecache
6443 * to back it. This way, we avoid allocating a hugepage, and
6444 * the sparse dumpfile avoids allocating disk blocks, but its
6445 * huge holes still show up with zeroes where they need to be.
6447 if (absent && (flags & FOLL_DUMP) &&
6448 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6456 * We need call hugetlb_fault for both hugepages under migration
6457 * (in which case hugetlb_fault waits for the migration,) and
6458 * hwpoisoned hugepages (in which case we need to prevent the
6459 * caller from accessing to them.) In order to do this, we use
6460 * here is_swap_pte instead of is_hugetlb_entry_migration and
6461 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6462 * both cases, and because we can't follow correct pages
6463 * directly from any kind of swap entries.
6466 __follow_hugetlb_must_fault(flags, pte, &unshare)) {
6468 unsigned int fault_flags = 0;
6472 if (flags & FOLL_WRITE)
6473 fault_flags |= FAULT_FLAG_WRITE;
6475 fault_flags |= FAULT_FLAG_UNSHARE;
6477 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6478 FAULT_FLAG_KILLABLE;
6479 if (flags & FOLL_NOWAIT)
6480 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6481 FAULT_FLAG_RETRY_NOWAIT;
6482 if (flags & FOLL_TRIED) {
6484 * Note: FAULT_FLAG_ALLOW_RETRY and
6485 * FAULT_FLAG_TRIED can co-exist
6487 fault_flags |= FAULT_FLAG_TRIED;
6489 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6490 if (ret & VM_FAULT_ERROR) {
6491 err = vm_fault_to_errno(ret, flags);
6495 if (ret & VM_FAULT_RETRY) {
6497 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6501 * VM_FAULT_RETRY must not return an
6502 * error, it will return zero
6505 * No need to update "position" as the
6506 * caller will not check it after
6507 * *nr_pages is set to 0.
6514 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6515 page = pte_page(huge_ptep_get(pte));
6517 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6518 !PageAnonExclusive(page), page);
6521 * If subpage information not requested, update counters
6522 * and skip the same_page loop below.
6524 if (!pages && !vmas && !pfn_offset &&
6525 (vaddr + huge_page_size(h) < vma->vm_end) &&
6526 (remainder >= pages_per_huge_page(h))) {
6527 vaddr += huge_page_size(h);
6528 remainder -= pages_per_huge_page(h);
6529 i += pages_per_huge_page(h);
6534 /* vaddr may not be aligned to PAGE_SIZE */
6535 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6536 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6539 record_subpages_vmas(nth_page(page, pfn_offset),
6541 likely(pages) ? pages + i : NULL,
6542 vmas ? vmas + i : NULL);
6546 * try_grab_folio() should always succeed here,
6547 * because: a) we hold the ptl lock, and b) we've just
6548 * checked that the huge page is present in the page
6549 * tables. If the huge page is present, then the tail
6550 * pages must also be present. The ptl prevents the
6551 * head page and tail pages from being rearranged in
6552 * any way. So this page must be available at this
6553 * point, unless the page refcount overflowed:
6555 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6564 vaddr += (refs << PAGE_SHIFT);
6570 *nr_pages = remainder;
6572 * setting position is actually required only if remainder is
6573 * not zero but it's faster not to add a "if (remainder)"
6581 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6582 unsigned long address, unsigned long end,
6583 pgprot_t newprot, unsigned long cp_flags)
6585 struct mm_struct *mm = vma->vm_mm;
6586 unsigned long start = address;
6589 struct hstate *h = hstate_vma(vma);
6590 unsigned long pages = 0, psize = huge_page_size(h);
6591 bool shared_pmd = false;
6592 struct mmu_notifier_range range;
6593 unsigned long last_addr_mask;
6594 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6595 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6598 * In the case of shared PMDs, the area to flush could be beyond
6599 * start/end. Set range.start/range.end to cover the maximum possible
6600 * range if PMD sharing is possible.
6602 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6603 0, vma, mm, start, end);
6604 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6606 BUG_ON(address >= end);
6607 flush_cache_range(vma, range.start, range.end);
6609 mmu_notifier_invalidate_range_start(&range);
6610 hugetlb_vma_lock_write(vma);
6611 i_mmap_lock_write(vma->vm_file->f_mapping);
6612 last_addr_mask = hugetlb_mask_last_page(h);
6613 for (; address < end; address += psize) {
6615 ptep = huge_pte_offset(mm, address, psize);
6618 address |= last_addr_mask;
6622 * Userfaultfd wr-protect requires pgtable
6623 * pre-allocations to install pte markers.
6625 ptep = huge_pte_alloc(mm, vma, address, psize);
6629 ptl = huge_pte_lock(h, mm, ptep);
6630 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6632 * When uffd-wp is enabled on the vma, unshare
6633 * shouldn't happen at all. Warn about it if it
6634 * happened due to some reason.
6636 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6640 address |= last_addr_mask;
6643 pte = huge_ptep_get(ptep);
6644 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6645 /* Nothing to do. */
6646 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6647 swp_entry_t entry = pte_to_swp_entry(pte);
6648 struct page *page = pfn_swap_entry_to_page(entry);
6651 if (is_writable_migration_entry(entry)) {
6653 entry = make_readable_exclusive_migration_entry(
6656 entry = make_readable_migration_entry(
6658 newpte = swp_entry_to_pte(entry);
6663 newpte = pte_swp_mkuffd_wp(newpte);
6664 else if (uffd_wp_resolve)
6665 newpte = pte_swp_clear_uffd_wp(newpte);
6666 if (!pte_same(pte, newpte))
6667 set_huge_pte_at(mm, address, ptep, newpte);
6668 } else if (unlikely(is_pte_marker(pte))) {
6669 /* No other markers apply for now. */
6670 WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
6671 if (uffd_wp_resolve)
6672 /* Safe to modify directly (non-present->none). */
6673 huge_pte_clear(mm, address, ptep, psize);
6674 } else if (!huge_pte_none(pte)) {
6676 unsigned int shift = huge_page_shift(hstate_vma(vma));
6678 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6679 pte = huge_pte_modify(old_pte, newprot);
6680 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6682 pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte));
6683 else if (uffd_wp_resolve)
6684 pte = huge_pte_clear_uffd_wp(pte);
6685 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6689 if (unlikely(uffd_wp))
6690 /* Safe to modify directly (none->non-present). */
6691 set_huge_pte_at(mm, address, ptep,
6692 make_pte_marker(PTE_MARKER_UFFD_WP));
6697 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6698 * may have cleared our pud entry and done put_page on the page table:
6699 * once we release i_mmap_rwsem, another task can do the final put_page
6700 * and that page table be reused and filled with junk. If we actually
6701 * did unshare a page of pmds, flush the range corresponding to the pud.
6704 flush_hugetlb_tlb_range(vma, range.start, range.end);
6706 flush_hugetlb_tlb_range(vma, start, end);
6708 * No need to call mmu_notifier_invalidate_range() we are downgrading
6709 * page table protection not changing it to point to a new page.
6711 * See Documentation/mm/mmu_notifier.rst
6713 i_mmap_unlock_write(vma->vm_file->f_mapping);
6714 hugetlb_vma_unlock_write(vma);
6715 mmu_notifier_invalidate_range_end(&range);
6717 return pages << h->order;
6720 /* Return true if reservation was successful, false otherwise. */
6721 bool hugetlb_reserve_pages(struct inode *inode,
6723 struct vm_area_struct *vma,
6724 vm_flags_t vm_flags)
6727 struct hstate *h = hstate_inode(inode);
6728 struct hugepage_subpool *spool = subpool_inode(inode);
6729 struct resv_map *resv_map;
6730 struct hugetlb_cgroup *h_cg = NULL;
6731 long gbl_reserve, regions_needed = 0;
6733 /* This should never happen */
6735 VM_WARN(1, "%s called with a negative range\n", __func__);
6740 * vma specific semaphore used for pmd sharing and fault/truncation
6743 hugetlb_vma_lock_alloc(vma);
6746 * Only apply hugepage reservation if asked. At fault time, an
6747 * attempt will be made for VM_NORESERVE to allocate a page
6748 * without using reserves
6750 if (vm_flags & VM_NORESERVE)
6754 * Shared mappings base their reservation on the number of pages that
6755 * are already allocated on behalf of the file. Private mappings need
6756 * to reserve the full area even if read-only as mprotect() may be
6757 * called to make the mapping read-write. Assume !vma is a shm mapping
6759 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6761 * resv_map can not be NULL as hugetlb_reserve_pages is only
6762 * called for inodes for which resv_maps were created (see
6763 * hugetlbfs_get_inode).
6765 resv_map = inode_resv_map(inode);
6767 chg = region_chg(resv_map, from, to, ®ions_needed);
6769 /* Private mapping. */
6770 resv_map = resv_map_alloc();
6776 set_vma_resv_map(vma, resv_map);
6777 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6783 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6784 chg * pages_per_huge_page(h), &h_cg) < 0)
6787 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6788 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6791 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6795 * There must be enough pages in the subpool for the mapping. If
6796 * the subpool has a minimum size, there may be some global
6797 * reservations already in place (gbl_reserve).
6799 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6800 if (gbl_reserve < 0)
6801 goto out_uncharge_cgroup;
6804 * Check enough hugepages are available for the reservation.
6805 * Hand the pages back to the subpool if there are not
6807 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6811 * Account for the reservations made. Shared mappings record regions
6812 * that have reservations as they are shared by multiple VMAs.
6813 * When the last VMA disappears, the region map says how much
6814 * the reservation was and the page cache tells how much of
6815 * the reservation was consumed. Private mappings are per-VMA and
6816 * only the consumed reservations are tracked. When the VMA
6817 * disappears, the original reservation is the VMA size and the
6818 * consumed reservations are stored in the map. Hence, nothing
6819 * else has to be done for private mappings here
6821 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6822 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6824 if (unlikely(add < 0)) {
6825 hugetlb_acct_memory(h, -gbl_reserve);
6827 } else if (unlikely(chg > add)) {
6829 * pages in this range were added to the reserve
6830 * map between region_chg and region_add. This
6831 * indicates a race with alloc_huge_page. Adjust
6832 * the subpool and reserve counts modified above
6833 * based on the difference.
6838 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6839 * reference to h_cg->css. See comment below for detail.
6841 hugetlb_cgroup_uncharge_cgroup_rsvd(
6843 (chg - add) * pages_per_huge_page(h), h_cg);
6845 rsv_adjust = hugepage_subpool_put_pages(spool,
6847 hugetlb_acct_memory(h, -rsv_adjust);
6850 * The file_regions will hold their own reference to
6851 * h_cg->css. So we should release the reference held
6852 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6855 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6861 /* put back original number of pages, chg */
6862 (void)hugepage_subpool_put_pages(spool, chg);
6863 out_uncharge_cgroup:
6864 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6865 chg * pages_per_huge_page(h), h_cg);
6867 hugetlb_vma_lock_free(vma);
6868 if (!vma || vma->vm_flags & VM_MAYSHARE)
6869 /* Only call region_abort if the region_chg succeeded but the
6870 * region_add failed or didn't run.
6872 if (chg >= 0 && add < 0)
6873 region_abort(resv_map, from, to, regions_needed);
6874 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6875 kref_put(&resv_map->refs, resv_map_release);
6879 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6882 struct hstate *h = hstate_inode(inode);
6883 struct resv_map *resv_map = inode_resv_map(inode);
6885 struct hugepage_subpool *spool = subpool_inode(inode);
6889 * Since this routine can be called in the evict inode path for all
6890 * hugetlbfs inodes, resv_map could be NULL.
6893 chg = region_del(resv_map, start, end);
6895 * region_del() can fail in the rare case where a region
6896 * must be split and another region descriptor can not be
6897 * allocated. If end == LONG_MAX, it will not fail.
6903 spin_lock(&inode->i_lock);
6904 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6905 spin_unlock(&inode->i_lock);
6908 * If the subpool has a minimum size, the number of global
6909 * reservations to be released may be adjusted.
6911 * Note that !resv_map implies freed == 0. So (chg - freed)
6912 * won't go negative.
6914 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6915 hugetlb_acct_memory(h, -gbl_reserve);
6920 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6921 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6922 struct vm_area_struct *vma,
6923 unsigned long addr, pgoff_t idx)
6925 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6927 unsigned long sbase = saddr & PUD_MASK;
6928 unsigned long s_end = sbase + PUD_SIZE;
6930 /* Allow segments to share if only one is marked locked */
6931 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6932 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6935 * match the virtual addresses, permission and the alignment of the
6938 * Also, vma_lock (vm_private_data) is required for sharing.
6940 if (pmd_index(addr) != pmd_index(saddr) ||
6941 vm_flags != svm_flags ||
6942 !range_in_vma(svma, sbase, s_end) ||
6943 !svma->vm_private_data)
6949 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6951 unsigned long start = addr & PUD_MASK;
6952 unsigned long end = start + PUD_SIZE;
6954 #ifdef CONFIG_USERFAULTFD
6955 if (uffd_disable_huge_pmd_share(vma))
6959 * check on proper vm_flags and page table alignment
6961 if (!(vma->vm_flags & VM_MAYSHARE))
6963 if (!vma->vm_private_data) /* vma lock required for sharing */
6965 if (!range_in_vma(vma, start, end))
6971 * Determine if start,end range within vma could be mapped by shared pmd.
6972 * If yes, adjust start and end to cover range associated with possible
6973 * shared pmd mappings.
6975 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6976 unsigned long *start, unsigned long *end)
6978 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6979 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6982 * vma needs to span at least one aligned PUD size, and the range
6983 * must be at least partially within in.
6985 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6986 (*end <= v_start) || (*start >= v_end))
6989 /* Extend the range to be PUD aligned for a worst case scenario */
6990 if (*start > v_start)
6991 *start = ALIGN_DOWN(*start, PUD_SIZE);
6994 *end = ALIGN(*end, PUD_SIZE);
6998 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6999 * and returns the corresponding pte. While this is not necessary for the
7000 * !shared pmd case because we can allocate the pmd later as well, it makes the
7001 * code much cleaner. pmd allocation is essential for the shared case because
7002 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7003 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7004 * bad pmd for sharing.
7006 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7007 unsigned long addr, pud_t *pud)
7009 struct address_space *mapping = vma->vm_file->f_mapping;
7010 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7012 struct vm_area_struct *svma;
7013 unsigned long saddr;
7018 i_mmap_lock_read(mapping);
7019 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7023 saddr = page_table_shareable(svma, vma, addr, idx);
7025 spte = huge_pte_offset(svma->vm_mm, saddr,
7026 vma_mmu_pagesize(svma));
7028 get_page(virt_to_page(spte));
7037 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
7038 if (pud_none(*pud)) {
7039 pud_populate(mm, pud,
7040 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7043 put_page(virt_to_page(spte));
7047 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7048 i_mmap_unlock_read(mapping);
7053 * unmap huge page backed by shared pte.
7055 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7056 * indicated by page_count > 1, unmap is achieved by clearing pud and
7057 * decrementing the ref count. If count == 1, the pte page is not shared.
7059 * Called with page table lock held.
7061 * returns: 1 successfully unmapped a shared pte page
7062 * 0 the underlying pte page is not shared, or it is the last user
7064 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7065 unsigned long addr, pte_t *ptep)
7067 pgd_t *pgd = pgd_offset(mm, addr);
7068 p4d_t *p4d = p4d_offset(pgd, addr);
7069 pud_t *pud = pud_offset(p4d, addr);
7071 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7072 hugetlb_vma_assert_locked(vma);
7073 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7074 if (page_count(virt_to_page(ptep)) == 1)
7078 put_page(virt_to_page(ptep));
7083 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7085 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7086 unsigned long addr, pud_t *pud)
7091 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7092 unsigned long addr, pte_t *ptep)
7097 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7098 unsigned long *start, unsigned long *end)
7102 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7106 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7108 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7109 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7110 unsigned long addr, unsigned long sz)
7117 pgd = pgd_offset(mm, addr);
7118 p4d = p4d_alloc(mm, pgd, addr);
7121 pud = pud_alloc(mm, p4d, addr);
7123 if (sz == PUD_SIZE) {
7126 BUG_ON(sz != PMD_SIZE);
7127 if (want_pmd_share(vma, addr) && pud_none(*pud))
7128 pte = huge_pmd_share(mm, vma, addr, pud);
7130 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7133 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
7139 * huge_pte_offset() - Walk the page table to resolve the hugepage
7140 * entry at address @addr
7142 * Return: Pointer to page table entry (PUD or PMD) for
7143 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7144 * size @sz doesn't match the hugepage size at this level of the page
7147 pte_t *huge_pte_offset(struct mm_struct *mm,
7148 unsigned long addr, unsigned long sz)
7155 pgd = pgd_offset(mm, addr);
7156 if (!pgd_present(*pgd))
7158 p4d = p4d_offset(pgd, addr);
7159 if (!p4d_present(*p4d))
7162 pud = pud_offset(p4d, addr);
7164 /* must be pud huge, non-present or none */
7165 return (pte_t *)pud;
7166 if (!pud_present(*pud))
7168 /* must have a valid entry and size to go further */
7170 pmd = pmd_offset(pud, addr);
7171 /* must be pmd huge, non-present or none */
7172 return (pte_t *)pmd;
7176 * Return a mask that can be used to update an address to the last huge
7177 * page in a page table page mapping size. Used to skip non-present
7178 * page table entries when linearly scanning address ranges. Architectures
7179 * with unique huge page to page table relationships can define their own
7180 * version of this routine.
7182 unsigned long hugetlb_mask_last_page(struct hstate *h)
7184 unsigned long hp_size = huge_page_size(h);
7186 if (hp_size == PUD_SIZE)
7187 return P4D_SIZE - PUD_SIZE;
7188 else if (hp_size == PMD_SIZE)
7189 return PUD_SIZE - PMD_SIZE;
7196 /* See description above. Architectures can provide their own version. */
7197 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7199 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7200 if (huge_page_size(h) == PMD_SIZE)
7201 return PUD_SIZE - PMD_SIZE;
7206 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7209 * These functions are overwritable if your architecture needs its own
7212 struct page * __weak
7213 follow_huge_addr(struct mm_struct *mm, unsigned long address,
7216 return ERR_PTR(-EINVAL);
7219 struct page * __weak
7220 follow_huge_pd(struct vm_area_struct *vma,
7221 unsigned long address, hugepd_t hpd, int flags, int pdshift)
7223 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
7227 struct page * __weak
7228 follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags)
7230 struct hstate *h = hstate_vma(vma);
7231 struct mm_struct *mm = vma->vm_mm;
7232 struct page *page = NULL;
7237 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
7238 * follow_hugetlb_page().
7240 if (WARN_ON_ONCE(flags & FOLL_PIN))
7244 ptep = huge_pte_offset(mm, address, huge_page_size(h));
7248 ptl = huge_pte_lock(h, mm, ptep);
7249 pte = huge_ptep_get(ptep);
7250 if (pte_present(pte)) {
7251 page = pte_page(pte) +
7252 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
7254 * try_grab_page() should always succeed here, because: a) we
7255 * hold the pmd (ptl) lock, and b) we've just checked that the
7256 * huge pmd (head) page is present in the page tables. The ptl
7257 * prevents the head page and tail pages from being rearranged
7258 * in any way. So this page must be available at this point,
7259 * unless the page refcount overflowed:
7261 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
7266 if (is_hugetlb_entry_migration(pte)) {
7268 __migration_entry_wait_huge(ptep, ptl);
7272 * hwpoisoned entry is treated as no_page_table in
7273 * follow_page_mask().
7281 struct page * __weak
7282 follow_huge_pud(struct mm_struct *mm, unsigned long address,
7283 pud_t *pud, int flags)
7285 struct page *page = NULL;
7289 if (WARN_ON_ONCE(flags & FOLL_PIN))
7293 ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud);
7294 if (!pud_huge(*pud))
7296 pte = huge_ptep_get((pte_t *)pud);
7297 if (pte_present(pte)) {
7298 page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
7299 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
7304 if (is_hugetlb_entry_migration(pte)) {
7306 __migration_entry_wait(mm, (pte_t *)pud, ptl);
7310 * hwpoisoned entry is treated as no_page_table in
7311 * follow_page_mask().
7319 struct page * __weak
7320 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
7322 if (flags & (FOLL_GET | FOLL_PIN))
7325 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
7328 int isolate_hugetlb(struct page *page, struct list_head *list)
7332 spin_lock_irq(&hugetlb_lock);
7333 if (!PageHeadHuge(page) ||
7334 !HPageMigratable(page) ||
7335 !get_page_unless_zero(page)) {
7339 ClearHPageMigratable(page);
7340 list_move_tail(&page->lru, list);
7342 spin_unlock_irq(&hugetlb_lock);
7346 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
7351 spin_lock_irq(&hugetlb_lock);
7352 if (PageHeadHuge(page)) {
7354 if (HPageFreed(page))
7356 else if (HPageMigratable(page))
7357 ret = get_page_unless_zero(page);
7361 spin_unlock_irq(&hugetlb_lock);
7365 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
7369 spin_lock_irq(&hugetlb_lock);
7370 ret = __get_huge_page_for_hwpoison(pfn, flags);
7371 spin_unlock_irq(&hugetlb_lock);
7375 void putback_active_hugepage(struct page *page)
7377 spin_lock_irq(&hugetlb_lock);
7378 SetHPageMigratable(page);
7379 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7380 spin_unlock_irq(&hugetlb_lock);
7384 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
7386 struct hstate *h = page_hstate(oldpage);
7388 hugetlb_cgroup_migrate(oldpage, newpage);
7389 set_page_owner_migrate_reason(newpage, reason);
7392 * transfer temporary state of the new huge page. This is
7393 * reverse to other transitions because the newpage is going to
7394 * be final while the old one will be freed so it takes over
7395 * the temporary status.
7397 * Also note that we have to transfer the per-node surplus state
7398 * here as well otherwise the global surplus count will not match
7401 if (HPageTemporary(newpage)) {
7402 int old_nid = page_to_nid(oldpage);
7403 int new_nid = page_to_nid(newpage);
7405 SetHPageTemporary(oldpage);
7406 ClearHPageTemporary(newpage);
7409 * There is no need to transfer the per-node surplus state
7410 * when we do not cross the node.
7412 if (new_nid == old_nid)
7414 spin_lock_irq(&hugetlb_lock);
7415 if (h->surplus_huge_pages_node[old_nid]) {
7416 h->surplus_huge_pages_node[old_nid]--;
7417 h->surplus_huge_pages_node[new_nid]++;
7419 spin_unlock_irq(&hugetlb_lock);
7423 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7424 unsigned long start,
7427 struct hstate *h = hstate_vma(vma);
7428 unsigned long sz = huge_page_size(h);
7429 struct mm_struct *mm = vma->vm_mm;
7430 struct mmu_notifier_range range;
7431 unsigned long address;
7435 if (!(vma->vm_flags & VM_MAYSHARE))
7441 flush_cache_range(vma, start, end);
7443 * No need to call adjust_range_if_pmd_sharing_possible(), because
7444 * we have already done the PUD_SIZE alignment.
7446 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7448 mmu_notifier_invalidate_range_start(&range);
7449 hugetlb_vma_lock_write(vma);
7450 i_mmap_lock_write(vma->vm_file->f_mapping);
7451 for (address = start; address < end; address += PUD_SIZE) {
7452 ptep = huge_pte_offset(mm, address, sz);
7455 ptl = huge_pte_lock(h, mm, ptep);
7456 huge_pmd_unshare(mm, vma, address, ptep);
7459 flush_hugetlb_tlb_range(vma, start, end);
7460 i_mmap_unlock_write(vma->vm_file->f_mapping);
7461 hugetlb_vma_unlock_write(vma);
7463 * No need to call mmu_notifier_invalidate_range(), see
7464 * Documentation/mm/mmu_notifier.rst.
7466 mmu_notifier_invalidate_range_end(&range);
7470 * This function will unconditionally remove all the shared pmd pgtable entries
7471 * within the specific vma for a hugetlbfs memory range.
7473 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7475 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7476 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7480 static bool cma_reserve_called __initdata;
7482 static int __init cmdline_parse_hugetlb_cma(char *p)
7489 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7492 if (s[count] == ':') {
7493 if (tmp >= MAX_NUMNODES)
7495 nid = array_index_nospec(tmp, MAX_NUMNODES);
7498 tmp = memparse(s, &s);
7499 hugetlb_cma_size_in_node[nid] = tmp;
7500 hugetlb_cma_size += tmp;
7503 * Skip the separator if have one, otherwise
7504 * break the parsing.
7511 hugetlb_cma_size = memparse(p, &p);
7519 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7521 void __init hugetlb_cma_reserve(int order)
7523 unsigned long size, reserved, per_node;
7524 bool node_specific_cma_alloc = false;
7527 cma_reserve_called = true;
7529 if (!hugetlb_cma_size)
7532 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7533 if (hugetlb_cma_size_in_node[nid] == 0)
7536 if (!node_online(nid)) {
7537 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7538 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7539 hugetlb_cma_size_in_node[nid] = 0;
7543 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7544 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7545 nid, (PAGE_SIZE << order) / SZ_1M);
7546 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7547 hugetlb_cma_size_in_node[nid] = 0;
7549 node_specific_cma_alloc = true;
7553 /* Validate the CMA size again in case some invalid nodes specified. */
7554 if (!hugetlb_cma_size)
7557 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7558 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7559 (PAGE_SIZE << order) / SZ_1M);
7560 hugetlb_cma_size = 0;
7564 if (!node_specific_cma_alloc) {
7566 * If 3 GB area is requested on a machine with 4 numa nodes,
7567 * let's allocate 1 GB on first three nodes and ignore the last one.
7569 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7570 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7571 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7575 for_each_online_node(nid) {
7577 char name[CMA_MAX_NAME];
7579 if (node_specific_cma_alloc) {
7580 if (hugetlb_cma_size_in_node[nid] == 0)
7583 size = hugetlb_cma_size_in_node[nid];
7585 size = min(per_node, hugetlb_cma_size - reserved);
7588 size = round_up(size, PAGE_SIZE << order);
7590 snprintf(name, sizeof(name), "hugetlb%d", nid);
7592 * Note that 'order per bit' is based on smallest size that
7593 * may be returned to CMA allocator in the case of
7594 * huge page demotion.
7596 res = cma_declare_contiguous_nid(0, size, 0,
7597 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7599 &hugetlb_cma[nid], nid);
7601 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7607 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7610 if (reserved >= hugetlb_cma_size)
7616 * hugetlb_cma_size is used to determine if allocations from
7617 * cma are possible. Set to zero if no cma regions are set up.
7619 hugetlb_cma_size = 0;
7622 static void __init hugetlb_cma_check(void)
7624 if (!hugetlb_cma_size || cma_reserve_called)
7627 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7630 #endif /* CONFIG_CMA */