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/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
31 #include <linux/cma.h>
34 #include <asm/pgtable.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/node.h>
41 #include <linux/userfaultfd_k.h>
42 #include <linux/page_owner.h>
45 int hugetlb_max_hstate __read_mostly;
46 unsigned int default_hstate_idx;
47 struct hstate hstates[HUGE_MAX_HSTATE];
49 static struct cma *hugetlb_cma[MAX_NUMNODES];
52 * Minimum page order among possible hugepage sizes, set to a proper value
55 static unsigned int minimum_order __read_mostly = UINT_MAX;
57 __initdata LIST_HEAD(huge_boot_pages);
59 /* for command line parsing */
60 static struct hstate * __initdata parsed_hstate;
61 static unsigned long __initdata default_hstate_max_huge_pages;
62 static bool __initdata parsed_valid_hugepagesz = true;
63 static bool __initdata parsed_default_hugepagesz;
66 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
67 * free_huge_pages, and surplus_huge_pages.
69 DEFINE_SPINLOCK(hugetlb_lock);
72 * Serializes faults on the same logical page. This is used to
73 * prevent spurious OOMs when the hugepage pool is fully utilized.
75 static int num_fault_mutexes;
76 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
78 /* Forward declaration */
79 static int hugetlb_acct_memory(struct hstate *h, long delta);
81 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
83 bool free = (spool->count == 0) && (spool->used_hpages == 0);
85 spin_unlock(&spool->lock);
87 /* If no pages are used, and no other handles to the subpool
88 * remain, give up any reservations based on minimum size and
91 if (spool->min_hpages != -1)
92 hugetlb_acct_memory(spool->hstate,
98 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
101 struct hugepage_subpool *spool;
103 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
107 spin_lock_init(&spool->lock);
109 spool->max_hpages = max_hpages;
111 spool->min_hpages = min_hpages;
113 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
117 spool->rsv_hpages = min_hpages;
122 void hugepage_put_subpool(struct hugepage_subpool *spool)
124 spin_lock(&spool->lock);
125 BUG_ON(!spool->count);
127 unlock_or_release_subpool(spool);
131 * Subpool accounting for allocating and reserving pages.
132 * Return -ENOMEM if there are not enough resources to satisfy the
133 * the request. Otherwise, return the number of pages by which the
134 * global pools must be adjusted (upward). The returned value may
135 * only be different than the passed value (delta) in the case where
136 * a subpool minimum size must be maintained.
138 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
146 spin_lock(&spool->lock);
148 if (spool->max_hpages != -1) { /* maximum size accounting */
149 if ((spool->used_hpages + delta) <= spool->max_hpages)
150 spool->used_hpages += delta;
157 /* minimum size accounting */
158 if (spool->min_hpages != -1 && spool->rsv_hpages) {
159 if (delta > spool->rsv_hpages) {
161 * Asking for more reserves than those already taken on
162 * behalf of subpool. Return difference.
164 ret = delta - spool->rsv_hpages;
165 spool->rsv_hpages = 0;
167 ret = 0; /* reserves already accounted for */
168 spool->rsv_hpages -= delta;
173 spin_unlock(&spool->lock);
178 * Subpool accounting for freeing and unreserving pages.
179 * Return the number of global page reservations that must be dropped.
180 * The return value may only be different than the passed value (delta)
181 * in the case where a subpool minimum size must be maintained.
183 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
191 spin_lock(&spool->lock);
193 if (spool->max_hpages != -1) /* maximum size accounting */
194 spool->used_hpages -= delta;
196 /* minimum size accounting */
197 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
198 if (spool->rsv_hpages + delta <= spool->min_hpages)
201 ret = spool->rsv_hpages + delta - spool->min_hpages;
203 spool->rsv_hpages += delta;
204 if (spool->rsv_hpages > spool->min_hpages)
205 spool->rsv_hpages = spool->min_hpages;
209 * If hugetlbfs_put_super couldn't free spool due to an outstanding
210 * quota reference, free it now.
212 unlock_or_release_subpool(spool);
217 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
219 return HUGETLBFS_SB(inode->i_sb)->spool;
222 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
224 return subpool_inode(file_inode(vma->vm_file));
227 /* Helper that removes a struct file_region from the resv_map cache and returns
230 static struct file_region *
231 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
233 struct file_region *nrg = NULL;
235 VM_BUG_ON(resv->region_cache_count <= 0);
237 resv->region_cache_count--;
238 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
240 list_del(&nrg->link);
248 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
249 struct file_region *rg)
251 #ifdef CONFIG_CGROUP_HUGETLB
252 nrg->reservation_counter = rg->reservation_counter;
259 /* Helper that records hugetlb_cgroup uncharge info. */
260 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
262 struct resv_map *resv,
263 struct file_region *nrg)
265 #ifdef CONFIG_CGROUP_HUGETLB
267 nrg->reservation_counter =
268 &h_cg->rsvd_hugepage[hstate_index(h)];
269 nrg->css = &h_cg->css;
270 if (!resv->pages_per_hpage)
271 resv->pages_per_hpage = pages_per_huge_page(h);
272 /* pages_per_hpage should be the same for all entries in
275 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
277 nrg->reservation_counter = NULL;
283 static bool has_same_uncharge_info(struct file_region *rg,
284 struct file_region *org)
286 #ifdef CONFIG_CGROUP_HUGETLB
288 rg->reservation_counter == org->reservation_counter &&
296 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
298 struct file_region *nrg = NULL, *prg = NULL;
300 prg = list_prev_entry(rg, link);
301 if (&prg->link != &resv->regions && prg->to == rg->from &&
302 has_same_uncharge_info(prg, rg)) {
308 coalesce_file_region(resv, prg);
312 nrg = list_next_entry(rg, link);
313 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
314 has_same_uncharge_info(nrg, rg)) {
315 nrg->from = rg->from;
320 coalesce_file_region(resv, nrg);
325 /* Must be called with resv->lock held. Calling this with count_only == true
326 * will count the number of pages to be added but will not modify the linked
327 * list. If regions_needed != NULL and count_only == true, then regions_needed
328 * will indicate the number of file_regions needed in the cache to carry out to
329 * add the regions for this range.
331 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
332 struct hugetlb_cgroup *h_cg,
333 struct hstate *h, long *regions_needed,
337 struct list_head *head = &resv->regions;
338 long last_accounted_offset = f;
339 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
344 /* In this loop, we essentially handle an entry for the range
345 * [last_accounted_offset, rg->from), at every iteration, with some
348 list_for_each_entry_safe(rg, trg, head, link) {
349 /* Skip irrelevant regions that start before our range. */
351 /* If this region ends after the last accounted offset,
352 * then we need to update last_accounted_offset.
354 if (rg->to > last_accounted_offset)
355 last_accounted_offset = rg->to;
359 /* When we find a region that starts beyond our range, we've
365 /* Add an entry for last_accounted_offset -> rg->from, and
366 * update last_accounted_offset.
368 if (rg->from > last_accounted_offset) {
369 add += rg->from - last_accounted_offset;
371 nrg = get_file_region_entry_from_cache(
372 resv, last_accounted_offset, rg->from);
373 record_hugetlb_cgroup_uncharge_info(h_cg, h,
375 list_add(&nrg->link, rg->link.prev);
376 coalesce_file_region(resv, nrg);
377 } else if (regions_needed)
378 *regions_needed += 1;
381 last_accounted_offset = rg->to;
384 /* Handle the case where our range extends beyond
385 * last_accounted_offset.
387 if (last_accounted_offset < t) {
388 add += t - last_accounted_offset;
390 nrg = get_file_region_entry_from_cache(
391 resv, last_accounted_offset, t);
392 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
393 list_add(&nrg->link, rg->link.prev);
394 coalesce_file_region(resv, nrg);
395 } else if (regions_needed)
396 *regions_needed += 1;
403 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
405 static int allocate_file_region_entries(struct resv_map *resv,
407 __must_hold(&resv->lock)
409 struct list_head allocated_regions;
410 int to_allocate = 0, i = 0;
411 struct file_region *trg = NULL, *rg = NULL;
413 VM_BUG_ON(regions_needed < 0);
415 INIT_LIST_HEAD(&allocated_regions);
418 * Check for sufficient descriptors in the cache to accommodate
419 * the number of in progress add operations plus regions_needed.
421 * This is a while loop because when we drop the lock, some other call
422 * to region_add or region_del may have consumed some region_entries,
423 * so we keep looping here until we finally have enough entries for
424 * (adds_in_progress + regions_needed).
426 while (resv->region_cache_count <
427 (resv->adds_in_progress + regions_needed)) {
428 to_allocate = resv->adds_in_progress + regions_needed -
429 resv->region_cache_count;
431 /* At this point, we should have enough entries in the cache
432 * for all the existings adds_in_progress. We should only be
433 * needing to allocate for regions_needed.
435 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
437 spin_unlock(&resv->lock);
438 for (i = 0; i < to_allocate; i++) {
439 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
442 list_add(&trg->link, &allocated_regions);
445 spin_lock(&resv->lock);
447 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
449 list_add(&rg->link, &resv->region_cache);
450 resv->region_cache_count++;
457 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
465 * Add the huge page range represented by [f, t) to the reserve
466 * map. Regions will be taken from the cache to fill in this range.
467 * Sufficient regions should exist in the cache due to the previous
468 * call to region_chg with the same range, but in some cases the cache will not
469 * have sufficient entries due to races with other code doing region_add or
470 * region_del. The extra needed entries will be allocated.
472 * regions_needed is the out value provided by a previous call to region_chg.
474 * Return the number of new huge pages added to the map. This number is greater
475 * than or equal to zero. If file_region entries needed to be allocated for
476 * this operation and we were not able to allocate, it returns -ENOMEM.
477 * region_add of regions of length 1 never allocate file_regions and cannot
478 * fail; region_chg will always allocate at least 1 entry and a region_add for
479 * 1 page will only require at most 1 entry.
481 static long region_add(struct resv_map *resv, long f, long t,
482 long in_regions_needed, struct hstate *h,
483 struct hugetlb_cgroup *h_cg)
485 long add = 0, actual_regions_needed = 0;
487 spin_lock(&resv->lock);
490 /* Count how many regions are actually needed to execute this add. */
491 add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed,
495 * Check for sufficient descriptors in the cache to accommodate
496 * this add operation. Note that actual_regions_needed may be greater
497 * than in_regions_needed, as the resv_map may have been modified since
498 * the region_chg call. In this case, we need to make sure that we
499 * allocate extra entries, such that we have enough for all the
500 * existing adds_in_progress, plus the excess needed for this
503 if (actual_regions_needed > in_regions_needed &&
504 resv->region_cache_count <
505 resv->adds_in_progress +
506 (actual_regions_needed - in_regions_needed)) {
507 /* region_add operation of range 1 should never need to
508 * allocate file_region entries.
510 VM_BUG_ON(t - f <= 1);
512 if (allocate_file_region_entries(
513 resv, actual_regions_needed - in_regions_needed)) {
520 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false);
522 resv->adds_in_progress -= in_regions_needed;
524 spin_unlock(&resv->lock);
530 * Examine the existing reserve map and determine how many
531 * huge pages in the specified range [f, t) are NOT currently
532 * represented. This routine is called before a subsequent
533 * call to region_add that will actually modify the reserve
534 * map to add the specified range [f, t). region_chg does
535 * not change the number of huge pages represented by the
536 * map. A number of new file_region structures is added to the cache as a
537 * placeholder, for the subsequent region_add call to use. At least 1
538 * file_region structure is added.
540 * out_regions_needed is the number of regions added to the
541 * resv->adds_in_progress. This value needs to be provided to a follow up call
542 * to region_add or region_abort for proper accounting.
544 * Returns the number of huge pages that need to be added to the existing
545 * reservation map for the range [f, t). This number is greater or equal to
546 * zero. -ENOMEM is returned if a new file_region structure or cache entry
547 * is needed and can not be allocated.
549 static long region_chg(struct resv_map *resv, long f, long t,
550 long *out_regions_needed)
554 spin_lock(&resv->lock);
556 /* Count how many hugepages in this range are NOT respresented. */
557 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
558 out_regions_needed, true);
560 if (*out_regions_needed == 0)
561 *out_regions_needed = 1;
563 if (allocate_file_region_entries(resv, *out_regions_needed))
566 resv->adds_in_progress += *out_regions_needed;
568 spin_unlock(&resv->lock);
573 * Abort the in progress add operation. The adds_in_progress field
574 * of the resv_map keeps track of the operations in progress between
575 * calls to region_chg and region_add. Operations are sometimes
576 * aborted after the call to region_chg. In such cases, region_abort
577 * is called to decrement the adds_in_progress counter. regions_needed
578 * is the value returned by the region_chg call, it is used to decrement
579 * the adds_in_progress counter.
581 * NOTE: The range arguments [f, t) are not needed or used in this
582 * routine. They are kept to make reading the calling code easier as
583 * arguments will match the associated region_chg call.
585 static void region_abort(struct resv_map *resv, long f, long t,
588 spin_lock(&resv->lock);
589 VM_BUG_ON(!resv->region_cache_count);
590 resv->adds_in_progress -= regions_needed;
591 spin_unlock(&resv->lock);
595 * Delete the specified range [f, t) from the reserve map. If the
596 * t parameter is LONG_MAX, this indicates that ALL regions after f
597 * should be deleted. Locate the regions which intersect [f, t)
598 * and either trim, delete or split the existing regions.
600 * Returns the number of huge pages deleted from the reserve map.
601 * In the normal case, the return value is zero or more. In the
602 * case where a region must be split, a new region descriptor must
603 * be allocated. If the allocation fails, -ENOMEM will be returned.
604 * NOTE: If the parameter t == LONG_MAX, then we will never split
605 * a region and possibly return -ENOMEM. Callers specifying
606 * t == LONG_MAX do not need to check for -ENOMEM error.
608 static long region_del(struct resv_map *resv, long f, long t)
610 struct list_head *head = &resv->regions;
611 struct file_region *rg, *trg;
612 struct file_region *nrg = NULL;
616 spin_lock(&resv->lock);
617 list_for_each_entry_safe(rg, trg, head, link) {
619 * Skip regions before the range to be deleted. file_region
620 * ranges are normally of the form [from, to). However, there
621 * may be a "placeholder" entry in the map which is of the form
622 * (from, to) with from == to. Check for placeholder entries
623 * at the beginning of the range to be deleted.
625 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
631 if (f > rg->from && t < rg->to) { /* Must split region */
633 * Check for an entry in the cache before dropping
634 * lock and attempting allocation.
637 resv->region_cache_count > resv->adds_in_progress) {
638 nrg = list_first_entry(&resv->region_cache,
641 list_del(&nrg->link);
642 resv->region_cache_count--;
646 spin_unlock(&resv->lock);
647 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
655 /* New entry for end of split region */
659 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
661 INIT_LIST_HEAD(&nrg->link);
663 /* Original entry is trimmed */
666 hugetlb_cgroup_uncharge_file_region(
667 resv, rg, nrg->to - nrg->from);
669 list_add(&nrg->link, &rg->link);
674 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
675 del += rg->to - rg->from;
676 hugetlb_cgroup_uncharge_file_region(resv, rg,
683 if (f <= rg->from) { /* Trim beginning of region */
687 hugetlb_cgroup_uncharge_file_region(resv, rg,
689 } else { /* Trim end of region */
693 hugetlb_cgroup_uncharge_file_region(resv, rg,
698 spin_unlock(&resv->lock);
704 * A rare out of memory error was encountered which prevented removal of
705 * the reserve map region for a page. The huge page itself was free'ed
706 * and removed from the page cache. This routine will adjust the subpool
707 * usage count, and the global reserve count if needed. By incrementing
708 * these counts, the reserve map entry which could not be deleted will
709 * appear as a "reserved" entry instead of simply dangling with incorrect
712 void hugetlb_fix_reserve_counts(struct inode *inode)
714 struct hugepage_subpool *spool = subpool_inode(inode);
717 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
719 struct hstate *h = hstate_inode(inode);
721 hugetlb_acct_memory(h, 1);
726 * Count and return the number of huge pages in the reserve map
727 * that intersect with the range [f, t).
729 static long region_count(struct resv_map *resv, long f, long t)
731 struct list_head *head = &resv->regions;
732 struct file_region *rg;
735 spin_lock(&resv->lock);
736 /* Locate each segment we overlap with, and count that overlap. */
737 list_for_each_entry(rg, head, link) {
746 seg_from = max(rg->from, f);
747 seg_to = min(rg->to, t);
749 chg += seg_to - seg_from;
751 spin_unlock(&resv->lock);
757 * Convert the address within this vma to the page offset within
758 * the mapping, in pagecache page units; huge pages here.
760 static pgoff_t vma_hugecache_offset(struct hstate *h,
761 struct vm_area_struct *vma, unsigned long address)
763 return ((address - vma->vm_start) >> huge_page_shift(h)) +
764 (vma->vm_pgoff >> huge_page_order(h));
767 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
768 unsigned long address)
770 return vma_hugecache_offset(hstate_vma(vma), vma, address);
772 EXPORT_SYMBOL_GPL(linear_hugepage_index);
775 * Return the size of the pages allocated when backing a VMA. In the majority
776 * cases this will be same size as used by the page table entries.
778 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
780 if (vma->vm_ops && vma->vm_ops->pagesize)
781 return vma->vm_ops->pagesize(vma);
784 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
787 * Return the page size being used by the MMU to back a VMA. In the majority
788 * of cases, the page size used by the kernel matches the MMU size. On
789 * architectures where it differs, an architecture-specific 'strong'
790 * version of this symbol is required.
792 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
794 return vma_kernel_pagesize(vma);
798 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
799 * bits of the reservation map pointer, which are always clear due to
802 #define HPAGE_RESV_OWNER (1UL << 0)
803 #define HPAGE_RESV_UNMAPPED (1UL << 1)
804 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
807 * These helpers are used to track how many pages are reserved for
808 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
809 * is guaranteed to have their future faults succeed.
811 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
812 * the reserve counters are updated with the hugetlb_lock held. It is safe
813 * to reset the VMA at fork() time as it is not in use yet and there is no
814 * chance of the global counters getting corrupted as a result of the values.
816 * The private mapping reservation is represented in a subtly different
817 * manner to a shared mapping. A shared mapping has a region map associated
818 * with the underlying file, this region map represents the backing file
819 * pages which have ever had a reservation assigned which this persists even
820 * after the page is instantiated. A private mapping has a region map
821 * associated with the original mmap which is attached to all VMAs which
822 * reference it, this region map represents those offsets which have consumed
823 * reservation ie. where pages have been instantiated.
825 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
827 return (unsigned long)vma->vm_private_data;
830 static void set_vma_private_data(struct vm_area_struct *vma,
833 vma->vm_private_data = (void *)value;
837 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
838 struct hugetlb_cgroup *h_cg,
841 #ifdef CONFIG_CGROUP_HUGETLB
843 resv_map->reservation_counter = NULL;
844 resv_map->pages_per_hpage = 0;
845 resv_map->css = NULL;
847 resv_map->reservation_counter =
848 &h_cg->rsvd_hugepage[hstate_index(h)];
849 resv_map->pages_per_hpage = pages_per_huge_page(h);
850 resv_map->css = &h_cg->css;
855 struct resv_map *resv_map_alloc(void)
857 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
858 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
860 if (!resv_map || !rg) {
866 kref_init(&resv_map->refs);
867 spin_lock_init(&resv_map->lock);
868 INIT_LIST_HEAD(&resv_map->regions);
870 resv_map->adds_in_progress = 0;
872 * Initialize these to 0. On shared mappings, 0's here indicate these
873 * fields don't do cgroup accounting. On private mappings, these will be
874 * re-initialized to the proper values, to indicate that hugetlb cgroup
875 * reservations are to be un-charged from here.
877 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
879 INIT_LIST_HEAD(&resv_map->region_cache);
880 list_add(&rg->link, &resv_map->region_cache);
881 resv_map->region_cache_count = 1;
886 void resv_map_release(struct kref *ref)
888 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
889 struct list_head *head = &resv_map->region_cache;
890 struct file_region *rg, *trg;
892 /* Clear out any active regions before we release the map. */
893 region_del(resv_map, 0, LONG_MAX);
895 /* ... and any entries left in the cache */
896 list_for_each_entry_safe(rg, trg, head, link) {
901 VM_BUG_ON(resv_map->adds_in_progress);
906 static inline struct resv_map *inode_resv_map(struct inode *inode)
909 * At inode evict time, i_mapping may not point to the original
910 * address space within the inode. This original address space
911 * contains the pointer to the resv_map. So, always use the
912 * address space embedded within the inode.
913 * The VERY common case is inode->mapping == &inode->i_data but,
914 * this may not be true for device special inodes.
916 return (struct resv_map *)(&inode->i_data)->private_data;
919 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
921 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
922 if (vma->vm_flags & VM_MAYSHARE) {
923 struct address_space *mapping = vma->vm_file->f_mapping;
924 struct inode *inode = mapping->host;
926 return inode_resv_map(inode);
929 return (struct resv_map *)(get_vma_private_data(vma) &
934 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
936 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
937 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
939 set_vma_private_data(vma, (get_vma_private_data(vma) &
940 HPAGE_RESV_MASK) | (unsigned long)map);
943 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
945 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
946 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
948 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
951 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
953 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
955 return (get_vma_private_data(vma) & flag) != 0;
958 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
959 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
961 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
962 if (!(vma->vm_flags & VM_MAYSHARE))
963 vma->vm_private_data = (void *)0;
966 /* Returns true if the VMA has associated reserve pages */
967 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
969 if (vma->vm_flags & VM_NORESERVE) {
971 * This address is already reserved by other process(chg == 0),
972 * so, we should decrement reserved count. Without decrementing,
973 * reserve count remains after releasing inode, because this
974 * allocated page will go into page cache and is regarded as
975 * coming from reserved pool in releasing step. Currently, we
976 * don't have any other solution to deal with this situation
977 * properly, so add work-around here.
979 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
985 /* Shared mappings always use reserves */
986 if (vma->vm_flags & VM_MAYSHARE) {
988 * We know VM_NORESERVE is not set. Therefore, there SHOULD
989 * be a region map for all pages. The only situation where
990 * there is no region map is if a hole was punched via
991 * fallocate. In this case, there really are no reserves to
992 * use. This situation is indicated if chg != 0.
1001 * Only the process that called mmap() has reserves for
1004 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1006 * Like the shared case above, a hole punch or truncate
1007 * could have been performed on the private mapping.
1008 * Examine the value of chg to determine if reserves
1009 * actually exist or were previously consumed.
1010 * Very Subtle - The value of chg comes from a previous
1011 * call to vma_needs_reserves(). The reserve map for
1012 * private mappings has different (opposite) semantics
1013 * than that of shared mappings. vma_needs_reserves()
1014 * has already taken this difference in semantics into
1015 * account. Therefore, the meaning of chg is the same
1016 * as in the shared case above. Code could easily be
1017 * combined, but keeping it separate draws attention to
1018 * subtle differences.
1029 static void enqueue_huge_page(struct hstate *h, struct page *page)
1031 int nid = page_to_nid(page);
1032 list_move(&page->lru, &h->hugepage_freelists[nid]);
1033 h->free_huge_pages++;
1034 h->free_huge_pages_node[nid]++;
1037 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1041 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
1042 if (!PageHWPoison(page))
1045 * if 'non-isolated free hugepage' not found on the list,
1046 * the allocation fails.
1048 if (&h->hugepage_freelists[nid] == &page->lru)
1050 list_move(&page->lru, &h->hugepage_activelist);
1051 set_page_refcounted(page);
1052 h->free_huge_pages--;
1053 h->free_huge_pages_node[nid]--;
1057 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1060 unsigned int cpuset_mems_cookie;
1061 struct zonelist *zonelist;
1064 int node = NUMA_NO_NODE;
1066 zonelist = node_zonelist(nid, gfp_mask);
1069 cpuset_mems_cookie = read_mems_allowed_begin();
1070 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1073 if (!cpuset_zone_allowed(zone, gfp_mask))
1076 * no need to ask again on the same node. Pool is node rather than
1079 if (zone_to_nid(zone) == node)
1081 node = zone_to_nid(zone);
1083 page = dequeue_huge_page_node_exact(h, node);
1087 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1093 /* Movability of hugepages depends on migration support. */
1094 static inline gfp_t htlb_alloc_mask(struct hstate *h)
1096 if (hugepage_movable_supported(h))
1097 return GFP_HIGHUSER_MOVABLE;
1099 return GFP_HIGHUSER;
1102 static struct page *dequeue_huge_page_vma(struct hstate *h,
1103 struct vm_area_struct *vma,
1104 unsigned long address, int avoid_reserve,
1108 struct mempolicy *mpol;
1110 nodemask_t *nodemask;
1114 * A child process with MAP_PRIVATE mappings created by their parent
1115 * have no page reserves. This check ensures that reservations are
1116 * not "stolen". The child may still get SIGKILLed
1118 if (!vma_has_reserves(vma, chg) &&
1119 h->free_huge_pages - h->resv_huge_pages == 0)
1122 /* If reserves cannot be used, ensure enough pages are in the pool */
1123 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1126 gfp_mask = htlb_alloc_mask(h);
1127 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1128 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1129 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1130 SetPagePrivate(page);
1131 h->resv_huge_pages--;
1134 mpol_cond_put(mpol);
1142 * common helper functions for hstate_next_node_to_{alloc|free}.
1143 * We may have allocated or freed a huge page based on a different
1144 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1145 * be outside of *nodes_allowed. Ensure that we use an allowed
1146 * node for alloc or free.
1148 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1150 nid = next_node_in(nid, *nodes_allowed);
1151 VM_BUG_ON(nid >= MAX_NUMNODES);
1156 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1158 if (!node_isset(nid, *nodes_allowed))
1159 nid = next_node_allowed(nid, nodes_allowed);
1164 * returns the previously saved node ["this node"] from which to
1165 * allocate a persistent huge page for the pool and advance the
1166 * next node from which to allocate, handling wrap at end of node
1169 static int hstate_next_node_to_alloc(struct hstate *h,
1170 nodemask_t *nodes_allowed)
1174 VM_BUG_ON(!nodes_allowed);
1176 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1177 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1183 * helper for free_pool_huge_page() - return the previously saved
1184 * node ["this node"] from which to free a huge page. Advance the
1185 * next node id whether or not we find a free huge page to free so
1186 * that the next attempt to free addresses the next node.
1188 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1192 VM_BUG_ON(!nodes_allowed);
1194 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1195 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1200 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1201 for (nr_nodes = nodes_weight(*mask); \
1203 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1206 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1207 for (nr_nodes = nodes_weight(*mask); \
1209 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1212 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1213 static void destroy_compound_gigantic_page(struct page *page,
1217 int nr_pages = 1 << order;
1218 struct page *p = page + 1;
1220 atomic_set(compound_mapcount_ptr(page), 0);
1221 if (hpage_pincount_available(page))
1222 atomic_set(compound_pincount_ptr(page), 0);
1224 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1225 clear_compound_head(p);
1226 set_page_refcounted(p);
1229 set_compound_order(page, 0);
1230 __ClearPageHead(page);
1233 static void free_gigantic_page(struct page *page, unsigned int order)
1236 * If the page isn't allocated using the cma allocator,
1237 * cma_release() returns false.
1239 if (IS_ENABLED(CONFIG_CMA) &&
1240 cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1243 free_contig_range(page_to_pfn(page), 1 << order);
1246 #ifdef CONFIG_CONTIG_ALLOC
1247 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1248 int nid, nodemask_t *nodemask)
1250 unsigned long nr_pages = 1UL << huge_page_order(h);
1252 if (IS_ENABLED(CONFIG_CMA)) {
1256 for_each_node_mask(node, *nodemask) {
1257 if (!hugetlb_cma[node])
1260 page = cma_alloc(hugetlb_cma[node], nr_pages,
1261 huge_page_order(h), true);
1267 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1270 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1271 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1272 #else /* !CONFIG_CONTIG_ALLOC */
1273 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1274 int nid, nodemask_t *nodemask)
1278 #endif /* CONFIG_CONTIG_ALLOC */
1280 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1281 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1282 int nid, nodemask_t *nodemask)
1286 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1287 static inline void destroy_compound_gigantic_page(struct page *page,
1288 unsigned int order) { }
1291 static void update_and_free_page(struct hstate *h, struct page *page)
1295 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1299 h->nr_huge_pages_node[page_to_nid(page)]--;
1300 for (i = 0; i < pages_per_huge_page(h); i++) {
1301 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1302 1 << PG_referenced | 1 << PG_dirty |
1303 1 << PG_active | 1 << PG_private |
1306 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1307 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1308 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1309 set_page_refcounted(page);
1310 if (hstate_is_gigantic(h)) {
1312 * Temporarily drop the hugetlb_lock, because
1313 * we might block in free_gigantic_page().
1315 spin_unlock(&hugetlb_lock);
1316 destroy_compound_gigantic_page(page, huge_page_order(h));
1317 free_gigantic_page(page, huge_page_order(h));
1318 spin_lock(&hugetlb_lock);
1320 __free_pages(page, huge_page_order(h));
1324 struct hstate *size_to_hstate(unsigned long size)
1328 for_each_hstate(h) {
1329 if (huge_page_size(h) == size)
1336 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1337 * to hstate->hugepage_activelist.)
1339 * This function can be called for tail pages, but never returns true for them.
1341 bool page_huge_active(struct page *page)
1343 VM_BUG_ON_PAGE(!PageHuge(page), page);
1344 return PageHead(page) && PagePrivate(&page[1]);
1347 /* never called for tail page */
1348 static void set_page_huge_active(struct page *page)
1350 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1351 SetPagePrivate(&page[1]);
1354 static void clear_page_huge_active(struct page *page)
1356 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1357 ClearPagePrivate(&page[1]);
1361 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1364 static inline bool PageHugeTemporary(struct page *page)
1366 if (!PageHuge(page))
1369 return (unsigned long)page[2].mapping == -1U;
1372 static inline void SetPageHugeTemporary(struct page *page)
1374 page[2].mapping = (void *)-1U;
1377 static inline void ClearPageHugeTemporary(struct page *page)
1379 page[2].mapping = NULL;
1382 static void __free_huge_page(struct page *page)
1385 * Can't pass hstate in here because it is called from the
1386 * compound page destructor.
1388 struct hstate *h = page_hstate(page);
1389 int nid = page_to_nid(page);
1390 struct hugepage_subpool *spool =
1391 (struct hugepage_subpool *)page_private(page);
1392 bool restore_reserve;
1394 VM_BUG_ON_PAGE(page_count(page), page);
1395 VM_BUG_ON_PAGE(page_mapcount(page), page);
1397 set_page_private(page, 0);
1398 page->mapping = NULL;
1399 restore_reserve = PagePrivate(page);
1400 ClearPagePrivate(page);
1403 * If PagePrivate() was set on page, page allocation consumed a
1404 * reservation. If the page was associated with a subpool, there
1405 * would have been a page reserved in the subpool before allocation
1406 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1407 * reservtion, do not call hugepage_subpool_put_pages() as this will
1408 * remove the reserved page from the subpool.
1410 if (!restore_reserve) {
1412 * A return code of zero implies that the subpool will be
1413 * under its minimum size if the reservation is not restored
1414 * after page is free. Therefore, force restore_reserve
1417 if (hugepage_subpool_put_pages(spool, 1) == 0)
1418 restore_reserve = true;
1421 spin_lock(&hugetlb_lock);
1422 clear_page_huge_active(page);
1423 hugetlb_cgroup_uncharge_page(hstate_index(h),
1424 pages_per_huge_page(h), page);
1425 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1426 pages_per_huge_page(h), page);
1427 if (restore_reserve)
1428 h->resv_huge_pages++;
1430 if (PageHugeTemporary(page)) {
1431 list_del(&page->lru);
1432 ClearPageHugeTemporary(page);
1433 update_and_free_page(h, page);
1434 } else if (h->surplus_huge_pages_node[nid]) {
1435 /* remove the page from active list */
1436 list_del(&page->lru);
1437 update_and_free_page(h, page);
1438 h->surplus_huge_pages--;
1439 h->surplus_huge_pages_node[nid]--;
1441 arch_clear_hugepage_flags(page);
1442 enqueue_huge_page(h, page);
1444 spin_unlock(&hugetlb_lock);
1448 * As free_huge_page() can be called from a non-task context, we have
1449 * to defer the actual freeing in a workqueue to prevent potential
1450 * hugetlb_lock deadlock.
1452 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1453 * be freed and frees them one-by-one. As the page->mapping pointer is
1454 * going to be cleared in __free_huge_page() anyway, it is reused as the
1455 * llist_node structure of a lockless linked list of huge pages to be freed.
1457 static LLIST_HEAD(hpage_freelist);
1459 static void free_hpage_workfn(struct work_struct *work)
1461 struct llist_node *node;
1464 node = llist_del_all(&hpage_freelist);
1467 page = container_of((struct address_space **)node,
1468 struct page, mapping);
1470 __free_huge_page(page);
1473 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1475 void free_huge_page(struct page *page)
1478 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1482 * Only call schedule_work() if hpage_freelist is previously
1483 * empty. Otherwise, schedule_work() had been called but the
1484 * workfn hasn't retrieved the list yet.
1486 if (llist_add((struct llist_node *)&page->mapping,
1488 schedule_work(&free_hpage_work);
1492 __free_huge_page(page);
1495 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1497 INIT_LIST_HEAD(&page->lru);
1498 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1499 spin_lock(&hugetlb_lock);
1500 set_hugetlb_cgroup(page, NULL);
1501 set_hugetlb_cgroup_rsvd(page, NULL);
1503 h->nr_huge_pages_node[nid]++;
1504 spin_unlock(&hugetlb_lock);
1507 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1510 int nr_pages = 1 << order;
1511 struct page *p = page + 1;
1513 /* we rely on prep_new_huge_page to set the destructor */
1514 set_compound_order(page, order);
1515 __ClearPageReserved(page);
1516 __SetPageHead(page);
1517 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1519 * For gigantic hugepages allocated through bootmem at
1520 * boot, it's safer to be consistent with the not-gigantic
1521 * hugepages and clear the PG_reserved bit from all tail pages
1522 * too. Otherwise drivers using get_user_pages() to access tail
1523 * pages may get the reference counting wrong if they see
1524 * PG_reserved set on a tail page (despite the head page not
1525 * having PG_reserved set). Enforcing this consistency between
1526 * head and tail pages allows drivers to optimize away a check
1527 * on the head page when they need know if put_page() is needed
1528 * after get_user_pages().
1530 __ClearPageReserved(p);
1531 set_page_count(p, 0);
1532 set_compound_head(p, page);
1534 atomic_set(compound_mapcount_ptr(page), -1);
1536 if (hpage_pincount_available(page))
1537 atomic_set(compound_pincount_ptr(page), 0);
1541 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1542 * transparent huge pages. See the PageTransHuge() documentation for more
1545 int PageHuge(struct page *page)
1547 if (!PageCompound(page))
1550 page = compound_head(page);
1551 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1553 EXPORT_SYMBOL_GPL(PageHuge);
1556 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1557 * normal or transparent huge pages.
1559 int PageHeadHuge(struct page *page_head)
1561 if (!PageHead(page_head))
1564 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1568 * Find address_space associated with hugetlbfs page.
1569 * Upon entry page is locked and page 'was' mapped although mapped state
1570 * could change. If necessary, use anon_vma to find vma and associated
1571 * address space. The returned mapping may be stale, but it can not be
1572 * invalid as page lock (which is held) is required to destroy mapping.
1574 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1576 struct anon_vma *anon_vma;
1577 pgoff_t pgoff_start, pgoff_end;
1578 struct anon_vma_chain *avc;
1579 struct address_space *mapping = page_mapping(hpage);
1581 /* Simple file based mapping */
1586 * Even anonymous hugetlbfs mappings are associated with an
1587 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1588 * code). Find a vma associated with the anonymous vma, and
1589 * use the file pointer to get address_space.
1591 anon_vma = page_lock_anon_vma_read(hpage);
1593 return mapping; /* NULL */
1595 /* Use first found vma */
1596 pgoff_start = page_to_pgoff(hpage);
1597 pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1;
1598 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1599 pgoff_start, pgoff_end) {
1600 struct vm_area_struct *vma = avc->vma;
1602 mapping = vma->vm_file->f_mapping;
1606 anon_vma_unlock_read(anon_vma);
1611 * Find and lock address space (mapping) in write mode.
1613 * Upon entry, the page is locked which allows us to find the mapping
1614 * even in the case of an anon page. However, locking order dictates
1615 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1616 * specific. So, we first try to lock the sema while still holding the
1617 * page lock. If this works, great! If not, then we need to drop the
1618 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1619 * course, need to revalidate state along the way.
1621 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1623 struct address_space *mapping, *mapping2;
1625 mapping = _get_hugetlb_page_mapping(hpage);
1631 * If no contention, take lock and return
1633 if (i_mmap_trylock_write(mapping))
1637 * Must drop page lock and wait on mapping sema.
1638 * Note: Once page lock is dropped, mapping could become invalid.
1639 * As a hack, increase map count until we lock page again.
1641 atomic_inc(&hpage->_mapcount);
1643 i_mmap_lock_write(mapping);
1645 atomic_add_negative(-1, &hpage->_mapcount);
1647 /* verify page is still mapped */
1648 if (!page_mapped(hpage)) {
1649 i_mmap_unlock_write(mapping);
1654 * Get address space again and verify it is the same one
1655 * we locked. If not, drop lock and retry.
1657 mapping2 = _get_hugetlb_page_mapping(hpage);
1658 if (mapping2 != mapping) {
1659 i_mmap_unlock_write(mapping);
1667 pgoff_t __basepage_index(struct page *page)
1669 struct page *page_head = compound_head(page);
1670 pgoff_t index = page_index(page_head);
1671 unsigned long compound_idx;
1673 if (!PageHuge(page_head))
1674 return page_index(page);
1676 if (compound_order(page_head) >= MAX_ORDER)
1677 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1679 compound_idx = page - page_head;
1681 return (index << compound_order(page_head)) + compound_idx;
1684 static struct page *alloc_buddy_huge_page(struct hstate *h,
1685 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1686 nodemask_t *node_alloc_noretry)
1688 int order = huge_page_order(h);
1690 bool alloc_try_hard = true;
1693 * By default we always try hard to allocate the page with
1694 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1695 * a loop (to adjust global huge page counts) and previous allocation
1696 * failed, do not continue to try hard on the same node. Use the
1697 * node_alloc_noretry bitmap to manage this state information.
1699 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1700 alloc_try_hard = false;
1701 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1703 gfp_mask |= __GFP_RETRY_MAYFAIL;
1704 if (nid == NUMA_NO_NODE)
1705 nid = numa_mem_id();
1706 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1708 __count_vm_event(HTLB_BUDDY_PGALLOC);
1710 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1713 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1714 * indicates an overall state change. Clear bit so that we resume
1715 * normal 'try hard' allocations.
1717 if (node_alloc_noretry && page && !alloc_try_hard)
1718 node_clear(nid, *node_alloc_noretry);
1721 * If we tried hard to get a page but failed, set bit so that
1722 * subsequent attempts will not try as hard until there is an
1723 * overall state change.
1725 if (node_alloc_noretry && !page && alloc_try_hard)
1726 node_set(nid, *node_alloc_noretry);
1732 * Common helper to allocate a fresh hugetlb page. All specific allocators
1733 * should use this function to get new hugetlb pages
1735 static struct page *alloc_fresh_huge_page(struct hstate *h,
1736 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1737 nodemask_t *node_alloc_noretry)
1741 if (hstate_is_gigantic(h))
1742 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1744 page = alloc_buddy_huge_page(h, gfp_mask,
1745 nid, nmask, node_alloc_noretry);
1749 if (hstate_is_gigantic(h))
1750 prep_compound_gigantic_page(page, huge_page_order(h));
1751 prep_new_huge_page(h, page, page_to_nid(page));
1757 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1760 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1761 nodemask_t *node_alloc_noretry)
1765 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1767 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1768 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1769 node_alloc_noretry);
1777 put_page(page); /* free it into the hugepage allocator */
1783 * Free huge page from pool from next node to free.
1784 * Attempt to keep persistent huge pages more or less
1785 * balanced over allowed nodes.
1786 * Called with hugetlb_lock locked.
1788 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1794 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1796 * If we're returning unused surplus pages, only examine
1797 * nodes with surplus pages.
1799 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1800 !list_empty(&h->hugepage_freelists[node])) {
1802 list_entry(h->hugepage_freelists[node].next,
1804 list_del(&page->lru);
1805 h->free_huge_pages--;
1806 h->free_huge_pages_node[node]--;
1808 h->surplus_huge_pages--;
1809 h->surplus_huge_pages_node[node]--;
1811 update_and_free_page(h, page);
1821 * Dissolve a given free hugepage into free buddy pages. This function does
1822 * nothing for in-use hugepages and non-hugepages.
1823 * This function returns values like below:
1825 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1826 * (allocated or reserved.)
1827 * 0: successfully dissolved free hugepages or the page is not a
1828 * hugepage (considered as already dissolved)
1830 int dissolve_free_huge_page(struct page *page)
1834 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1835 if (!PageHuge(page))
1838 spin_lock(&hugetlb_lock);
1839 if (!PageHuge(page)) {
1844 if (!page_count(page)) {
1845 struct page *head = compound_head(page);
1846 struct hstate *h = page_hstate(head);
1847 int nid = page_to_nid(head);
1848 if (h->free_huge_pages - h->resv_huge_pages == 0)
1851 * Move PageHWPoison flag from head page to the raw error page,
1852 * which makes any subpages rather than the error page reusable.
1854 if (PageHWPoison(head) && page != head) {
1855 SetPageHWPoison(page);
1856 ClearPageHWPoison(head);
1858 list_del(&head->lru);
1859 h->free_huge_pages--;
1860 h->free_huge_pages_node[nid]--;
1861 h->max_huge_pages--;
1862 update_and_free_page(h, head);
1866 spin_unlock(&hugetlb_lock);
1871 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1872 * make specified memory blocks removable from the system.
1873 * Note that this will dissolve a free gigantic hugepage completely, if any
1874 * part of it lies within the given range.
1875 * Also note that if dissolve_free_huge_page() returns with an error, all
1876 * free hugepages that were dissolved before that error are lost.
1878 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1884 if (!hugepages_supported())
1887 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1888 page = pfn_to_page(pfn);
1889 rc = dissolve_free_huge_page(page);
1898 * Allocates a fresh surplus page from the page allocator.
1900 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1901 int nid, nodemask_t *nmask)
1903 struct page *page = NULL;
1905 if (hstate_is_gigantic(h))
1908 spin_lock(&hugetlb_lock);
1909 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1911 spin_unlock(&hugetlb_lock);
1913 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1917 spin_lock(&hugetlb_lock);
1919 * We could have raced with the pool size change.
1920 * Double check that and simply deallocate the new page
1921 * if we would end up overcommiting the surpluses. Abuse
1922 * temporary page to workaround the nasty free_huge_page
1925 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1926 SetPageHugeTemporary(page);
1927 spin_unlock(&hugetlb_lock);
1931 h->surplus_huge_pages++;
1932 h->surplus_huge_pages_node[page_to_nid(page)]++;
1936 spin_unlock(&hugetlb_lock);
1941 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1942 int nid, nodemask_t *nmask)
1946 if (hstate_is_gigantic(h))
1949 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1954 * We do not account these pages as surplus because they are only
1955 * temporary and will be released properly on the last reference
1957 SetPageHugeTemporary(page);
1963 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1966 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1967 struct vm_area_struct *vma, unsigned long addr)
1970 struct mempolicy *mpol;
1971 gfp_t gfp_mask = htlb_alloc_mask(h);
1973 nodemask_t *nodemask;
1975 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1976 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1977 mpol_cond_put(mpol);
1982 /* page migration callback function */
1983 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1985 gfp_t gfp_mask = htlb_alloc_mask(h);
1986 struct page *page = NULL;
1988 if (nid != NUMA_NO_NODE)
1989 gfp_mask |= __GFP_THISNODE;
1991 spin_lock(&hugetlb_lock);
1992 if (h->free_huge_pages - h->resv_huge_pages > 0)
1993 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1994 spin_unlock(&hugetlb_lock);
1997 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
2002 /* page migration callback function */
2003 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2006 gfp_t gfp_mask = htlb_alloc_mask(h);
2008 spin_lock(&hugetlb_lock);
2009 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2012 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2014 spin_unlock(&hugetlb_lock);
2018 spin_unlock(&hugetlb_lock);
2020 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2023 /* mempolicy aware migration callback */
2024 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2025 unsigned long address)
2027 struct mempolicy *mpol;
2028 nodemask_t *nodemask;
2033 gfp_mask = htlb_alloc_mask(h);
2034 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2035 page = alloc_huge_page_nodemask(h, node, nodemask);
2036 mpol_cond_put(mpol);
2042 * Increase the hugetlb pool such that it can accommodate a reservation
2045 static int gather_surplus_pages(struct hstate *h, int delta)
2046 __must_hold(&hugetlb_lock)
2048 struct list_head surplus_list;
2049 struct page *page, *tmp;
2051 int needed, allocated;
2052 bool alloc_ok = true;
2054 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2056 h->resv_huge_pages += delta;
2061 INIT_LIST_HEAD(&surplus_list);
2065 spin_unlock(&hugetlb_lock);
2066 for (i = 0; i < needed; i++) {
2067 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2068 NUMA_NO_NODE, NULL);
2073 list_add(&page->lru, &surplus_list);
2079 * After retaking hugetlb_lock, we need to recalculate 'needed'
2080 * because either resv_huge_pages or free_huge_pages may have changed.
2082 spin_lock(&hugetlb_lock);
2083 needed = (h->resv_huge_pages + delta) -
2084 (h->free_huge_pages + allocated);
2089 * We were not able to allocate enough pages to
2090 * satisfy the entire reservation so we free what
2091 * we've allocated so far.
2096 * The surplus_list now contains _at_least_ the number of extra pages
2097 * needed to accommodate the reservation. Add the appropriate number
2098 * of pages to the hugetlb pool and free the extras back to the buddy
2099 * allocator. Commit the entire reservation here to prevent another
2100 * process from stealing the pages as they are added to the pool but
2101 * before they are reserved.
2103 needed += allocated;
2104 h->resv_huge_pages += delta;
2107 /* Free the needed pages to the hugetlb pool */
2108 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2112 * This page is now managed by the hugetlb allocator and has
2113 * no users -- drop the buddy allocator's reference.
2115 put_page_testzero(page);
2116 VM_BUG_ON_PAGE(page_count(page), page);
2117 enqueue_huge_page(h, page);
2120 spin_unlock(&hugetlb_lock);
2122 /* Free unnecessary surplus pages to the buddy allocator */
2123 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2125 spin_lock(&hugetlb_lock);
2131 * This routine has two main purposes:
2132 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2133 * in unused_resv_pages. This corresponds to the prior adjustments made
2134 * to the associated reservation map.
2135 * 2) Free any unused surplus pages that may have been allocated to satisfy
2136 * the reservation. As many as unused_resv_pages may be freed.
2138 * Called with hugetlb_lock held. However, the lock could be dropped (and
2139 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2140 * we must make sure nobody else can claim pages we are in the process of
2141 * freeing. Do this by ensuring resv_huge_page always is greater than the
2142 * number of huge pages we plan to free when dropping the lock.
2144 static void return_unused_surplus_pages(struct hstate *h,
2145 unsigned long unused_resv_pages)
2147 unsigned long nr_pages;
2149 /* Cannot return gigantic pages currently */
2150 if (hstate_is_gigantic(h))
2154 * Part (or even all) of the reservation could have been backed
2155 * by pre-allocated pages. Only free surplus pages.
2157 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2160 * We want to release as many surplus pages as possible, spread
2161 * evenly across all nodes with memory. Iterate across these nodes
2162 * until we can no longer free unreserved surplus pages. This occurs
2163 * when the nodes with surplus pages have no free pages.
2164 * free_pool_huge_page() will balance the the freed pages across the
2165 * on-line nodes with memory and will handle the hstate accounting.
2167 * Note that we decrement resv_huge_pages as we free the pages. If
2168 * we drop the lock, resv_huge_pages will still be sufficiently large
2169 * to cover subsequent pages we may free.
2171 while (nr_pages--) {
2172 h->resv_huge_pages--;
2173 unused_resv_pages--;
2174 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2176 cond_resched_lock(&hugetlb_lock);
2180 /* Fully uncommit the reservation */
2181 h->resv_huge_pages -= unused_resv_pages;
2186 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2187 * are used by the huge page allocation routines to manage reservations.
2189 * vma_needs_reservation is called to determine if the huge page at addr
2190 * within the vma has an associated reservation. If a reservation is
2191 * needed, the value 1 is returned. The caller is then responsible for
2192 * managing the global reservation and subpool usage counts. After
2193 * the huge page has been allocated, vma_commit_reservation is called
2194 * to add the page to the reservation map. If the page allocation fails,
2195 * the reservation must be ended instead of committed. vma_end_reservation
2196 * is called in such cases.
2198 * In the normal case, vma_commit_reservation returns the same value
2199 * as the preceding vma_needs_reservation call. The only time this
2200 * is not the case is if a reserve map was changed between calls. It
2201 * is the responsibility of the caller to notice the difference and
2202 * take appropriate action.
2204 * vma_add_reservation is used in error paths where a reservation must
2205 * be restored when a newly allocated huge page must be freed. It is
2206 * to be called after calling vma_needs_reservation to determine if a
2207 * reservation exists.
2209 enum vma_resv_mode {
2215 static long __vma_reservation_common(struct hstate *h,
2216 struct vm_area_struct *vma, unsigned long addr,
2217 enum vma_resv_mode mode)
2219 struct resv_map *resv;
2222 long dummy_out_regions_needed;
2224 resv = vma_resv_map(vma);
2228 idx = vma_hugecache_offset(h, vma, addr);
2230 case VMA_NEEDS_RESV:
2231 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2232 /* We assume that vma_reservation_* routines always operate on
2233 * 1 page, and that adding to resv map a 1 page entry can only
2234 * ever require 1 region.
2236 VM_BUG_ON(dummy_out_regions_needed != 1);
2238 case VMA_COMMIT_RESV:
2239 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2240 /* region_add calls of range 1 should never fail. */
2244 region_abort(resv, idx, idx + 1, 1);
2248 if (vma->vm_flags & VM_MAYSHARE) {
2249 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2250 /* region_add calls of range 1 should never fail. */
2253 region_abort(resv, idx, idx + 1, 1);
2254 ret = region_del(resv, idx, idx + 1);
2261 if (vma->vm_flags & VM_MAYSHARE)
2263 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2265 * In most cases, reserves always exist for private mappings.
2266 * However, a file associated with mapping could have been
2267 * hole punched or truncated after reserves were consumed.
2268 * As subsequent fault on such a range will not use reserves.
2269 * Subtle - The reserve map for private mappings has the
2270 * opposite meaning than that of shared mappings. If NO
2271 * entry is in the reserve map, it means a reservation exists.
2272 * If an entry exists in the reserve map, it means the
2273 * reservation has already been consumed. As a result, the
2274 * return value of this routine is the opposite of the
2275 * value returned from reserve map manipulation routines above.
2283 return ret < 0 ? ret : 0;
2286 static long vma_needs_reservation(struct hstate *h,
2287 struct vm_area_struct *vma, unsigned long addr)
2289 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2292 static long vma_commit_reservation(struct hstate *h,
2293 struct vm_area_struct *vma, unsigned long addr)
2295 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2298 static void vma_end_reservation(struct hstate *h,
2299 struct vm_area_struct *vma, unsigned long addr)
2301 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2304 static long vma_add_reservation(struct hstate *h,
2305 struct vm_area_struct *vma, unsigned long addr)
2307 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2311 * This routine is called to restore a reservation on error paths. In the
2312 * specific error paths, a huge page was allocated (via alloc_huge_page)
2313 * and is about to be freed. If a reservation for the page existed,
2314 * alloc_huge_page would have consumed the reservation and set PagePrivate
2315 * in the newly allocated page. When the page is freed via free_huge_page,
2316 * the global reservation count will be incremented if PagePrivate is set.
2317 * However, free_huge_page can not adjust the reserve map. Adjust the
2318 * reserve map here to be consistent with global reserve count adjustments
2319 * to be made by free_huge_page.
2321 static void restore_reserve_on_error(struct hstate *h,
2322 struct vm_area_struct *vma, unsigned long address,
2325 if (unlikely(PagePrivate(page))) {
2326 long rc = vma_needs_reservation(h, vma, address);
2328 if (unlikely(rc < 0)) {
2330 * Rare out of memory condition in reserve map
2331 * manipulation. Clear PagePrivate so that
2332 * global reserve count will not be incremented
2333 * by free_huge_page. This will make it appear
2334 * as though the reservation for this page was
2335 * consumed. This may prevent the task from
2336 * faulting in the page at a later time. This
2337 * is better than inconsistent global huge page
2338 * accounting of reserve counts.
2340 ClearPagePrivate(page);
2342 rc = vma_add_reservation(h, vma, address);
2343 if (unlikely(rc < 0))
2345 * See above comment about rare out of
2348 ClearPagePrivate(page);
2350 vma_end_reservation(h, vma, address);
2354 struct page *alloc_huge_page(struct vm_area_struct *vma,
2355 unsigned long addr, int avoid_reserve)
2357 struct hugepage_subpool *spool = subpool_vma(vma);
2358 struct hstate *h = hstate_vma(vma);
2360 long map_chg, map_commit;
2363 struct hugetlb_cgroup *h_cg;
2364 bool deferred_reserve;
2366 idx = hstate_index(h);
2368 * Examine the region/reserve map to determine if the process
2369 * has a reservation for the page to be allocated. A return
2370 * code of zero indicates a reservation exists (no change).
2372 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2374 return ERR_PTR(-ENOMEM);
2377 * Processes that did not create the mapping will have no
2378 * reserves as indicated by the region/reserve map. Check
2379 * that the allocation will not exceed the subpool limit.
2380 * Allocations for MAP_NORESERVE mappings also need to be
2381 * checked against any subpool limit.
2383 if (map_chg || avoid_reserve) {
2384 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2386 vma_end_reservation(h, vma, addr);
2387 return ERR_PTR(-ENOSPC);
2391 * Even though there was no reservation in the region/reserve
2392 * map, there could be reservations associated with the
2393 * subpool that can be used. This would be indicated if the
2394 * return value of hugepage_subpool_get_pages() is zero.
2395 * However, if avoid_reserve is specified we still avoid even
2396 * the subpool reservations.
2402 /* If this allocation is not consuming a reservation, charge it now.
2404 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2405 if (deferred_reserve) {
2406 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2407 idx, pages_per_huge_page(h), &h_cg);
2409 goto out_subpool_put;
2412 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2414 goto out_uncharge_cgroup_reservation;
2416 spin_lock(&hugetlb_lock);
2418 * glb_chg is passed to indicate whether or not a page must be taken
2419 * from the global free pool (global change). gbl_chg == 0 indicates
2420 * a reservation exists for the allocation.
2422 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2424 spin_unlock(&hugetlb_lock);
2425 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2427 goto out_uncharge_cgroup;
2428 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2429 SetPagePrivate(page);
2430 h->resv_huge_pages--;
2432 spin_lock(&hugetlb_lock);
2433 list_move(&page->lru, &h->hugepage_activelist);
2436 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2437 /* If allocation is not consuming a reservation, also store the
2438 * hugetlb_cgroup pointer on the page.
2440 if (deferred_reserve) {
2441 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2445 spin_unlock(&hugetlb_lock);
2447 set_page_private(page, (unsigned long)spool);
2449 map_commit = vma_commit_reservation(h, vma, addr);
2450 if (unlikely(map_chg > map_commit)) {
2452 * The page was added to the reservation map between
2453 * vma_needs_reservation and vma_commit_reservation.
2454 * This indicates a race with hugetlb_reserve_pages.
2455 * Adjust for the subpool count incremented above AND
2456 * in hugetlb_reserve_pages for the same page. Also,
2457 * the reservation count added in hugetlb_reserve_pages
2458 * no longer applies.
2462 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2463 hugetlb_acct_memory(h, -rsv_adjust);
2467 out_uncharge_cgroup:
2468 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2469 out_uncharge_cgroup_reservation:
2470 if (deferred_reserve)
2471 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2474 if (map_chg || avoid_reserve)
2475 hugepage_subpool_put_pages(spool, 1);
2476 vma_end_reservation(h, vma, addr);
2477 return ERR_PTR(-ENOSPC);
2480 int alloc_bootmem_huge_page(struct hstate *h)
2481 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2482 int __alloc_bootmem_huge_page(struct hstate *h)
2484 struct huge_bootmem_page *m;
2487 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2490 addr = memblock_alloc_try_nid_raw(
2491 huge_page_size(h), huge_page_size(h),
2492 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2495 * Use the beginning of the huge page to store the
2496 * huge_bootmem_page struct (until gather_bootmem
2497 * puts them into the mem_map).
2506 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2507 /* Put them into a private list first because mem_map is not up yet */
2508 INIT_LIST_HEAD(&m->list);
2509 list_add(&m->list, &huge_boot_pages);
2514 static void __init prep_compound_huge_page(struct page *page,
2517 if (unlikely(order > (MAX_ORDER - 1)))
2518 prep_compound_gigantic_page(page, order);
2520 prep_compound_page(page, order);
2523 /* Put bootmem huge pages into the standard lists after mem_map is up */
2524 static void __init gather_bootmem_prealloc(void)
2526 struct huge_bootmem_page *m;
2528 list_for_each_entry(m, &huge_boot_pages, list) {
2529 struct page *page = virt_to_page(m);
2530 struct hstate *h = m->hstate;
2532 WARN_ON(page_count(page) != 1);
2533 prep_compound_huge_page(page, h->order);
2534 WARN_ON(PageReserved(page));
2535 prep_new_huge_page(h, page, page_to_nid(page));
2536 put_page(page); /* free it into the hugepage allocator */
2539 * If we had gigantic hugepages allocated at boot time, we need
2540 * to restore the 'stolen' pages to totalram_pages in order to
2541 * fix confusing memory reports from free(1) and another
2542 * side-effects, like CommitLimit going negative.
2544 if (hstate_is_gigantic(h))
2545 adjust_managed_page_count(page, 1 << h->order);
2550 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2553 nodemask_t *node_alloc_noretry;
2555 if (!hstate_is_gigantic(h)) {
2557 * Bit mask controlling how hard we retry per-node allocations.
2558 * Ignore errors as lower level routines can deal with
2559 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2560 * time, we are likely in bigger trouble.
2562 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2565 /* allocations done at boot time */
2566 node_alloc_noretry = NULL;
2569 /* bit mask controlling how hard we retry per-node allocations */
2570 if (node_alloc_noretry)
2571 nodes_clear(*node_alloc_noretry);
2573 for (i = 0; i < h->max_huge_pages; ++i) {
2574 if (hstate_is_gigantic(h)) {
2575 if (IS_ENABLED(CONFIG_CMA) && hugetlb_cma[0]) {
2576 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2579 if (!alloc_bootmem_huge_page(h))
2581 } else if (!alloc_pool_huge_page(h,
2582 &node_states[N_MEMORY],
2583 node_alloc_noretry))
2587 if (i < h->max_huge_pages) {
2590 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2591 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2592 h->max_huge_pages, buf, i);
2593 h->max_huge_pages = i;
2596 kfree(node_alloc_noretry);
2599 static void __init hugetlb_init_hstates(void)
2603 for_each_hstate(h) {
2604 if (minimum_order > huge_page_order(h))
2605 minimum_order = huge_page_order(h);
2607 /* oversize hugepages were init'ed in early boot */
2608 if (!hstate_is_gigantic(h))
2609 hugetlb_hstate_alloc_pages(h);
2611 VM_BUG_ON(minimum_order == UINT_MAX);
2614 static void __init report_hugepages(void)
2618 for_each_hstate(h) {
2621 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2622 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2623 buf, h->free_huge_pages);
2627 #ifdef CONFIG_HIGHMEM
2628 static void try_to_free_low(struct hstate *h, unsigned long count,
2629 nodemask_t *nodes_allowed)
2633 if (hstate_is_gigantic(h))
2636 for_each_node_mask(i, *nodes_allowed) {
2637 struct page *page, *next;
2638 struct list_head *freel = &h->hugepage_freelists[i];
2639 list_for_each_entry_safe(page, next, freel, lru) {
2640 if (count >= h->nr_huge_pages)
2642 if (PageHighMem(page))
2644 list_del(&page->lru);
2645 update_and_free_page(h, page);
2646 h->free_huge_pages--;
2647 h->free_huge_pages_node[page_to_nid(page)]--;
2652 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2653 nodemask_t *nodes_allowed)
2659 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2660 * balanced by operating on them in a round-robin fashion.
2661 * Returns 1 if an adjustment was made.
2663 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2668 VM_BUG_ON(delta != -1 && delta != 1);
2671 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2672 if (h->surplus_huge_pages_node[node])
2676 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2677 if (h->surplus_huge_pages_node[node] <
2678 h->nr_huge_pages_node[node])
2685 h->surplus_huge_pages += delta;
2686 h->surplus_huge_pages_node[node] += delta;
2690 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2691 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2692 nodemask_t *nodes_allowed)
2694 unsigned long min_count, ret;
2695 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2698 * Bit mask controlling how hard we retry per-node allocations.
2699 * If we can not allocate the bit mask, do not attempt to allocate
2700 * the requested huge pages.
2702 if (node_alloc_noretry)
2703 nodes_clear(*node_alloc_noretry);
2707 spin_lock(&hugetlb_lock);
2710 * Check for a node specific request.
2711 * Changing node specific huge page count may require a corresponding
2712 * change to the global count. In any case, the passed node mask
2713 * (nodes_allowed) will restrict alloc/free to the specified node.
2715 if (nid != NUMA_NO_NODE) {
2716 unsigned long old_count = count;
2718 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2720 * User may have specified a large count value which caused the
2721 * above calculation to overflow. In this case, they wanted
2722 * to allocate as many huge pages as possible. Set count to
2723 * largest possible value to align with their intention.
2725 if (count < old_count)
2730 * Gigantic pages runtime allocation depend on the capability for large
2731 * page range allocation.
2732 * If the system does not provide this feature, return an error when
2733 * the user tries to allocate gigantic pages but let the user free the
2734 * boottime allocated gigantic pages.
2736 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2737 if (count > persistent_huge_pages(h)) {
2738 spin_unlock(&hugetlb_lock);
2739 NODEMASK_FREE(node_alloc_noretry);
2742 /* Fall through to decrease pool */
2746 * Increase the pool size
2747 * First take pages out of surplus state. Then make up the
2748 * remaining difference by allocating fresh huge pages.
2750 * We might race with alloc_surplus_huge_page() here and be unable
2751 * to convert a surplus huge page to a normal huge page. That is
2752 * not critical, though, it just means the overall size of the
2753 * pool might be one hugepage larger than it needs to be, but
2754 * within all the constraints specified by the sysctls.
2756 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2757 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2761 while (count > persistent_huge_pages(h)) {
2763 * If this allocation races such that we no longer need the
2764 * page, free_huge_page will handle it by freeing the page
2765 * and reducing the surplus.
2767 spin_unlock(&hugetlb_lock);
2769 /* yield cpu to avoid soft lockup */
2772 ret = alloc_pool_huge_page(h, nodes_allowed,
2773 node_alloc_noretry);
2774 spin_lock(&hugetlb_lock);
2778 /* Bail for signals. Probably ctrl-c from user */
2779 if (signal_pending(current))
2784 * Decrease the pool size
2785 * First return free pages to the buddy allocator (being careful
2786 * to keep enough around to satisfy reservations). Then place
2787 * pages into surplus state as needed so the pool will shrink
2788 * to the desired size as pages become free.
2790 * By placing pages into the surplus state independent of the
2791 * overcommit value, we are allowing the surplus pool size to
2792 * exceed overcommit. There are few sane options here. Since
2793 * alloc_surplus_huge_page() is checking the global counter,
2794 * though, we'll note that we're not allowed to exceed surplus
2795 * and won't grow the pool anywhere else. Not until one of the
2796 * sysctls are changed, or the surplus pages go out of use.
2798 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2799 min_count = max(count, min_count);
2800 try_to_free_low(h, min_count, nodes_allowed);
2801 while (min_count < persistent_huge_pages(h)) {
2802 if (!free_pool_huge_page(h, nodes_allowed, 0))
2804 cond_resched_lock(&hugetlb_lock);
2806 while (count < persistent_huge_pages(h)) {
2807 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2811 h->max_huge_pages = persistent_huge_pages(h);
2812 spin_unlock(&hugetlb_lock);
2814 NODEMASK_FREE(node_alloc_noretry);
2819 #define HSTATE_ATTR_RO(_name) \
2820 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2822 #define HSTATE_ATTR(_name) \
2823 static struct kobj_attribute _name##_attr = \
2824 __ATTR(_name, 0644, _name##_show, _name##_store)
2826 static struct kobject *hugepages_kobj;
2827 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2829 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2831 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2835 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2836 if (hstate_kobjs[i] == kobj) {
2838 *nidp = NUMA_NO_NODE;
2842 return kobj_to_node_hstate(kobj, nidp);
2845 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2846 struct kobj_attribute *attr, char *buf)
2849 unsigned long nr_huge_pages;
2852 h = kobj_to_hstate(kobj, &nid);
2853 if (nid == NUMA_NO_NODE)
2854 nr_huge_pages = h->nr_huge_pages;
2856 nr_huge_pages = h->nr_huge_pages_node[nid];
2858 return sprintf(buf, "%lu\n", nr_huge_pages);
2861 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2862 struct hstate *h, int nid,
2863 unsigned long count, size_t len)
2866 nodemask_t nodes_allowed, *n_mask;
2868 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2871 if (nid == NUMA_NO_NODE) {
2873 * global hstate attribute
2875 if (!(obey_mempolicy &&
2876 init_nodemask_of_mempolicy(&nodes_allowed)))
2877 n_mask = &node_states[N_MEMORY];
2879 n_mask = &nodes_allowed;
2882 * Node specific request. count adjustment happens in
2883 * set_max_huge_pages() after acquiring hugetlb_lock.
2885 init_nodemask_of_node(&nodes_allowed, nid);
2886 n_mask = &nodes_allowed;
2889 err = set_max_huge_pages(h, count, nid, n_mask);
2891 return err ? err : len;
2894 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2895 struct kobject *kobj, const char *buf,
2899 unsigned long count;
2903 err = kstrtoul(buf, 10, &count);
2907 h = kobj_to_hstate(kobj, &nid);
2908 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2911 static ssize_t nr_hugepages_show(struct kobject *kobj,
2912 struct kobj_attribute *attr, char *buf)
2914 return nr_hugepages_show_common(kobj, attr, buf);
2917 static ssize_t nr_hugepages_store(struct kobject *kobj,
2918 struct kobj_attribute *attr, const char *buf, size_t len)
2920 return nr_hugepages_store_common(false, kobj, buf, len);
2922 HSTATE_ATTR(nr_hugepages);
2927 * hstate attribute for optionally mempolicy-based constraint on persistent
2928 * huge page alloc/free.
2930 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2931 struct kobj_attribute *attr, char *buf)
2933 return nr_hugepages_show_common(kobj, attr, buf);
2936 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2937 struct kobj_attribute *attr, const char *buf, size_t len)
2939 return nr_hugepages_store_common(true, kobj, buf, len);
2941 HSTATE_ATTR(nr_hugepages_mempolicy);
2945 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2946 struct kobj_attribute *attr, char *buf)
2948 struct hstate *h = kobj_to_hstate(kobj, NULL);
2949 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2952 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2953 struct kobj_attribute *attr, const char *buf, size_t count)
2956 unsigned long input;
2957 struct hstate *h = kobj_to_hstate(kobj, NULL);
2959 if (hstate_is_gigantic(h))
2962 err = kstrtoul(buf, 10, &input);
2966 spin_lock(&hugetlb_lock);
2967 h->nr_overcommit_huge_pages = input;
2968 spin_unlock(&hugetlb_lock);
2972 HSTATE_ATTR(nr_overcommit_hugepages);
2974 static ssize_t free_hugepages_show(struct kobject *kobj,
2975 struct kobj_attribute *attr, char *buf)
2978 unsigned long free_huge_pages;
2981 h = kobj_to_hstate(kobj, &nid);
2982 if (nid == NUMA_NO_NODE)
2983 free_huge_pages = h->free_huge_pages;
2985 free_huge_pages = h->free_huge_pages_node[nid];
2987 return sprintf(buf, "%lu\n", free_huge_pages);
2989 HSTATE_ATTR_RO(free_hugepages);
2991 static ssize_t resv_hugepages_show(struct kobject *kobj,
2992 struct kobj_attribute *attr, char *buf)
2994 struct hstate *h = kobj_to_hstate(kobj, NULL);
2995 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2997 HSTATE_ATTR_RO(resv_hugepages);
2999 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3000 struct kobj_attribute *attr, char *buf)
3003 unsigned long surplus_huge_pages;
3006 h = kobj_to_hstate(kobj, &nid);
3007 if (nid == NUMA_NO_NODE)
3008 surplus_huge_pages = h->surplus_huge_pages;
3010 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3012 return sprintf(buf, "%lu\n", surplus_huge_pages);
3014 HSTATE_ATTR_RO(surplus_hugepages);
3016 static struct attribute *hstate_attrs[] = {
3017 &nr_hugepages_attr.attr,
3018 &nr_overcommit_hugepages_attr.attr,
3019 &free_hugepages_attr.attr,
3020 &resv_hugepages_attr.attr,
3021 &surplus_hugepages_attr.attr,
3023 &nr_hugepages_mempolicy_attr.attr,
3028 static const struct attribute_group hstate_attr_group = {
3029 .attrs = hstate_attrs,
3032 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3033 struct kobject **hstate_kobjs,
3034 const struct attribute_group *hstate_attr_group)
3037 int hi = hstate_index(h);
3039 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3040 if (!hstate_kobjs[hi])
3043 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3045 kobject_put(hstate_kobjs[hi]);
3050 static void __init hugetlb_sysfs_init(void)
3055 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3056 if (!hugepages_kobj)
3059 for_each_hstate(h) {
3060 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3061 hstate_kobjs, &hstate_attr_group);
3063 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3070 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3071 * with node devices in node_devices[] using a parallel array. The array
3072 * index of a node device or _hstate == node id.
3073 * This is here to avoid any static dependency of the node device driver, in
3074 * the base kernel, on the hugetlb module.
3076 struct node_hstate {
3077 struct kobject *hugepages_kobj;
3078 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3080 static struct node_hstate node_hstates[MAX_NUMNODES];
3083 * A subset of global hstate attributes for node devices
3085 static struct attribute *per_node_hstate_attrs[] = {
3086 &nr_hugepages_attr.attr,
3087 &free_hugepages_attr.attr,
3088 &surplus_hugepages_attr.attr,
3092 static const struct attribute_group per_node_hstate_attr_group = {
3093 .attrs = per_node_hstate_attrs,
3097 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3098 * Returns node id via non-NULL nidp.
3100 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3104 for (nid = 0; nid < nr_node_ids; nid++) {
3105 struct node_hstate *nhs = &node_hstates[nid];
3107 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3108 if (nhs->hstate_kobjs[i] == kobj) {
3120 * Unregister hstate attributes from a single node device.
3121 * No-op if no hstate attributes attached.
3123 static void hugetlb_unregister_node(struct node *node)
3126 struct node_hstate *nhs = &node_hstates[node->dev.id];
3128 if (!nhs->hugepages_kobj)
3129 return; /* no hstate attributes */
3131 for_each_hstate(h) {
3132 int idx = hstate_index(h);
3133 if (nhs->hstate_kobjs[idx]) {
3134 kobject_put(nhs->hstate_kobjs[idx]);
3135 nhs->hstate_kobjs[idx] = NULL;
3139 kobject_put(nhs->hugepages_kobj);
3140 nhs->hugepages_kobj = NULL;
3145 * Register hstate attributes for a single node device.
3146 * No-op if attributes already registered.
3148 static void hugetlb_register_node(struct node *node)
3151 struct node_hstate *nhs = &node_hstates[node->dev.id];
3154 if (nhs->hugepages_kobj)
3155 return; /* already allocated */
3157 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3159 if (!nhs->hugepages_kobj)
3162 for_each_hstate(h) {
3163 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3165 &per_node_hstate_attr_group);
3167 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3168 h->name, node->dev.id);
3169 hugetlb_unregister_node(node);
3176 * hugetlb init time: register hstate attributes for all registered node
3177 * devices of nodes that have memory. All on-line nodes should have
3178 * registered their associated device by this time.
3180 static void __init hugetlb_register_all_nodes(void)
3184 for_each_node_state(nid, N_MEMORY) {
3185 struct node *node = node_devices[nid];
3186 if (node->dev.id == nid)
3187 hugetlb_register_node(node);
3191 * Let the node device driver know we're here so it can
3192 * [un]register hstate attributes on node hotplug.
3194 register_hugetlbfs_with_node(hugetlb_register_node,
3195 hugetlb_unregister_node);
3197 #else /* !CONFIG_NUMA */
3199 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3207 static void hugetlb_register_all_nodes(void) { }
3211 static int __init hugetlb_init(void)
3215 if (!hugepages_supported()) {
3216 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3217 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3222 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3223 * architectures depend on setup being done here.
3225 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3226 if (!parsed_default_hugepagesz) {
3228 * If we did not parse a default huge page size, set
3229 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3230 * number of huge pages for this default size was implicitly
3231 * specified, set that here as well.
3232 * Note that the implicit setting will overwrite an explicit
3233 * setting. A warning will be printed in this case.
3235 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3236 if (default_hstate_max_huge_pages) {
3237 if (default_hstate.max_huge_pages) {
3240 string_get_size(huge_page_size(&default_hstate),
3241 1, STRING_UNITS_2, buf, 32);
3242 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3243 default_hstate.max_huge_pages, buf);
3244 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3245 default_hstate_max_huge_pages);
3247 default_hstate.max_huge_pages =
3248 default_hstate_max_huge_pages;
3252 hugetlb_cma_check();
3253 hugetlb_init_hstates();
3254 gather_bootmem_prealloc();
3257 hugetlb_sysfs_init();
3258 hugetlb_register_all_nodes();
3259 hugetlb_cgroup_file_init();
3262 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3264 num_fault_mutexes = 1;
3266 hugetlb_fault_mutex_table =
3267 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3269 BUG_ON(!hugetlb_fault_mutex_table);
3271 for (i = 0; i < num_fault_mutexes; i++)
3272 mutex_init(&hugetlb_fault_mutex_table[i]);
3275 subsys_initcall(hugetlb_init);
3277 /* Overwritten by architectures with more huge page sizes */
3278 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3280 return size == HPAGE_SIZE;
3283 void __init hugetlb_add_hstate(unsigned int order)
3288 if (size_to_hstate(PAGE_SIZE << order)) {
3291 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3293 h = &hstates[hugetlb_max_hstate++];
3295 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3296 h->nr_huge_pages = 0;
3297 h->free_huge_pages = 0;
3298 for (i = 0; i < MAX_NUMNODES; ++i)
3299 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3300 INIT_LIST_HEAD(&h->hugepage_activelist);
3301 h->next_nid_to_alloc = first_memory_node;
3302 h->next_nid_to_free = first_memory_node;
3303 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3304 huge_page_size(h)/1024);
3310 * hugepages command line processing
3311 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3312 * specification. If not, ignore the hugepages value. hugepages can also
3313 * be the first huge page command line option in which case it implicitly
3314 * specifies the number of huge pages for the default size.
3316 static int __init hugepages_setup(char *s)
3319 static unsigned long *last_mhp;
3321 if (!parsed_valid_hugepagesz) {
3322 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3323 parsed_valid_hugepagesz = true;
3328 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3329 * yet, so this hugepages= parameter goes to the "default hstate".
3330 * Otherwise, it goes with the previously parsed hugepagesz or
3331 * default_hugepagesz.
3333 else if (!hugetlb_max_hstate)
3334 mhp = &default_hstate_max_huge_pages;
3336 mhp = &parsed_hstate->max_huge_pages;
3338 if (mhp == last_mhp) {
3339 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3343 if (sscanf(s, "%lu", mhp) <= 0)
3347 * Global state is always initialized later in hugetlb_init.
3348 * But we need to allocate >= MAX_ORDER hstates here early to still
3349 * use the bootmem allocator.
3351 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3352 hugetlb_hstate_alloc_pages(parsed_hstate);
3358 __setup("hugepages=", hugepages_setup);
3361 * hugepagesz command line processing
3362 * A specific huge page size can only be specified once with hugepagesz.
3363 * hugepagesz is followed by hugepages on the command line. The global
3364 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3365 * hugepagesz argument was valid.
3367 static int __init hugepagesz_setup(char *s)
3372 parsed_valid_hugepagesz = false;
3373 size = (unsigned long)memparse(s, NULL);
3375 if (!arch_hugetlb_valid_size(size)) {
3376 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3380 h = size_to_hstate(size);
3383 * hstate for this size already exists. This is normally
3384 * an error, but is allowed if the existing hstate is the
3385 * default hstate. More specifically, it is only allowed if
3386 * the number of huge pages for the default hstate was not
3387 * previously specified.
3389 if (!parsed_default_hugepagesz || h != &default_hstate ||
3390 default_hstate.max_huge_pages) {
3391 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3396 * No need to call hugetlb_add_hstate() as hstate already
3397 * exists. But, do set parsed_hstate so that a following
3398 * hugepages= parameter will be applied to this hstate.
3401 parsed_valid_hugepagesz = true;
3405 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3406 parsed_valid_hugepagesz = true;
3409 __setup("hugepagesz=", hugepagesz_setup);
3412 * default_hugepagesz command line input
3413 * Only one instance of default_hugepagesz allowed on command line.
3415 static int __init default_hugepagesz_setup(char *s)
3419 parsed_valid_hugepagesz = false;
3420 if (parsed_default_hugepagesz) {
3421 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3425 size = (unsigned long)memparse(s, NULL);
3427 if (!arch_hugetlb_valid_size(size)) {
3428 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3432 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3433 parsed_valid_hugepagesz = true;
3434 parsed_default_hugepagesz = true;
3435 default_hstate_idx = hstate_index(size_to_hstate(size));
3438 * The number of default huge pages (for this size) could have been
3439 * specified as the first hugetlb parameter: hugepages=X. If so,
3440 * then default_hstate_max_huge_pages is set. If the default huge
3441 * page size is gigantic (>= MAX_ORDER), then the pages must be
3442 * allocated here from bootmem allocator.
3444 if (default_hstate_max_huge_pages) {
3445 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3446 if (hstate_is_gigantic(&default_hstate))
3447 hugetlb_hstate_alloc_pages(&default_hstate);
3448 default_hstate_max_huge_pages = 0;
3453 __setup("default_hugepagesz=", default_hugepagesz_setup);
3455 static unsigned int cpuset_mems_nr(unsigned int *array)
3458 unsigned int nr = 0;
3460 for_each_node_mask(node, cpuset_current_mems_allowed)
3466 #ifdef CONFIG_SYSCTL
3467 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3468 struct ctl_table *table, int write,
3469 void *buffer, size_t *length, loff_t *ppos)
3471 struct hstate *h = &default_hstate;
3472 unsigned long tmp = h->max_huge_pages;
3475 if (!hugepages_supported())
3479 table->maxlen = sizeof(unsigned long);
3480 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3485 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3486 NUMA_NO_NODE, tmp, *length);
3491 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3492 void *buffer, size_t *length, loff_t *ppos)
3495 return hugetlb_sysctl_handler_common(false, table, write,
3496 buffer, length, ppos);
3500 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3501 void *buffer, size_t *length, loff_t *ppos)
3503 return hugetlb_sysctl_handler_common(true, table, write,
3504 buffer, length, ppos);
3506 #endif /* CONFIG_NUMA */
3508 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3509 void *buffer, size_t *length, loff_t *ppos)
3511 struct hstate *h = &default_hstate;
3515 if (!hugepages_supported())
3518 tmp = h->nr_overcommit_huge_pages;
3520 if (write && hstate_is_gigantic(h))
3524 table->maxlen = sizeof(unsigned long);
3525 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3530 spin_lock(&hugetlb_lock);
3531 h->nr_overcommit_huge_pages = tmp;
3532 spin_unlock(&hugetlb_lock);
3538 #endif /* CONFIG_SYSCTL */
3540 void hugetlb_report_meminfo(struct seq_file *m)
3543 unsigned long total = 0;
3545 if (!hugepages_supported())
3548 for_each_hstate(h) {
3549 unsigned long count = h->nr_huge_pages;
3551 total += (PAGE_SIZE << huge_page_order(h)) * count;
3553 if (h == &default_hstate)
3555 "HugePages_Total: %5lu\n"
3556 "HugePages_Free: %5lu\n"
3557 "HugePages_Rsvd: %5lu\n"
3558 "HugePages_Surp: %5lu\n"
3559 "Hugepagesize: %8lu kB\n",
3563 h->surplus_huge_pages,
3564 (PAGE_SIZE << huge_page_order(h)) / 1024);
3567 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3570 int hugetlb_report_node_meminfo(int nid, char *buf)
3572 struct hstate *h = &default_hstate;
3573 if (!hugepages_supported())
3576 "Node %d HugePages_Total: %5u\n"
3577 "Node %d HugePages_Free: %5u\n"
3578 "Node %d HugePages_Surp: %5u\n",
3579 nid, h->nr_huge_pages_node[nid],
3580 nid, h->free_huge_pages_node[nid],
3581 nid, h->surplus_huge_pages_node[nid]);
3584 void hugetlb_show_meminfo(void)
3589 if (!hugepages_supported())
3592 for_each_node_state(nid, N_MEMORY)
3594 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3596 h->nr_huge_pages_node[nid],
3597 h->free_huge_pages_node[nid],
3598 h->surplus_huge_pages_node[nid],
3599 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3602 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3604 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3605 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3608 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3609 unsigned long hugetlb_total_pages(void)
3612 unsigned long nr_total_pages = 0;
3615 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3616 return nr_total_pages;
3619 static int hugetlb_acct_memory(struct hstate *h, long delta)
3623 spin_lock(&hugetlb_lock);
3625 * When cpuset is configured, it breaks the strict hugetlb page
3626 * reservation as the accounting is done on a global variable. Such
3627 * reservation is completely rubbish in the presence of cpuset because
3628 * the reservation is not checked against page availability for the
3629 * current cpuset. Application can still potentially OOM'ed by kernel
3630 * with lack of free htlb page in cpuset that the task is in.
3631 * Attempt to enforce strict accounting with cpuset is almost
3632 * impossible (or too ugly) because cpuset is too fluid that
3633 * task or memory node can be dynamically moved between cpusets.
3635 * The change of semantics for shared hugetlb mapping with cpuset is
3636 * undesirable. However, in order to preserve some of the semantics,
3637 * we fall back to check against current free page availability as
3638 * a best attempt and hopefully to minimize the impact of changing
3639 * semantics that cpuset has.
3642 if (gather_surplus_pages(h, delta) < 0)
3645 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3646 return_unused_surplus_pages(h, delta);
3653 return_unused_surplus_pages(h, (unsigned long) -delta);
3656 spin_unlock(&hugetlb_lock);
3660 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3662 struct resv_map *resv = vma_resv_map(vma);
3665 * This new VMA should share its siblings reservation map if present.
3666 * The VMA will only ever have a valid reservation map pointer where
3667 * it is being copied for another still existing VMA. As that VMA
3668 * has a reference to the reservation map it cannot disappear until
3669 * after this open call completes. It is therefore safe to take a
3670 * new reference here without additional locking.
3672 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3673 kref_get(&resv->refs);
3676 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3678 struct hstate *h = hstate_vma(vma);
3679 struct resv_map *resv = vma_resv_map(vma);
3680 struct hugepage_subpool *spool = subpool_vma(vma);
3681 unsigned long reserve, start, end;
3684 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3687 start = vma_hugecache_offset(h, vma, vma->vm_start);
3688 end = vma_hugecache_offset(h, vma, vma->vm_end);
3690 reserve = (end - start) - region_count(resv, start, end);
3691 hugetlb_cgroup_uncharge_counter(resv, start, end);
3694 * Decrement reserve counts. The global reserve count may be
3695 * adjusted if the subpool has a minimum size.
3697 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3698 hugetlb_acct_memory(h, -gbl_reserve);
3701 kref_put(&resv->refs, resv_map_release);
3704 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3706 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3711 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3713 struct hstate *hstate = hstate_vma(vma);
3715 return 1UL << huge_page_shift(hstate);
3719 * We cannot handle pagefaults against hugetlb pages at all. They cause
3720 * handle_mm_fault() to try to instantiate regular-sized pages in the
3721 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3724 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3731 * When a new function is introduced to vm_operations_struct and added
3732 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3733 * This is because under System V memory model, mappings created via
3734 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3735 * their original vm_ops are overwritten with shm_vm_ops.
3737 const struct vm_operations_struct hugetlb_vm_ops = {
3738 .fault = hugetlb_vm_op_fault,
3739 .open = hugetlb_vm_op_open,
3740 .close = hugetlb_vm_op_close,
3741 .split = hugetlb_vm_op_split,
3742 .pagesize = hugetlb_vm_op_pagesize,
3745 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3751 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3752 vma->vm_page_prot)));
3754 entry = huge_pte_wrprotect(mk_huge_pte(page,
3755 vma->vm_page_prot));
3757 entry = pte_mkyoung(entry);
3758 entry = pte_mkhuge(entry);
3759 entry = arch_make_huge_pte(entry, vma, page, writable);
3764 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3765 unsigned long address, pte_t *ptep)
3769 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3770 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3771 update_mmu_cache(vma, address, ptep);
3774 bool is_hugetlb_entry_migration(pte_t pte)
3778 if (huge_pte_none(pte) || pte_present(pte))
3780 swp = pte_to_swp_entry(pte);
3781 if (non_swap_entry(swp) && is_migration_entry(swp))
3787 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3791 if (huge_pte_none(pte) || pte_present(pte))
3793 swp = pte_to_swp_entry(pte);
3794 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3800 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3801 struct vm_area_struct *vma)
3803 pte_t *src_pte, *dst_pte, entry, dst_entry;
3804 struct page *ptepage;
3807 struct hstate *h = hstate_vma(vma);
3808 unsigned long sz = huge_page_size(h);
3809 struct address_space *mapping = vma->vm_file->f_mapping;
3810 struct mmu_notifier_range range;
3813 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3816 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3819 mmu_notifier_invalidate_range_start(&range);
3822 * For shared mappings i_mmap_rwsem must be held to call
3823 * huge_pte_alloc, otherwise the returned ptep could go
3824 * away if part of a shared pmd and another thread calls
3827 i_mmap_lock_read(mapping);
3830 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3831 spinlock_t *src_ptl, *dst_ptl;
3832 src_pte = huge_pte_offset(src, addr, sz);
3835 dst_pte = huge_pte_alloc(dst, addr, sz);
3842 * If the pagetables are shared don't copy or take references.
3843 * dst_pte == src_pte is the common case of src/dest sharing.
3845 * However, src could have 'unshared' and dst shares with
3846 * another vma. If dst_pte !none, this implies sharing.
3847 * Check here before taking page table lock, and once again
3848 * after taking the lock below.
3850 dst_entry = huge_ptep_get(dst_pte);
3851 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3854 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3855 src_ptl = huge_pte_lockptr(h, src, src_pte);
3856 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3857 entry = huge_ptep_get(src_pte);
3858 dst_entry = huge_ptep_get(dst_pte);
3859 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3861 * Skip if src entry none. Also, skip in the
3862 * unlikely case dst entry !none as this implies
3863 * sharing with another vma.
3866 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3867 is_hugetlb_entry_hwpoisoned(entry))) {
3868 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3870 if (is_write_migration_entry(swp_entry) && cow) {
3872 * COW mappings require pages in both
3873 * parent and child to be set to read.
3875 make_migration_entry_read(&swp_entry);
3876 entry = swp_entry_to_pte(swp_entry);
3877 set_huge_swap_pte_at(src, addr, src_pte,
3880 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3884 * No need to notify as we are downgrading page
3885 * table protection not changing it to point
3888 * See Documentation/vm/mmu_notifier.rst
3890 huge_ptep_set_wrprotect(src, addr, src_pte);
3892 entry = huge_ptep_get(src_pte);
3893 ptepage = pte_page(entry);
3895 page_dup_rmap(ptepage, true);
3896 set_huge_pte_at(dst, addr, dst_pte, entry);
3897 hugetlb_count_add(pages_per_huge_page(h), dst);
3899 spin_unlock(src_ptl);
3900 spin_unlock(dst_ptl);
3904 mmu_notifier_invalidate_range_end(&range);
3906 i_mmap_unlock_read(mapping);
3911 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3912 unsigned long start, unsigned long end,
3913 struct page *ref_page)
3915 struct mm_struct *mm = vma->vm_mm;
3916 unsigned long address;
3921 struct hstate *h = hstate_vma(vma);
3922 unsigned long sz = huge_page_size(h);
3923 struct mmu_notifier_range range;
3925 WARN_ON(!is_vm_hugetlb_page(vma));
3926 BUG_ON(start & ~huge_page_mask(h));
3927 BUG_ON(end & ~huge_page_mask(h));
3930 * This is a hugetlb vma, all the pte entries should point
3933 tlb_change_page_size(tlb, sz);
3934 tlb_start_vma(tlb, vma);
3937 * If sharing possible, alert mmu notifiers of worst case.
3939 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3941 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3942 mmu_notifier_invalidate_range_start(&range);
3944 for (; address < end; address += sz) {
3945 ptep = huge_pte_offset(mm, address, sz);
3949 ptl = huge_pte_lock(h, mm, ptep);
3950 if (huge_pmd_unshare(mm, &address, ptep)) {
3953 * We just unmapped a page of PMDs by clearing a PUD.
3954 * The caller's TLB flush range should cover this area.
3959 pte = huge_ptep_get(ptep);
3960 if (huge_pte_none(pte)) {
3966 * Migrating hugepage or HWPoisoned hugepage is already
3967 * unmapped and its refcount is dropped, so just clear pte here.
3969 if (unlikely(!pte_present(pte))) {
3970 huge_pte_clear(mm, address, ptep, sz);
3975 page = pte_page(pte);
3977 * If a reference page is supplied, it is because a specific
3978 * page is being unmapped, not a range. Ensure the page we
3979 * are about to unmap is the actual page of interest.
3982 if (page != ref_page) {
3987 * Mark the VMA as having unmapped its page so that
3988 * future faults in this VMA will fail rather than
3989 * looking like data was lost
3991 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3994 pte = huge_ptep_get_and_clear(mm, address, ptep);
3995 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3996 if (huge_pte_dirty(pte))
3997 set_page_dirty(page);
3999 hugetlb_count_sub(pages_per_huge_page(h), mm);
4000 page_remove_rmap(page, true);
4003 tlb_remove_page_size(tlb, page, huge_page_size(h));
4005 * Bail out after unmapping reference page if supplied
4010 mmu_notifier_invalidate_range_end(&range);
4011 tlb_end_vma(tlb, vma);
4014 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4015 struct vm_area_struct *vma, unsigned long start,
4016 unsigned long end, struct page *ref_page)
4018 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4021 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4022 * test will fail on a vma being torn down, and not grab a page table
4023 * on its way out. We're lucky that the flag has such an appropriate
4024 * name, and can in fact be safely cleared here. We could clear it
4025 * before the __unmap_hugepage_range above, but all that's necessary
4026 * is to clear it before releasing the i_mmap_rwsem. This works
4027 * because in the context this is called, the VMA is about to be
4028 * destroyed and the i_mmap_rwsem is held.
4030 vma->vm_flags &= ~VM_MAYSHARE;
4033 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4034 unsigned long end, struct page *ref_page)
4036 struct mm_struct *mm;
4037 struct mmu_gather tlb;
4038 unsigned long tlb_start = start;
4039 unsigned long tlb_end = end;
4042 * If shared PMDs were possibly used within this vma range, adjust
4043 * start/end for worst case tlb flushing.
4044 * Note that we can not be sure if PMDs are shared until we try to
4045 * unmap pages. However, we want to make sure TLB flushing covers
4046 * the largest possible range.
4048 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4052 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4053 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4054 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4058 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4059 * mappping it owns the reserve page for. The intention is to unmap the page
4060 * from other VMAs and let the children be SIGKILLed if they are faulting the
4063 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4064 struct page *page, unsigned long address)
4066 struct hstate *h = hstate_vma(vma);
4067 struct vm_area_struct *iter_vma;
4068 struct address_space *mapping;
4072 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4073 * from page cache lookup which is in HPAGE_SIZE units.
4075 address = address & huge_page_mask(h);
4076 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4078 mapping = vma->vm_file->f_mapping;
4081 * Take the mapping lock for the duration of the table walk. As
4082 * this mapping should be shared between all the VMAs,
4083 * __unmap_hugepage_range() is called as the lock is already held
4085 i_mmap_lock_write(mapping);
4086 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4087 /* Do not unmap the current VMA */
4088 if (iter_vma == vma)
4092 * Shared VMAs have their own reserves and do not affect
4093 * MAP_PRIVATE accounting but it is possible that a shared
4094 * VMA is using the same page so check and skip such VMAs.
4096 if (iter_vma->vm_flags & VM_MAYSHARE)
4100 * Unmap the page from other VMAs without their own reserves.
4101 * They get marked to be SIGKILLed if they fault in these
4102 * areas. This is because a future no-page fault on this VMA
4103 * could insert a zeroed page instead of the data existing
4104 * from the time of fork. This would look like data corruption
4106 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4107 unmap_hugepage_range(iter_vma, address,
4108 address + huge_page_size(h), page);
4110 i_mmap_unlock_write(mapping);
4114 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4115 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4116 * cannot race with other handlers or page migration.
4117 * Keep the pte_same checks anyway to make transition from the mutex easier.
4119 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4120 unsigned long address, pte_t *ptep,
4121 struct page *pagecache_page, spinlock_t *ptl)
4124 struct hstate *h = hstate_vma(vma);
4125 struct page *old_page, *new_page;
4126 int outside_reserve = 0;
4128 unsigned long haddr = address & huge_page_mask(h);
4129 struct mmu_notifier_range range;
4131 pte = huge_ptep_get(ptep);
4132 old_page = pte_page(pte);
4135 /* If no-one else is actually using this page, avoid the copy
4136 * and just make the page writable */
4137 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4138 page_move_anon_rmap(old_page, vma);
4139 set_huge_ptep_writable(vma, haddr, ptep);
4144 * If the process that created a MAP_PRIVATE mapping is about to
4145 * perform a COW due to a shared page count, attempt to satisfy
4146 * the allocation without using the existing reserves. The pagecache
4147 * page is used to determine if the reserve at this address was
4148 * consumed or not. If reserves were used, a partial faulted mapping
4149 * at the time of fork() could consume its reserves on COW instead
4150 * of the full address range.
4152 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4153 old_page != pagecache_page)
4154 outside_reserve = 1;
4159 * Drop page table lock as buddy allocator may be called. It will
4160 * be acquired again before returning to the caller, as expected.
4163 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4165 if (IS_ERR(new_page)) {
4167 * If a process owning a MAP_PRIVATE mapping fails to COW,
4168 * it is due to references held by a child and an insufficient
4169 * huge page pool. To guarantee the original mappers
4170 * reliability, unmap the page from child processes. The child
4171 * may get SIGKILLed if it later faults.
4173 if (outside_reserve) {
4175 BUG_ON(huge_pte_none(pte));
4176 unmap_ref_private(mm, vma, old_page, haddr);
4177 BUG_ON(huge_pte_none(pte));
4179 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4181 pte_same(huge_ptep_get(ptep), pte)))
4182 goto retry_avoidcopy;
4184 * race occurs while re-acquiring page table
4185 * lock, and our job is done.
4190 ret = vmf_error(PTR_ERR(new_page));
4191 goto out_release_old;
4195 * When the original hugepage is shared one, it does not have
4196 * anon_vma prepared.
4198 if (unlikely(anon_vma_prepare(vma))) {
4200 goto out_release_all;
4203 copy_user_huge_page(new_page, old_page, address, vma,
4204 pages_per_huge_page(h));
4205 __SetPageUptodate(new_page);
4207 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4208 haddr + huge_page_size(h));
4209 mmu_notifier_invalidate_range_start(&range);
4212 * Retake the page table lock to check for racing updates
4213 * before the page tables are altered
4216 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4217 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4218 ClearPagePrivate(new_page);
4221 huge_ptep_clear_flush(vma, haddr, ptep);
4222 mmu_notifier_invalidate_range(mm, range.start, range.end);
4223 set_huge_pte_at(mm, haddr, ptep,
4224 make_huge_pte(vma, new_page, 1));
4225 page_remove_rmap(old_page, true);
4226 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4227 set_page_huge_active(new_page);
4228 /* Make the old page be freed below */
4229 new_page = old_page;
4232 mmu_notifier_invalidate_range_end(&range);
4234 restore_reserve_on_error(h, vma, haddr, new_page);
4239 spin_lock(ptl); /* Caller expects lock to be held */
4243 /* Return the pagecache page at a given address within a VMA */
4244 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4245 struct vm_area_struct *vma, unsigned long address)
4247 struct address_space *mapping;
4250 mapping = vma->vm_file->f_mapping;
4251 idx = vma_hugecache_offset(h, vma, address);
4253 return find_lock_page(mapping, idx);
4257 * Return whether there is a pagecache page to back given address within VMA.
4258 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4260 static bool hugetlbfs_pagecache_present(struct hstate *h,
4261 struct vm_area_struct *vma, unsigned long address)
4263 struct address_space *mapping;
4267 mapping = vma->vm_file->f_mapping;
4268 idx = vma_hugecache_offset(h, vma, address);
4270 page = find_get_page(mapping, idx);
4273 return page != NULL;
4276 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4279 struct inode *inode = mapping->host;
4280 struct hstate *h = hstate_inode(inode);
4281 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4285 ClearPagePrivate(page);
4288 * set page dirty so that it will not be removed from cache/file
4289 * by non-hugetlbfs specific code paths.
4291 set_page_dirty(page);
4293 spin_lock(&inode->i_lock);
4294 inode->i_blocks += blocks_per_huge_page(h);
4295 spin_unlock(&inode->i_lock);
4299 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4300 struct vm_area_struct *vma,
4301 struct address_space *mapping, pgoff_t idx,
4302 unsigned long address, pte_t *ptep, unsigned int flags)
4304 struct hstate *h = hstate_vma(vma);
4305 vm_fault_t ret = VM_FAULT_SIGBUS;
4311 unsigned long haddr = address & huge_page_mask(h);
4312 bool new_page = false;
4315 * Currently, we are forced to kill the process in the event the
4316 * original mapper has unmapped pages from the child due to a failed
4317 * COW. Warn that such a situation has occurred as it may not be obvious
4319 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4320 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4326 * We can not race with truncation due to holding i_mmap_rwsem.
4327 * i_size is modified when holding i_mmap_rwsem, so check here
4328 * once for faults beyond end of file.
4330 size = i_size_read(mapping->host) >> huge_page_shift(h);
4335 page = find_lock_page(mapping, idx);
4338 * Check for page in userfault range
4340 if (userfaultfd_missing(vma)) {
4342 struct vm_fault vmf = {
4347 * Hard to debug if it ends up being
4348 * used by a callee that assumes
4349 * something about the other
4350 * uninitialized fields... same as in
4356 * hugetlb_fault_mutex and i_mmap_rwsem must be
4357 * dropped before handling userfault. Reacquire
4358 * after handling fault to make calling code simpler.
4360 hash = hugetlb_fault_mutex_hash(mapping, idx);
4361 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4362 i_mmap_unlock_read(mapping);
4363 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4364 i_mmap_lock_read(mapping);
4365 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4369 page = alloc_huge_page(vma, haddr, 0);
4372 * Returning error will result in faulting task being
4373 * sent SIGBUS. The hugetlb fault mutex prevents two
4374 * tasks from racing to fault in the same page which
4375 * could result in false unable to allocate errors.
4376 * Page migration does not take the fault mutex, but
4377 * does a clear then write of pte's under page table
4378 * lock. Page fault code could race with migration,
4379 * notice the clear pte and try to allocate a page
4380 * here. Before returning error, get ptl and make
4381 * sure there really is no pte entry.
4383 ptl = huge_pte_lock(h, mm, ptep);
4384 if (!huge_pte_none(huge_ptep_get(ptep))) {
4390 ret = vmf_error(PTR_ERR(page));
4393 clear_huge_page(page, address, pages_per_huge_page(h));
4394 __SetPageUptodate(page);
4397 if (vma->vm_flags & VM_MAYSHARE) {
4398 int err = huge_add_to_page_cache(page, mapping, idx);
4407 if (unlikely(anon_vma_prepare(vma))) {
4409 goto backout_unlocked;
4415 * If memory error occurs between mmap() and fault, some process
4416 * don't have hwpoisoned swap entry for errored virtual address.
4417 * So we need to block hugepage fault by PG_hwpoison bit check.
4419 if (unlikely(PageHWPoison(page))) {
4420 ret = VM_FAULT_HWPOISON |
4421 VM_FAULT_SET_HINDEX(hstate_index(h));
4422 goto backout_unlocked;
4427 * If we are going to COW a private mapping later, we examine the
4428 * pending reservations for this page now. This will ensure that
4429 * any allocations necessary to record that reservation occur outside
4432 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4433 if (vma_needs_reservation(h, vma, haddr) < 0) {
4435 goto backout_unlocked;
4437 /* Just decrements count, does not deallocate */
4438 vma_end_reservation(h, vma, haddr);
4441 ptl = huge_pte_lock(h, mm, ptep);
4443 if (!huge_pte_none(huge_ptep_get(ptep)))
4447 ClearPagePrivate(page);
4448 hugepage_add_new_anon_rmap(page, vma, haddr);
4450 page_dup_rmap(page, true);
4451 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4452 && (vma->vm_flags & VM_SHARED)));
4453 set_huge_pte_at(mm, haddr, ptep, new_pte);
4455 hugetlb_count_add(pages_per_huge_page(h), mm);
4456 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4457 /* Optimization, do the COW without a second fault */
4458 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4464 * Only make newly allocated pages active. Existing pages found
4465 * in the pagecache could be !page_huge_active() if they have been
4466 * isolated for migration.
4469 set_page_huge_active(page);
4479 restore_reserve_on_error(h, vma, haddr, page);
4485 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4487 unsigned long key[2];
4490 key[0] = (unsigned long) mapping;
4493 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4495 return hash & (num_fault_mutexes - 1);
4499 * For uniprocesor systems we always use a single mutex, so just
4500 * return 0 and avoid the hashing overhead.
4502 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4508 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4509 unsigned long address, unsigned int flags)
4516 struct page *page = NULL;
4517 struct page *pagecache_page = NULL;
4518 struct hstate *h = hstate_vma(vma);
4519 struct address_space *mapping;
4520 int need_wait_lock = 0;
4521 unsigned long haddr = address & huge_page_mask(h);
4523 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4526 * Since we hold no locks, ptep could be stale. That is
4527 * OK as we are only making decisions based on content and
4528 * not actually modifying content here.
4530 entry = huge_ptep_get(ptep);
4531 if (unlikely(is_hugetlb_entry_migration(entry))) {
4532 migration_entry_wait_huge(vma, mm, ptep);
4534 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4535 return VM_FAULT_HWPOISON_LARGE |
4536 VM_FAULT_SET_HINDEX(hstate_index(h));
4538 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4540 return VM_FAULT_OOM;
4544 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4545 * until finished with ptep. This serves two purposes:
4546 * 1) It prevents huge_pmd_unshare from being called elsewhere
4547 * and making the ptep no longer valid.
4548 * 2) It synchronizes us with i_size modifications during truncation.
4550 * ptep could have already be assigned via huge_pte_offset. That
4551 * is OK, as huge_pte_alloc will return the same value unless
4552 * something has changed.
4554 mapping = vma->vm_file->f_mapping;
4555 i_mmap_lock_read(mapping);
4556 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4558 i_mmap_unlock_read(mapping);
4559 return VM_FAULT_OOM;
4563 * Serialize hugepage allocation and instantiation, so that we don't
4564 * get spurious allocation failures if two CPUs race to instantiate
4565 * the same page in the page cache.
4567 idx = vma_hugecache_offset(h, vma, haddr);
4568 hash = hugetlb_fault_mutex_hash(mapping, idx);
4569 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4571 entry = huge_ptep_get(ptep);
4572 if (huge_pte_none(entry)) {
4573 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4580 * entry could be a migration/hwpoison entry at this point, so this
4581 * check prevents the kernel from going below assuming that we have
4582 * an active hugepage in pagecache. This goto expects the 2nd page
4583 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4584 * properly handle it.
4586 if (!pte_present(entry))
4590 * If we are going to COW the mapping later, we examine the pending
4591 * reservations for this page now. This will ensure that any
4592 * allocations necessary to record that reservation occur outside the
4593 * spinlock. For private mappings, we also lookup the pagecache
4594 * page now as it is used to determine if a reservation has been
4597 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4598 if (vma_needs_reservation(h, vma, haddr) < 0) {
4602 /* Just decrements count, does not deallocate */
4603 vma_end_reservation(h, vma, haddr);
4605 if (!(vma->vm_flags & VM_MAYSHARE))
4606 pagecache_page = hugetlbfs_pagecache_page(h,
4610 ptl = huge_pte_lock(h, mm, ptep);
4612 /* Check for a racing update before calling hugetlb_cow */
4613 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4617 * hugetlb_cow() requires page locks of pte_page(entry) and
4618 * pagecache_page, so here we need take the former one
4619 * when page != pagecache_page or !pagecache_page.
4621 page = pte_page(entry);
4622 if (page != pagecache_page)
4623 if (!trylock_page(page)) {
4630 if (flags & FAULT_FLAG_WRITE) {
4631 if (!huge_pte_write(entry)) {
4632 ret = hugetlb_cow(mm, vma, address, ptep,
4633 pagecache_page, ptl);
4636 entry = huge_pte_mkdirty(entry);
4638 entry = pte_mkyoung(entry);
4639 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4640 flags & FAULT_FLAG_WRITE))
4641 update_mmu_cache(vma, haddr, ptep);
4643 if (page != pagecache_page)
4649 if (pagecache_page) {
4650 unlock_page(pagecache_page);
4651 put_page(pagecache_page);
4654 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4655 i_mmap_unlock_read(mapping);
4657 * Generally it's safe to hold refcount during waiting page lock. But
4658 * here we just wait to defer the next page fault to avoid busy loop and
4659 * the page is not used after unlocked before returning from the current
4660 * page fault. So we are safe from accessing freed page, even if we wait
4661 * here without taking refcount.
4664 wait_on_page_locked(page);
4669 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4670 * modifications for huge pages.
4672 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4674 struct vm_area_struct *dst_vma,
4675 unsigned long dst_addr,
4676 unsigned long src_addr,
4677 struct page **pagep)
4679 struct address_space *mapping;
4682 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4683 struct hstate *h = hstate_vma(dst_vma);
4691 page = alloc_huge_page(dst_vma, dst_addr, 0);
4695 ret = copy_huge_page_from_user(page,
4696 (const void __user *) src_addr,
4697 pages_per_huge_page(h), false);
4699 /* fallback to copy_from_user outside mmap_sem */
4700 if (unlikely(ret)) {
4703 /* don't free the page */
4712 * The memory barrier inside __SetPageUptodate makes sure that
4713 * preceding stores to the page contents become visible before
4714 * the set_pte_at() write.
4716 __SetPageUptodate(page);
4718 mapping = dst_vma->vm_file->f_mapping;
4719 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4722 * If shared, add to page cache
4725 size = i_size_read(mapping->host) >> huge_page_shift(h);
4728 goto out_release_nounlock;
4731 * Serialization between remove_inode_hugepages() and
4732 * huge_add_to_page_cache() below happens through the
4733 * hugetlb_fault_mutex_table that here must be hold by
4736 ret = huge_add_to_page_cache(page, mapping, idx);
4738 goto out_release_nounlock;
4741 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4745 * Recheck the i_size after holding PT lock to make sure not
4746 * to leave any page mapped (as page_mapped()) beyond the end
4747 * of the i_size (remove_inode_hugepages() is strict about
4748 * enforcing that). If we bail out here, we'll also leave a
4749 * page in the radix tree in the vm_shared case beyond the end
4750 * of the i_size, but remove_inode_hugepages() will take care
4751 * of it as soon as we drop the hugetlb_fault_mutex_table.
4753 size = i_size_read(mapping->host) >> huge_page_shift(h);
4756 goto out_release_unlock;
4759 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4760 goto out_release_unlock;
4763 page_dup_rmap(page, true);
4765 ClearPagePrivate(page);
4766 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4769 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4770 if (dst_vma->vm_flags & VM_WRITE)
4771 _dst_pte = huge_pte_mkdirty(_dst_pte);
4772 _dst_pte = pte_mkyoung(_dst_pte);
4774 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4776 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4777 dst_vma->vm_flags & VM_WRITE);
4778 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4780 /* No need to invalidate - it was non-present before */
4781 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4784 set_page_huge_active(page);
4794 out_release_nounlock:
4799 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4800 struct page **pages, struct vm_area_struct **vmas,
4801 unsigned long *position, unsigned long *nr_pages,
4802 long i, unsigned int flags, int *locked)
4804 unsigned long pfn_offset;
4805 unsigned long vaddr = *position;
4806 unsigned long remainder = *nr_pages;
4807 struct hstate *h = hstate_vma(vma);
4810 while (vaddr < vma->vm_end && remainder) {
4812 spinlock_t *ptl = NULL;
4817 * If we have a pending SIGKILL, don't keep faulting pages and
4818 * potentially allocating memory.
4820 if (fatal_signal_pending(current)) {
4826 * Some archs (sparc64, sh*) have multiple pte_ts to
4827 * each hugepage. We have to make sure we get the
4828 * first, for the page indexing below to work.
4830 * Note that page table lock is not held when pte is null.
4832 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4835 ptl = huge_pte_lock(h, mm, pte);
4836 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4839 * When coredumping, it suits get_dump_page if we just return
4840 * an error where there's an empty slot with no huge pagecache
4841 * to back it. This way, we avoid allocating a hugepage, and
4842 * the sparse dumpfile avoids allocating disk blocks, but its
4843 * huge holes still show up with zeroes where they need to be.
4845 if (absent && (flags & FOLL_DUMP) &&
4846 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4854 * We need call hugetlb_fault for both hugepages under migration
4855 * (in which case hugetlb_fault waits for the migration,) and
4856 * hwpoisoned hugepages (in which case we need to prevent the
4857 * caller from accessing to them.) In order to do this, we use
4858 * here is_swap_pte instead of is_hugetlb_entry_migration and
4859 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4860 * both cases, and because we can't follow correct pages
4861 * directly from any kind of swap entries.
4863 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4864 ((flags & FOLL_WRITE) &&
4865 !huge_pte_write(huge_ptep_get(pte)))) {
4867 unsigned int fault_flags = 0;
4871 if (flags & FOLL_WRITE)
4872 fault_flags |= FAULT_FLAG_WRITE;
4874 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4875 FAULT_FLAG_KILLABLE;
4876 if (flags & FOLL_NOWAIT)
4877 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4878 FAULT_FLAG_RETRY_NOWAIT;
4879 if (flags & FOLL_TRIED) {
4881 * Note: FAULT_FLAG_ALLOW_RETRY and
4882 * FAULT_FLAG_TRIED can co-exist
4884 fault_flags |= FAULT_FLAG_TRIED;
4886 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4887 if (ret & VM_FAULT_ERROR) {
4888 err = vm_fault_to_errno(ret, flags);
4892 if (ret & VM_FAULT_RETRY) {
4894 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4898 * VM_FAULT_RETRY must not return an
4899 * error, it will return zero
4902 * No need to update "position" as the
4903 * caller will not check it after
4904 * *nr_pages is set to 0.
4911 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4912 page = pte_page(huge_ptep_get(pte));
4915 * If subpage information not requested, update counters
4916 * and skip the same_page loop below.
4918 if (!pages && !vmas && !pfn_offset &&
4919 (vaddr + huge_page_size(h) < vma->vm_end) &&
4920 (remainder >= pages_per_huge_page(h))) {
4921 vaddr += huge_page_size(h);
4922 remainder -= pages_per_huge_page(h);
4923 i += pages_per_huge_page(h);
4930 pages[i] = mem_map_offset(page, pfn_offset);
4932 * try_grab_page() should always succeed here, because:
4933 * a) we hold the ptl lock, and b) we've just checked
4934 * that the huge page is present in the page tables. If
4935 * the huge page is present, then the tail pages must
4936 * also be present. The ptl prevents the head page and
4937 * tail pages from being rearranged in any way. So this
4938 * page must be available at this point, unless the page
4939 * refcount overflowed:
4941 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4956 if (vaddr < vma->vm_end && remainder &&
4957 pfn_offset < pages_per_huge_page(h)) {
4959 * We use pfn_offset to avoid touching the pageframes
4960 * of this compound page.
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, &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);
5188 } else if (unlikely(chg > add)) {
5190 * pages in this range were added to the reserve
5191 * map between region_chg and region_add. This
5192 * indicates a race with alloc_huge_page. Adjust
5193 * the subpool and reserve counts modified above
5194 * based on the difference.
5198 hugetlb_cgroup_uncharge_cgroup_rsvd(
5200 (chg - add) * pages_per_huge_page(h), h_cg);
5202 rsv_adjust = hugepage_subpool_put_pages(spool,
5204 hugetlb_acct_memory(h, -rsv_adjust);
5209 /* put back original number of pages, chg */
5210 (void)hugepage_subpool_put_pages(spool, chg);
5211 out_uncharge_cgroup:
5212 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5213 chg * pages_per_huge_page(h), h_cg);
5215 if (!vma || vma->vm_flags & VM_MAYSHARE)
5216 /* Only call region_abort if the region_chg succeeded but the
5217 * region_add failed or didn't run.
5219 if (chg >= 0 && add < 0)
5220 region_abort(resv_map, from, to, regions_needed);
5221 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5222 kref_put(&resv_map->refs, resv_map_release);
5226 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5229 struct hstate *h = hstate_inode(inode);
5230 struct resv_map *resv_map = inode_resv_map(inode);
5232 struct hugepage_subpool *spool = subpool_inode(inode);
5236 * Since this routine can be called in the evict inode path for all
5237 * hugetlbfs inodes, resv_map could be NULL.
5240 chg = region_del(resv_map, start, end);
5242 * region_del() can fail in the rare case where a region
5243 * must be split and another region descriptor can not be
5244 * allocated. If end == LONG_MAX, it will not fail.
5250 spin_lock(&inode->i_lock);
5251 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5252 spin_unlock(&inode->i_lock);
5255 * If the subpool has a minimum size, the number of global
5256 * reservations to be released may be adjusted.
5258 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5259 hugetlb_acct_memory(h, -gbl_reserve);
5264 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5265 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5266 struct vm_area_struct *vma,
5267 unsigned long addr, pgoff_t idx)
5269 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5271 unsigned long sbase = saddr & PUD_MASK;
5272 unsigned long s_end = sbase + PUD_SIZE;
5274 /* Allow segments to share if only one is marked locked */
5275 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5276 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5279 * match the virtual addresses, permission and the alignment of the
5282 if (pmd_index(addr) != pmd_index(saddr) ||
5283 vm_flags != svm_flags ||
5284 sbase < svma->vm_start || svma->vm_end < s_end)
5290 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5292 unsigned long base = addr & PUD_MASK;
5293 unsigned long end = base + PUD_SIZE;
5296 * check on proper vm_flags and page table alignment
5298 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5304 * Determine if start,end range within vma could be mapped by shared pmd.
5305 * If yes, adjust start and end to cover range associated with possible
5306 * shared pmd mappings.
5308 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5309 unsigned long *start, unsigned long *end)
5311 unsigned long check_addr;
5313 if (!(vma->vm_flags & VM_MAYSHARE))
5316 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
5317 unsigned long a_start = check_addr & PUD_MASK;
5318 unsigned long a_end = a_start + PUD_SIZE;
5321 * If sharing is possible, adjust start/end if necessary.
5323 if (range_in_vma(vma, a_start, a_end)) {
5324 if (a_start < *start)
5333 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5334 * and returns the corresponding pte. While this is not necessary for the
5335 * !shared pmd case because we can allocate the pmd later as well, it makes the
5336 * code much cleaner.
5338 * This routine must be called with i_mmap_rwsem held in at least read mode.
5339 * For hugetlbfs, this prevents removal of any page table entries associated
5340 * with the address space. This is important as we are setting up sharing
5341 * based on existing page table entries (mappings).
5343 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5345 struct vm_area_struct *vma = find_vma(mm, addr);
5346 struct address_space *mapping = vma->vm_file->f_mapping;
5347 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5349 struct vm_area_struct *svma;
5350 unsigned long saddr;
5355 if (!vma_shareable(vma, addr))
5356 return (pte_t *)pmd_alloc(mm, pud, addr);
5358 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5362 saddr = page_table_shareable(svma, vma, addr, idx);
5364 spte = huge_pte_offset(svma->vm_mm, saddr,
5365 vma_mmu_pagesize(svma));
5367 get_page(virt_to_page(spte));
5376 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5377 if (pud_none(*pud)) {
5378 pud_populate(mm, pud,
5379 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5382 put_page(virt_to_page(spte));
5386 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5391 * unmap huge page backed by shared pte.
5393 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5394 * indicated by page_count > 1, unmap is achieved by clearing pud and
5395 * decrementing the ref count. If count == 1, the pte page is not shared.
5397 * Called with page table lock held and i_mmap_rwsem held in write mode.
5399 * returns: 1 successfully unmapped a shared pte page
5400 * 0 the underlying pte page is not shared, or it is the last user
5402 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5404 pgd_t *pgd = pgd_offset(mm, *addr);
5405 p4d_t *p4d = p4d_offset(pgd, *addr);
5406 pud_t *pud = pud_offset(p4d, *addr);
5408 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5409 if (page_count(virt_to_page(ptep)) == 1)
5413 put_page(virt_to_page(ptep));
5415 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5418 #define want_pmd_share() (1)
5419 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5420 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5425 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5430 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5431 unsigned long *start, unsigned long *end)
5434 #define want_pmd_share() (0)
5435 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5437 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5438 pte_t *huge_pte_alloc(struct mm_struct *mm,
5439 unsigned long addr, unsigned long sz)
5446 pgd = pgd_offset(mm, addr);
5447 p4d = p4d_alloc(mm, pgd, addr);
5450 pud = pud_alloc(mm, p4d, addr);
5452 if (sz == PUD_SIZE) {
5455 BUG_ON(sz != PMD_SIZE);
5456 if (want_pmd_share() && pud_none(*pud))
5457 pte = huge_pmd_share(mm, addr, pud);
5459 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5462 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5468 * huge_pte_offset() - Walk the page table to resolve the hugepage
5469 * entry at address @addr
5471 * Return: Pointer to page table entry (PUD or PMD) for
5472 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5473 * size @sz doesn't match the hugepage size at this level of the page
5476 pte_t *huge_pte_offset(struct mm_struct *mm,
5477 unsigned long addr, unsigned long sz)
5484 pgd = pgd_offset(mm, addr);
5485 if (!pgd_present(*pgd))
5487 p4d = p4d_offset(pgd, addr);
5488 if (!p4d_present(*p4d))
5491 pud = pud_offset(p4d, addr);
5493 /* must be pud huge, non-present or none */
5494 return (pte_t *)pud;
5495 if (!pud_present(*pud))
5497 /* must have a valid entry and size to go further */
5499 pmd = pmd_offset(pud, addr);
5500 /* must be pmd huge, non-present or none */
5501 return (pte_t *)pmd;
5504 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5507 * These functions are overwritable if your architecture needs its own
5510 struct page * __weak
5511 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5514 return ERR_PTR(-EINVAL);
5517 struct page * __weak
5518 follow_huge_pd(struct vm_area_struct *vma,
5519 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5521 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5525 struct page * __weak
5526 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5527 pmd_t *pmd, int flags)
5529 struct page *page = NULL;
5533 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5534 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5535 (FOLL_PIN | FOLL_GET)))
5539 ptl = pmd_lockptr(mm, pmd);
5542 * make sure that the address range covered by this pmd is not
5543 * unmapped from other threads.
5545 if (!pmd_huge(*pmd))
5547 pte = huge_ptep_get((pte_t *)pmd);
5548 if (pte_present(pte)) {
5549 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5551 * try_grab_page() should always succeed here, because: a) we
5552 * hold the pmd (ptl) lock, and b) we've just checked that the
5553 * huge pmd (head) page is present in the page tables. The ptl
5554 * prevents the head page and tail pages from being rearranged
5555 * in any way. So this page must be available at this point,
5556 * unless the page refcount overflowed:
5558 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5563 if (is_hugetlb_entry_migration(pte)) {
5565 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5569 * hwpoisoned entry is treated as no_page_table in
5570 * follow_page_mask().
5578 struct page * __weak
5579 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5580 pud_t *pud, int flags)
5582 if (flags & (FOLL_GET | FOLL_PIN))
5585 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5588 struct page * __weak
5589 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5591 if (flags & (FOLL_GET | FOLL_PIN))
5594 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5597 bool isolate_huge_page(struct page *page, struct list_head *list)
5601 VM_BUG_ON_PAGE(!PageHead(page), page);
5602 spin_lock(&hugetlb_lock);
5603 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5607 clear_page_huge_active(page);
5608 list_move_tail(&page->lru, list);
5610 spin_unlock(&hugetlb_lock);
5614 void putback_active_hugepage(struct page *page)
5616 VM_BUG_ON_PAGE(!PageHead(page), page);
5617 spin_lock(&hugetlb_lock);
5618 set_page_huge_active(page);
5619 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5620 spin_unlock(&hugetlb_lock);
5624 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5626 struct hstate *h = page_hstate(oldpage);
5628 hugetlb_cgroup_migrate(oldpage, newpage);
5629 set_page_owner_migrate_reason(newpage, reason);
5632 * transfer temporary state of the new huge page. This is
5633 * reverse to other transitions because the newpage is going to
5634 * be final while the old one will be freed so it takes over
5635 * the temporary status.
5637 * Also note that we have to transfer the per-node surplus state
5638 * here as well otherwise the global surplus count will not match
5641 if (PageHugeTemporary(newpage)) {
5642 int old_nid = page_to_nid(oldpage);
5643 int new_nid = page_to_nid(newpage);
5645 SetPageHugeTemporary(oldpage);
5646 ClearPageHugeTemporary(newpage);
5648 spin_lock(&hugetlb_lock);
5649 if (h->surplus_huge_pages_node[old_nid]) {
5650 h->surplus_huge_pages_node[old_nid]--;
5651 h->surplus_huge_pages_node[new_nid]++;
5653 spin_unlock(&hugetlb_lock);
5658 static unsigned long hugetlb_cma_size __initdata;
5659 static bool cma_reserve_called __initdata;
5661 static int __init cmdline_parse_hugetlb_cma(char *p)
5663 hugetlb_cma_size = memparse(p, &p);
5667 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5669 void __init hugetlb_cma_reserve(int order)
5671 unsigned long size, reserved, per_node;
5674 cma_reserve_called = true;
5676 if (!hugetlb_cma_size)
5679 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5680 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5681 (PAGE_SIZE << order) / SZ_1M);
5686 * If 3 GB area is requested on a machine with 4 numa nodes,
5687 * let's allocate 1 GB on first three nodes and ignore the last one.
5689 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5690 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5691 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5694 for_each_node_state(nid, N_ONLINE) {
5697 size = min(per_node, hugetlb_cma_size - reserved);
5698 size = round_up(size, PAGE_SIZE << order);
5700 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5701 0, false, "hugetlb",
5702 &hugetlb_cma[nid], nid);
5704 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5710 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5713 if (reserved >= hugetlb_cma_size)
5718 void __init hugetlb_cma_check(void)
5720 if (!hugetlb_cma_size || cma_reserve_called)
5723 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5726 #endif /* CONFIG_CMA */