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>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 static inline bool PageHugeFreed(struct page *head)
84 return page_private(head + 4) == -1UL;
87 static inline void SetPageHugeFreed(struct page *head)
89 set_page_private(head + 4, -1UL);
92 static inline void ClearPageHugeFreed(struct page *head)
94 set_page_private(head + 4, 0);
97 /* Forward declaration */
98 static int hugetlb_acct_memory(struct hstate *h, long delta);
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)
114 spin_unlock(&spool->lock);
116 /* If no pages are used, and no other handles to the subpool
117 * remain, give up any reservations based on minimum size and
118 * free the subpool */
119 if (subpool_is_free(spool)) {
120 if (spool->min_hpages != -1)
121 hugetlb_acct_memory(spool->hstate,
127 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
130 struct hugepage_subpool *spool;
132 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
136 spin_lock_init(&spool->lock);
138 spool->max_hpages = max_hpages;
140 spool->min_hpages = min_hpages;
142 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
146 spool->rsv_hpages = min_hpages;
151 void hugepage_put_subpool(struct hugepage_subpool *spool)
153 spin_lock(&spool->lock);
154 BUG_ON(!spool->count);
156 unlock_or_release_subpool(spool);
160 * Subpool accounting for allocating and reserving pages.
161 * Return -ENOMEM if there are not enough resources to satisfy the
162 * request. Otherwise, return the number of pages by which the
163 * global pools must be adjusted (upward). The returned value may
164 * only be different than the passed value (delta) in the case where
165 * a subpool minimum size must be maintained.
167 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
175 spin_lock(&spool->lock);
177 if (spool->max_hpages != -1) { /* maximum size accounting */
178 if ((spool->used_hpages + delta) <= spool->max_hpages)
179 spool->used_hpages += delta;
186 /* minimum size accounting */
187 if (spool->min_hpages != -1 && spool->rsv_hpages) {
188 if (delta > spool->rsv_hpages) {
190 * Asking for more reserves than those already taken on
191 * behalf of subpool. Return difference.
193 ret = delta - spool->rsv_hpages;
194 spool->rsv_hpages = 0;
196 ret = 0; /* reserves already accounted for */
197 spool->rsv_hpages -= delta;
202 spin_unlock(&spool->lock);
207 * Subpool accounting for freeing and unreserving pages.
208 * Return the number of global page reservations that must be dropped.
209 * The return value may only be different than the passed value (delta)
210 * in the case where a subpool minimum size must be maintained.
212 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
220 spin_lock(&spool->lock);
222 if (spool->max_hpages != -1) /* maximum size accounting */
223 spool->used_hpages -= delta;
225 /* minimum size accounting */
226 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
227 if (spool->rsv_hpages + delta <= spool->min_hpages)
230 ret = spool->rsv_hpages + delta - spool->min_hpages;
232 spool->rsv_hpages += delta;
233 if (spool->rsv_hpages > spool->min_hpages)
234 spool->rsv_hpages = spool->min_hpages;
238 * If hugetlbfs_put_super couldn't free spool due to an outstanding
239 * quota reference, free it now.
241 unlock_or_release_subpool(spool);
246 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
248 return HUGETLBFS_SB(inode->i_sb)->spool;
251 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
253 return subpool_inode(file_inode(vma->vm_file));
256 /* Helper that removes a struct file_region from the resv_map cache and returns
259 static struct file_region *
260 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
262 struct file_region *nrg = NULL;
264 VM_BUG_ON(resv->region_cache_count <= 0);
266 resv->region_cache_count--;
267 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
268 list_del(&nrg->link);
276 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
277 struct file_region *rg)
279 #ifdef CONFIG_CGROUP_HUGETLB
280 nrg->reservation_counter = rg->reservation_counter;
287 /* Helper that records hugetlb_cgroup uncharge info. */
288 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
290 struct resv_map *resv,
291 struct file_region *nrg)
293 #ifdef CONFIG_CGROUP_HUGETLB
295 nrg->reservation_counter =
296 &h_cg->rsvd_hugepage[hstate_index(h)];
297 nrg->css = &h_cg->css;
298 if (!resv->pages_per_hpage)
299 resv->pages_per_hpage = pages_per_huge_page(h);
300 /* pages_per_hpage should be the same for all entries in
303 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305 nrg->reservation_counter = NULL;
311 static bool has_same_uncharge_info(struct file_region *rg,
312 struct file_region *org)
314 #ifdef CONFIG_CGROUP_HUGETLB
316 rg->reservation_counter == org->reservation_counter &&
324 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
326 struct file_region *nrg = NULL, *prg = NULL;
328 prg = list_prev_entry(rg, link);
329 if (&prg->link != &resv->regions && prg->to == rg->from &&
330 has_same_uncharge_info(prg, rg)) {
339 nrg = list_next_entry(rg, link);
340 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
341 has_same_uncharge_info(nrg, rg)) {
342 nrg->from = rg->from;
350 * Must be called with resv->lock held.
352 * Calling this with regions_needed != NULL will count the number of pages
353 * to be added but will not modify the linked list. And regions_needed will
354 * indicate the number of file_regions needed in the cache to carry out to add
355 * the regions for this range.
357 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
358 struct hugetlb_cgroup *h_cg,
359 struct hstate *h, long *regions_needed)
362 struct list_head *head = &resv->regions;
363 long last_accounted_offset = f;
364 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
369 /* In this loop, we essentially handle an entry for the range
370 * [last_accounted_offset, rg->from), at every iteration, with some
373 list_for_each_entry_safe(rg, trg, head, link) {
374 /* Skip irrelevant regions that start before our range. */
376 /* If this region ends after the last accounted offset,
377 * then we need to update last_accounted_offset.
379 if (rg->to > last_accounted_offset)
380 last_accounted_offset = rg->to;
384 /* When we find a region that starts beyond our range, we've
390 /* Add an entry for last_accounted_offset -> rg->from, and
391 * update last_accounted_offset.
393 if (rg->from > last_accounted_offset) {
394 add += rg->from - last_accounted_offset;
395 if (!regions_needed) {
396 nrg = get_file_region_entry_from_cache(
397 resv, last_accounted_offset, rg->from);
398 record_hugetlb_cgroup_uncharge_info(h_cg, h,
400 list_add(&nrg->link, rg->link.prev);
401 coalesce_file_region(resv, nrg);
403 *regions_needed += 1;
406 last_accounted_offset = rg->to;
409 /* Handle the case where our range extends beyond
410 * last_accounted_offset.
412 if (last_accounted_offset < t) {
413 add += t - last_accounted_offset;
414 if (!regions_needed) {
415 nrg = get_file_region_entry_from_cache(
416 resv, last_accounted_offset, t);
417 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
418 list_add(&nrg->link, rg->link.prev);
419 coalesce_file_region(resv, nrg);
421 *regions_needed += 1;
428 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
430 static int allocate_file_region_entries(struct resv_map *resv,
432 __must_hold(&resv->lock)
434 struct list_head allocated_regions;
435 int to_allocate = 0, i = 0;
436 struct file_region *trg = NULL, *rg = NULL;
438 VM_BUG_ON(regions_needed < 0);
440 INIT_LIST_HEAD(&allocated_regions);
443 * Check for sufficient descriptors in the cache to accommodate
444 * the number of in progress add operations plus regions_needed.
446 * This is a while loop because when we drop the lock, some other call
447 * to region_add or region_del may have consumed some region_entries,
448 * so we keep looping here until we finally have enough entries for
449 * (adds_in_progress + regions_needed).
451 while (resv->region_cache_count <
452 (resv->adds_in_progress + regions_needed)) {
453 to_allocate = resv->adds_in_progress + regions_needed -
454 resv->region_cache_count;
456 /* At this point, we should have enough entries in the cache
457 * for all the existings adds_in_progress. We should only be
458 * needing to allocate for regions_needed.
460 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
462 spin_unlock(&resv->lock);
463 for (i = 0; i < to_allocate; i++) {
464 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
467 list_add(&trg->link, &allocated_regions);
470 spin_lock(&resv->lock);
472 list_splice(&allocated_regions, &resv->region_cache);
473 resv->region_cache_count += to_allocate;
479 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
487 * Add the huge page range represented by [f, t) to the reserve
488 * map. Regions will be taken from the cache to fill in this range.
489 * Sufficient regions should exist in the cache due to the previous
490 * call to region_chg with the same range, but in some cases the cache will not
491 * have sufficient entries due to races with other code doing region_add or
492 * region_del. The extra needed entries will be allocated.
494 * regions_needed is the out value provided by a previous call to region_chg.
496 * Return the number of new huge pages added to the map. This number is greater
497 * than or equal to zero. If file_region entries needed to be allocated for
498 * this operation and we were not able to allocate, it returns -ENOMEM.
499 * region_add of regions of length 1 never allocate file_regions and cannot
500 * fail; region_chg will always allocate at least 1 entry and a region_add for
501 * 1 page will only require at most 1 entry.
503 static long region_add(struct resv_map *resv, long f, long t,
504 long in_regions_needed, struct hstate *h,
505 struct hugetlb_cgroup *h_cg)
507 long add = 0, actual_regions_needed = 0;
509 spin_lock(&resv->lock);
512 /* Count how many regions are actually needed to execute this add. */
513 add_reservation_in_range(resv, f, t, NULL, NULL,
514 &actual_regions_needed);
517 * Check for sufficient descriptors in the cache to accommodate
518 * this add operation. Note that actual_regions_needed may be greater
519 * than in_regions_needed, as the resv_map may have been modified since
520 * the region_chg call. In this case, we need to make sure that we
521 * allocate extra entries, such that we have enough for all the
522 * existing adds_in_progress, plus the excess needed for this
525 if (actual_regions_needed > in_regions_needed &&
526 resv->region_cache_count <
527 resv->adds_in_progress +
528 (actual_regions_needed - in_regions_needed)) {
529 /* region_add operation of range 1 should never need to
530 * allocate file_region entries.
532 VM_BUG_ON(t - f <= 1);
534 if (allocate_file_region_entries(
535 resv, actual_regions_needed - in_regions_needed)) {
542 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
544 resv->adds_in_progress -= in_regions_needed;
546 spin_unlock(&resv->lock);
552 * Examine the existing reserve map and determine how many
553 * huge pages in the specified range [f, t) are NOT currently
554 * represented. This routine is called before a subsequent
555 * call to region_add that will actually modify the reserve
556 * map to add the specified range [f, t). region_chg does
557 * not change the number of huge pages represented by the
558 * map. A number of new file_region structures is added to the cache as a
559 * placeholder, for the subsequent region_add call to use. At least 1
560 * file_region structure is added.
562 * out_regions_needed is the number of regions added to the
563 * resv->adds_in_progress. This value needs to be provided to a follow up call
564 * to region_add or region_abort for proper accounting.
566 * Returns the number of huge pages that need to be added to the existing
567 * reservation map for the range [f, t). This number is greater or equal to
568 * zero. -ENOMEM is returned if a new file_region structure or cache entry
569 * is needed and can not be allocated.
571 static long region_chg(struct resv_map *resv, long f, long t,
572 long *out_regions_needed)
576 spin_lock(&resv->lock);
578 /* Count how many hugepages in this range are NOT represented. */
579 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
582 if (*out_regions_needed == 0)
583 *out_regions_needed = 1;
585 if (allocate_file_region_entries(resv, *out_regions_needed))
588 resv->adds_in_progress += *out_regions_needed;
590 spin_unlock(&resv->lock);
595 * Abort the in progress add operation. The adds_in_progress field
596 * of the resv_map keeps track of the operations in progress between
597 * calls to region_chg and region_add. Operations are sometimes
598 * aborted after the call to region_chg. In such cases, region_abort
599 * is called to decrement the adds_in_progress counter. regions_needed
600 * is the value returned by the region_chg call, it is used to decrement
601 * the adds_in_progress counter.
603 * NOTE: The range arguments [f, t) are not needed or used in this
604 * routine. They are kept to make reading the calling code easier as
605 * arguments will match the associated region_chg call.
607 static void region_abort(struct resv_map *resv, long f, long t,
610 spin_lock(&resv->lock);
611 VM_BUG_ON(!resv->region_cache_count);
612 resv->adds_in_progress -= regions_needed;
613 spin_unlock(&resv->lock);
617 * Delete the specified range [f, t) from the reserve map. If the
618 * t parameter is LONG_MAX, this indicates that ALL regions after f
619 * should be deleted. Locate the regions which intersect [f, t)
620 * and either trim, delete or split the existing regions.
622 * Returns the number of huge pages deleted from the reserve map.
623 * In the normal case, the return value is zero or more. In the
624 * case where a region must be split, a new region descriptor must
625 * be allocated. If the allocation fails, -ENOMEM will be returned.
626 * NOTE: If the parameter t == LONG_MAX, then we will never split
627 * a region and possibly return -ENOMEM. Callers specifying
628 * t == LONG_MAX do not need to check for -ENOMEM error.
630 static long region_del(struct resv_map *resv, long f, long t)
632 struct list_head *head = &resv->regions;
633 struct file_region *rg, *trg;
634 struct file_region *nrg = NULL;
638 spin_lock(&resv->lock);
639 list_for_each_entry_safe(rg, trg, head, link) {
641 * Skip regions before the range to be deleted. file_region
642 * ranges are normally of the form [from, to). However, there
643 * may be a "placeholder" entry in the map which is of the form
644 * (from, to) with from == to. Check for placeholder entries
645 * at the beginning of the range to be deleted.
647 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
653 if (f > rg->from && t < rg->to) { /* Must split region */
655 * Check for an entry in the cache before dropping
656 * lock and attempting allocation.
659 resv->region_cache_count > resv->adds_in_progress) {
660 nrg = list_first_entry(&resv->region_cache,
663 list_del(&nrg->link);
664 resv->region_cache_count--;
668 spin_unlock(&resv->lock);
669 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
676 hugetlb_cgroup_uncharge_file_region(
679 /* New entry for end of split region */
683 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
685 INIT_LIST_HEAD(&nrg->link);
687 /* Original entry is trimmed */
690 list_add(&nrg->link, &rg->link);
695 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
696 del += rg->to - rg->from;
697 hugetlb_cgroup_uncharge_file_region(resv, rg,
704 if (f <= rg->from) { /* Trim beginning of region */
705 hugetlb_cgroup_uncharge_file_region(resv, rg,
710 } else { /* Trim end of region */
711 hugetlb_cgroup_uncharge_file_region(resv, rg,
719 spin_unlock(&resv->lock);
725 * A rare out of memory error was encountered which prevented removal of
726 * the reserve map region for a page. The huge page itself was free'ed
727 * and removed from the page cache. This routine will adjust the subpool
728 * usage count, and the global reserve count if needed. By incrementing
729 * these counts, the reserve map entry which could not be deleted will
730 * appear as a "reserved" entry instead of simply dangling with incorrect
733 void hugetlb_fix_reserve_counts(struct inode *inode)
735 struct hugepage_subpool *spool = subpool_inode(inode);
738 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
740 struct hstate *h = hstate_inode(inode);
742 hugetlb_acct_memory(h, 1);
747 * Count and return the number of huge pages in the reserve map
748 * that intersect with the range [f, t).
750 static long region_count(struct resv_map *resv, long f, long t)
752 struct list_head *head = &resv->regions;
753 struct file_region *rg;
756 spin_lock(&resv->lock);
757 /* Locate each segment we overlap with, and count that overlap. */
758 list_for_each_entry(rg, head, link) {
767 seg_from = max(rg->from, f);
768 seg_to = min(rg->to, t);
770 chg += seg_to - seg_from;
772 spin_unlock(&resv->lock);
778 * Convert the address within this vma to the page offset within
779 * the mapping, in pagecache page units; huge pages here.
781 static pgoff_t vma_hugecache_offset(struct hstate *h,
782 struct vm_area_struct *vma, unsigned long address)
784 return ((address - vma->vm_start) >> huge_page_shift(h)) +
785 (vma->vm_pgoff >> huge_page_order(h));
788 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
789 unsigned long address)
791 return vma_hugecache_offset(hstate_vma(vma), vma, address);
793 EXPORT_SYMBOL_GPL(linear_hugepage_index);
796 * Return the size of the pages allocated when backing a VMA. In the majority
797 * cases this will be same size as used by the page table entries.
799 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
801 if (vma->vm_ops && vma->vm_ops->pagesize)
802 return vma->vm_ops->pagesize(vma);
805 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
808 * Return the page size being used by the MMU to back a VMA. In the majority
809 * of cases, the page size used by the kernel matches the MMU size. On
810 * architectures where it differs, an architecture-specific 'strong'
811 * version of this symbol is required.
813 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
815 return vma_kernel_pagesize(vma);
819 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
820 * bits of the reservation map pointer, which are always clear due to
823 #define HPAGE_RESV_OWNER (1UL << 0)
824 #define HPAGE_RESV_UNMAPPED (1UL << 1)
825 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
828 * These helpers are used to track how many pages are reserved for
829 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
830 * is guaranteed to have their future faults succeed.
832 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
833 * the reserve counters are updated with the hugetlb_lock held. It is safe
834 * to reset the VMA at fork() time as it is not in use yet and there is no
835 * chance of the global counters getting corrupted as a result of the values.
837 * The private mapping reservation is represented in a subtly different
838 * manner to a shared mapping. A shared mapping has a region map associated
839 * with the underlying file, this region map represents the backing file
840 * pages which have ever had a reservation assigned which this persists even
841 * after the page is instantiated. A private mapping has a region map
842 * associated with the original mmap which is attached to all VMAs which
843 * reference it, this region map represents those offsets which have consumed
844 * reservation ie. where pages have been instantiated.
846 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
848 return (unsigned long)vma->vm_private_data;
851 static void set_vma_private_data(struct vm_area_struct *vma,
854 vma->vm_private_data = (void *)value;
858 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
859 struct hugetlb_cgroup *h_cg,
862 #ifdef CONFIG_CGROUP_HUGETLB
864 resv_map->reservation_counter = NULL;
865 resv_map->pages_per_hpage = 0;
866 resv_map->css = NULL;
868 resv_map->reservation_counter =
869 &h_cg->rsvd_hugepage[hstate_index(h)];
870 resv_map->pages_per_hpage = pages_per_huge_page(h);
871 resv_map->css = &h_cg->css;
876 struct resv_map *resv_map_alloc(void)
878 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
879 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
881 if (!resv_map || !rg) {
887 kref_init(&resv_map->refs);
888 spin_lock_init(&resv_map->lock);
889 INIT_LIST_HEAD(&resv_map->regions);
891 resv_map->adds_in_progress = 0;
893 * Initialize these to 0. On shared mappings, 0's here indicate these
894 * fields don't do cgroup accounting. On private mappings, these will be
895 * re-initialized to the proper values, to indicate that hugetlb cgroup
896 * reservations are to be un-charged from here.
898 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
900 INIT_LIST_HEAD(&resv_map->region_cache);
901 list_add(&rg->link, &resv_map->region_cache);
902 resv_map->region_cache_count = 1;
907 void resv_map_release(struct kref *ref)
909 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
910 struct list_head *head = &resv_map->region_cache;
911 struct file_region *rg, *trg;
913 /* Clear out any active regions before we release the map. */
914 region_del(resv_map, 0, LONG_MAX);
916 /* ... and any entries left in the cache */
917 list_for_each_entry_safe(rg, trg, head, link) {
922 VM_BUG_ON(resv_map->adds_in_progress);
927 static inline struct resv_map *inode_resv_map(struct inode *inode)
930 * At inode evict time, i_mapping may not point to the original
931 * address space within the inode. This original address space
932 * contains the pointer to the resv_map. So, always use the
933 * address space embedded within the inode.
934 * The VERY common case is inode->mapping == &inode->i_data but,
935 * this may not be true for device special inodes.
937 return (struct resv_map *)(&inode->i_data)->private_data;
940 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
942 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
943 if (vma->vm_flags & VM_MAYSHARE) {
944 struct address_space *mapping = vma->vm_file->f_mapping;
945 struct inode *inode = mapping->host;
947 return inode_resv_map(inode);
950 return (struct resv_map *)(get_vma_private_data(vma) &
955 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
957 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
960 set_vma_private_data(vma, (get_vma_private_data(vma) &
961 HPAGE_RESV_MASK) | (unsigned long)map);
964 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
966 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
967 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
969 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
972 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
974 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
976 return (get_vma_private_data(vma) & flag) != 0;
979 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
980 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
982 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
983 if (!(vma->vm_flags & VM_MAYSHARE))
984 vma->vm_private_data = (void *)0;
987 /* Returns true if the VMA has associated reserve pages */
988 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
990 if (vma->vm_flags & VM_NORESERVE) {
992 * This address is already reserved by other process(chg == 0),
993 * so, we should decrement reserved count. Without decrementing,
994 * reserve count remains after releasing inode, because this
995 * allocated page will go into page cache and is regarded as
996 * coming from reserved pool in releasing step. Currently, we
997 * don't have any other solution to deal with this situation
998 * properly, so add work-around here.
1000 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1006 /* Shared mappings always use reserves */
1007 if (vma->vm_flags & VM_MAYSHARE) {
1009 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1010 * be a region map for all pages. The only situation where
1011 * there is no region map is if a hole was punched via
1012 * fallocate. In this case, there really are no reserves to
1013 * use. This situation is indicated if chg != 0.
1022 * Only the process that called mmap() has reserves for
1025 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1027 * Like the shared case above, a hole punch or truncate
1028 * could have been performed on the private mapping.
1029 * Examine the value of chg to determine if reserves
1030 * actually exist or were previously consumed.
1031 * Very Subtle - The value of chg comes from a previous
1032 * call to vma_needs_reserves(). The reserve map for
1033 * private mappings has different (opposite) semantics
1034 * than that of shared mappings. vma_needs_reserves()
1035 * has already taken this difference in semantics into
1036 * account. Therefore, the meaning of chg is the same
1037 * as in the shared case above. Code could easily be
1038 * combined, but keeping it separate draws attention to
1039 * subtle differences.
1050 static void enqueue_huge_page(struct hstate *h, struct page *page)
1052 int nid = page_to_nid(page);
1053 list_move(&page->lru, &h->hugepage_freelists[nid]);
1054 h->free_huge_pages++;
1055 h->free_huge_pages_node[nid]++;
1056 SetPageHugeFreed(page);
1059 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1062 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1064 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1065 if (nocma && is_migrate_cma_page(page))
1068 if (PageHWPoison(page))
1071 list_move(&page->lru, &h->hugepage_activelist);
1072 set_page_refcounted(page);
1073 ClearPageHugeFreed(page);
1074 h->free_huge_pages--;
1075 h->free_huge_pages_node[nid]--;
1082 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1085 unsigned int cpuset_mems_cookie;
1086 struct zonelist *zonelist;
1089 int node = NUMA_NO_NODE;
1091 zonelist = node_zonelist(nid, gfp_mask);
1094 cpuset_mems_cookie = read_mems_allowed_begin();
1095 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1098 if (!cpuset_zone_allowed(zone, gfp_mask))
1101 * no need to ask again on the same node. Pool is node rather than
1104 if (zone_to_nid(zone) == node)
1106 node = zone_to_nid(zone);
1108 page = dequeue_huge_page_node_exact(h, node);
1112 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1118 static struct page *dequeue_huge_page_vma(struct hstate *h,
1119 struct vm_area_struct *vma,
1120 unsigned long address, int avoid_reserve,
1124 struct mempolicy *mpol;
1126 nodemask_t *nodemask;
1130 * A child process with MAP_PRIVATE mappings created by their parent
1131 * have no page reserves. This check ensures that reservations are
1132 * not "stolen". The child may still get SIGKILLed
1134 if (!vma_has_reserves(vma, chg) &&
1135 h->free_huge_pages - h->resv_huge_pages == 0)
1138 /* If reserves cannot be used, ensure enough pages are in the pool */
1139 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1142 gfp_mask = htlb_alloc_mask(h);
1143 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1144 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1145 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1146 SetPagePrivate(page);
1147 h->resv_huge_pages--;
1150 mpol_cond_put(mpol);
1158 * common helper functions for hstate_next_node_to_{alloc|free}.
1159 * We may have allocated or freed a huge page based on a different
1160 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1161 * be outside of *nodes_allowed. Ensure that we use an allowed
1162 * node for alloc or free.
1164 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1166 nid = next_node_in(nid, *nodes_allowed);
1167 VM_BUG_ON(nid >= MAX_NUMNODES);
1172 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1174 if (!node_isset(nid, *nodes_allowed))
1175 nid = next_node_allowed(nid, nodes_allowed);
1180 * returns the previously saved node ["this node"] from which to
1181 * allocate a persistent huge page for the pool and advance the
1182 * next node from which to allocate, handling wrap at end of node
1185 static int hstate_next_node_to_alloc(struct hstate *h,
1186 nodemask_t *nodes_allowed)
1190 VM_BUG_ON(!nodes_allowed);
1192 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1193 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1199 * helper for free_pool_huge_page() - return the previously saved
1200 * node ["this node"] from which to free a huge page. Advance the
1201 * next node id whether or not we find a free huge page to free so
1202 * that the next attempt to free addresses the next node.
1204 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1208 VM_BUG_ON(!nodes_allowed);
1210 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1211 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1216 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1217 for (nr_nodes = nodes_weight(*mask); \
1219 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1222 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1223 for (nr_nodes = nodes_weight(*mask); \
1225 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1228 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1229 static void destroy_compound_gigantic_page(struct page *page,
1233 int nr_pages = 1 << order;
1234 struct page *p = page + 1;
1236 atomic_set(compound_mapcount_ptr(page), 0);
1237 atomic_set(compound_pincount_ptr(page), 0);
1239 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1240 clear_compound_head(p);
1241 set_page_refcounted(p);
1244 set_compound_order(page, 0);
1245 page[1].compound_nr = 0;
1246 __ClearPageHead(page);
1249 static void free_gigantic_page(struct page *page, unsigned int order)
1252 * If the page isn't allocated using the cma allocator,
1253 * cma_release() returns false.
1256 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1260 free_contig_range(page_to_pfn(page), 1 << order);
1263 #ifdef CONFIG_CONTIG_ALLOC
1264 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1265 int nid, nodemask_t *nodemask)
1267 unsigned long nr_pages = 1UL << huge_page_order(h);
1268 if (nid == NUMA_NO_NODE)
1269 nid = numa_mem_id();
1276 if (hugetlb_cma[nid]) {
1277 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1278 huge_page_order(h), true);
1283 if (!(gfp_mask & __GFP_THISNODE)) {
1284 for_each_node_mask(node, *nodemask) {
1285 if (node == nid || !hugetlb_cma[node])
1288 page = cma_alloc(hugetlb_cma[node], nr_pages,
1289 huge_page_order(h), true);
1297 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1300 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1301 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1302 #else /* !CONFIG_CONTIG_ALLOC */
1303 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1304 int nid, nodemask_t *nodemask)
1308 #endif /* CONFIG_CONTIG_ALLOC */
1310 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1311 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1312 int nid, nodemask_t *nodemask)
1316 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1317 static inline void destroy_compound_gigantic_page(struct page *page,
1318 unsigned int order) { }
1321 static void update_and_free_page(struct hstate *h, struct page *page)
1325 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1329 h->nr_huge_pages_node[page_to_nid(page)]--;
1330 for (i = 0; i < pages_per_huge_page(h); i++) {
1331 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1332 1 << PG_referenced | 1 << PG_dirty |
1333 1 << PG_active | 1 << PG_private |
1336 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1337 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1338 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1339 set_page_refcounted(page);
1340 if (hstate_is_gigantic(h)) {
1342 * Temporarily drop the hugetlb_lock, because
1343 * we might block in free_gigantic_page().
1345 spin_unlock(&hugetlb_lock);
1346 destroy_compound_gigantic_page(page, huge_page_order(h));
1347 free_gigantic_page(page, huge_page_order(h));
1348 spin_lock(&hugetlb_lock);
1350 __free_pages(page, huge_page_order(h));
1354 struct hstate *size_to_hstate(unsigned long size)
1358 for_each_hstate(h) {
1359 if (huge_page_size(h) == size)
1366 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1367 * to hstate->hugepage_activelist.)
1369 * This function can be called for tail pages, but never returns true for them.
1371 bool page_huge_active(struct page *page)
1373 return PageHeadHuge(page) && PagePrivate(&page[1]);
1376 /* never called for tail page */
1377 void set_page_huge_active(struct page *page)
1379 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1380 SetPagePrivate(&page[1]);
1383 static void clear_page_huge_active(struct page *page)
1385 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1386 ClearPagePrivate(&page[1]);
1390 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1393 static inline bool PageHugeTemporary(struct page *page)
1395 if (!PageHuge(page))
1398 return (unsigned long)page[2].mapping == -1U;
1401 static inline void SetPageHugeTemporary(struct page *page)
1403 page[2].mapping = (void *)-1U;
1406 static inline void ClearPageHugeTemporary(struct page *page)
1408 page[2].mapping = NULL;
1411 static void __free_huge_page(struct page *page)
1414 * Can't pass hstate in here because it is called from the
1415 * compound page destructor.
1417 struct hstate *h = page_hstate(page);
1418 int nid = page_to_nid(page);
1419 struct hugepage_subpool *spool =
1420 (struct hugepage_subpool *)page_private(page);
1421 bool restore_reserve;
1423 VM_BUG_ON_PAGE(page_count(page), page);
1424 VM_BUG_ON_PAGE(page_mapcount(page), page);
1426 set_page_private(page, 0);
1427 page->mapping = NULL;
1428 restore_reserve = PagePrivate(page);
1429 ClearPagePrivate(page);
1432 * If PagePrivate() was set on page, page allocation consumed a
1433 * reservation. If the page was associated with a subpool, there
1434 * would have been a page reserved in the subpool before allocation
1435 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1436 * reservation, do not call hugepage_subpool_put_pages() as this will
1437 * remove the reserved page from the subpool.
1439 if (!restore_reserve) {
1441 * A return code of zero implies that the subpool will be
1442 * under its minimum size if the reservation is not restored
1443 * after page is free. Therefore, force restore_reserve
1446 if (hugepage_subpool_put_pages(spool, 1) == 0)
1447 restore_reserve = true;
1450 spin_lock(&hugetlb_lock);
1451 clear_page_huge_active(page);
1452 hugetlb_cgroup_uncharge_page(hstate_index(h),
1453 pages_per_huge_page(h), page);
1454 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1455 pages_per_huge_page(h), page);
1456 if (restore_reserve)
1457 h->resv_huge_pages++;
1459 if (PageHugeTemporary(page)) {
1460 list_del(&page->lru);
1461 ClearPageHugeTemporary(page);
1462 update_and_free_page(h, page);
1463 } else if (h->surplus_huge_pages_node[nid]) {
1464 /* remove the page from active list */
1465 list_del(&page->lru);
1466 update_and_free_page(h, page);
1467 h->surplus_huge_pages--;
1468 h->surplus_huge_pages_node[nid]--;
1470 arch_clear_hugepage_flags(page);
1471 enqueue_huge_page(h, page);
1473 spin_unlock(&hugetlb_lock);
1477 * As free_huge_page() can be called from a non-task context, we have
1478 * to defer the actual freeing in a workqueue to prevent potential
1479 * hugetlb_lock deadlock.
1481 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1482 * be freed and frees them one-by-one. As the page->mapping pointer is
1483 * going to be cleared in __free_huge_page() anyway, it is reused as the
1484 * llist_node structure of a lockless linked list of huge pages to be freed.
1486 static LLIST_HEAD(hpage_freelist);
1488 static void free_hpage_workfn(struct work_struct *work)
1490 struct llist_node *node;
1493 node = llist_del_all(&hpage_freelist);
1496 page = container_of((struct address_space **)node,
1497 struct page, mapping);
1499 __free_huge_page(page);
1502 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1504 void free_huge_page(struct page *page)
1507 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1511 * Only call schedule_work() if hpage_freelist is previously
1512 * empty. Otherwise, schedule_work() had been called but the
1513 * workfn hasn't retrieved the list yet.
1515 if (llist_add((struct llist_node *)&page->mapping,
1517 schedule_work(&free_hpage_work);
1521 __free_huge_page(page);
1524 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1526 INIT_LIST_HEAD(&page->lru);
1527 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1528 set_hugetlb_cgroup(page, NULL);
1529 set_hugetlb_cgroup_rsvd(page, NULL);
1530 spin_lock(&hugetlb_lock);
1532 h->nr_huge_pages_node[nid]++;
1533 ClearPageHugeFreed(page);
1534 spin_unlock(&hugetlb_lock);
1537 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1540 int nr_pages = 1 << order;
1541 struct page *p = page + 1;
1543 /* we rely on prep_new_huge_page to set the destructor */
1544 set_compound_order(page, order);
1545 __ClearPageReserved(page);
1546 __SetPageHead(page);
1547 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1549 * For gigantic hugepages allocated through bootmem at
1550 * boot, it's safer to be consistent with the not-gigantic
1551 * hugepages and clear the PG_reserved bit from all tail pages
1552 * too. Otherwise drivers using get_user_pages() to access tail
1553 * pages may get the reference counting wrong if they see
1554 * PG_reserved set on a tail page (despite the head page not
1555 * having PG_reserved set). Enforcing this consistency between
1556 * head and tail pages allows drivers to optimize away a check
1557 * on the head page when they need know if put_page() is needed
1558 * after get_user_pages().
1560 __ClearPageReserved(p);
1561 set_page_count(p, 0);
1562 set_compound_head(p, page);
1564 atomic_set(compound_mapcount_ptr(page), -1);
1565 atomic_set(compound_pincount_ptr(page), 0);
1569 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1570 * transparent huge pages. See the PageTransHuge() documentation for more
1573 int PageHuge(struct page *page)
1575 if (!PageCompound(page))
1578 page = compound_head(page);
1579 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1581 EXPORT_SYMBOL_GPL(PageHuge);
1584 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1585 * normal or transparent huge pages.
1587 int PageHeadHuge(struct page *page_head)
1589 if (!PageHead(page_head))
1592 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1596 * Find and lock address space (mapping) in write mode.
1598 * Upon entry, the page is locked which means that page_mapping() is
1599 * stable. Due to locking order, we can only trylock_write. If we can
1600 * not get the lock, simply return NULL to caller.
1602 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1604 struct address_space *mapping = page_mapping(hpage);
1609 if (i_mmap_trylock_write(mapping))
1615 pgoff_t __basepage_index(struct page *page)
1617 struct page *page_head = compound_head(page);
1618 pgoff_t index = page_index(page_head);
1619 unsigned long compound_idx;
1621 if (!PageHuge(page_head))
1622 return page_index(page);
1624 if (compound_order(page_head) >= MAX_ORDER)
1625 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1627 compound_idx = page - page_head;
1629 return (index << compound_order(page_head)) + compound_idx;
1632 static struct page *alloc_buddy_huge_page(struct hstate *h,
1633 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1634 nodemask_t *node_alloc_noretry)
1636 int order = huge_page_order(h);
1638 bool alloc_try_hard = true;
1641 * By default we always try hard to allocate the page with
1642 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1643 * a loop (to adjust global huge page counts) and previous allocation
1644 * failed, do not continue to try hard on the same node. Use the
1645 * node_alloc_noretry bitmap to manage this state information.
1647 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1648 alloc_try_hard = false;
1649 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1651 gfp_mask |= __GFP_RETRY_MAYFAIL;
1652 if (nid == NUMA_NO_NODE)
1653 nid = numa_mem_id();
1654 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1656 __count_vm_event(HTLB_BUDDY_PGALLOC);
1658 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1661 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1662 * indicates an overall state change. Clear bit so that we resume
1663 * normal 'try hard' allocations.
1665 if (node_alloc_noretry && page && !alloc_try_hard)
1666 node_clear(nid, *node_alloc_noretry);
1669 * If we tried hard to get a page but failed, set bit so that
1670 * subsequent attempts will not try as hard until there is an
1671 * overall state change.
1673 if (node_alloc_noretry && !page && alloc_try_hard)
1674 node_set(nid, *node_alloc_noretry);
1680 * Common helper to allocate a fresh hugetlb page. All specific allocators
1681 * should use this function to get new hugetlb pages
1683 static struct page *alloc_fresh_huge_page(struct hstate *h,
1684 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1685 nodemask_t *node_alloc_noretry)
1689 if (hstate_is_gigantic(h))
1690 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1692 page = alloc_buddy_huge_page(h, gfp_mask,
1693 nid, nmask, node_alloc_noretry);
1697 if (hstate_is_gigantic(h))
1698 prep_compound_gigantic_page(page, huge_page_order(h));
1699 prep_new_huge_page(h, page, page_to_nid(page));
1705 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1708 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1709 nodemask_t *node_alloc_noretry)
1713 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1715 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1716 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1717 node_alloc_noretry);
1725 put_page(page); /* free it into the hugepage allocator */
1731 * Free huge page from pool from next node to free.
1732 * Attempt to keep persistent huge pages more or less
1733 * balanced over allowed nodes.
1734 * Called with hugetlb_lock locked.
1736 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1742 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1744 * If we're returning unused surplus pages, only examine
1745 * nodes with surplus pages.
1747 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1748 !list_empty(&h->hugepage_freelists[node])) {
1750 list_entry(h->hugepage_freelists[node].next,
1752 list_del(&page->lru);
1753 h->free_huge_pages--;
1754 h->free_huge_pages_node[node]--;
1756 h->surplus_huge_pages--;
1757 h->surplus_huge_pages_node[node]--;
1759 update_and_free_page(h, page);
1769 * Dissolve a given free hugepage into free buddy pages. This function does
1770 * nothing for in-use hugepages and non-hugepages.
1771 * This function returns values like below:
1773 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1774 * (allocated or reserved.)
1775 * 0: successfully dissolved free hugepages or the page is not a
1776 * hugepage (considered as already dissolved)
1778 int dissolve_free_huge_page(struct page *page)
1783 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1784 if (!PageHuge(page))
1787 spin_lock(&hugetlb_lock);
1788 if (!PageHuge(page)) {
1793 if (!page_count(page)) {
1794 struct page *head = compound_head(page);
1795 struct hstate *h = page_hstate(head);
1796 int nid = page_to_nid(head);
1797 if (h->free_huge_pages - h->resv_huge_pages == 0)
1801 * We should make sure that the page is already on the free list
1802 * when it is dissolved.
1804 if (unlikely(!PageHugeFreed(head))) {
1805 spin_unlock(&hugetlb_lock);
1809 * Theoretically, we should return -EBUSY when we
1810 * encounter this race. In fact, we have a chance
1811 * to successfully dissolve the page if we do a
1812 * retry. Because the race window is quite small.
1813 * If we seize this opportunity, it is an optimization
1814 * for increasing the success rate of dissolving page.
1820 * Move PageHWPoison flag from head page to the raw error page,
1821 * which makes any subpages rather than the error page reusable.
1823 if (PageHWPoison(head) && page != head) {
1824 SetPageHWPoison(page);
1825 ClearPageHWPoison(head);
1827 list_del(&head->lru);
1828 h->free_huge_pages--;
1829 h->free_huge_pages_node[nid]--;
1830 h->max_huge_pages--;
1831 update_and_free_page(h, head);
1835 spin_unlock(&hugetlb_lock);
1840 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1841 * make specified memory blocks removable from the system.
1842 * Note that this will dissolve a free gigantic hugepage completely, if any
1843 * part of it lies within the given range.
1844 * Also note that if dissolve_free_huge_page() returns with an error, all
1845 * free hugepages that were dissolved before that error are lost.
1847 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1853 if (!hugepages_supported())
1856 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1857 page = pfn_to_page(pfn);
1858 rc = dissolve_free_huge_page(page);
1867 * Allocates a fresh surplus page from the page allocator.
1869 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1870 int nid, nodemask_t *nmask)
1872 struct page *page = NULL;
1874 if (hstate_is_gigantic(h))
1877 spin_lock(&hugetlb_lock);
1878 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1880 spin_unlock(&hugetlb_lock);
1882 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1886 spin_lock(&hugetlb_lock);
1888 * We could have raced with the pool size change.
1889 * Double check that and simply deallocate the new page
1890 * if we would end up overcommiting the surpluses. Abuse
1891 * temporary page to workaround the nasty free_huge_page
1894 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1895 SetPageHugeTemporary(page);
1896 spin_unlock(&hugetlb_lock);
1900 h->surplus_huge_pages++;
1901 h->surplus_huge_pages_node[page_to_nid(page)]++;
1905 spin_unlock(&hugetlb_lock);
1910 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1911 int nid, nodemask_t *nmask)
1915 if (hstate_is_gigantic(h))
1918 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1923 * We do not account these pages as surplus because they are only
1924 * temporary and will be released properly on the last reference
1926 SetPageHugeTemporary(page);
1932 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1935 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1936 struct vm_area_struct *vma, unsigned long addr)
1939 struct mempolicy *mpol;
1940 gfp_t gfp_mask = htlb_alloc_mask(h);
1942 nodemask_t *nodemask;
1944 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1945 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1946 mpol_cond_put(mpol);
1951 /* page migration callback function */
1952 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1953 nodemask_t *nmask, gfp_t gfp_mask)
1955 spin_lock(&hugetlb_lock);
1956 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1959 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1961 spin_unlock(&hugetlb_lock);
1965 spin_unlock(&hugetlb_lock);
1967 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1970 /* mempolicy aware migration callback */
1971 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1972 unsigned long address)
1974 struct mempolicy *mpol;
1975 nodemask_t *nodemask;
1980 gfp_mask = htlb_alloc_mask(h);
1981 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1982 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1983 mpol_cond_put(mpol);
1989 * Increase the hugetlb pool such that it can accommodate a reservation
1992 static int gather_surplus_pages(struct hstate *h, long delta)
1993 __must_hold(&hugetlb_lock)
1995 struct list_head surplus_list;
1996 struct page *page, *tmp;
1999 long needed, allocated;
2000 bool alloc_ok = true;
2002 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2004 h->resv_huge_pages += delta;
2009 INIT_LIST_HEAD(&surplus_list);
2013 spin_unlock(&hugetlb_lock);
2014 for (i = 0; i < needed; i++) {
2015 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2016 NUMA_NO_NODE, NULL);
2021 list_add(&page->lru, &surplus_list);
2027 * After retaking hugetlb_lock, we need to recalculate 'needed'
2028 * because either resv_huge_pages or free_huge_pages may have changed.
2030 spin_lock(&hugetlb_lock);
2031 needed = (h->resv_huge_pages + delta) -
2032 (h->free_huge_pages + allocated);
2037 * We were not able to allocate enough pages to
2038 * satisfy the entire reservation so we free what
2039 * we've allocated so far.
2044 * The surplus_list now contains _at_least_ the number of extra pages
2045 * needed to accommodate the reservation. Add the appropriate number
2046 * of pages to the hugetlb pool and free the extras back to the buddy
2047 * allocator. Commit the entire reservation here to prevent another
2048 * process from stealing the pages as they are added to the pool but
2049 * before they are reserved.
2051 needed += allocated;
2052 h->resv_huge_pages += delta;
2055 /* Free the needed pages to the hugetlb pool */
2056 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2062 * This page is now managed by the hugetlb allocator and has
2063 * no users -- drop the buddy allocator's reference.
2065 zeroed = put_page_testzero(page);
2066 VM_BUG_ON_PAGE(!zeroed, page);
2067 enqueue_huge_page(h, page);
2070 spin_unlock(&hugetlb_lock);
2072 /* Free unnecessary surplus pages to the buddy allocator */
2073 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2075 spin_lock(&hugetlb_lock);
2081 * This routine has two main purposes:
2082 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2083 * in unused_resv_pages. This corresponds to the prior adjustments made
2084 * to the associated reservation map.
2085 * 2) Free any unused surplus pages that may have been allocated to satisfy
2086 * the reservation. As many as unused_resv_pages may be freed.
2088 * Called with hugetlb_lock held. However, the lock could be dropped (and
2089 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2090 * we must make sure nobody else can claim pages we are in the process of
2091 * freeing. Do this by ensuring resv_huge_page always is greater than the
2092 * number of huge pages we plan to free when dropping the lock.
2094 static void return_unused_surplus_pages(struct hstate *h,
2095 unsigned long unused_resv_pages)
2097 unsigned long nr_pages;
2099 /* Cannot return gigantic pages currently */
2100 if (hstate_is_gigantic(h))
2104 * Part (or even all) of the reservation could have been backed
2105 * by pre-allocated pages. Only free surplus pages.
2107 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2110 * We want to release as many surplus pages as possible, spread
2111 * evenly across all nodes with memory. Iterate across these nodes
2112 * until we can no longer free unreserved surplus pages. This occurs
2113 * when the nodes with surplus pages have no free pages.
2114 * free_pool_huge_page() will balance the freed pages across the
2115 * on-line nodes with memory and will handle the hstate accounting.
2117 * Note that we decrement resv_huge_pages as we free the pages. If
2118 * we drop the lock, resv_huge_pages will still be sufficiently large
2119 * to cover subsequent pages we may free.
2121 while (nr_pages--) {
2122 h->resv_huge_pages--;
2123 unused_resv_pages--;
2124 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2126 cond_resched_lock(&hugetlb_lock);
2130 /* Fully uncommit the reservation */
2131 h->resv_huge_pages -= unused_resv_pages;
2136 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2137 * are used by the huge page allocation routines to manage reservations.
2139 * vma_needs_reservation is called to determine if the huge page at addr
2140 * within the vma has an associated reservation. If a reservation is
2141 * needed, the value 1 is returned. The caller is then responsible for
2142 * managing the global reservation and subpool usage counts. After
2143 * the huge page has been allocated, vma_commit_reservation is called
2144 * to add the page to the reservation map. If the page allocation fails,
2145 * the reservation must be ended instead of committed. vma_end_reservation
2146 * is called in such cases.
2148 * In the normal case, vma_commit_reservation returns the same value
2149 * as the preceding vma_needs_reservation call. The only time this
2150 * is not the case is if a reserve map was changed between calls. It
2151 * is the responsibility of the caller to notice the difference and
2152 * take appropriate action.
2154 * vma_add_reservation is used in error paths where a reservation must
2155 * be restored when a newly allocated huge page must be freed. It is
2156 * to be called after calling vma_needs_reservation to determine if a
2157 * reservation exists.
2159 enum vma_resv_mode {
2165 static long __vma_reservation_common(struct hstate *h,
2166 struct vm_area_struct *vma, unsigned long addr,
2167 enum vma_resv_mode mode)
2169 struct resv_map *resv;
2172 long dummy_out_regions_needed;
2174 resv = vma_resv_map(vma);
2178 idx = vma_hugecache_offset(h, vma, addr);
2180 case VMA_NEEDS_RESV:
2181 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2182 /* We assume that vma_reservation_* routines always operate on
2183 * 1 page, and that adding to resv map a 1 page entry can only
2184 * ever require 1 region.
2186 VM_BUG_ON(dummy_out_regions_needed != 1);
2188 case VMA_COMMIT_RESV:
2189 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2190 /* region_add calls of range 1 should never fail. */
2194 region_abort(resv, idx, idx + 1, 1);
2198 if (vma->vm_flags & VM_MAYSHARE) {
2199 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2200 /* region_add calls of range 1 should never fail. */
2203 region_abort(resv, idx, idx + 1, 1);
2204 ret = region_del(resv, idx, idx + 1);
2211 if (vma->vm_flags & VM_MAYSHARE)
2213 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2215 * In most cases, reserves always exist for private mappings.
2216 * However, a file associated with mapping could have been
2217 * hole punched or truncated after reserves were consumed.
2218 * As subsequent fault on such a range will not use reserves.
2219 * Subtle - The reserve map for private mappings has the
2220 * opposite meaning than that of shared mappings. If NO
2221 * entry is in the reserve map, it means a reservation exists.
2222 * If an entry exists in the reserve map, it means the
2223 * reservation has already been consumed. As a result, the
2224 * return value of this routine is the opposite of the
2225 * value returned from reserve map manipulation routines above.
2233 return ret < 0 ? ret : 0;
2236 static long vma_needs_reservation(struct hstate *h,
2237 struct vm_area_struct *vma, unsigned long addr)
2239 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2242 static long vma_commit_reservation(struct hstate *h,
2243 struct vm_area_struct *vma, unsigned long addr)
2245 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2248 static void vma_end_reservation(struct hstate *h,
2249 struct vm_area_struct *vma, unsigned long addr)
2251 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2254 static long vma_add_reservation(struct hstate *h,
2255 struct vm_area_struct *vma, unsigned long addr)
2257 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2261 * This routine is called to restore a reservation on error paths. In the
2262 * specific error paths, a huge page was allocated (via alloc_huge_page)
2263 * and is about to be freed. If a reservation for the page existed,
2264 * alloc_huge_page would have consumed the reservation and set PagePrivate
2265 * in the newly allocated page. When the page is freed via free_huge_page,
2266 * the global reservation count will be incremented if PagePrivate is set.
2267 * However, free_huge_page can not adjust the reserve map. Adjust the
2268 * reserve map here to be consistent with global reserve count adjustments
2269 * to be made by free_huge_page.
2271 static void restore_reserve_on_error(struct hstate *h,
2272 struct vm_area_struct *vma, unsigned long address,
2275 if (unlikely(PagePrivate(page))) {
2276 long rc = vma_needs_reservation(h, vma, address);
2278 if (unlikely(rc < 0)) {
2280 * Rare out of memory condition in reserve map
2281 * manipulation. Clear PagePrivate so that
2282 * global reserve count will not be incremented
2283 * by free_huge_page. This will make it appear
2284 * as though the reservation for this page was
2285 * consumed. This may prevent the task from
2286 * faulting in the page at a later time. This
2287 * is better than inconsistent global huge page
2288 * accounting of reserve counts.
2290 ClearPagePrivate(page);
2292 rc = vma_add_reservation(h, vma, address);
2293 if (unlikely(rc < 0))
2295 * See above comment about rare out of
2298 ClearPagePrivate(page);
2300 vma_end_reservation(h, vma, address);
2304 struct page *alloc_huge_page(struct vm_area_struct *vma,
2305 unsigned long addr, int avoid_reserve)
2307 struct hugepage_subpool *spool = subpool_vma(vma);
2308 struct hstate *h = hstate_vma(vma);
2310 long map_chg, map_commit;
2313 struct hugetlb_cgroup *h_cg;
2314 bool deferred_reserve;
2316 idx = hstate_index(h);
2318 * Examine the region/reserve map to determine if the process
2319 * has a reservation for the page to be allocated. A return
2320 * code of zero indicates a reservation exists (no change).
2322 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2324 return ERR_PTR(-ENOMEM);
2327 * Processes that did not create the mapping will have no
2328 * reserves as indicated by the region/reserve map. Check
2329 * that the allocation will not exceed the subpool limit.
2330 * Allocations for MAP_NORESERVE mappings also need to be
2331 * checked against any subpool limit.
2333 if (map_chg || avoid_reserve) {
2334 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2336 vma_end_reservation(h, vma, addr);
2337 return ERR_PTR(-ENOSPC);
2341 * Even though there was no reservation in the region/reserve
2342 * map, there could be reservations associated with the
2343 * subpool that can be used. This would be indicated if the
2344 * return value of hugepage_subpool_get_pages() is zero.
2345 * However, if avoid_reserve is specified we still avoid even
2346 * the subpool reservations.
2352 /* If this allocation is not consuming a reservation, charge it now.
2354 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2355 if (deferred_reserve) {
2356 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2357 idx, pages_per_huge_page(h), &h_cg);
2359 goto out_subpool_put;
2362 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2364 goto out_uncharge_cgroup_reservation;
2366 spin_lock(&hugetlb_lock);
2368 * glb_chg is passed to indicate whether or not a page must be taken
2369 * from the global free pool (global change). gbl_chg == 0 indicates
2370 * a reservation exists for the allocation.
2372 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2374 spin_unlock(&hugetlb_lock);
2375 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2377 goto out_uncharge_cgroup;
2378 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2379 SetPagePrivate(page);
2380 h->resv_huge_pages--;
2382 spin_lock(&hugetlb_lock);
2383 list_add(&page->lru, &h->hugepage_activelist);
2386 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2387 /* If allocation is not consuming a reservation, also store the
2388 * hugetlb_cgroup pointer on the page.
2390 if (deferred_reserve) {
2391 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2395 spin_unlock(&hugetlb_lock);
2397 set_page_private(page, (unsigned long)spool);
2399 map_commit = vma_commit_reservation(h, vma, addr);
2400 if (unlikely(map_chg > map_commit)) {
2402 * The page was added to the reservation map between
2403 * vma_needs_reservation and vma_commit_reservation.
2404 * This indicates a race with hugetlb_reserve_pages.
2405 * Adjust for the subpool count incremented above AND
2406 * in hugetlb_reserve_pages for the same page. Also,
2407 * the reservation count added in hugetlb_reserve_pages
2408 * no longer applies.
2412 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2413 hugetlb_acct_memory(h, -rsv_adjust);
2414 if (deferred_reserve)
2415 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2416 pages_per_huge_page(h), page);
2420 out_uncharge_cgroup:
2421 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2422 out_uncharge_cgroup_reservation:
2423 if (deferred_reserve)
2424 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2427 if (map_chg || avoid_reserve)
2428 hugepage_subpool_put_pages(spool, 1);
2429 vma_end_reservation(h, vma, addr);
2430 return ERR_PTR(-ENOSPC);
2433 int alloc_bootmem_huge_page(struct hstate *h)
2434 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2435 int __alloc_bootmem_huge_page(struct hstate *h)
2437 struct huge_bootmem_page *m;
2440 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2443 addr = memblock_alloc_try_nid_raw(
2444 huge_page_size(h), huge_page_size(h),
2445 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2448 * Use the beginning of the huge page to store the
2449 * huge_bootmem_page struct (until gather_bootmem
2450 * puts them into the mem_map).
2459 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2460 /* Put them into a private list first because mem_map is not up yet */
2461 INIT_LIST_HEAD(&m->list);
2462 list_add(&m->list, &huge_boot_pages);
2467 static void __init prep_compound_huge_page(struct page *page,
2470 if (unlikely(order > (MAX_ORDER - 1)))
2471 prep_compound_gigantic_page(page, order);
2473 prep_compound_page(page, order);
2476 /* Put bootmem huge pages into the standard lists after mem_map is up */
2477 static void __init gather_bootmem_prealloc(void)
2479 struct huge_bootmem_page *m;
2481 list_for_each_entry(m, &huge_boot_pages, list) {
2482 struct page *page = virt_to_page(m);
2483 struct hstate *h = m->hstate;
2485 WARN_ON(page_count(page) != 1);
2486 prep_compound_huge_page(page, huge_page_order(h));
2487 WARN_ON(PageReserved(page));
2488 prep_new_huge_page(h, page, page_to_nid(page));
2489 put_page(page); /* free it into the hugepage allocator */
2492 * If we had gigantic hugepages allocated at boot time, we need
2493 * to restore the 'stolen' pages to totalram_pages in order to
2494 * fix confusing memory reports from free(1) and another
2495 * side-effects, like CommitLimit going negative.
2497 if (hstate_is_gigantic(h))
2498 adjust_managed_page_count(page, pages_per_huge_page(h));
2503 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2506 nodemask_t *node_alloc_noretry;
2508 if (!hstate_is_gigantic(h)) {
2510 * Bit mask controlling how hard we retry per-node allocations.
2511 * Ignore errors as lower level routines can deal with
2512 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2513 * time, we are likely in bigger trouble.
2515 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2518 /* allocations done at boot time */
2519 node_alloc_noretry = NULL;
2522 /* bit mask controlling how hard we retry per-node allocations */
2523 if (node_alloc_noretry)
2524 nodes_clear(*node_alloc_noretry);
2526 for (i = 0; i < h->max_huge_pages; ++i) {
2527 if (hstate_is_gigantic(h)) {
2528 if (hugetlb_cma_size) {
2529 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2532 if (!alloc_bootmem_huge_page(h))
2534 } else if (!alloc_pool_huge_page(h,
2535 &node_states[N_MEMORY],
2536 node_alloc_noretry))
2540 if (i < h->max_huge_pages) {
2543 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2544 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2545 h->max_huge_pages, buf, i);
2546 h->max_huge_pages = i;
2549 kfree(node_alloc_noretry);
2552 static void __init hugetlb_init_hstates(void)
2556 for_each_hstate(h) {
2557 if (minimum_order > huge_page_order(h))
2558 minimum_order = huge_page_order(h);
2560 /* oversize hugepages were init'ed in early boot */
2561 if (!hstate_is_gigantic(h))
2562 hugetlb_hstate_alloc_pages(h);
2564 VM_BUG_ON(minimum_order == UINT_MAX);
2567 static void __init report_hugepages(void)
2571 for_each_hstate(h) {
2574 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2575 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2576 buf, h->free_huge_pages);
2580 #ifdef CONFIG_HIGHMEM
2581 static void try_to_free_low(struct hstate *h, unsigned long count,
2582 nodemask_t *nodes_allowed)
2586 if (hstate_is_gigantic(h))
2589 for_each_node_mask(i, *nodes_allowed) {
2590 struct page *page, *next;
2591 struct list_head *freel = &h->hugepage_freelists[i];
2592 list_for_each_entry_safe(page, next, freel, lru) {
2593 if (count >= h->nr_huge_pages)
2595 if (PageHighMem(page))
2597 list_del(&page->lru);
2598 update_and_free_page(h, page);
2599 h->free_huge_pages--;
2600 h->free_huge_pages_node[page_to_nid(page)]--;
2605 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2606 nodemask_t *nodes_allowed)
2612 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2613 * balanced by operating on them in a round-robin fashion.
2614 * Returns 1 if an adjustment was made.
2616 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2621 VM_BUG_ON(delta != -1 && delta != 1);
2624 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2625 if (h->surplus_huge_pages_node[node])
2629 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2630 if (h->surplus_huge_pages_node[node] <
2631 h->nr_huge_pages_node[node])
2638 h->surplus_huge_pages += delta;
2639 h->surplus_huge_pages_node[node] += delta;
2643 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2644 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2645 nodemask_t *nodes_allowed)
2647 unsigned long min_count, ret;
2648 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2651 * Bit mask controlling how hard we retry per-node allocations.
2652 * If we can not allocate the bit mask, do not attempt to allocate
2653 * the requested huge pages.
2655 if (node_alloc_noretry)
2656 nodes_clear(*node_alloc_noretry);
2660 spin_lock(&hugetlb_lock);
2663 * Check for a node specific request.
2664 * Changing node specific huge page count may require a corresponding
2665 * change to the global count. In any case, the passed node mask
2666 * (nodes_allowed) will restrict alloc/free to the specified node.
2668 if (nid != NUMA_NO_NODE) {
2669 unsigned long old_count = count;
2671 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2673 * User may have specified a large count value which caused the
2674 * above calculation to overflow. In this case, they wanted
2675 * to allocate as many huge pages as possible. Set count to
2676 * largest possible value to align with their intention.
2678 if (count < old_count)
2683 * Gigantic pages runtime allocation depend on the capability for large
2684 * page range allocation.
2685 * If the system does not provide this feature, return an error when
2686 * the user tries to allocate gigantic pages but let the user free the
2687 * boottime allocated gigantic pages.
2689 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2690 if (count > persistent_huge_pages(h)) {
2691 spin_unlock(&hugetlb_lock);
2692 NODEMASK_FREE(node_alloc_noretry);
2695 /* Fall through to decrease pool */
2699 * Increase the pool size
2700 * First take pages out of surplus state. Then make up the
2701 * remaining difference by allocating fresh huge pages.
2703 * We might race with alloc_surplus_huge_page() here and be unable
2704 * to convert a surplus huge page to a normal huge page. That is
2705 * not critical, though, it just means the overall size of the
2706 * pool might be one hugepage larger than it needs to be, but
2707 * within all the constraints specified by the sysctls.
2709 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2710 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2714 while (count > persistent_huge_pages(h)) {
2716 * If this allocation races such that we no longer need the
2717 * page, free_huge_page will handle it by freeing the page
2718 * and reducing the surplus.
2720 spin_unlock(&hugetlb_lock);
2722 /* yield cpu to avoid soft lockup */
2725 ret = alloc_pool_huge_page(h, nodes_allowed,
2726 node_alloc_noretry);
2727 spin_lock(&hugetlb_lock);
2731 /* Bail for signals. Probably ctrl-c from user */
2732 if (signal_pending(current))
2737 * Decrease the pool size
2738 * First return free pages to the buddy allocator (being careful
2739 * to keep enough around to satisfy reservations). Then place
2740 * pages into surplus state as needed so the pool will shrink
2741 * to the desired size as pages become free.
2743 * By placing pages into the surplus state independent of the
2744 * overcommit value, we are allowing the surplus pool size to
2745 * exceed overcommit. There are few sane options here. Since
2746 * alloc_surplus_huge_page() is checking the global counter,
2747 * though, we'll note that we're not allowed to exceed surplus
2748 * and won't grow the pool anywhere else. Not until one of the
2749 * sysctls are changed, or the surplus pages go out of use.
2751 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2752 min_count = max(count, min_count);
2753 try_to_free_low(h, min_count, nodes_allowed);
2754 while (min_count < persistent_huge_pages(h)) {
2755 if (!free_pool_huge_page(h, nodes_allowed, 0))
2757 cond_resched_lock(&hugetlb_lock);
2759 while (count < persistent_huge_pages(h)) {
2760 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2764 h->max_huge_pages = persistent_huge_pages(h);
2765 spin_unlock(&hugetlb_lock);
2767 NODEMASK_FREE(node_alloc_noretry);
2772 #define HSTATE_ATTR_RO(_name) \
2773 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2775 #define HSTATE_ATTR(_name) \
2776 static struct kobj_attribute _name##_attr = \
2777 __ATTR(_name, 0644, _name##_show, _name##_store)
2779 static struct kobject *hugepages_kobj;
2780 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2782 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2784 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2788 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2789 if (hstate_kobjs[i] == kobj) {
2791 *nidp = NUMA_NO_NODE;
2795 return kobj_to_node_hstate(kobj, nidp);
2798 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2799 struct kobj_attribute *attr, char *buf)
2802 unsigned long nr_huge_pages;
2805 h = kobj_to_hstate(kobj, &nid);
2806 if (nid == NUMA_NO_NODE)
2807 nr_huge_pages = h->nr_huge_pages;
2809 nr_huge_pages = h->nr_huge_pages_node[nid];
2811 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2814 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2815 struct hstate *h, int nid,
2816 unsigned long count, size_t len)
2819 nodemask_t nodes_allowed, *n_mask;
2821 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2824 if (nid == NUMA_NO_NODE) {
2826 * global hstate attribute
2828 if (!(obey_mempolicy &&
2829 init_nodemask_of_mempolicy(&nodes_allowed)))
2830 n_mask = &node_states[N_MEMORY];
2832 n_mask = &nodes_allowed;
2835 * Node specific request. count adjustment happens in
2836 * set_max_huge_pages() after acquiring hugetlb_lock.
2838 init_nodemask_of_node(&nodes_allowed, nid);
2839 n_mask = &nodes_allowed;
2842 err = set_max_huge_pages(h, count, nid, n_mask);
2844 return err ? err : len;
2847 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2848 struct kobject *kobj, const char *buf,
2852 unsigned long count;
2856 err = kstrtoul(buf, 10, &count);
2860 h = kobj_to_hstate(kobj, &nid);
2861 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2864 static ssize_t nr_hugepages_show(struct kobject *kobj,
2865 struct kobj_attribute *attr, char *buf)
2867 return nr_hugepages_show_common(kobj, attr, buf);
2870 static ssize_t nr_hugepages_store(struct kobject *kobj,
2871 struct kobj_attribute *attr, const char *buf, size_t len)
2873 return nr_hugepages_store_common(false, kobj, buf, len);
2875 HSTATE_ATTR(nr_hugepages);
2880 * hstate attribute for optionally mempolicy-based constraint on persistent
2881 * huge page alloc/free.
2883 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2884 struct kobj_attribute *attr,
2887 return nr_hugepages_show_common(kobj, attr, buf);
2890 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2891 struct kobj_attribute *attr, const char *buf, size_t len)
2893 return nr_hugepages_store_common(true, kobj, buf, len);
2895 HSTATE_ATTR(nr_hugepages_mempolicy);
2899 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2900 struct kobj_attribute *attr, char *buf)
2902 struct hstate *h = kobj_to_hstate(kobj, NULL);
2903 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2906 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2907 struct kobj_attribute *attr, const char *buf, size_t count)
2910 unsigned long input;
2911 struct hstate *h = kobj_to_hstate(kobj, NULL);
2913 if (hstate_is_gigantic(h))
2916 err = kstrtoul(buf, 10, &input);
2920 spin_lock(&hugetlb_lock);
2921 h->nr_overcommit_huge_pages = input;
2922 spin_unlock(&hugetlb_lock);
2926 HSTATE_ATTR(nr_overcommit_hugepages);
2928 static ssize_t free_hugepages_show(struct kobject *kobj,
2929 struct kobj_attribute *attr, char *buf)
2932 unsigned long free_huge_pages;
2935 h = kobj_to_hstate(kobj, &nid);
2936 if (nid == NUMA_NO_NODE)
2937 free_huge_pages = h->free_huge_pages;
2939 free_huge_pages = h->free_huge_pages_node[nid];
2941 return sysfs_emit(buf, "%lu\n", free_huge_pages);
2943 HSTATE_ATTR_RO(free_hugepages);
2945 static ssize_t resv_hugepages_show(struct kobject *kobj,
2946 struct kobj_attribute *attr, char *buf)
2948 struct hstate *h = kobj_to_hstate(kobj, NULL);
2949 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2951 HSTATE_ATTR_RO(resv_hugepages);
2953 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2954 struct kobj_attribute *attr, char *buf)
2957 unsigned long surplus_huge_pages;
2960 h = kobj_to_hstate(kobj, &nid);
2961 if (nid == NUMA_NO_NODE)
2962 surplus_huge_pages = h->surplus_huge_pages;
2964 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2966 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2968 HSTATE_ATTR_RO(surplus_hugepages);
2970 static struct attribute *hstate_attrs[] = {
2971 &nr_hugepages_attr.attr,
2972 &nr_overcommit_hugepages_attr.attr,
2973 &free_hugepages_attr.attr,
2974 &resv_hugepages_attr.attr,
2975 &surplus_hugepages_attr.attr,
2977 &nr_hugepages_mempolicy_attr.attr,
2982 static const struct attribute_group hstate_attr_group = {
2983 .attrs = hstate_attrs,
2986 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2987 struct kobject **hstate_kobjs,
2988 const struct attribute_group *hstate_attr_group)
2991 int hi = hstate_index(h);
2993 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2994 if (!hstate_kobjs[hi])
2997 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2999 kobject_put(hstate_kobjs[hi]);
3000 hstate_kobjs[hi] = NULL;
3006 static void __init hugetlb_sysfs_init(void)
3011 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3012 if (!hugepages_kobj)
3015 for_each_hstate(h) {
3016 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3017 hstate_kobjs, &hstate_attr_group);
3019 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3026 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3027 * with node devices in node_devices[] using a parallel array. The array
3028 * index of a node device or _hstate == node id.
3029 * This is here to avoid any static dependency of the node device driver, in
3030 * the base kernel, on the hugetlb module.
3032 struct node_hstate {
3033 struct kobject *hugepages_kobj;
3034 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3036 static struct node_hstate node_hstates[MAX_NUMNODES];
3039 * A subset of global hstate attributes for node devices
3041 static struct attribute *per_node_hstate_attrs[] = {
3042 &nr_hugepages_attr.attr,
3043 &free_hugepages_attr.attr,
3044 &surplus_hugepages_attr.attr,
3048 static const struct attribute_group per_node_hstate_attr_group = {
3049 .attrs = per_node_hstate_attrs,
3053 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3054 * Returns node id via non-NULL nidp.
3056 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3060 for (nid = 0; nid < nr_node_ids; nid++) {
3061 struct node_hstate *nhs = &node_hstates[nid];
3063 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3064 if (nhs->hstate_kobjs[i] == kobj) {
3076 * Unregister hstate attributes from a single node device.
3077 * No-op if no hstate attributes attached.
3079 static void hugetlb_unregister_node(struct node *node)
3082 struct node_hstate *nhs = &node_hstates[node->dev.id];
3084 if (!nhs->hugepages_kobj)
3085 return; /* no hstate attributes */
3087 for_each_hstate(h) {
3088 int idx = hstate_index(h);
3089 if (nhs->hstate_kobjs[idx]) {
3090 kobject_put(nhs->hstate_kobjs[idx]);
3091 nhs->hstate_kobjs[idx] = NULL;
3095 kobject_put(nhs->hugepages_kobj);
3096 nhs->hugepages_kobj = NULL;
3101 * Register hstate attributes for a single node device.
3102 * No-op if attributes already registered.
3104 static void hugetlb_register_node(struct node *node)
3107 struct node_hstate *nhs = &node_hstates[node->dev.id];
3110 if (nhs->hugepages_kobj)
3111 return; /* already allocated */
3113 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3115 if (!nhs->hugepages_kobj)
3118 for_each_hstate(h) {
3119 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3121 &per_node_hstate_attr_group);
3123 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3124 h->name, node->dev.id);
3125 hugetlb_unregister_node(node);
3132 * hugetlb init time: register hstate attributes for all registered node
3133 * devices of nodes that have memory. All on-line nodes should have
3134 * registered their associated device by this time.
3136 static void __init hugetlb_register_all_nodes(void)
3140 for_each_node_state(nid, N_MEMORY) {
3141 struct node *node = node_devices[nid];
3142 if (node->dev.id == nid)
3143 hugetlb_register_node(node);
3147 * Let the node device driver know we're here so it can
3148 * [un]register hstate attributes on node hotplug.
3150 register_hugetlbfs_with_node(hugetlb_register_node,
3151 hugetlb_unregister_node);
3153 #else /* !CONFIG_NUMA */
3155 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3163 static void hugetlb_register_all_nodes(void) { }
3167 static int __init hugetlb_init(void)
3171 if (!hugepages_supported()) {
3172 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3173 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3178 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3179 * architectures depend on setup being done here.
3181 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3182 if (!parsed_default_hugepagesz) {
3184 * If we did not parse a default huge page size, set
3185 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3186 * number of huge pages for this default size was implicitly
3187 * specified, set that here as well.
3188 * Note that the implicit setting will overwrite an explicit
3189 * setting. A warning will be printed in this case.
3191 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3192 if (default_hstate_max_huge_pages) {
3193 if (default_hstate.max_huge_pages) {
3196 string_get_size(huge_page_size(&default_hstate),
3197 1, STRING_UNITS_2, buf, 32);
3198 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3199 default_hstate.max_huge_pages, buf);
3200 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3201 default_hstate_max_huge_pages);
3203 default_hstate.max_huge_pages =
3204 default_hstate_max_huge_pages;
3208 hugetlb_cma_check();
3209 hugetlb_init_hstates();
3210 gather_bootmem_prealloc();
3213 hugetlb_sysfs_init();
3214 hugetlb_register_all_nodes();
3215 hugetlb_cgroup_file_init();
3218 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3220 num_fault_mutexes = 1;
3222 hugetlb_fault_mutex_table =
3223 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3225 BUG_ON(!hugetlb_fault_mutex_table);
3227 for (i = 0; i < num_fault_mutexes; i++)
3228 mutex_init(&hugetlb_fault_mutex_table[i]);
3231 subsys_initcall(hugetlb_init);
3233 /* Overwritten by architectures with more huge page sizes */
3234 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3236 return size == HPAGE_SIZE;
3239 void __init hugetlb_add_hstate(unsigned int order)
3244 if (size_to_hstate(PAGE_SIZE << order)) {
3247 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3249 h = &hstates[hugetlb_max_hstate++];
3251 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3252 for (i = 0; i < MAX_NUMNODES; ++i)
3253 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3254 INIT_LIST_HEAD(&h->hugepage_activelist);
3255 h->next_nid_to_alloc = first_memory_node;
3256 h->next_nid_to_free = first_memory_node;
3257 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3258 huge_page_size(h)/1024);
3264 * hugepages command line processing
3265 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3266 * specification. If not, ignore the hugepages value. hugepages can also
3267 * be the first huge page command line option in which case it implicitly
3268 * specifies the number of huge pages for the default size.
3270 static int __init hugepages_setup(char *s)
3273 static unsigned long *last_mhp;
3275 if (!parsed_valid_hugepagesz) {
3276 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3277 parsed_valid_hugepagesz = true;
3282 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3283 * yet, so this hugepages= parameter goes to the "default hstate".
3284 * Otherwise, it goes with the previously parsed hugepagesz or
3285 * default_hugepagesz.
3287 else if (!hugetlb_max_hstate)
3288 mhp = &default_hstate_max_huge_pages;
3290 mhp = &parsed_hstate->max_huge_pages;
3292 if (mhp == last_mhp) {
3293 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3297 if (sscanf(s, "%lu", mhp) <= 0)
3301 * Global state is always initialized later in hugetlb_init.
3302 * But we need to allocate >= MAX_ORDER hstates here early to still
3303 * use the bootmem allocator.
3305 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3306 hugetlb_hstate_alloc_pages(parsed_hstate);
3312 __setup("hugepages=", hugepages_setup);
3315 * hugepagesz command line processing
3316 * A specific huge page size can only be specified once with hugepagesz.
3317 * hugepagesz is followed by hugepages on the command line. The global
3318 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3319 * hugepagesz argument was valid.
3321 static int __init hugepagesz_setup(char *s)
3326 parsed_valid_hugepagesz = false;
3327 size = (unsigned long)memparse(s, NULL);
3329 if (!arch_hugetlb_valid_size(size)) {
3330 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3334 h = size_to_hstate(size);
3337 * hstate for this size already exists. This is normally
3338 * an error, but is allowed if the existing hstate is the
3339 * default hstate. More specifically, it is only allowed if
3340 * the number of huge pages for the default hstate was not
3341 * previously specified.
3343 if (!parsed_default_hugepagesz || h != &default_hstate ||
3344 default_hstate.max_huge_pages) {
3345 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3350 * No need to call hugetlb_add_hstate() as hstate already
3351 * exists. But, do set parsed_hstate so that a following
3352 * hugepages= parameter will be applied to this hstate.
3355 parsed_valid_hugepagesz = true;
3359 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3360 parsed_valid_hugepagesz = true;
3363 __setup("hugepagesz=", hugepagesz_setup);
3366 * default_hugepagesz command line input
3367 * Only one instance of default_hugepagesz allowed on command line.
3369 static int __init default_hugepagesz_setup(char *s)
3373 parsed_valid_hugepagesz = false;
3374 if (parsed_default_hugepagesz) {
3375 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3379 size = (unsigned long)memparse(s, NULL);
3381 if (!arch_hugetlb_valid_size(size)) {
3382 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3386 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3387 parsed_valid_hugepagesz = true;
3388 parsed_default_hugepagesz = true;
3389 default_hstate_idx = hstate_index(size_to_hstate(size));
3392 * The number of default huge pages (for this size) could have been
3393 * specified as the first hugetlb parameter: hugepages=X. If so,
3394 * then default_hstate_max_huge_pages is set. If the default huge
3395 * page size is gigantic (>= MAX_ORDER), then the pages must be
3396 * allocated here from bootmem allocator.
3398 if (default_hstate_max_huge_pages) {
3399 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3400 if (hstate_is_gigantic(&default_hstate))
3401 hugetlb_hstate_alloc_pages(&default_hstate);
3402 default_hstate_max_huge_pages = 0;
3407 __setup("default_hugepagesz=", default_hugepagesz_setup);
3409 static unsigned int allowed_mems_nr(struct hstate *h)
3412 unsigned int nr = 0;
3413 nodemask_t *mpol_allowed;
3414 unsigned int *array = h->free_huge_pages_node;
3415 gfp_t gfp_mask = htlb_alloc_mask(h);
3417 mpol_allowed = policy_nodemask_current(gfp_mask);
3419 for_each_node_mask(node, cpuset_current_mems_allowed) {
3420 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3427 #ifdef CONFIG_SYSCTL
3428 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3429 void *buffer, size_t *length,
3430 loff_t *ppos, unsigned long *out)
3432 struct ctl_table dup_table;
3435 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3436 * can duplicate the @table and alter the duplicate of it.
3439 dup_table.data = out;
3441 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3444 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3445 struct ctl_table *table, int write,
3446 void *buffer, size_t *length, loff_t *ppos)
3448 struct hstate *h = &default_hstate;
3449 unsigned long tmp = h->max_huge_pages;
3452 if (!hugepages_supported())
3455 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3461 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3462 NUMA_NO_NODE, tmp, *length);
3467 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3468 void *buffer, size_t *length, loff_t *ppos)
3471 return hugetlb_sysctl_handler_common(false, table, write,
3472 buffer, length, ppos);
3476 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3477 void *buffer, size_t *length, loff_t *ppos)
3479 return hugetlb_sysctl_handler_common(true, table, write,
3480 buffer, length, ppos);
3482 #endif /* CONFIG_NUMA */
3484 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3485 void *buffer, size_t *length, loff_t *ppos)
3487 struct hstate *h = &default_hstate;
3491 if (!hugepages_supported())
3494 tmp = h->nr_overcommit_huge_pages;
3496 if (write && hstate_is_gigantic(h))
3499 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3505 spin_lock(&hugetlb_lock);
3506 h->nr_overcommit_huge_pages = tmp;
3507 spin_unlock(&hugetlb_lock);
3513 #endif /* CONFIG_SYSCTL */
3515 void hugetlb_report_meminfo(struct seq_file *m)
3518 unsigned long total = 0;
3520 if (!hugepages_supported())
3523 for_each_hstate(h) {
3524 unsigned long count = h->nr_huge_pages;
3526 total += (PAGE_SIZE << huge_page_order(h)) * count;
3528 if (h == &default_hstate)
3530 "HugePages_Total: %5lu\n"
3531 "HugePages_Free: %5lu\n"
3532 "HugePages_Rsvd: %5lu\n"
3533 "HugePages_Surp: %5lu\n"
3534 "Hugepagesize: %8lu kB\n",
3538 h->surplus_huge_pages,
3539 (PAGE_SIZE << huge_page_order(h)) / 1024);
3542 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3545 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3547 struct hstate *h = &default_hstate;
3549 if (!hugepages_supported())
3552 return sysfs_emit_at(buf, len,
3553 "Node %d HugePages_Total: %5u\n"
3554 "Node %d HugePages_Free: %5u\n"
3555 "Node %d HugePages_Surp: %5u\n",
3556 nid, h->nr_huge_pages_node[nid],
3557 nid, h->free_huge_pages_node[nid],
3558 nid, h->surplus_huge_pages_node[nid]);
3561 void hugetlb_show_meminfo(void)
3566 if (!hugepages_supported())
3569 for_each_node_state(nid, N_MEMORY)
3571 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3573 h->nr_huge_pages_node[nid],
3574 h->free_huge_pages_node[nid],
3575 h->surplus_huge_pages_node[nid],
3576 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3579 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3581 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3582 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3585 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3586 unsigned long hugetlb_total_pages(void)
3589 unsigned long nr_total_pages = 0;
3592 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3593 return nr_total_pages;
3596 static int hugetlb_acct_memory(struct hstate *h, long delta)
3603 spin_lock(&hugetlb_lock);
3605 * When cpuset is configured, it breaks the strict hugetlb page
3606 * reservation as the accounting is done on a global variable. Such
3607 * reservation is completely rubbish in the presence of cpuset because
3608 * the reservation is not checked against page availability for the
3609 * current cpuset. Application can still potentially OOM'ed by kernel
3610 * with lack of free htlb page in cpuset that the task is in.
3611 * Attempt to enforce strict accounting with cpuset is almost
3612 * impossible (or too ugly) because cpuset is too fluid that
3613 * task or memory node can be dynamically moved between cpusets.
3615 * The change of semantics for shared hugetlb mapping with cpuset is
3616 * undesirable. However, in order to preserve some of the semantics,
3617 * we fall back to check against current free page availability as
3618 * a best attempt and hopefully to minimize the impact of changing
3619 * semantics that cpuset has.
3621 * Apart from cpuset, we also have memory policy mechanism that
3622 * also determines from which node the kernel will allocate memory
3623 * in a NUMA system. So similar to cpuset, we also should consider
3624 * the memory policy of the current task. Similar to the description
3628 if (gather_surplus_pages(h, delta) < 0)
3631 if (delta > allowed_mems_nr(h)) {
3632 return_unused_surplus_pages(h, delta);
3639 return_unused_surplus_pages(h, (unsigned long) -delta);
3642 spin_unlock(&hugetlb_lock);
3646 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3648 struct resv_map *resv = vma_resv_map(vma);
3651 * This new VMA should share its siblings reservation map if present.
3652 * The VMA will only ever have a valid reservation map pointer where
3653 * it is being copied for another still existing VMA. As that VMA
3654 * has a reference to the reservation map it cannot disappear until
3655 * after this open call completes. It is therefore safe to take a
3656 * new reference here without additional locking.
3658 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3659 kref_get(&resv->refs);
3662 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3664 struct hstate *h = hstate_vma(vma);
3665 struct resv_map *resv = vma_resv_map(vma);
3666 struct hugepage_subpool *spool = subpool_vma(vma);
3667 unsigned long reserve, start, end;
3670 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3673 start = vma_hugecache_offset(h, vma, vma->vm_start);
3674 end = vma_hugecache_offset(h, vma, vma->vm_end);
3676 reserve = (end - start) - region_count(resv, start, end);
3677 hugetlb_cgroup_uncharge_counter(resv, start, end);
3680 * Decrement reserve counts. The global reserve count may be
3681 * adjusted if the subpool has a minimum size.
3683 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3684 hugetlb_acct_memory(h, -gbl_reserve);
3687 kref_put(&resv->refs, resv_map_release);
3690 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3692 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3697 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3699 struct hstate *hstate = hstate_vma(vma);
3701 return 1UL << huge_page_shift(hstate);
3705 * We cannot handle pagefaults against hugetlb pages at all. They cause
3706 * handle_mm_fault() to try to instantiate regular-sized pages in the
3707 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3710 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3717 * When a new function is introduced to vm_operations_struct and added
3718 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3719 * This is because under System V memory model, mappings created via
3720 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3721 * their original vm_ops are overwritten with shm_vm_ops.
3723 const struct vm_operations_struct hugetlb_vm_ops = {
3724 .fault = hugetlb_vm_op_fault,
3725 .open = hugetlb_vm_op_open,
3726 .close = hugetlb_vm_op_close,
3727 .may_split = hugetlb_vm_op_split,
3728 .pagesize = hugetlb_vm_op_pagesize,
3731 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3737 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3738 vma->vm_page_prot)));
3740 entry = huge_pte_wrprotect(mk_huge_pte(page,
3741 vma->vm_page_prot));
3743 entry = pte_mkyoung(entry);
3744 entry = pte_mkhuge(entry);
3745 entry = arch_make_huge_pte(entry, vma, page, writable);
3750 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3751 unsigned long address, pte_t *ptep)
3755 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3756 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3757 update_mmu_cache(vma, address, ptep);
3760 bool is_hugetlb_entry_migration(pte_t pte)
3764 if (huge_pte_none(pte) || pte_present(pte))
3766 swp = pte_to_swp_entry(pte);
3767 if (is_migration_entry(swp))
3773 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3777 if (huge_pte_none(pte) || pte_present(pte))
3779 swp = pte_to_swp_entry(pte);
3780 if (is_hwpoison_entry(swp))
3786 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3787 struct vm_area_struct *vma)
3789 pte_t *src_pte, *dst_pte, entry, dst_entry;
3790 struct page *ptepage;
3793 struct hstate *h = hstate_vma(vma);
3794 unsigned long sz = huge_page_size(h);
3795 struct address_space *mapping = vma->vm_file->f_mapping;
3796 struct mmu_notifier_range range;
3799 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3802 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3805 mmu_notifier_invalidate_range_start(&range);
3808 * For shared mappings i_mmap_rwsem must be held to call
3809 * huge_pte_alloc, otherwise the returned ptep could go
3810 * away if part of a shared pmd and another thread calls
3813 i_mmap_lock_read(mapping);
3816 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3817 spinlock_t *src_ptl, *dst_ptl;
3818 src_pte = huge_pte_offset(src, addr, sz);
3821 dst_pte = huge_pte_alloc(dst, addr, sz);
3828 * If the pagetables are shared don't copy or take references.
3829 * dst_pte == src_pte is the common case of src/dest sharing.
3831 * However, src could have 'unshared' and dst shares with
3832 * another vma. If dst_pte !none, this implies sharing.
3833 * Check here before taking page table lock, and once again
3834 * after taking the lock below.
3836 dst_entry = huge_ptep_get(dst_pte);
3837 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3840 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3841 src_ptl = huge_pte_lockptr(h, src, src_pte);
3842 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3843 entry = huge_ptep_get(src_pte);
3844 dst_entry = huge_ptep_get(dst_pte);
3845 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3847 * Skip if src entry none. Also, skip in the
3848 * unlikely case dst entry !none as this implies
3849 * sharing with another vma.
3852 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3853 is_hugetlb_entry_hwpoisoned(entry))) {
3854 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3856 if (is_write_migration_entry(swp_entry) && cow) {
3858 * COW mappings require pages in both
3859 * parent and child to be set to read.
3861 make_migration_entry_read(&swp_entry);
3862 entry = swp_entry_to_pte(swp_entry);
3863 set_huge_swap_pte_at(src, addr, src_pte,
3866 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3870 * No need to notify as we are downgrading page
3871 * table protection not changing it to point
3874 * See Documentation/vm/mmu_notifier.rst
3876 huge_ptep_set_wrprotect(src, addr, src_pte);
3878 entry = huge_ptep_get(src_pte);
3879 ptepage = pte_page(entry);
3881 page_dup_rmap(ptepage, true);
3882 set_huge_pte_at(dst, addr, dst_pte, entry);
3883 hugetlb_count_add(pages_per_huge_page(h), dst);
3885 spin_unlock(src_ptl);
3886 spin_unlock(dst_ptl);
3890 mmu_notifier_invalidate_range_end(&range);
3892 i_mmap_unlock_read(mapping);
3897 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3898 unsigned long start, unsigned long end,
3899 struct page *ref_page)
3901 struct mm_struct *mm = vma->vm_mm;
3902 unsigned long address;
3907 struct hstate *h = hstate_vma(vma);
3908 unsigned long sz = huge_page_size(h);
3909 struct mmu_notifier_range range;
3911 WARN_ON(!is_vm_hugetlb_page(vma));
3912 BUG_ON(start & ~huge_page_mask(h));
3913 BUG_ON(end & ~huge_page_mask(h));
3916 * This is a hugetlb vma, all the pte entries should point
3919 tlb_change_page_size(tlb, sz);
3920 tlb_start_vma(tlb, vma);
3923 * If sharing possible, alert mmu notifiers of worst case.
3925 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3927 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3928 mmu_notifier_invalidate_range_start(&range);
3930 for (; address < end; address += sz) {
3931 ptep = huge_pte_offset(mm, address, sz);
3935 ptl = huge_pte_lock(h, mm, ptep);
3936 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3939 * We just unmapped a page of PMDs by clearing a PUD.
3940 * The caller's TLB flush range should cover this area.
3945 pte = huge_ptep_get(ptep);
3946 if (huge_pte_none(pte)) {
3952 * Migrating hugepage or HWPoisoned hugepage is already
3953 * unmapped and its refcount is dropped, so just clear pte here.
3955 if (unlikely(!pte_present(pte))) {
3956 huge_pte_clear(mm, address, ptep, sz);
3961 page = pte_page(pte);
3963 * If a reference page is supplied, it is because a specific
3964 * page is being unmapped, not a range. Ensure the page we
3965 * are about to unmap is the actual page of interest.
3968 if (page != ref_page) {
3973 * Mark the VMA as having unmapped its page so that
3974 * future faults in this VMA will fail rather than
3975 * looking like data was lost
3977 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3980 pte = huge_ptep_get_and_clear(mm, address, ptep);
3981 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3982 if (huge_pte_dirty(pte))
3983 set_page_dirty(page);
3985 hugetlb_count_sub(pages_per_huge_page(h), mm);
3986 page_remove_rmap(page, true);
3989 tlb_remove_page_size(tlb, page, huge_page_size(h));
3991 * Bail out after unmapping reference page if supplied
3996 mmu_notifier_invalidate_range_end(&range);
3997 tlb_end_vma(tlb, vma);
4000 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4001 struct vm_area_struct *vma, unsigned long start,
4002 unsigned long end, struct page *ref_page)
4004 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4007 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4008 * test will fail on a vma being torn down, and not grab a page table
4009 * on its way out. We're lucky that the flag has such an appropriate
4010 * name, and can in fact be safely cleared here. We could clear it
4011 * before the __unmap_hugepage_range above, but all that's necessary
4012 * is to clear it before releasing the i_mmap_rwsem. This works
4013 * because in the context this is called, the VMA is about to be
4014 * destroyed and the i_mmap_rwsem is held.
4016 vma->vm_flags &= ~VM_MAYSHARE;
4019 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4020 unsigned long end, struct page *ref_page)
4022 struct mmu_gather tlb;
4024 tlb_gather_mmu(&tlb, vma->vm_mm);
4025 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4026 tlb_finish_mmu(&tlb);
4030 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4031 * mapping it owns the reserve page for. The intention is to unmap the page
4032 * from other VMAs and let the children be SIGKILLed if they are faulting the
4035 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4036 struct page *page, unsigned long address)
4038 struct hstate *h = hstate_vma(vma);
4039 struct vm_area_struct *iter_vma;
4040 struct address_space *mapping;
4044 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4045 * from page cache lookup which is in HPAGE_SIZE units.
4047 address = address & huge_page_mask(h);
4048 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4050 mapping = vma->vm_file->f_mapping;
4053 * Take the mapping lock for the duration of the table walk. As
4054 * this mapping should be shared between all the VMAs,
4055 * __unmap_hugepage_range() is called as the lock is already held
4057 i_mmap_lock_write(mapping);
4058 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4059 /* Do not unmap the current VMA */
4060 if (iter_vma == vma)
4064 * Shared VMAs have their own reserves and do not affect
4065 * MAP_PRIVATE accounting but it is possible that a shared
4066 * VMA is using the same page so check and skip such VMAs.
4068 if (iter_vma->vm_flags & VM_MAYSHARE)
4072 * Unmap the page from other VMAs without their own reserves.
4073 * They get marked to be SIGKILLed if they fault in these
4074 * areas. This is because a future no-page fault on this VMA
4075 * could insert a zeroed page instead of the data existing
4076 * from the time of fork. This would look like data corruption
4078 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4079 unmap_hugepage_range(iter_vma, address,
4080 address + huge_page_size(h), page);
4082 i_mmap_unlock_write(mapping);
4086 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4087 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4088 * cannot race with other handlers or page migration.
4089 * Keep the pte_same checks anyway to make transition from the mutex easier.
4091 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4092 unsigned long address, pte_t *ptep,
4093 struct page *pagecache_page, spinlock_t *ptl)
4096 struct hstate *h = hstate_vma(vma);
4097 struct page *old_page, *new_page;
4098 int outside_reserve = 0;
4100 unsigned long haddr = address & huge_page_mask(h);
4101 struct mmu_notifier_range range;
4103 pte = huge_ptep_get(ptep);
4104 old_page = pte_page(pte);
4107 /* If no-one else is actually using this page, avoid the copy
4108 * and just make the page writable */
4109 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4110 page_move_anon_rmap(old_page, vma);
4111 set_huge_ptep_writable(vma, haddr, ptep);
4116 * If the process that created a MAP_PRIVATE mapping is about to
4117 * perform a COW due to a shared page count, attempt to satisfy
4118 * the allocation without using the existing reserves. The pagecache
4119 * page is used to determine if the reserve at this address was
4120 * consumed or not. If reserves were used, a partial faulted mapping
4121 * at the time of fork() could consume its reserves on COW instead
4122 * of the full address range.
4124 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4125 old_page != pagecache_page)
4126 outside_reserve = 1;
4131 * Drop page table lock as buddy allocator may be called. It will
4132 * be acquired again before returning to the caller, as expected.
4135 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4137 if (IS_ERR(new_page)) {
4139 * If a process owning a MAP_PRIVATE mapping fails to COW,
4140 * it is due to references held by a child and an insufficient
4141 * huge page pool. To guarantee the original mappers
4142 * reliability, unmap the page from child processes. The child
4143 * may get SIGKILLed if it later faults.
4145 if (outside_reserve) {
4146 struct address_space *mapping = vma->vm_file->f_mapping;
4151 BUG_ON(huge_pte_none(pte));
4153 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4154 * unmapping. unmapping needs to hold i_mmap_rwsem
4155 * in write mode. Dropping i_mmap_rwsem in read mode
4156 * here is OK as COW mappings do not interact with
4159 * Reacquire both after unmap operation.
4161 idx = vma_hugecache_offset(h, vma, haddr);
4162 hash = hugetlb_fault_mutex_hash(mapping, idx);
4163 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4164 i_mmap_unlock_read(mapping);
4166 unmap_ref_private(mm, vma, old_page, haddr);
4168 i_mmap_lock_read(mapping);
4169 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4171 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4173 pte_same(huge_ptep_get(ptep), pte)))
4174 goto retry_avoidcopy;
4176 * race occurs while re-acquiring page table
4177 * lock, and our job is done.
4182 ret = vmf_error(PTR_ERR(new_page));
4183 goto out_release_old;
4187 * When the original hugepage is shared one, it does not have
4188 * anon_vma prepared.
4190 if (unlikely(anon_vma_prepare(vma))) {
4192 goto out_release_all;
4195 copy_user_huge_page(new_page, old_page, address, vma,
4196 pages_per_huge_page(h));
4197 __SetPageUptodate(new_page);
4199 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4200 haddr + huge_page_size(h));
4201 mmu_notifier_invalidate_range_start(&range);
4204 * Retake the page table lock to check for racing updates
4205 * before the page tables are altered
4208 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4209 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4210 ClearPagePrivate(new_page);
4213 huge_ptep_clear_flush(vma, haddr, ptep);
4214 mmu_notifier_invalidate_range(mm, range.start, range.end);
4215 set_huge_pte_at(mm, haddr, ptep,
4216 make_huge_pte(vma, new_page, 1));
4217 page_remove_rmap(old_page, true);
4218 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4219 set_page_huge_active(new_page);
4220 /* Make the old page be freed below */
4221 new_page = old_page;
4224 mmu_notifier_invalidate_range_end(&range);
4226 restore_reserve_on_error(h, vma, haddr, new_page);
4231 spin_lock(ptl); /* Caller expects lock to be held */
4235 /* Return the pagecache page at a given address within a VMA */
4236 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4237 struct vm_area_struct *vma, unsigned long address)
4239 struct address_space *mapping;
4242 mapping = vma->vm_file->f_mapping;
4243 idx = vma_hugecache_offset(h, vma, address);
4245 return find_lock_page(mapping, idx);
4249 * Return whether there is a pagecache page to back given address within VMA.
4250 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4252 static bool hugetlbfs_pagecache_present(struct hstate *h,
4253 struct vm_area_struct *vma, unsigned long address)
4255 struct address_space *mapping;
4259 mapping = vma->vm_file->f_mapping;
4260 idx = vma_hugecache_offset(h, vma, address);
4262 page = find_get_page(mapping, idx);
4265 return page != NULL;
4268 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4271 struct inode *inode = mapping->host;
4272 struct hstate *h = hstate_inode(inode);
4273 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4277 ClearPagePrivate(page);
4280 * set page dirty so that it will not be removed from cache/file
4281 * by non-hugetlbfs specific code paths.
4283 set_page_dirty(page);
4285 spin_lock(&inode->i_lock);
4286 inode->i_blocks += blocks_per_huge_page(h);
4287 spin_unlock(&inode->i_lock);
4291 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4292 struct vm_area_struct *vma,
4293 struct address_space *mapping, pgoff_t idx,
4294 unsigned long address, pte_t *ptep, unsigned int flags)
4296 struct hstate *h = hstate_vma(vma);
4297 vm_fault_t ret = VM_FAULT_SIGBUS;
4303 unsigned long haddr = address & huge_page_mask(h);
4304 bool new_page = false;
4307 * Currently, we are forced to kill the process in the event the
4308 * original mapper has unmapped pages from the child due to a failed
4309 * COW. Warn that such a situation has occurred as it may not be obvious
4311 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4312 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4318 * We can not race with truncation due to holding i_mmap_rwsem.
4319 * i_size is modified when holding i_mmap_rwsem, so check here
4320 * once for faults beyond end of file.
4322 size = i_size_read(mapping->host) >> huge_page_shift(h);
4327 page = find_lock_page(mapping, idx);
4330 * Check for page in userfault range
4332 if (userfaultfd_missing(vma)) {
4334 struct vm_fault vmf = {
4339 * Hard to debug if it ends up being
4340 * used by a callee that assumes
4341 * something about the other
4342 * uninitialized fields... same as in
4348 * hugetlb_fault_mutex and i_mmap_rwsem must be
4349 * dropped before handling userfault. Reacquire
4350 * after handling fault to make calling code simpler.
4352 hash = hugetlb_fault_mutex_hash(mapping, idx);
4353 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4354 i_mmap_unlock_read(mapping);
4355 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4356 i_mmap_lock_read(mapping);
4357 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4361 page = alloc_huge_page(vma, haddr, 0);
4364 * Returning error will result in faulting task being
4365 * sent SIGBUS. The hugetlb fault mutex prevents two
4366 * tasks from racing to fault in the same page which
4367 * could result in false unable to allocate errors.
4368 * Page migration does not take the fault mutex, but
4369 * does a clear then write of pte's under page table
4370 * lock. Page fault code could race with migration,
4371 * notice the clear pte and try to allocate a page
4372 * here. Before returning error, get ptl and make
4373 * sure there really is no pte entry.
4375 ptl = huge_pte_lock(h, mm, ptep);
4376 if (!huge_pte_none(huge_ptep_get(ptep))) {
4382 ret = vmf_error(PTR_ERR(page));
4385 clear_huge_page(page, address, pages_per_huge_page(h));
4386 __SetPageUptodate(page);
4389 if (vma->vm_flags & VM_MAYSHARE) {
4390 int err = huge_add_to_page_cache(page, mapping, idx);
4399 if (unlikely(anon_vma_prepare(vma))) {
4401 goto backout_unlocked;
4407 * If memory error occurs between mmap() and fault, some process
4408 * don't have hwpoisoned swap entry for errored virtual address.
4409 * So we need to block hugepage fault by PG_hwpoison bit check.
4411 if (unlikely(PageHWPoison(page))) {
4412 ret = VM_FAULT_HWPOISON_LARGE |
4413 VM_FAULT_SET_HINDEX(hstate_index(h));
4414 goto backout_unlocked;
4419 * If we are going to COW a private mapping later, we examine the
4420 * pending reservations for this page now. This will ensure that
4421 * any allocations necessary to record that reservation occur outside
4424 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4425 if (vma_needs_reservation(h, vma, haddr) < 0) {
4427 goto backout_unlocked;
4429 /* Just decrements count, does not deallocate */
4430 vma_end_reservation(h, vma, haddr);
4433 ptl = huge_pte_lock(h, mm, ptep);
4435 if (!huge_pte_none(huge_ptep_get(ptep)))
4439 ClearPagePrivate(page);
4440 hugepage_add_new_anon_rmap(page, vma, haddr);
4442 page_dup_rmap(page, true);
4443 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4444 && (vma->vm_flags & VM_SHARED)));
4445 set_huge_pte_at(mm, haddr, ptep, new_pte);
4447 hugetlb_count_add(pages_per_huge_page(h), mm);
4448 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4449 /* Optimization, do the COW without a second fault */
4450 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4456 * Only make newly allocated pages active. Existing pages found
4457 * in the pagecache could be !page_huge_active() if they have been
4458 * isolated for migration.
4461 set_page_huge_active(page);
4471 restore_reserve_on_error(h, vma, haddr, page);
4477 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4479 unsigned long key[2];
4482 key[0] = (unsigned long) mapping;
4485 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4487 return hash & (num_fault_mutexes - 1);
4491 * For uniprocessor systems we always use a single mutex, so just
4492 * return 0 and avoid the hashing overhead.
4494 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4500 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4501 unsigned long address, unsigned int flags)
4508 struct page *page = NULL;
4509 struct page *pagecache_page = NULL;
4510 struct hstate *h = hstate_vma(vma);
4511 struct address_space *mapping;
4512 int need_wait_lock = 0;
4513 unsigned long haddr = address & huge_page_mask(h);
4515 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4518 * Since we hold no locks, ptep could be stale. That is
4519 * OK as we are only making decisions based on content and
4520 * not actually modifying content here.
4522 entry = huge_ptep_get(ptep);
4523 if (unlikely(is_hugetlb_entry_migration(entry))) {
4524 migration_entry_wait_huge(vma, mm, ptep);
4526 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4527 return VM_FAULT_HWPOISON_LARGE |
4528 VM_FAULT_SET_HINDEX(hstate_index(h));
4532 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4533 * until finished with ptep. This serves two purposes:
4534 * 1) It prevents huge_pmd_unshare from being called elsewhere
4535 * and making the ptep no longer valid.
4536 * 2) It synchronizes us with i_size modifications during truncation.
4538 * ptep could have already be assigned via huge_pte_offset. That
4539 * is OK, as huge_pte_alloc will return the same value unless
4540 * something has changed.
4542 mapping = vma->vm_file->f_mapping;
4543 i_mmap_lock_read(mapping);
4544 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4546 i_mmap_unlock_read(mapping);
4547 return VM_FAULT_OOM;
4551 * Serialize hugepage allocation and instantiation, so that we don't
4552 * get spurious allocation failures if two CPUs race to instantiate
4553 * the same page in the page cache.
4555 idx = vma_hugecache_offset(h, vma, haddr);
4556 hash = hugetlb_fault_mutex_hash(mapping, idx);
4557 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4559 entry = huge_ptep_get(ptep);
4560 if (huge_pte_none(entry)) {
4561 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4568 * entry could be a migration/hwpoison entry at this point, so this
4569 * check prevents the kernel from going below assuming that we have
4570 * an active hugepage in pagecache. This goto expects the 2nd page
4571 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4572 * properly handle it.
4574 if (!pte_present(entry))
4578 * If we are going to COW the mapping later, we examine the pending
4579 * reservations for this page now. This will ensure that any
4580 * allocations necessary to record that reservation occur outside the
4581 * spinlock. For private mappings, we also lookup the pagecache
4582 * page now as it is used to determine if a reservation has been
4585 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4586 if (vma_needs_reservation(h, vma, haddr) < 0) {
4590 /* Just decrements count, does not deallocate */
4591 vma_end_reservation(h, vma, haddr);
4593 if (!(vma->vm_flags & VM_MAYSHARE))
4594 pagecache_page = hugetlbfs_pagecache_page(h,
4598 ptl = huge_pte_lock(h, mm, ptep);
4600 /* Check for a racing update before calling hugetlb_cow */
4601 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4605 * hugetlb_cow() requires page locks of pte_page(entry) and
4606 * pagecache_page, so here we need take the former one
4607 * when page != pagecache_page or !pagecache_page.
4609 page = pte_page(entry);
4610 if (page != pagecache_page)
4611 if (!trylock_page(page)) {
4618 if (flags & FAULT_FLAG_WRITE) {
4619 if (!huge_pte_write(entry)) {
4620 ret = hugetlb_cow(mm, vma, address, ptep,
4621 pagecache_page, ptl);
4624 entry = huge_pte_mkdirty(entry);
4626 entry = pte_mkyoung(entry);
4627 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4628 flags & FAULT_FLAG_WRITE))
4629 update_mmu_cache(vma, haddr, ptep);
4631 if (page != pagecache_page)
4637 if (pagecache_page) {
4638 unlock_page(pagecache_page);
4639 put_page(pagecache_page);
4642 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4643 i_mmap_unlock_read(mapping);
4645 * Generally it's safe to hold refcount during waiting page lock. But
4646 * here we just wait to defer the next page fault to avoid busy loop and
4647 * the page is not used after unlocked before returning from the current
4648 * page fault. So we are safe from accessing freed page, even if we wait
4649 * here without taking refcount.
4652 wait_on_page_locked(page);
4657 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4658 * modifications for huge pages.
4660 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4662 struct vm_area_struct *dst_vma,
4663 unsigned long dst_addr,
4664 unsigned long src_addr,
4665 struct page **pagep)
4667 struct address_space *mapping;
4670 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4671 struct hstate *h = hstate_vma(dst_vma);
4679 page = alloc_huge_page(dst_vma, dst_addr, 0);
4683 ret = copy_huge_page_from_user(page,
4684 (const void __user *) src_addr,
4685 pages_per_huge_page(h), false);
4687 /* fallback to copy_from_user outside mmap_lock */
4688 if (unlikely(ret)) {
4691 /* don't free the page */
4700 * The memory barrier inside __SetPageUptodate makes sure that
4701 * preceding stores to the page contents become visible before
4702 * the set_pte_at() write.
4704 __SetPageUptodate(page);
4706 mapping = dst_vma->vm_file->f_mapping;
4707 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4710 * If shared, add to page cache
4713 size = i_size_read(mapping->host) >> huge_page_shift(h);
4716 goto out_release_nounlock;
4719 * Serialization between remove_inode_hugepages() and
4720 * huge_add_to_page_cache() below happens through the
4721 * hugetlb_fault_mutex_table that here must be hold by
4724 ret = huge_add_to_page_cache(page, mapping, idx);
4726 goto out_release_nounlock;
4729 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4733 * Recheck the i_size after holding PT lock to make sure not
4734 * to leave any page mapped (as page_mapped()) beyond the end
4735 * of the i_size (remove_inode_hugepages() is strict about
4736 * enforcing that). If we bail out here, we'll also leave a
4737 * page in the radix tree in the vm_shared case beyond the end
4738 * of the i_size, but remove_inode_hugepages() will take care
4739 * of it as soon as we drop the hugetlb_fault_mutex_table.
4741 size = i_size_read(mapping->host) >> huge_page_shift(h);
4744 goto out_release_unlock;
4747 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4748 goto out_release_unlock;
4751 page_dup_rmap(page, true);
4753 ClearPagePrivate(page);
4754 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4757 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4758 if (dst_vma->vm_flags & VM_WRITE)
4759 _dst_pte = huge_pte_mkdirty(_dst_pte);
4760 _dst_pte = pte_mkyoung(_dst_pte);
4762 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4764 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4765 dst_vma->vm_flags & VM_WRITE);
4766 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4768 /* No need to invalidate - it was non-present before */
4769 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4772 set_page_huge_active(page);
4782 out_release_nounlock:
4787 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4788 int refs, struct page **pages,
4789 struct vm_area_struct **vmas)
4793 for (nr = 0; nr < refs; nr++) {
4795 pages[nr] = mem_map_offset(page, nr);
4801 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4802 struct page **pages, struct vm_area_struct **vmas,
4803 unsigned long *position, unsigned long *nr_pages,
4804 long i, unsigned int flags, int *locked)
4806 unsigned long pfn_offset;
4807 unsigned long vaddr = *position;
4808 unsigned long remainder = *nr_pages;
4809 struct hstate *h = hstate_vma(vma);
4810 int err = -EFAULT, refs;
4812 while (vaddr < vma->vm_end && remainder) {
4814 spinlock_t *ptl = NULL;
4819 * If we have a pending SIGKILL, don't keep faulting pages and
4820 * potentially allocating memory.
4822 if (fatal_signal_pending(current)) {
4828 * Some archs (sparc64, sh*) have multiple pte_ts to
4829 * each hugepage. We have to make sure we get the
4830 * first, for the page indexing below to work.
4832 * Note that page table lock is not held when pte is null.
4834 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4837 ptl = huge_pte_lock(h, mm, pte);
4838 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4841 * When coredumping, it suits get_dump_page if we just return
4842 * an error where there's an empty slot with no huge pagecache
4843 * to back it. This way, we avoid allocating a hugepage, and
4844 * the sparse dumpfile avoids allocating disk blocks, but its
4845 * huge holes still show up with zeroes where they need to be.
4847 if (absent && (flags & FOLL_DUMP) &&
4848 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4856 * We need call hugetlb_fault for both hugepages under migration
4857 * (in which case hugetlb_fault waits for the migration,) and
4858 * hwpoisoned hugepages (in which case we need to prevent the
4859 * caller from accessing to them.) In order to do this, we use
4860 * here is_swap_pte instead of is_hugetlb_entry_migration and
4861 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4862 * both cases, and because we can't follow correct pages
4863 * directly from any kind of swap entries.
4865 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4866 ((flags & FOLL_WRITE) &&
4867 !huge_pte_write(huge_ptep_get(pte)))) {
4869 unsigned int fault_flags = 0;
4873 if (flags & FOLL_WRITE)
4874 fault_flags |= FAULT_FLAG_WRITE;
4876 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4877 FAULT_FLAG_KILLABLE;
4878 if (flags & FOLL_NOWAIT)
4879 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4880 FAULT_FLAG_RETRY_NOWAIT;
4881 if (flags & FOLL_TRIED) {
4883 * Note: FAULT_FLAG_ALLOW_RETRY and
4884 * FAULT_FLAG_TRIED can co-exist
4886 fault_flags |= FAULT_FLAG_TRIED;
4888 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4889 if (ret & VM_FAULT_ERROR) {
4890 err = vm_fault_to_errno(ret, flags);
4894 if (ret & VM_FAULT_RETRY) {
4896 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4900 * VM_FAULT_RETRY must not return an
4901 * error, it will return zero
4904 * No need to update "position" as the
4905 * caller will not check it after
4906 * *nr_pages is set to 0.
4913 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4914 page = pte_page(huge_ptep_get(pte));
4917 * If subpage information not requested, update counters
4918 * and skip the same_page loop below.
4920 if (!pages && !vmas && !pfn_offset &&
4921 (vaddr + huge_page_size(h) < vma->vm_end) &&
4922 (remainder >= pages_per_huge_page(h))) {
4923 vaddr += huge_page_size(h);
4924 remainder -= pages_per_huge_page(h);
4925 i += pages_per_huge_page(h);
4930 refs = min3(pages_per_huge_page(h) - pfn_offset,
4931 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4934 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4936 likely(pages) ? pages + i : NULL,
4937 vmas ? vmas + i : NULL);
4941 * try_grab_compound_head() should always succeed here,
4942 * because: a) we hold the ptl lock, and b) we've just
4943 * checked that the huge page is present in the page
4944 * tables. If the huge page is present, then the tail
4945 * pages must also be present. The ptl prevents the
4946 * head page and tail pages from being rearranged in
4947 * any way. So this page must be available at this
4948 * point, unless the page refcount overflowed:
4950 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
4960 vaddr += (refs << PAGE_SHIFT);
4966 *nr_pages = remainder;
4968 * setting position is actually required only if remainder is
4969 * not zero but it's faster not to add a "if (remainder)"
4977 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4979 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4982 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4985 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4986 unsigned long address, unsigned long end, pgprot_t newprot)
4988 struct mm_struct *mm = vma->vm_mm;
4989 unsigned long start = address;
4992 struct hstate *h = hstate_vma(vma);
4993 unsigned long pages = 0;
4994 bool shared_pmd = false;
4995 struct mmu_notifier_range range;
4998 * In the case of shared PMDs, the area to flush could be beyond
4999 * start/end. Set range.start/range.end to cover the maximum possible
5000 * range if PMD sharing is possible.
5002 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5003 0, vma, mm, start, end);
5004 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5006 BUG_ON(address >= end);
5007 flush_cache_range(vma, range.start, range.end);
5009 mmu_notifier_invalidate_range_start(&range);
5010 i_mmap_lock_write(vma->vm_file->f_mapping);
5011 for (; address < end; address += huge_page_size(h)) {
5013 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5016 ptl = huge_pte_lock(h, mm, ptep);
5017 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5023 pte = huge_ptep_get(ptep);
5024 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5028 if (unlikely(is_hugetlb_entry_migration(pte))) {
5029 swp_entry_t entry = pte_to_swp_entry(pte);
5031 if (is_write_migration_entry(entry)) {
5034 make_migration_entry_read(&entry);
5035 newpte = swp_entry_to_pte(entry);
5036 set_huge_swap_pte_at(mm, address, ptep,
5037 newpte, huge_page_size(h));
5043 if (!huge_pte_none(pte)) {
5046 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5047 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5048 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5049 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5055 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5056 * may have cleared our pud entry and done put_page on the page table:
5057 * once we release i_mmap_rwsem, another task can do the final put_page
5058 * and that page table be reused and filled with junk. If we actually
5059 * did unshare a page of pmds, flush the range corresponding to the pud.
5062 flush_hugetlb_tlb_range(vma, range.start, range.end);
5064 flush_hugetlb_tlb_range(vma, start, end);
5066 * No need to call mmu_notifier_invalidate_range() we are downgrading
5067 * page table protection not changing it to point to a new page.
5069 * See Documentation/vm/mmu_notifier.rst
5071 i_mmap_unlock_write(vma->vm_file->f_mapping);
5072 mmu_notifier_invalidate_range_end(&range);
5074 return pages << h->order;
5077 int hugetlb_reserve_pages(struct inode *inode,
5079 struct vm_area_struct *vma,
5080 vm_flags_t vm_flags)
5082 long ret, chg, add = -1;
5083 struct hstate *h = hstate_inode(inode);
5084 struct hugepage_subpool *spool = subpool_inode(inode);
5085 struct resv_map *resv_map;
5086 struct hugetlb_cgroup *h_cg = NULL;
5087 long gbl_reserve, regions_needed = 0;
5089 /* This should never happen */
5091 VM_WARN(1, "%s called with a negative range\n", __func__);
5096 * Only apply hugepage reservation if asked. At fault time, an
5097 * attempt will be made for VM_NORESERVE to allocate a page
5098 * without using reserves
5100 if (vm_flags & VM_NORESERVE)
5104 * Shared mappings base their reservation on the number of pages that
5105 * are already allocated on behalf of the file. Private mappings need
5106 * to reserve the full area even if read-only as mprotect() may be
5107 * called to make the mapping read-write. Assume !vma is a shm mapping
5109 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5111 * resv_map can not be NULL as hugetlb_reserve_pages is only
5112 * called for inodes for which resv_maps were created (see
5113 * hugetlbfs_get_inode).
5115 resv_map = inode_resv_map(inode);
5117 chg = region_chg(resv_map, from, to, ®ions_needed);
5120 /* Private mapping. */
5121 resv_map = resv_map_alloc();
5127 set_vma_resv_map(vma, resv_map);
5128 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5136 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5137 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5144 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5145 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5148 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5152 * There must be enough pages in the subpool for the mapping. If
5153 * the subpool has a minimum size, there may be some global
5154 * reservations already in place (gbl_reserve).
5156 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5157 if (gbl_reserve < 0) {
5159 goto out_uncharge_cgroup;
5163 * Check enough hugepages are available for the reservation.
5164 * Hand the pages back to the subpool if there are not
5166 ret = hugetlb_acct_memory(h, gbl_reserve);
5172 * Account for the reservations made. Shared mappings record regions
5173 * that have reservations as they are shared by multiple VMAs.
5174 * When the last VMA disappears, the region map says how much
5175 * the reservation was and the page cache tells how much of
5176 * the reservation was consumed. Private mappings are per-VMA and
5177 * only the consumed reservations are tracked. When the VMA
5178 * disappears, the original reservation is the VMA size and the
5179 * consumed reservations are stored in the map. Hence, nothing
5180 * else has to be done for private mappings here
5182 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5183 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5185 if (unlikely(add < 0)) {
5186 hugetlb_acct_memory(h, -gbl_reserve);
5189 } else if (unlikely(chg > add)) {
5191 * pages in this range were added to the reserve
5192 * map between region_chg and region_add. This
5193 * indicates a race with alloc_huge_page. Adjust
5194 * the subpool and reserve counts modified above
5195 * based on the difference.
5199 hugetlb_cgroup_uncharge_cgroup_rsvd(
5201 (chg - add) * pages_per_huge_page(h), h_cg);
5203 rsv_adjust = hugepage_subpool_put_pages(spool,
5205 hugetlb_acct_memory(h, -rsv_adjust);
5210 /* put back original number of pages, chg */
5211 (void)hugepage_subpool_put_pages(spool, chg);
5212 out_uncharge_cgroup:
5213 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5214 chg * pages_per_huge_page(h), h_cg);
5216 if (!vma || vma->vm_flags & VM_MAYSHARE)
5217 /* Only call region_abort if the region_chg succeeded but the
5218 * region_add failed or didn't run.
5220 if (chg >= 0 && add < 0)
5221 region_abort(resv_map, from, to, regions_needed);
5222 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5223 kref_put(&resv_map->refs, resv_map_release);
5227 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5230 struct hstate *h = hstate_inode(inode);
5231 struct resv_map *resv_map = inode_resv_map(inode);
5233 struct hugepage_subpool *spool = subpool_inode(inode);
5237 * Since this routine can be called in the evict inode path for all
5238 * hugetlbfs inodes, resv_map could be NULL.
5241 chg = region_del(resv_map, start, end);
5243 * region_del() can fail in the rare case where a region
5244 * must be split and another region descriptor can not be
5245 * allocated. If end == LONG_MAX, it will not fail.
5251 spin_lock(&inode->i_lock);
5252 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5253 spin_unlock(&inode->i_lock);
5256 * If the subpool has a minimum size, the number of global
5257 * reservations to be released may be adjusted.
5259 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5260 hugetlb_acct_memory(h, -gbl_reserve);
5265 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5266 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5267 struct vm_area_struct *vma,
5268 unsigned long addr, pgoff_t idx)
5270 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5272 unsigned long sbase = saddr & PUD_MASK;
5273 unsigned long s_end = sbase + PUD_SIZE;
5275 /* Allow segments to share if only one is marked locked */
5276 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5277 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5280 * match the virtual addresses, permission and the alignment of the
5283 if (pmd_index(addr) != pmd_index(saddr) ||
5284 vm_flags != svm_flags ||
5285 !range_in_vma(svma, sbase, s_end))
5291 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5293 unsigned long base = addr & PUD_MASK;
5294 unsigned long end = base + PUD_SIZE;
5297 * check on proper vm_flags and page table alignment
5299 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5305 * Determine if start,end range within vma could be mapped by shared pmd.
5306 * If yes, adjust start and end to cover range associated with possible
5307 * shared pmd mappings.
5309 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5310 unsigned long *start, unsigned long *end)
5312 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5313 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5316 * vma need span at least one aligned PUD size and the start,end range
5317 * must at least partialy within it.
5319 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5320 (*end <= v_start) || (*start >= v_end))
5323 /* Extend the range to be PUD aligned for a worst case scenario */
5324 if (*start > v_start)
5325 *start = ALIGN_DOWN(*start, PUD_SIZE);
5328 *end = ALIGN(*end, PUD_SIZE);
5332 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5333 * and returns the corresponding pte. While this is not necessary for the
5334 * !shared pmd case because we can allocate the pmd later as well, it makes the
5335 * code much cleaner.
5337 * This routine must be called with i_mmap_rwsem held in at least read mode if
5338 * sharing is possible. For hugetlbfs, this prevents removal of any page
5339 * table entries associated with the address space. This is important as we
5340 * are setting up sharing based on existing page table entries (mappings).
5342 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5343 * huge_pte_alloc know that sharing is not possible and do not take
5344 * i_mmap_rwsem as a performance optimization. This is handled by the
5345 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5346 * only required for subsequent processing.
5348 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5350 struct vm_area_struct *vma = find_vma(mm, addr);
5351 struct address_space *mapping = vma->vm_file->f_mapping;
5352 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5354 struct vm_area_struct *svma;
5355 unsigned long saddr;
5360 if (!vma_shareable(vma, addr))
5361 return (pte_t *)pmd_alloc(mm, pud, addr);
5363 i_mmap_assert_locked(mapping);
5364 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5368 saddr = page_table_shareable(svma, vma, addr, idx);
5370 spte = huge_pte_offset(svma->vm_mm, saddr,
5371 vma_mmu_pagesize(svma));
5373 get_page(virt_to_page(spte));
5382 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5383 if (pud_none(*pud)) {
5384 pud_populate(mm, pud,
5385 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5388 put_page(virt_to_page(spte));
5392 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5397 * unmap huge page backed by shared pte.
5399 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5400 * indicated by page_count > 1, unmap is achieved by clearing pud and
5401 * decrementing the ref count. If count == 1, the pte page is not shared.
5403 * Called with page table lock held and i_mmap_rwsem held in write mode.
5405 * returns: 1 successfully unmapped a shared pte page
5406 * 0 the underlying pte page is not shared, or it is the last user
5408 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5409 unsigned long *addr, pte_t *ptep)
5411 pgd_t *pgd = pgd_offset(mm, *addr);
5412 p4d_t *p4d = p4d_offset(pgd, *addr);
5413 pud_t *pud = pud_offset(p4d, *addr);
5415 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5416 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5417 if (page_count(virt_to_page(ptep)) == 1)
5421 put_page(virt_to_page(ptep));
5423 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5426 #define want_pmd_share() (1)
5427 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5428 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5433 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5434 unsigned long *addr, pte_t *ptep)
5439 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5440 unsigned long *start, unsigned long *end)
5443 #define want_pmd_share() (0)
5444 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5446 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5447 pte_t *huge_pte_alloc(struct mm_struct *mm,
5448 unsigned long addr, unsigned long sz)
5455 pgd = pgd_offset(mm, addr);
5456 p4d = p4d_alloc(mm, pgd, addr);
5459 pud = pud_alloc(mm, p4d, addr);
5461 if (sz == PUD_SIZE) {
5464 BUG_ON(sz != PMD_SIZE);
5465 if (want_pmd_share() && pud_none(*pud))
5466 pte = huge_pmd_share(mm, addr, pud);
5468 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5471 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5477 * huge_pte_offset() - Walk the page table to resolve the hugepage
5478 * entry at address @addr
5480 * Return: Pointer to page table entry (PUD or PMD) for
5481 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5482 * size @sz doesn't match the hugepage size at this level of the page
5485 pte_t *huge_pte_offset(struct mm_struct *mm,
5486 unsigned long addr, unsigned long sz)
5493 pgd = pgd_offset(mm, addr);
5494 if (!pgd_present(*pgd))
5496 p4d = p4d_offset(pgd, addr);
5497 if (!p4d_present(*p4d))
5500 pud = pud_offset(p4d, addr);
5502 /* must be pud huge, non-present or none */
5503 return (pte_t *)pud;
5504 if (!pud_present(*pud))
5506 /* must have a valid entry and size to go further */
5508 pmd = pmd_offset(pud, addr);
5509 /* must be pmd huge, non-present or none */
5510 return (pte_t *)pmd;
5513 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5516 * These functions are overwritable if your architecture needs its own
5519 struct page * __weak
5520 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5523 return ERR_PTR(-EINVAL);
5526 struct page * __weak
5527 follow_huge_pd(struct vm_area_struct *vma,
5528 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5530 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5534 struct page * __weak
5535 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5536 pmd_t *pmd, int flags)
5538 struct page *page = NULL;
5542 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5543 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5544 (FOLL_PIN | FOLL_GET)))
5548 ptl = pmd_lockptr(mm, pmd);
5551 * make sure that the address range covered by this pmd is not
5552 * unmapped from other threads.
5554 if (!pmd_huge(*pmd))
5556 pte = huge_ptep_get((pte_t *)pmd);
5557 if (pte_present(pte)) {
5558 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5560 * try_grab_page() should always succeed here, because: a) we
5561 * hold the pmd (ptl) lock, and b) we've just checked that the
5562 * huge pmd (head) page is present in the page tables. The ptl
5563 * prevents the head page and tail pages from being rearranged
5564 * in any way. So this page must be available at this point,
5565 * unless the page refcount overflowed:
5567 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5572 if (is_hugetlb_entry_migration(pte)) {
5574 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5578 * hwpoisoned entry is treated as no_page_table in
5579 * follow_page_mask().
5587 struct page * __weak
5588 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5589 pud_t *pud, int flags)
5591 if (flags & (FOLL_GET | FOLL_PIN))
5594 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5597 struct page * __weak
5598 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5600 if (flags & (FOLL_GET | FOLL_PIN))
5603 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5606 bool isolate_huge_page(struct page *page, struct list_head *list)
5610 spin_lock(&hugetlb_lock);
5611 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5612 !get_page_unless_zero(page)) {
5616 clear_page_huge_active(page);
5617 list_move_tail(&page->lru, list);
5619 spin_unlock(&hugetlb_lock);
5623 void putback_active_hugepage(struct page *page)
5625 VM_BUG_ON_PAGE(!PageHead(page), page);
5626 spin_lock(&hugetlb_lock);
5627 set_page_huge_active(page);
5628 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5629 spin_unlock(&hugetlb_lock);
5633 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5635 struct hstate *h = page_hstate(oldpage);
5637 hugetlb_cgroup_migrate(oldpage, newpage);
5638 set_page_owner_migrate_reason(newpage, reason);
5641 * transfer temporary state of the new huge page. This is
5642 * reverse to other transitions because the newpage is going to
5643 * be final while the old one will be freed so it takes over
5644 * the temporary status.
5646 * Also note that we have to transfer the per-node surplus state
5647 * here as well otherwise the global surplus count will not match
5650 if (PageHugeTemporary(newpage)) {
5651 int old_nid = page_to_nid(oldpage);
5652 int new_nid = page_to_nid(newpage);
5654 SetPageHugeTemporary(oldpage);
5655 ClearPageHugeTemporary(newpage);
5657 spin_lock(&hugetlb_lock);
5658 if (h->surplus_huge_pages_node[old_nid]) {
5659 h->surplus_huge_pages_node[old_nid]--;
5660 h->surplus_huge_pages_node[new_nid]++;
5662 spin_unlock(&hugetlb_lock);
5667 static bool cma_reserve_called __initdata;
5669 static int __init cmdline_parse_hugetlb_cma(char *p)
5671 hugetlb_cma_size = memparse(p, &p);
5675 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5677 void __init hugetlb_cma_reserve(int order)
5679 unsigned long size, reserved, per_node;
5682 cma_reserve_called = true;
5684 if (!hugetlb_cma_size)
5687 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5688 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5689 (PAGE_SIZE << order) / SZ_1M);
5694 * If 3 GB area is requested on a machine with 4 numa nodes,
5695 * let's allocate 1 GB on first three nodes and ignore the last one.
5697 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5698 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5699 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5702 for_each_node_state(nid, N_ONLINE) {
5704 char name[CMA_MAX_NAME];
5706 size = min(per_node, hugetlb_cma_size - reserved);
5707 size = round_up(size, PAGE_SIZE << order);
5709 snprintf(name, sizeof(name), "hugetlb%d", nid);
5710 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5712 &hugetlb_cma[nid], nid);
5714 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5720 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5723 if (reserved >= hugetlb_cma_size)
5728 void __init hugetlb_cma_check(void)
5730 if (!hugetlb_cma_size || cma_reserve_called)
5733 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5736 #endif /* CONFIG_CMA */