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>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/node.h>
40 #include <linux/userfaultfd_k.h>
41 #include <linux/page_owner.h>
44 int hugetlb_max_hstate __read_mostly;
45 unsigned int default_hstate_idx;
46 struct hstate hstates[HUGE_MAX_HSTATE];
49 static struct cma *hugetlb_cma[MAX_NUMNODES];
51 static unsigned long hugetlb_cma_size __initdata;
54 * Minimum page order among possible hugepage sizes, set to a proper value
57 static unsigned int minimum_order __read_mostly = UINT_MAX;
59 __initdata LIST_HEAD(huge_boot_pages);
61 /* for command line parsing */
62 static struct hstate * __initdata parsed_hstate;
63 static unsigned long __initdata default_hstate_max_huge_pages;
64 static bool __initdata parsed_valid_hugepagesz = true;
65 static bool __initdata parsed_default_hugepagesz;
68 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
69 * free_huge_pages, and surplus_huge_pages.
71 DEFINE_SPINLOCK(hugetlb_lock);
74 * Serializes faults on the same logical page. This is used to
75 * prevent spurious OOMs when the hugepage pool is fully utilized.
77 static int num_fault_mutexes;
78 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
80 /* Forward declaration */
81 static int hugetlb_acct_memory(struct hstate *h, long delta);
83 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
85 bool free = (spool->count == 0) && (spool->used_hpages == 0);
87 spin_unlock(&spool->lock);
89 /* If no pages are used, and no other handles to the subpool
90 * remain, give up any reservations based on minimum size and
93 if (spool->min_hpages != -1)
94 hugetlb_acct_memory(spool->hstate,
100 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
103 struct hugepage_subpool *spool;
105 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
109 spin_lock_init(&spool->lock);
111 spool->max_hpages = max_hpages;
113 spool->min_hpages = min_hpages;
115 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
119 spool->rsv_hpages = min_hpages;
124 void hugepage_put_subpool(struct hugepage_subpool *spool)
126 spin_lock(&spool->lock);
127 BUG_ON(!spool->count);
129 unlock_or_release_subpool(spool);
133 * Subpool accounting for allocating and reserving pages.
134 * Return -ENOMEM if there are not enough resources to satisfy the
135 * the request. Otherwise, return the number of pages by which the
136 * global pools must be adjusted (upward). The returned value may
137 * only be different than the passed value (delta) in the case where
138 * a subpool minimum size must be maintained.
140 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
148 spin_lock(&spool->lock);
150 if (spool->max_hpages != -1) { /* maximum size accounting */
151 if ((spool->used_hpages + delta) <= spool->max_hpages)
152 spool->used_hpages += delta;
159 /* minimum size accounting */
160 if (spool->min_hpages != -1 && spool->rsv_hpages) {
161 if (delta > spool->rsv_hpages) {
163 * Asking for more reserves than those already taken on
164 * behalf of subpool. Return difference.
166 ret = delta - spool->rsv_hpages;
167 spool->rsv_hpages = 0;
169 ret = 0; /* reserves already accounted for */
170 spool->rsv_hpages -= delta;
175 spin_unlock(&spool->lock);
180 * Subpool accounting for freeing and unreserving pages.
181 * Return the number of global page reservations that must be dropped.
182 * The return value may only be different than the passed value (delta)
183 * in the case where a subpool minimum size must be maintained.
185 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
193 spin_lock(&spool->lock);
195 if (spool->max_hpages != -1) /* maximum size accounting */
196 spool->used_hpages -= delta;
198 /* minimum size accounting */
199 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
200 if (spool->rsv_hpages + delta <= spool->min_hpages)
203 ret = spool->rsv_hpages + delta - spool->min_hpages;
205 spool->rsv_hpages += delta;
206 if (spool->rsv_hpages > spool->min_hpages)
207 spool->rsv_hpages = spool->min_hpages;
211 * If hugetlbfs_put_super couldn't free spool due to an outstanding
212 * quota reference, free it now.
214 unlock_or_release_subpool(spool);
219 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
221 return HUGETLBFS_SB(inode->i_sb)->spool;
224 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
226 return subpool_inode(file_inode(vma->vm_file));
229 /* Helper that removes a struct file_region from the resv_map cache and returns
232 static struct file_region *
233 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
235 struct file_region *nrg = NULL;
237 VM_BUG_ON(resv->region_cache_count <= 0);
239 resv->region_cache_count--;
240 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
242 list_del(&nrg->link);
250 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
251 struct file_region *rg)
253 #ifdef CONFIG_CGROUP_HUGETLB
254 nrg->reservation_counter = rg->reservation_counter;
261 /* Helper that records hugetlb_cgroup uncharge info. */
262 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
264 struct resv_map *resv,
265 struct file_region *nrg)
267 #ifdef CONFIG_CGROUP_HUGETLB
269 nrg->reservation_counter =
270 &h_cg->rsvd_hugepage[hstate_index(h)];
271 nrg->css = &h_cg->css;
272 if (!resv->pages_per_hpage)
273 resv->pages_per_hpage = pages_per_huge_page(h);
274 /* pages_per_hpage should be the same for all entries in
277 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
279 nrg->reservation_counter = NULL;
285 static bool has_same_uncharge_info(struct file_region *rg,
286 struct file_region *org)
288 #ifdef CONFIG_CGROUP_HUGETLB
290 rg->reservation_counter == org->reservation_counter &&
298 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
300 struct file_region *nrg = NULL, *prg = NULL;
302 prg = list_prev_entry(rg, link);
303 if (&prg->link != &resv->regions && prg->to == rg->from &&
304 has_same_uncharge_info(prg, rg)) {
310 coalesce_file_region(resv, prg);
314 nrg = list_next_entry(rg, link);
315 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
316 has_same_uncharge_info(nrg, rg)) {
317 nrg->from = rg->from;
322 coalesce_file_region(resv, nrg);
327 /* Must be called with resv->lock held. Calling this with count_only == true
328 * will count the number of pages to be added but will not modify the linked
329 * list. If regions_needed != NULL and count_only == true, then regions_needed
330 * will indicate the number of file_regions needed in the cache to carry out to
331 * add the regions for this range.
333 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
334 struct hugetlb_cgroup *h_cg,
335 struct hstate *h, long *regions_needed,
339 struct list_head *head = &resv->regions;
340 long last_accounted_offset = f;
341 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
346 /* In this loop, we essentially handle an entry for the range
347 * [last_accounted_offset, rg->from), at every iteration, with some
350 list_for_each_entry_safe(rg, trg, head, link) {
351 /* Skip irrelevant regions that start before our range. */
353 /* If this region ends after the last accounted offset,
354 * then we need to update last_accounted_offset.
356 if (rg->to > last_accounted_offset)
357 last_accounted_offset = rg->to;
361 /* When we find a region that starts beyond our range, we've
367 /* Add an entry for last_accounted_offset -> rg->from, and
368 * update last_accounted_offset.
370 if (rg->from > last_accounted_offset) {
371 add += rg->from - last_accounted_offset;
373 nrg = get_file_region_entry_from_cache(
374 resv, last_accounted_offset, rg->from);
375 record_hugetlb_cgroup_uncharge_info(h_cg, h,
377 list_add(&nrg->link, rg->link.prev);
378 coalesce_file_region(resv, nrg);
379 } else if (regions_needed)
380 *regions_needed += 1;
383 last_accounted_offset = rg->to;
386 /* Handle the case where our range extends beyond
387 * last_accounted_offset.
389 if (last_accounted_offset < t) {
390 add += t - last_accounted_offset;
392 nrg = get_file_region_entry_from_cache(
393 resv, last_accounted_offset, t);
394 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
395 list_add(&nrg->link, rg->link.prev);
396 coalesce_file_region(resv, nrg);
397 } else if (regions_needed)
398 *regions_needed += 1;
405 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
407 static int allocate_file_region_entries(struct resv_map *resv,
409 __must_hold(&resv->lock)
411 struct list_head allocated_regions;
412 int to_allocate = 0, i = 0;
413 struct file_region *trg = NULL, *rg = NULL;
415 VM_BUG_ON(regions_needed < 0);
417 INIT_LIST_HEAD(&allocated_regions);
420 * Check for sufficient descriptors in the cache to accommodate
421 * the number of in progress add operations plus regions_needed.
423 * This is a while loop because when we drop the lock, some other call
424 * to region_add or region_del may have consumed some region_entries,
425 * so we keep looping here until we finally have enough entries for
426 * (adds_in_progress + regions_needed).
428 while (resv->region_cache_count <
429 (resv->adds_in_progress + regions_needed)) {
430 to_allocate = resv->adds_in_progress + regions_needed -
431 resv->region_cache_count;
433 /* At this point, we should have enough entries in the cache
434 * for all the existings adds_in_progress. We should only be
435 * needing to allocate for regions_needed.
437 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
439 spin_unlock(&resv->lock);
440 for (i = 0; i < to_allocate; i++) {
441 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
444 list_add(&trg->link, &allocated_regions);
447 spin_lock(&resv->lock);
449 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
451 list_add(&rg->link, &resv->region_cache);
452 resv->region_cache_count++;
459 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
467 * Add the huge page range represented by [f, t) to the reserve
468 * map. Regions will be taken from the cache to fill in this range.
469 * Sufficient regions should exist in the cache due to the previous
470 * call to region_chg with the same range, but in some cases the cache will not
471 * have sufficient entries due to races with other code doing region_add or
472 * region_del. The extra needed entries will be allocated.
474 * regions_needed is the out value provided by a previous call to region_chg.
476 * Return the number of new huge pages added to the map. This number is greater
477 * than or equal to zero. If file_region entries needed to be allocated for
478 * this operation and we were not able to allocate, it returns -ENOMEM.
479 * region_add of regions of length 1 never allocate file_regions and cannot
480 * fail; region_chg will always allocate at least 1 entry and a region_add for
481 * 1 page will only require at most 1 entry.
483 static long region_add(struct resv_map *resv, long f, long t,
484 long in_regions_needed, struct hstate *h,
485 struct hugetlb_cgroup *h_cg)
487 long add = 0, actual_regions_needed = 0;
489 spin_lock(&resv->lock);
492 /* Count how many regions are actually needed to execute this add. */
493 add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed,
497 * Check for sufficient descriptors in the cache to accommodate
498 * this add operation. Note that actual_regions_needed may be greater
499 * than in_regions_needed, as the resv_map may have been modified since
500 * the region_chg call. In this case, we need to make sure that we
501 * allocate extra entries, such that we have enough for all the
502 * existing adds_in_progress, plus the excess needed for this
505 if (actual_regions_needed > in_regions_needed &&
506 resv->region_cache_count <
507 resv->adds_in_progress +
508 (actual_regions_needed - in_regions_needed)) {
509 /* region_add operation of range 1 should never need to
510 * allocate file_region entries.
512 VM_BUG_ON(t - f <= 1);
514 if (allocate_file_region_entries(
515 resv, actual_regions_needed - in_regions_needed)) {
522 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false);
524 resv->adds_in_progress -= in_regions_needed;
526 spin_unlock(&resv->lock);
532 * Examine the existing reserve map and determine how many
533 * huge pages in the specified range [f, t) are NOT currently
534 * represented. This routine is called before a subsequent
535 * call to region_add that will actually modify the reserve
536 * map to add the specified range [f, t). region_chg does
537 * not change the number of huge pages represented by the
538 * map. A number of new file_region structures is added to the cache as a
539 * placeholder, for the subsequent region_add call to use. At least 1
540 * file_region structure is added.
542 * out_regions_needed is the number of regions added to the
543 * resv->adds_in_progress. This value needs to be provided to a follow up call
544 * to region_add or region_abort for proper accounting.
546 * Returns the number of huge pages that need to be added to the existing
547 * reservation map for the range [f, t). This number is greater or equal to
548 * zero. -ENOMEM is returned if a new file_region structure or cache entry
549 * is needed and can not be allocated.
551 static long region_chg(struct resv_map *resv, long f, long t,
552 long *out_regions_needed)
556 spin_lock(&resv->lock);
558 /* Count how many hugepages in this range are NOT respresented. */
559 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
560 out_regions_needed, true);
562 if (*out_regions_needed == 0)
563 *out_regions_needed = 1;
565 if (allocate_file_region_entries(resv, *out_regions_needed))
568 resv->adds_in_progress += *out_regions_needed;
570 spin_unlock(&resv->lock);
575 * Abort the in progress add operation. The adds_in_progress field
576 * of the resv_map keeps track of the operations in progress between
577 * calls to region_chg and region_add. Operations are sometimes
578 * aborted after the call to region_chg. In such cases, region_abort
579 * is called to decrement the adds_in_progress counter. regions_needed
580 * is the value returned by the region_chg call, it is used to decrement
581 * the adds_in_progress counter.
583 * NOTE: The range arguments [f, t) are not needed or used in this
584 * routine. They are kept to make reading the calling code easier as
585 * arguments will match the associated region_chg call.
587 static void region_abort(struct resv_map *resv, long f, long t,
590 spin_lock(&resv->lock);
591 VM_BUG_ON(!resv->region_cache_count);
592 resv->adds_in_progress -= regions_needed;
593 spin_unlock(&resv->lock);
597 * Delete the specified range [f, t) from the reserve map. If the
598 * t parameter is LONG_MAX, this indicates that ALL regions after f
599 * should be deleted. Locate the regions which intersect [f, t)
600 * and either trim, delete or split the existing regions.
602 * Returns the number of huge pages deleted from the reserve map.
603 * In the normal case, the return value is zero or more. In the
604 * case where a region must be split, a new region descriptor must
605 * be allocated. If the allocation fails, -ENOMEM will be returned.
606 * NOTE: If the parameter t == LONG_MAX, then we will never split
607 * a region and possibly return -ENOMEM. Callers specifying
608 * t == LONG_MAX do not need to check for -ENOMEM error.
610 static long region_del(struct resv_map *resv, long f, long t)
612 struct list_head *head = &resv->regions;
613 struct file_region *rg, *trg;
614 struct file_region *nrg = NULL;
618 spin_lock(&resv->lock);
619 list_for_each_entry_safe(rg, trg, head, link) {
621 * Skip regions before the range to be deleted. file_region
622 * ranges are normally of the form [from, to). However, there
623 * may be a "placeholder" entry in the map which is of the form
624 * (from, to) with from == to. Check for placeholder entries
625 * at the beginning of the range to be deleted.
627 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
633 if (f > rg->from && t < rg->to) { /* Must split region */
635 * Check for an entry in the cache before dropping
636 * lock and attempting allocation.
639 resv->region_cache_count > resv->adds_in_progress) {
640 nrg = list_first_entry(&resv->region_cache,
643 list_del(&nrg->link);
644 resv->region_cache_count--;
648 spin_unlock(&resv->lock);
649 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
657 /* New entry for end of split region */
661 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
663 INIT_LIST_HEAD(&nrg->link);
665 /* Original entry is trimmed */
668 hugetlb_cgroup_uncharge_file_region(
669 resv, rg, nrg->to - nrg->from);
671 list_add(&nrg->link, &rg->link);
676 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
677 del += rg->to - rg->from;
678 hugetlb_cgroup_uncharge_file_region(resv, rg,
685 if (f <= rg->from) { /* Trim beginning of region */
689 hugetlb_cgroup_uncharge_file_region(resv, rg,
691 } else { /* Trim end of region */
695 hugetlb_cgroup_uncharge_file_region(resv, rg,
700 spin_unlock(&resv->lock);
706 * A rare out of memory error was encountered which prevented removal of
707 * the reserve map region for a page. The huge page itself was free'ed
708 * and removed from the page cache. This routine will adjust the subpool
709 * usage count, and the global reserve count if needed. By incrementing
710 * these counts, the reserve map entry which could not be deleted will
711 * appear as a "reserved" entry instead of simply dangling with incorrect
714 void hugetlb_fix_reserve_counts(struct inode *inode)
716 struct hugepage_subpool *spool = subpool_inode(inode);
719 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
721 struct hstate *h = hstate_inode(inode);
723 hugetlb_acct_memory(h, 1);
728 * Count and return the number of huge pages in the reserve map
729 * that intersect with the range [f, t).
731 static long region_count(struct resv_map *resv, long f, long t)
733 struct list_head *head = &resv->regions;
734 struct file_region *rg;
737 spin_lock(&resv->lock);
738 /* Locate each segment we overlap with, and count that overlap. */
739 list_for_each_entry(rg, head, link) {
748 seg_from = max(rg->from, f);
749 seg_to = min(rg->to, t);
751 chg += seg_to - seg_from;
753 spin_unlock(&resv->lock);
759 * Convert the address within this vma to the page offset within
760 * the mapping, in pagecache page units; huge pages here.
762 static pgoff_t vma_hugecache_offset(struct hstate *h,
763 struct vm_area_struct *vma, unsigned long address)
765 return ((address - vma->vm_start) >> huge_page_shift(h)) +
766 (vma->vm_pgoff >> huge_page_order(h));
769 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
770 unsigned long address)
772 return vma_hugecache_offset(hstate_vma(vma), vma, address);
774 EXPORT_SYMBOL_GPL(linear_hugepage_index);
777 * Return the size of the pages allocated when backing a VMA. In the majority
778 * cases this will be same size as used by the page table entries.
780 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
782 if (vma->vm_ops && vma->vm_ops->pagesize)
783 return vma->vm_ops->pagesize(vma);
786 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
789 * Return the page size being used by the MMU to back a VMA. In the majority
790 * of cases, the page size used by the kernel matches the MMU size. On
791 * architectures where it differs, an architecture-specific 'strong'
792 * version of this symbol is required.
794 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
796 return vma_kernel_pagesize(vma);
800 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
801 * bits of the reservation map pointer, which are always clear due to
804 #define HPAGE_RESV_OWNER (1UL << 0)
805 #define HPAGE_RESV_UNMAPPED (1UL << 1)
806 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
809 * These helpers are used to track how many pages are reserved for
810 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
811 * is guaranteed to have their future faults succeed.
813 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
814 * the reserve counters are updated with the hugetlb_lock held. It is safe
815 * to reset the VMA at fork() time as it is not in use yet and there is no
816 * chance of the global counters getting corrupted as a result of the values.
818 * The private mapping reservation is represented in a subtly different
819 * manner to a shared mapping. A shared mapping has a region map associated
820 * with the underlying file, this region map represents the backing file
821 * pages which have ever had a reservation assigned which this persists even
822 * after the page is instantiated. A private mapping has a region map
823 * associated with the original mmap which is attached to all VMAs which
824 * reference it, this region map represents those offsets which have consumed
825 * reservation ie. where pages have been instantiated.
827 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
829 return (unsigned long)vma->vm_private_data;
832 static void set_vma_private_data(struct vm_area_struct *vma,
835 vma->vm_private_data = (void *)value;
839 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
840 struct hugetlb_cgroup *h_cg,
843 #ifdef CONFIG_CGROUP_HUGETLB
845 resv_map->reservation_counter = NULL;
846 resv_map->pages_per_hpage = 0;
847 resv_map->css = NULL;
849 resv_map->reservation_counter =
850 &h_cg->rsvd_hugepage[hstate_index(h)];
851 resv_map->pages_per_hpage = pages_per_huge_page(h);
852 resv_map->css = &h_cg->css;
857 struct resv_map *resv_map_alloc(void)
859 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
860 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
862 if (!resv_map || !rg) {
868 kref_init(&resv_map->refs);
869 spin_lock_init(&resv_map->lock);
870 INIT_LIST_HEAD(&resv_map->regions);
872 resv_map->adds_in_progress = 0;
874 * Initialize these to 0. On shared mappings, 0's here indicate these
875 * fields don't do cgroup accounting. On private mappings, these will be
876 * re-initialized to the proper values, to indicate that hugetlb cgroup
877 * reservations are to be un-charged from here.
879 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
881 INIT_LIST_HEAD(&resv_map->region_cache);
882 list_add(&rg->link, &resv_map->region_cache);
883 resv_map->region_cache_count = 1;
888 void resv_map_release(struct kref *ref)
890 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
891 struct list_head *head = &resv_map->region_cache;
892 struct file_region *rg, *trg;
894 /* Clear out any active regions before we release the map. */
895 region_del(resv_map, 0, LONG_MAX);
897 /* ... and any entries left in the cache */
898 list_for_each_entry_safe(rg, trg, head, link) {
903 VM_BUG_ON(resv_map->adds_in_progress);
908 static inline struct resv_map *inode_resv_map(struct inode *inode)
911 * At inode evict time, i_mapping may not point to the original
912 * address space within the inode. This original address space
913 * contains the pointer to the resv_map. So, always use the
914 * address space embedded within the inode.
915 * The VERY common case is inode->mapping == &inode->i_data but,
916 * this may not be true for device special inodes.
918 return (struct resv_map *)(&inode->i_data)->private_data;
921 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
923 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
924 if (vma->vm_flags & VM_MAYSHARE) {
925 struct address_space *mapping = vma->vm_file->f_mapping;
926 struct inode *inode = mapping->host;
928 return inode_resv_map(inode);
931 return (struct resv_map *)(get_vma_private_data(vma) &
936 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
938 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
939 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
941 set_vma_private_data(vma, (get_vma_private_data(vma) &
942 HPAGE_RESV_MASK) | (unsigned long)map);
945 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
947 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
948 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
950 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
953 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
955 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
957 return (get_vma_private_data(vma) & flag) != 0;
960 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
961 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
963 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
964 if (!(vma->vm_flags & VM_MAYSHARE))
965 vma->vm_private_data = (void *)0;
968 /* Returns true if the VMA has associated reserve pages */
969 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
971 if (vma->vm_flags & VM_NORESERVE) {
973 * This address is already reserved by other process(chg == 0),
974 * so, we should decrement reserved count. Without decrementing,
975 * reserve count remains after releasing inode, because this
976 * allocated page will go into page cache and is regarded as
977 * coming from reserved pool in releasing step. Currently, we
978 * don't have any other solution to deal with this situation
979 * properly, so add work-around here.
981 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
987 /* Shared mappings always use reserves */
988 if (vma->vm_flags & VM_MAYSHARE) {
990 * We know VM_NORESERVE is not set. Therefore, there SHOULD
991 * be a region map for all pages. The only situation where
992 * there is no region map is if a hole was punched via
993 * fallocate. In this case, there really are no reserves to
994 * use. This situation is indicated if chg != 0.
1003 * Only the process that called mmap() has reserves for
1006 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1008 * Like the shared case above, a hole punch or truncate
1009 * could have been performed on the private mapping.
1010 * Examine the value of chg to determine if reserves
1011 * actually exist or were previously consumed.
1012 * Very Subtle - The value of chg comes from a previous
1013 * call to vma_needs_reserves(). The reserve map for
1014 * private mappings has different (opposite) semantics
1015 * than that of shared mappings. vma_needs_reserves()
1016 * has already taken this difference in semantics into
1017 * account. Therefore, the meaning of chg is the same
1018 * as in the shared case above. Code could easily be
1019 * combined, but keeping it separate draws attention to
1020 * subtle differences.
1031 static void enqueue_huge_page(struct hstate *h, struct page *page)
1033 int nid = page_to_nid(page);
1034 list_move(&page->lru, &h->hugepage_freelists[nid]);
1035 h->free_huge_pages++;
1036 h->free_huge_pages_node[nid]++;
1039 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1043 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
1044 if (!PageHWPoison(page))
1047 * if 'non-isolated free hugepage' not found on the list,
1048 * the allocation fails.
1050 if (&h->hugepage_freelists[nid] == &page->lru)
1052 list_move(&page->lru, &h->hugepage_activelist);
1053 set_page_refcounted(page);
1054 h->free_huge_pages--;
1055 h->free_huge_pages_node[nid]--;
1059 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1062 unsigned int cpuset_mems_cookie;
1063 struct zonelist *zonelist;
1066 int node = NUMA_NO_NODE;
1068 zonelist = node_zonelist(nid, gfp_mask);
1071 cpuset_mems_cookie = read_mems_allowed_begin();
1072 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1075 if (!cpuset_zone_allowed(zone, gfp_mask))
1078 * no need to ask again on the same node. Pool is node rather than
1081 if (zone_to_nid(zone) == node)
1083 node = zone_to_nid(zone);
1085 page = dequeue_huge_page_node_exact(h, node);
1089 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1095 /* Movability of hugepages depends on migration support. */
1096 static inline gfp_t htlb_alloc_mask(struct hstate *h)
1098 if (hugepage_movable_supported(h))
1099 return GFP_HIGHUSER_MOVABLE;
1101 return GFP_HIGHUSER;
1104 static struct page *dequeue_huge_page_vma(struct hstate *h,
1105 struct vm_area_struct *vma,
1106 unsigned long address, int avoid_reserve,
1110 struct mempolicy *mpol;
1112 nodemask_t *nodemask;
1116 * A child process with MAP_PRIVATE mappings created by their parent
1117 * have no page reserves. This check ensures that reservations are
1118 * not "stolen". The child may still get SIGKILLed
1120 if (!vma_has_reserves(vma, chg) &&
1121 h->free_huge_pages - h->resv_huge_pages == 0)
1124 /* If reserves cannot be used, ensure enough pages are in the pool */
1125 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1128 gfp_mask = htlb_alloc_mask(h);
1129 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1130 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1131 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1132 SetPagePrivate(page);
1133 h->resv_huge_pages--;
1136 mpol_cond_put(mpol);
1144 * common helper functions for hstate_next_node_to_{alloc|free}.
1145 * We may have allocated or freed a huge page based on a different
1146 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1147 * be outside of *nodes_allowed. Ensure that we use an allowed
1148 * node for alloc or free.
1150 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1152 nid = next_node_in(nid, *nodes_allowed);
1153 VM_BUG_ON(nid >= MAX_NUMNODES);
1158 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1160 if (!node_isset(nid, *nodes_allowed))
1161 nid = next_node_allowed(nid, nodes_allowed);
1166 * returns the previously saved node ["this node"] from which to
1167 * allocate a persistent huge page for the pool and advance the
1168 * next node from which to allocate, handling wrap at end of node
1171 static int hstate_next_node_to_alloc(struct hstate *h,
1172 nodemask_t *nodes_allowed)
1176 VM_BUG_ON(!nodes_allowed);
1178 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1179 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1185 * helper for free_pool_huge_page() - return the previously saved
1186 * node ["this node"] from which to free a huge page. Advance the
1187 * next node id whether or not we find a free huge page to free so
1188 * that the next attempt to free addresses the next node.
1190 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1194 VM_BUG_ON(!nodes_allowed);
1196 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1197 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1202 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1203 for (nr_nodes = nodes_weight(*mask); \
1205 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1208 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1209 for (nr_nodes = nodes_weight(*mask); \
1211 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1214 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1215 static void destroy_compound_gigantic_page(struct page *page,
1219 int nr_pages = 1 << order;
1220 struct page *p = page + 1;
1222 atomic_set(compound_mapcount_ptr(page), 0);
1223 if (hpage_pincount_available(page))
1224 atomic_set(compound_pincount_ptr(page), 0);
1226 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1227 clear_compound_head(p);
1228 set_page_refcounted(p);
1231 set_compound_order(page, 0);
1232 __ClearPageHead(page);
1235 static void free_gigantic_page(struct page *page, unsigned int order)
1238 * If the page isn't allocated using the cma allocator,
1239 * cma_release() returns false.
1242 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1246 free_contig_range(page_to_pfn(page), 1 << order);
1249 #ifdef CONFIG_CONTIG_ALLOC
1250 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1251 int nid, nodemask_t *nodemask)
1253 unsigned long nr_pages = 1UL << huge_page_order(h);
1260 for_each_node_mask(node, *nodemask) {
1261 if (!hugetlb_cma[node])
1264 page = cma_alloc(hugetlb_cma[node], nr_pages,
1265 huge_page_order(h), true);
1272 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1275 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1276 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1277 #else /* !CONFIG_CONTIG_ALLOC */
1278 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1279 int nid, nodemask_t *nodemask)
1283 #endif /* CONFIG_CONTIG_ALLOC */
1285 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1286 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1287 int nid, nodemask_t *nodemask)
1291 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1292 static inline void destroy_compound_gigantic_page(struct page *page,
1293 unsigned int order) { }
1296 static void update_and_free_page(struct hstate *h, struct page *page)
1300 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1304 h->nr_huge_pages_node[page_to_nid(page)]--;
1305 for (i = 0; i < pages_per_huge_page(h); i++) {
1306 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1307 1 << PG_referenced | 1 << PG_dirty |
1308 1 << PG_active | 1 << PG_private |
1311 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1312 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1313 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1314 set_page_refcounted(page);
1315 if (hstate_is_gigantic(h)) {
1317 * Temporarily drop the hugetlb_lock, because
1318 * we might block in free_gigantic_page().
1320 spin_unlock(&hugetlb_lock);
1321 destroy_compound_gigantic_page(page, huge_page_order(h));
1322 free_gigantic_page(page, huge_page_order(h));
1323 spin_lock(&hugetlb_lock);
1325 __free_pages(page, huge_page_order(h));
1329 struct hstate *size_to_hstate(unsigned long size)
1333 for_each_hstate(h) {
1334 if (huge_page_size(h) == size)
1341 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1342 * to hstate->hugepage_activelist.)
1344 * This function can be called for tail pages, but never returns true for them.
1346 bool page_huge_active(struct page *page)
1348 VM_BUG_ON_PAGE(!PageHuge(page), page);
1349 return PageHead(page) && PagePrivate(&page[1]);
1352 /* never called for tail page */
1353 static void set_page_huge_active(struct page *page)
1355 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1356 SetPagePrivate(&page[1]);
1359 static void clear_page_huge_active(struct page *page)
1361 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1362 ClearPagePrivate(&page[1]);
1366 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1369 static inline bool PageHugeTemporary(struct page *page)
1371 if (!PageHuge(page))
1374 return (unsigned long)page[2].mapping == -1U;
1377 static inline void SetPageHugeTemporary(struct page *page)
1379 page[2].mapping = (void *)-1U;
1382 static inline void ClearPageHugeTemporary(struct page *page)
1384 page[2].mapping = NULL;
1387 static void __free_huge_page(struct page *page)
1390 * Can't pass hstate in here because it is called from the
1391 * compound page destructor.
1393 struct hstate *h = page_hstate(page);
1394 int nid = page_to_nid(page);
1395 struct hugepage_subpool *spool =
1396 (struct hugepage_subpool *)page_private(page);
1397 bool restore_reserve;
1399 VM_BUG_ON_PAGE(page_count(page), page);
1400 VM_BUG_ON_PAGE(page_mapcount(page), page);
1402 set_page_private(page, 0);
1403 page->mapping = NULL;
1404 restore_reserve = PagePrivate(page);
1405 ClearPagePrivate(page);
1408 * If PagePrivate() was set on page, page allocation consumed a
1409 * reservation. If the page was associated with a subpool, there
1410 * would have been a page reserved in the subpool before allocation
1411 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1412 * reservtion, do not call hugepage_subpool_put_pages() as this will
1413 * remove the reserved page from the subpool.
1415 if (!restore_reserve) {
1417 * A return code of zero implies that the subpool will be
1418 * under its minimum size if the reservation is not restored
1419 * after page is free. Therefore, force restore_reserve
1422 if (hugepage_subpool_put_pages(spool, 1) == 0)
1423 restore_reserve = true;
1426 spin_lock(&hugetlb_lock);
1427 clear_page_huge_active(page);
1428 hugetlb_cgroup_uncharge_page(hstate_index(h),
1429 pages_per_huge_page(h), page);
1430 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1431 pages_per_huge_page(h), page);
1432 if (restore_reserve)
1433 h->resv_huge_pages++;
1435 if (PageHugeTemporary(page)) {
1436 list_del(&page->lru);
1437 ClearPageHugeTemporary(page);
1438 update_and_free_page(h, page);
1439 } else if (h->surplus_huge_pages_node[nid]) {
1440 /* remove the page from active list */
1441 list_del(&page->lru);
1442 update_and_free_page(h, page);
1443 h->surplus_huge_pages--;
1444 h->surplus_huge_pages_node[nid]--;
1446 arch_clear_hugepage_flags(page);
1447 enqueue_huge_page(h, page);
1449 spin_unlock(&hugetlb_lock);
1453 * As free_huge_page() can be called from a non-task context, we have
1454 * to defer the actual freeing in a workqueue to prevent potential
1455 * hugetlb_lock deadlock.
1457 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1458 * be freed and frees them one-by-one. As the page->mapping pointer is
1459 * going to be cleared in __free_huge_page() anyway, it is reused as the
1460 * llist_node structure of a lockless linked list of huge pages to be freed.
1462 static LLIST_HEAD(hpage_freelist);
1464 static void free_hpage_workfn(struct work_struct *work)
1466 struct llist_node *node;
1469 node = llist_del_all(&hpage_freelist);
1472 page = container_of((struct address_space **)node,
1473 struct page, mapping);
1475 __free_huge_page(page);
1478 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1480 void free_huge_page(struct page *page)
1483 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1487 * Only call schedule_work() if hpage_freelist is previously
1488 * empty. Otherwise, schedule_work() had been called but the
1489 * workfn hasn't retrieved the list yet.
1491 if (llist_add((struct llist_node *)&page->mapping,
1493 schedule_work(&free_hpage_work);
1497 __free_huge_page(page);
1500 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1502 INIT_LIST_HEAD(&page->lru);
1503 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1504 spin_lock(&hugetlb_lock);
1505 set_hugetlb_cgroup(page, NULL);
1506 set_hugetlb_cgroup_rsvd(page, NULL);
1508 h->nr_huge_pages_node[nid]++;
1509 spin_unlock(&hugetlb_lock);
1512 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1515 int nr_pages = 1 << order;
1516 struct page *p = page + 1;
1518 /* we rely on prep_new_huge_page to set the destructor */
1519 set_compound_order(page, order);
1520 __ClearPageReserved(page);
1521 __SetPageHead(page);
1522 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1524 * For gigantic hugepages allocated through bootmem at
1525 * boot, it's safer to be consistent with the not-gigantic
1526 * hugepages and clear the PG_reserved bit from all tail pages
1527 * too. Otherwise drivers using get_user_pages() to access tail
1528 * pages may get the reference counting wrong if they see
1529 * PG_reserved set on a tail page (despite the head page not
1530 * having PG_reserved set). Enforcing this consistency between
1531 * head and tail pages allows drivers to optimize away a check
1532 * on the head page when they need know if put_page() is needed
1533 * after get_user_pages().
1535 __ClearPageReserved(p);
1536 set_page_count(p, 0);
1537 set_compound_head(p, page);
1539 atomic_set(compound_mapcount_ptr(page), -1);
1541 if (hpage_pincount_available(page))
1542 atomic_set(compound_pincount_ptr(page), 0);
1546 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1547 * transparent huge pages. See the PageTransHuge() documentation for more
1550 int PageHuge(struct page *page)
1552 if (!PageCompound(page))
1555 page = compound_head(page);
1556 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1558 EXPORT_SYMBOL_GPL(PageHuge);
1561 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1562 * normal or transparent huge pages.
1564 int PageHeadHuge(struct page *page_head)
1566 if (!PageHead(page_head))
1569 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1573 * Find address_space associated with hugetlbfs page.
1574 * Upon entry page is locked and page 'was' mapped although mapped state
1575 * could change. If necessary, use anon_vma to find vma and associated
1576 * address space. The returned mapping may be stale, but it can not be
1577 * invalid as page lock (which is held) is required to destroy mapping.
1579 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1581 struct anon_vma *anon_vma;
1582 pgoff_t pgoff_start, pgoff_end;
1583 struct anon_vma_chain *avc;
1584 struct address_space *mapping = page_mapping(hpage);
1586 /* Simple file based mapping */
1591 * Even anonymous hugetlbfs mappings are associated with an
1592 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1593 * code). Find a vma associated with the anonymous vma, and
1594 * use the file pointer to get address_space.
1596 anon_vma = page_lock_anon_vma_read(hpage);
1598 return mapping; /* NULL */
1600 /* Use first found vma */
1601 pgoff_start = page_to_pgoff(hpage);
1602 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
1603 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1604 pgoff_start, pgoff_end) {
1605 struct vm_area_struct *vma = avc->vma;
1607 mapping = vma->vm_file->f_mapping;
1611 anon_vma_unlock_read(anon_vma);
1616 * Find and lock address space (mapping) in write mode.
1618 * Upon entry, the page is locked which allows us to find the mapping
1619 * even in the case of an anon page. However, locking order dictates
1620 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1621 * specific. So, we first try to lock the sema while still holding the
1622 * page lock. If this works, great! If not, then we need to drop the
1623 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1624 * course, need to revalidate state along the way.
1626 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1628 struct address_space *mapping, *mapping2;
1630 mapping = _get_hugetlb_page_mapping(hpage);
1636 * If no contention, take lock and return
1638 if (i_mmap_trylock_write(mapping))
1642 * Must drop page lock and wait on mapping sema.
1643 * Note: Once page lock is dropped, mapping could become invalid.
1644 * As a hack, increase map count until we lock page again.
1646 atomic_inc(&hpage->_mapcount);
1648 i_mmap_lock_write(mapping);
1650 atomic_add_negative(-1, &hpage->_mapcount);
1652 /* verify page is still mapped */
1653 if (!page_mapped(hpage)) {
1654 i_mmap_unlock_write(mapping);
1659 * Get address space again and verify it is the same one
1660 * we locked. If not, drop lock and retry.
1662 mapping2 = _get_hugetlb_page_mapping(hpage);
1663 if (mapping2 != mapping) {
1664 i_mmap_unlock_write(mapping);
1672 pgoff_t __basepage_index(struct page *page)
1674 struct page *page_head = compound_head(page);
1675 pgoff_t index = page_index(page_head);
1676 unsigned long compound_idx;
1678 if (!PageHuge(page_head))
1679 return page_index(page);
1681 if (compound_order(page_head) >= MAX_ORDER)
1682 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1684 compound_idx = page - page_head;
1686 return (index << compound_order(page_head)) + compound_idx;
1689 static struct page *alloc_buddy_huge_page(struct hstate *h,
1690 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1691 nodemask_t *node_alloc_noretry)
1693 int order = huge_page_order(h);
1695 bool alloc_try_hard = true;
1698 * By default we always try hard to allocate the page with
1699 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1700 * a loop (to adjust global huge page counts) and previous allocation
1701 * failed, do not continue to try hard on the same node. Use the
1702 * node_alloc_noretry bitmap to manage this state information.
1704 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1705 alloc_try_hard = false;
1706 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1708 gfp_mask |= __GFP_RETRY_MAYFAIL;
1709 if (nid == NUMA_NO_NODE)
1710 nid = numa_mem_id();
1711 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1713 __count_vm_event(HTLB_BUDDY_PGALLOC);
1715 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1718 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1719 * indicates an overall state change. Clear bit so that we resume
1720 * normal 'try hard' allocations.
1722 if (node_alloc_noretry && page && !alloc_try_hard)
1723 node_clear(nid, *node_alloc_noretry);
1726 * If we tried hard to get a page but failed, set bit so that
1727 * subsequent attempts will not try as hard until there is an
1728 * overall state change.
1730 if (node_alloc_noretry && !page && alloc_try_hard)
1731 node_set(nid, *node_alloc_noretry);
1737 * Common helper to allocate a fresh hugetlb page. All specific allocators
1738 * should use this function to get new hugetlb pages
1740 static struct page *alloc_fresh_huge_page(struct hstate *h,
1741 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1742 nodemask_t *node_alloc_noretry)
1746 if (hstate_is_gigantic(h))
1747 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1749 page = alloc_buddy_huge_page(h, gfp_mask,
1750 nid, nmask, node_alloc_noretry);
1754 if (hstate_is_gigantic(h))
1755 prep_compound_gigantic_page(page, huge_page_order(h));
1756 prep_new_huge_page(h, page, page_to_nid(page));
1762 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1765 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1766 nodemask_t *node_alloc_noretry)
1770 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1772 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1773 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1774 node_alloc_noretry);
1782 put_page(page); /* free it into the hugepage allocator */
1788 * Free huge page from pool from next node to free.
1789 * Attempt to keep persistent huge pages more or less
1790 * balanced over allowed nodes.
1791 * Called with hugetlb_lock locked.
1793 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1799 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1801 * If we're returning unused surplus pages, only examine
1802 * nodes with surplus pages.
1804 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1805 !list_empty(&h->hugepage_freelists[node])) {
1807 list_entry(h->hugepage_freelists[node].next,
1809 list_del(&page->lru);
1810 h->free_huge_pages--;
1811 h->free_huge_pages_node[node]--;
1813 h->surplus_huge_pages--;
1814 h->surplus_huge_pages_node[node]--;
1816 update_and_free_page(h, page);
1826 * Dissolve a given free hugepage into free buddy pages. This function does
1827 * nothing for in-use hugepages and non-hugepages.
1828 * This function returns values like below:
1830 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1831 * (allocated or reserved.)
1832 * 0: successfully dissolved free hugepages or the page is not a
1833 * hugepage (considered as already dissolved)
1835 int dissolve_free_huge_page(struct page *page)
1839 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1840 if (!PageHuge(page))
1843 spin_lock(&hugetlb_lock);
1844 if (!PageHuge(page)) {
1849 if (!page_count(page)) {
1850 struct page *head = compound_head(page);
1851 struct hstate *h = page_hstate(head);
1852 int nid = page_to_nid(head);
1853 if (h->free_huge_pages - h->resv_huge_pages == 0)
1856 * Move PageHWPoison flag from head page to the raw error page,
1857 * which makes any subpages rather than the error page reusable.
1859 if (PageHWPoison(head) && page != head) {
1860 SetPageHWPoison(page);
1861 ClearPageHWPoison(head);
1863 list_del(&head->lru);
1864 h->free_huge_pages--;
1865 h->free_huge_pages_node[nid]--;
1866 h->max_huge_pages--;
1867 update_and_free_page(h, head);
1871 spin_unlock(&hugetlb_lock);
1876 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1877 * make specified memory blocks removable from the system.
1878 * Note that this will dissolve a free gigantic hugepage completely, if any
1879 * part of it lies within the given range.
1880 * Also note that if dissolve_free_huge_page() returns with an error, all
1881 * free hugepages that were dissolved before that error are lost.
1883 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1889 if (!hugepages_supported())
1892 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1893 page = pfn_to_page(pfn);
1894 rc = dissolve_free_huge_page(page);
1903 * Allocates a fresh surplus page from the page allocator.
1905 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1906 int nid, nodemask_t *nmask)
1908 struct page *page = NULL;
1910 if (hstate_is_gigantic(h))
1913 spin_lock(&hugetlb_lock);
1914 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1916 spin_unlock(&hugetlb_lock);
1918 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1922 spin_lock(&hugetlb_lock);
1924 * We could have raced with the pool size change.
1925 * Double check that and simply deallocate the new page
1926 * if we would end up overcommiting the surpluses. Abuse
1927 * temporary page to workaround the nasty free_huge_page
1930 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1931 SetPageHugeTemporary(page);
1932 spin_unlock(&hugetlb_lock);
1936 h->surplus_huge_pages++;
1937 h->surplus_huge_pages_node[page_to_nid(page)]++;
1941 spin_unlock(&hugetlb_lock);
1946 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1947 int nid, nodemask_t *nmask)
1951 if (hstate_is_gigantic(h))
1954 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1959 * We do not account these pages as surplus because they are only
1960 * temporary and will be released properly on the last reference
1962 SetPageHugeTemporary(page);
1968 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1971 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1972 struct vm_area_struct *vma, unsigned long addr)
1975 struct mempolicy *mpol;
1976 gfp_t gfp_mask = htlb_alloc_mask(h);
1978 nodemask_t *nodemask;
1980 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1981 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1982 mpol_cond_put(mpol);
1987 /* page migration callback function */
1988 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1990 gfp_t gfp_mask = htlb_alloc_mask(h);
1991 struct page *page = NULL;
1993 if (nid != NUMA_NO_NODE)
1994 gfp_mask |= __GFP_THISNODE;
1996 spin_lock(&hugetlb_lock);
1997 if (h->free_huge_pages - h->resv_huge_pages > 0)
1998 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1999 spin_unlock(&hugetlb_lock);
2002 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
2007 /* page migration callback function */
2008 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2011 gfp_t gfp_mask = htlb_alloc_mask(h);
2013 spin_lock(&hugetlb_lock);
2014 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2017 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2019 spin_unlock(&hugetlb_lock);
2023 spin_unlock(&hugetlb_lock);
2025 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2028 /* mempolicy aware migration callback */
2029 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2030 unsigned long address)
2032 struct mempolicy *mpol;
2033 nodemask_t *nodemask;
2038 gfp_mask = htlb_alloc_mask(h);
2039 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2040 page = alloc_huge_page_nodemask(h, node, nodemask);
2041 mpol_cond_put(mpol);
2047 * Increase the hugetlb pool such that it can accommodate a reservation
2050 static int gather_surplus_pages(struct hstate *h, int delta)
2051 __must_hold(&hugetlb_lock)
2053 struct list_head surplus_list;
2054 struct page *page, *tmp;
2056 int needed, allocated;
2057 bool alloc_ok = true;
2059 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2061 h->resv_huge_pages += delta;
2066 INIT_LIST_HEAD(&surplus_list);
2070 spin_unlock(&hugetlb_lock);
2071 for (i = 0; i < needed; i++) {
2072 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2073 NUMA_NO_NODE, NULL);
2078 list_add(&page->lru, &surplus_list);
2084 * After retaking hugetlb_lock, we need to recalculate 'needed'
2085 * because either resv_huge_pages or free_huge_pages may have changed.
2087 spin_lock(&hugetlb_lock);
2088 needed = (h->resv_huge_pages + delta) -
2089 (h->free_huge_pages + allocated);
2094 * We were not able to allocate enough pages to
2095 * satisfy the entire reservation so we free what
2096 * we've allocated so far.
2101 * The surplus_list now contains _at_least_ the number of extra pages
2102 * needed to accommodate the reservation. Add the appropriate number
2103 * of pages to the hugetlb pool and free the extras back to the buddy
2104 * allocator. Commit the entire reservation here to prevent another
2105 * process from stealing the pages as they are added to the pool but
2106 * before they are reserved.
2108 needed += allocated;
2109 h->resv_huge_pages += delta;
2112 /* Free the needed pages to the hugetlb pool */
2113 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2117 * This page is now managed by the hugetlb allocator and has
2118 * no users -- drop the buddy allocator's reference.
2120 put_page_testzero(page);
2121 VM_BUG_ON_PAGE(page_count(page), page);
2122 enqueue_huge_page(h, page);
2125 spin_unlock(&hugetlb_lock);
2127 /* Free unnecessary surplus pages to the buddy allocator */
2128 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2130 spin_lock(&hugetlb_lock);
2136 * This routine has two main purposes:
2137 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2138 * in unused_resv_pages. This corresponds to the prior adjustments made
2139 * to the associated reservation map.
2140 * 2) Free any unused surplus pages that may have been allocated to satisfy
2141 * the reservation. As many as unused_resv_pages may be freed.
2143 * Called with hugetlb_lock held. However, the lock could be dropped (and
2144 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2145 * we must make sure nobody else can claim pages we are in the process of
2146 * freeing. Do this by ensuring resv_huge_page always is greater than the
2147 * number of huge pages we plan to free when dropping the lock.
2149 static void return_unused_surplus_pages(struct hstate *h,
2150 unsigned long unused_resv_pages)
2152 unsigned long nr_pages;
2154 /* Cannot return gigantic pages currently */
2155 if (hstate_is_gigantic(h))
2159 * Part (or even all) of the reservation could have been backed
2160 * by pre-allocated pages. Only free surplus pages.
2162 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2165 * We want to release as many surplus pages as possible, spread
2166 * evenly across all nodes with memory. Iterate across these nodes
2167 * until we can no longer free unreserved surplus pages. This occurs
2168 * when the nodes with surplus pages have no free pages.
2169 * free_pool_huge_page() will balance the the freed pages across the
2170 * on-line nodes with memory and will handle the hstate accounting.
2172 * Note that we decrement resv_huge_pages as we free the pages. If
2173 * we drop the lock, resv_huge_pages will still be sufficiently large
2174 * to cover subsequent pages we may free.
2176 while (nr_pages--) {
2177 h->resv_huge_pages--;
2178 unused_resv_pages--;
2179 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2181 cond_resched_lock(&hugetlb_lock);
2185 /* Fully uncommit the reservation */
2186 h->resv_huge_pages -= unused_resv_pages;
2191 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2192 * are used by the huge page allocation routines to manage reservations.
2194 * vma_needs_reservation is called to determine if the huge page at addr
2195 * within the vma has an associated reservation. If a reservation is
2196 * needed, the value 1 is returned. The caller is then responsible for
2197 * managing the global reservation and subpool usage counts. After
2198 * the huge page has been allocated, vma_commit_reservation is called
2199 * to add the page to the reservation map. If the page allocation fails,
2200 * the reservation must be ended instead of committed. vma_end_reservation
2201 * is called in such cases.
2203 * In the normal case, vma_commit_reservation returns the same value
2204 * as the preceding vma_needs_reservation call. The only time this
2205 * is not the case is if a reserve map was changed between calls. It
2206 * is the responsibility of the caller to notice the difference and
2207 * take appropriate action.
2209 * vma_add_reservation is used in error paths where a reservation must
2210 * be restored when a newly allocated huge page must be freed. It is
2211 * to be called after calling vma_needs_reservation to determine if a
2212 * reservation exists.
2214 enum vma_resv_mode {
2220 static long __vma_reservation_common(struct hstate *h,
2221 struct vm_area_struct *vma, unsigned long addr,
2222 enum vma_resv_mode mode)
2224 struct resv_map *resv;
2227 long dummy_out_regions_needed;
2229 resv = vma_resv_map(vma);
2233 idx = vma_hugecache_offset(h, vma, addr);
2235 case VMA_NEEDS_RESV:
2236 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2237 /* We assume that vma_reservation_* routines always operate on
2238 * 1 page, and that adding to resv map a 1 page entry can only
2239 * ever require 1 region.
2241 VM_BUG_ON(dummy_out_regions_needed != 1);
2243 case VMA_COMMIT_RESV:
2244 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2245 /* region_add calls of range 1 should never fail. */
2249 region_abort(resv, idx, idx + 1, 1);
2253 if (vma->vm_flags & VM_MAYSHARE) {
2254 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2255 /* region_add calls of range 1 should never fail. */
2258 region_abort(resv, idx, idx + 1, 1);
2259 ret = region_del(resv, idx, idx + 1);
2266 if (vma->vm_flags & VM_MAYSHARE)
2268 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2270 * In most cases, reserves always exist for private mappings.
2271 * However, a file associated with mapping could have been
2272 * hole punched or truncated after reserves were consumed.
2273 * As subsequent fault on such a range will not use reserves.
2274 * Subtle - The reserve map for private mappings has the
2275 * opposite meaning than that of shared mappings. If NO
2276 * entry is in the reserve map, it means a reservation exists.
2277 * If an entry exists in the reserve map, it means the
2278 * reservation has already been consumed. As a result, the
2279 * return value of this routine is the opposite of the
2280 * value returned from reserve map manipulation routines above.
2288 return ret < 0 ? ret : 0;
2291 static long vma_needs_reservation(struct hstate *h,
2292 struct vm_area_struct *vma, unsigned long addr)
2294 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2297 static long vma_commit_reservation(struct hstate *h,
2298 struct vm_area_struct *vma, unsigned long addr)
2300 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2303 static void vma_end_reservation(struct hstate *h,
2304 struct vm_area_struct *vma, unsigned long addr)
2306 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2309 static long vma_add_reservation(struct hstate *h,
2310 struct vm_area_struct *vma, unsigned long addr)
2312 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2316 * This routine is called to restore a reservation on error paths. In the
2317 * specific error paths, a huge page was allocated (via alloc_huge_page)
2318 * and is about to be freed. If a reservation for the page existed,
2319 * alloc_huge_page would have consumed the reservation and set PagePrivate
2320 * in the newly allocated page. When the page is freed via free_huge_page,
2321 * the global reservation count will be incremented if PagePrivate is set.
2322 * However, free_huge_page can not adjust the reserve map. Adjust the
2323 * reserve map here to be consistent with global reserve count adjustments
2324 * to be made by free_huge_page.
2326 static void restore_reserve_on_error(struct hstate *h,
2327 struct vm_area_struct *vma, unsigned long address,
2330 if (unlikely(PagePrivate(page))) {
2331 long rc = vma_needs_reservation(h, vma, address);
2333 if (unlikely(rc < 0)) {
2335 * Rare out of memory condition in reserve map
2336 * manipulation. Clear PagePrivate so that
2337 * global reserve count will not be incremented
2338 * by free_huge_page. This will make it appear
2339 * as though the reservation for this page was
2340 * consumed. This may prevent the task from
2341 * faulting in the page at a later time. This
2342 * is better than inconsistent global huge page
2343 * accounting of reserve counts.
2345 ClearPagePrivate(page);
2347 rc = vma_add_reservation(h, vma, address);
2348 if (unlikely(rc < 0))
2350 * See above comment about rare out of
2353 ClearPagePrivate(page);
2355 vma_end_reservation(h, vma, address);
2359 struct page *alloc_huge_page(struct vm_area_struct *vma,
2360 unsigned long addr, int avoid_reserve)
2362 struct hugepage_subpool *spool = subpool_vma(vma);
2363 struct hstate *h = hstate_vma(vma);
2365 long map_chg, map_commit;
2368 struct hugetlb_cgroup *h_cg;
2369 bool deferred_reserve;
2371 idx = hstate_index(h);
2373 * Examine the region/reserve map to determine if the process
2374 * has a reservation for the page to be allocated. A return
2375 * code of zero indicates a reservation exists (no change).
2377 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2379 return ERR_PTR(-ENOMEM);
2382 * Processes that did not create the mapping will have no
2383 * reserves as indicated by the region/reserve map. Check
2384 * that the allocation will not exceed the subpool limit.
2385 * Allocations for MAP_NORESERVE mappings also need to be
2386 * checked against any subpool limit.
2388 if (map_chg || avoid_reserve) {
2389 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2391 vma_end_reservation(h, vma, addr);
2392 return ERR_PTR(-ENOSPC);
2396 * Even though there was no reservation in the region/reserve
2397 * map, there could be reservations associated with the
2398 * subpool that can be used. This would be indicated if the
2399 * return value of hugepage_subpool_get_pages() is zero.
2400 * However, if avoid_reserve is specified we still avoid even
2401 * the subpool reservations.
2407 /* If this allocation is not consuming a reservation, charge it now.
2409 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2410 if (deferred_reserve) {
2411 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2412 idx, pages_per_huge_page(h), &h_cg);
2414 goto out_subpool_put;
2417 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2419 goto out_uncharge_cgroup_reservation;
2421 spin_lock(&hugetlb_lock);
2423 * glb_chg is passed to indicate whether or not a page must be taken
2424 * from the global free pool (global change). gbl_chg == 0 indicates
2425 * a reservation exists for the allocation.
2427 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2429 spin_unlock(&hugetlb_lock);
2430 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2432 goto out_uncharge_cgroup;
2433 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2434 SetPagePrivate(page);
2435 h->resv_huge_pages--;
2437 spin_lock(&hugetlb_lock);
2438 list_move(&page->lru, &h->hugepage_activelist);
2441 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2442 /* If allocation is not consuming a reservation, also store the
2443 * hugetlb_cgroup pointer on the page.
2445 if (deferred_reserve) {
2446 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2450 spin_unlock(&hugetlb_lock);
2452 set_page_private(page, (unsigned long)spool);
2454 map_commit = vma_commit_reservation(h, vma, addr);
2455 if (unlikely(map_chg > map_commit)) {
2457 * The page was added to the reservation map between
2458 * vma_needs_reservation and vma_commit_reservation.
2459 * This indicates a race with hugetlb_reserve_pages.
2460 * Adjust for the subpool count incremented above AND
2461 * in hugetlb_reserve_pages for the same page. Also,
2462 * the reservation count added in hugetlb_reserve_pages
2463 * no longer applies.
2467 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2468 hugetlb_acct_memory(h, -rsv_adjust);
2472 out_uncharge_cgroup:
2473 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2474 out_uncharge_cgroup_reservation:
2475 if (deferred_reserve)
2476 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2479 if (map_chg || avoid_reserve)
2480 hugepage_subpool_put_pages(spool, 1);
2481 vma_end_reservation(h, vma, addr);
2482 return ERR_PTR(-ENOSPC);
2485 int alloc_bootmem_huge_page(struct hstate *h)
2486 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2487 int __alloc_bootmem_huge_page(struct hstate *h)
2489 struct huge_bootmem_page *m;
2492 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2495 addr = memblock_alloc_try_nid_raw(
2496 huge_page_size(h), huge_page_size(h),
2497 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2500 * Use the beginning of the huge page to store the
2501 * huge_bootmem_page struct (until gather_bootmem
2502 * puts them into the mem_map).
2511 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2512 /* Put them into a private list first because mem_map is not up yet */
2513 INIT_LIST_HEAD(&m->list);
2514 list_add(&m->list, &huge_boot_pages);
2519 static void __init prep_compound_huge_page(struct page *page,
2522 if (unlikely(order > (MAX_ORDER - 1)))
2523 prep_compound_gigantic_page(page, order);
2525 prep_compound_page(page, order);
2528 /* Put bootmem huge pages into the standard lists after mem_map is up */
2529 static void __init gather_bootmem_prealloc(void)
2531 struct huge_bootmem_page *m;
2533 list_for_each_entry(m, &huge_boot_pages, list) {
2534 struct page *page = virt_to_page(m);
2535 struct hstate *h = m->hstate;
2537 WARN_ON(page_count(page) != 1);
2538 prep_compound_huge_page(page, h->order);
2539 WARN_ON(PageReserved(page));
2540 prep_new_huge_page(h, page, page_to_nid(page));
2541 put_page(page); /* free it into the hugepage allocator */
2544 * If we had gigantic hugepages allocated at boot time, we need
2545 * to restore the 'stolen' pages to totalram_pages in order to
2546 * fix confusing memory reports from free(1) and another
2547 * side-effects, like CommitLimit going negative.
2549 if (hstate_is_gigantic(h))
2550 adjust_managed_page_count(page, 1 << h->order);
2555 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2558 nodemask_t *node_alloc_noretry;
2560 if (!hstate_is_gigantic(h)) {
2562 * Bit mask controlling how hard we retry per-node allocations.
2563 * Ignore errors as lower level routines can deal with
2564 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2565 * time, we are likely in bigger trouble.
2567 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2570 /* allocations done at boot time */
2571 node_alloc_noretry = NULL;
2574 /* bit mask controlling how hard we retry per-node allocations */
2575 if (node_alloc_noretry)
2576 nodes_clear(*node_alloc_noretry);
2578 for (i = 0; i < h->max_huge_pages; ++i) {
2579 if (hstate_is_gigantic(h)) {
2580 if (hugetlb_cma_size) {
2581 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2584 if (!alloc_bootmem_huge_page(h))
2586 } else if (!alloc_pool_huge_page(h,
2587 &node_states[N_MEMORY],
2588 node_alloc_noretry))
2592 if (i < h->max_huge_pages) {
2595 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2596 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2597 h->max_huge_pages, buf, i);
2598 h->max_huge_pages = i;
2601 kfree(node_alloc_noretry);
2604 static void __init hugetlb_init_hstates(void)
2608 for_each_hstate(h) {
2609 if (minimum_order > huge_page_order(h))
2610 minimum_order = huge_page_order(h);
2612 /* oversize hugepages were init'ed in early boot */
2613 if (!hstate_is_gigantic(h))
2614 hugetlb_hstate_alloc_pages(h);
2616 VM_BUG_ON(minimum_order == UINT_MAX);
2619 static void __init report_hugepages(void)
2623 for_each_hstate(h) {
2626 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2627 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2628 buf, h->free_huge_pages);
2632 #ifdef CONFIG_HIGHMEM
2633 static void try_to_free_low(struct hstate *h, unsigned long count,
2634 nodemask_t *nodes_allowed)
2638 if (hstate_is_gigantic(h))
2641 for_each_node_mask(i, *nodes_allowed) {
2642 struct page *page, *next;
2643 struct list_head *freel = &h->hugepage_freelists[i];
2644 list_for_each_entry_safe(page, next, freel, lru) {
2645 if (count >= h->nr_huge_pages)
2647 if (PageHighMem(page))
2649 list_del(&page->lru);
2650 update_and_free_page(h, page);
2651 h->free_huge_pages--;
2652 h->free_huge_pages_node[page_to_nid(page)]--;
2657 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2658 nodemask_t *nodes_allowed)
2664 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2665 * balanced by operating on them in a round-robin fashion.
2666 * Returns 1 if an adjustment was made.
2668 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2673 VM_BUG_ON(delta != -1 && delta != 1);
2676 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2677 if (h->surplus_huge_pages_node[node])
2681 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2682 if (h->surplus_huge_pages_node[node] <
2683 h->nr_huge_pages_node[node])
2690 h->surplus_huge_pages += delta;
2691 h->surplus_huge_pages_node[node] += delta;
2695 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2696 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2697 nodemask_t *nodes_allowed)
2699 unsigned long min_count, ret;
2700 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2703 * Bit mask controlling how hard we retry per-node allocations.
2704 * If we can not allocate the bit mask, do not attempt to allocate
2705 * the requested huge pages.
2707 if (node_alloc_noretry)
2708 nodes_clear(*node_alloc_noretry);
2712 spin_lock(&hugetlb_lock);
2715 * Check for a node specific request.
2716 * Changing node specific huge page count may require a corresponding
2717 * change to the global count. In any case, the passed node mask
2718 * (nodes_allowed) will restrict alloc/free to the specified node.
2720 if (nid != NUMA_NO_NODE) {
2721 unsigned long old_count = count;
2723 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2725 * User may have specified a large count value which caused the
2726 * above calculation to overflow. In this case, they wanted
2727 * to allocate as many huge pages as possible. Set count to
2728 * largest possible value to align with their intention.
2730 if (count < old_count)
2735 * Gigantic pages runtime allocation depend on the capability for large
2736 * page range allocation.
2737 * If the system does not provide this feature, return an error when
2738 * the user tries to allocate gigantic pages but let the user free the
2739 * boottime allocated gigantic pages.
2741 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2742 if (count > persistent_huge_pages(h)) {
2743 spin_unlock(&hugetlb_lock);
2744 NODEMASK_FREE(node_alloc_noretry);
2747 /* Fall through to decrease pool */
2751 * Increase the pool size
2752 * First take pages out of surplus state. Then make up the
2753 * remaining difference by allocating fresh huge pages.
2755 * We might race with alloc_surplus_huge_page() here and be unable
2756 * to convert a surplus huge page to a normal huge page. That is
2757 * not critical, though, it just means the overall size of the
2758 * pool might be one hugepage larger than it needs to be, but
2759 * within all the constraints specified by the sysctls.
2761 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2762 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2766 while (count > persistent_huge_pages(h)) {
2768 * If this allocation races such that we no longer need the
2769 * page, free_huge_page will handle it by freeing the page
2770 * and reducing the surplus.
2772 spin_unlock(&hugetlb_lock);
2774 /* yield cpu to avoid soft lockup */
2777 ret = alloc_pool_huge_page(h, nodes_allowed,
2778 node_alloc_noretry);
2779 spin_lock(&hugetlb_lock);
2783 /* Bail for signals. Probably ctrl-c from user */
2784 if (signal_pending(current))
2789 * Decrease the pool size
2790 * First return free pages to the buddy allocator (being careful
2791 * to keep enough around to satisfy reservations). Then place
2792 * pages into surplus state as needed so the pool will shrink
2793 * to the desired size as pages become free.
2795 * By placing pages into the surplus state independent of the
2796 * overcommit value, we are allowing the surplus pool size to
2797 * exceed overcommit. There are few sane options here. Since
2798 * alloc_surplus_huge_page() is checking the global counter,
2799 * though, we'll note that we're not allowed to exceed surplus
2800 * and won't grow the pool anywhere else. Not until one of the
2801 * sysctls are changed, or the surplus pages go out of use.
2803 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2804 min_count = max(count, min_count);
2805 try_to_free_low(h, min_count, nodes_allowed);
2806 while (min_count < persistent_huge_pages(h)) {
2807 if (!free_pool_huge_page(h, nodes_allowed, 0))
2809 cond_resched_lock(&hugetlb_lock);
2811 while (count < persistent_huge_pages(h)) {
2812 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2816 h->max_huge_pages = persistent_huge_pages(h);
2817 spin_unlock(&hugetlb_lock);
2819 NODEMASK_FREE(node_alloc_noretry);
2824 #define HSTATE_ATTR_RO(_name) \
2825 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2827 #define HSTATE_ATTR(_name) \
2828 static struct kobj_attribute _name##_attr = \
2829 __ATTR(_name, 0644, _name##_show, _name##_store)
2831 static struct kobject *hugepages_kobj;
2832 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2834 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2836 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2840 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2841 if (hstate_kobjs[i] == kobj) {
2843 *nidp = NUMA_NO_NODE;
2847 return kobj_to_node_hstate(kobj, nidp);
2850 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2851 struct kobj_attribute *attr, char *buf)
2854 unsigned long nr_huge_pages;
2857 h = kobj_to_hstate(kobj, &nid);
2858 if (nid == NUMA_NO_NODE)
2859 nr_huge_pages = h->nr_huge_pages;
2861 nr_huge_pages = h->nr_huge_pages_node[nid];
2863 return sprintf(buf, "%lu\n", nr_huge_pages);
2866 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2867 struct hstate *h, int nid,
2868 unsigned long count, size_t len)
2871 nodemask_t nodes_allowed, *n_mask;
2873 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2876 if (nid == NUMA_NO_NODE) {
2878 * global hstate attribute
2880 if (!(obey_mempolicy &&
2881 init_nodemask_of_mempolicy(&nodes_allowed)))
2882 n_mask = &node_states[N_MEMORY];
2884 n_mask = &nodes_allowed;
2887 * Node specific request. count adjustment happens in
2888 * set_max_huge_pages() after acquiring hugetlb_lock.
2890 init_nodemask_of_node(&nodes_allowed, nid);
2891 n_mask = &nodes_allowed;
2894 err = set_max_huge_pages(h, count, nid, n_mask);
2896 return err ? err : len;
2899 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2900 struct kobject *kobj, const char *buf,
2904 unsigned long count;
2908 err = kstrtoul(buf, 10, &count);
2912 h = kobj_to_hstate(kobj, &nid);
2913 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2916 static ssize_t nr_hugepages_show(struct kobject *kobj,
2917 struct kobj_attribute *attr, char *buf)
2919 return nr_hugepages_show_common(kobj, attr, buf);
2922 static ssize_t nr_hugepages_store(struct kobject *kobj,
2923 struct kobj_attribute *attr, const char *buf, size_t len)
2925 return nr_hugepages_store_common(false, kobj, buf, len);
2927 HSTATE_ATTR(nr_hugepages);
2932 * hstate attribute for optionally mempolicy-based constraint on persistent
2933 * huge page alloc/free.
2935 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2936 struct kobj_attribute *attr, char *buf)
2938 return nr_hugepages_show_common(kobj, attr, buf);
2941 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2942 struct kobj_attribute *attr, const char *buf, size_t len)
2944 return nr_hugepages_store_common(true, kobj, buf, len);
2946 HSTATE_ATTR(nr_hugepages_mempolicy);
2950 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2951 struct kobj_attribute *attr, char *buf)
2953 struct hstate *h = kobj_to_hstate(kobj, NULL);
2954 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2957 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2958 struct kobj_attribute *attr, const char *buf, size_t count)
2961 unsigned long input;
2962 struct hstate *h = kobj_to_hstate(kobj, NULL);
2964 if (hstate_is_gigantic(h))
2967 err = kstrtoul(buf, 10, &input);
2971 spin_lock(&hugetlb_lock);
2972 h->nr_overcommit_huge_pages = input;
2973 spin_unlock(&hugetlb_lock);
2977 HSTATE_ATTR(nr_overcommit_hugepages);
2979 static ssize_t free_hugepages_show(struct kobject *kobj,
2980 struct kobj_attribute *attr, char *buf)
2983 unsigned long free_huge_pages;
2986 h = kobj_to_hstate(kobj, &nid);
2987 if (nid == NUMA_NO_NODE)
2988 free_huge_pages = h->free_huge_pages;
2990 free_huge_pages = h->free_huge_pages_node[nid];
2992 return sprintf(buf, "%lu\n", free_huge_pages);
2994 HSTATE_ATTR_RO(free_hugepages);
2996 static ssize_t resv_hugepages_show(struct kobject *kobj,
2997 struct kobj_attribute *attr, char *buf)
2999 struct hstate *h = kobj_to_hstate(kobj, NULL);
3000 return sprintf(buf, "%lu\n", h->resv_huge_pages);
3002 HSTATE_ATTR_RO(resv_hugepages);
3004 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3005 struct kobj_attribute *attr, char *buf)
3008 unsigned long surplus_huge_pages;
3011 h = kobj_to_hstate(kobj, &nid);
3012 if (nid == NUMA_NO_NODE)
3013 surplus_huge_pages = h->surplus_huge_pages;
3015 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3017 return sprintf(buf, "%lu\n", surplus_huge_pages);
3019 HSTATE_ATTR_RO(surplus_hugepages);
3021 static struct attribute *hstate_attrs[] = {
3022 &nr_hugepages_attr.attr,
3023 &nr_overcommit_hugepages_attr.attr,
3024 &free_hugepages_attr.attr,
3025 &resv_hugepages_attr.attr,
3026 &surplus_hugepages_attr.attr,
3028 &nr_hugepages_mempolicy_attr.attr,
3033 static const struct attribute_group hstate_attr_group = {
3034 .attrs = hstate_attrs,
3037 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3038 struct kobject **hstate_kobjs,
3039 const struct attribute_group *hstate_attr_group)
3042 int hi = hstate_index(h);
3044 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3045 if (!hstate_kobjs[hi])
3048 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3050 kobject_put(hstate_kobjs[hi]);
3055 static void __init hugetlb_sysfs_init(void)
3060 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3061 if (!hugepages_kobj)
3064 for_each_hstate(h) {
3065 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3066 hstate_kobjs, &hstate_attr_group);
3068 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3075 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3076 * with node devices in node_devices[] using a parallel array. The array
3077 * index of a node device or _hstate == node id.
3078 * This is here to avoid any static dependency of the node device driver, in
3079 * the base kernel, on the hugetlb module.
3081 struct node_hstate {
3082 struct kobject *hugepages_kobj;
3083 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3085 static struct node_hstate node_hstates[MAX_NUMNODES];
3088 * A subset of global hstate attributes for node devices
3090 static struct attribute *per_node_hstate_attrs[] = {
3091 &nr_hugepages_attr.attr,
3092 &free_hugepages_attr.attr,
3093 &surplus_hugepages_attr.attr,
3097 static const struct attribute_group per_node_hstate_attr_group = {
3098 .attrs = per_node_hstate_attrs,
3102 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3103 * Returns node id via non-NULL nidp.
3105 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3109 for (nid = 0; nid < nr_node_ids; nid++) {
3110 struct node_hstate *nhs = &node_hstates[nid];
3112 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3113 if (nhs->hstate_kobjs[i] == kobj) {
3125 * Unregister hstate attributes from a single node device.
3126 * No-op if no hstate attributes attached.
3128 static void hugetlb_unregister_node(struct node *node)
3131 struct node_hstate *nhs = &node_hstates[node->dev.id];
3133 if (!nhs->hugepages_kobj)
3134 return; /* no hstate attributes */
3136 for_each_hstate(h) {
3137 int idx = hstate_index(h);
3138 if (nhs->hstate_kobjs[idx]) {
3139 kobject_put(nhs->hstate_kobjs[idx]);
3140 nhs->hstate_kobjs[idx] = NULL;
3144 kobject_put(nhs->hugepages_kobj);
3145 nhs->hugepages_kobj = NULL;
3150 * Register hstate attributes for a single node device.
3151 * No-op if attributes already registered.
3153 static void hugetlb_register_node(struct node *node)
3156 struct node_hstate *nhs = &node_hstates[node->dev.id];
3159 if (nhs->hugepages_kobj)
3160 return; /* already allocated */
3162 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3164 if (!nhs->hugepages_kobj)
3167 for_each_hstate(h) {
3168 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3170 &per_node_hstate_attr_group);
3172 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3173 h->name, node->dev.id);
3174 hugetlb_unregister_node(node);
3181 * hugetlb init time: register hstate attributes for all registered node
3182 * devices of nodes that have memory. All on-line nodes should have
3183 * registered their associated device by this time.
3185 static void __init hugetlb_register_all_nodes(void)
3189 for_each_node_state(nid, N_MEMORY) {
3190 struct node *node = node_devices[nid];
3191 if (node->dev.id == nid)
3192 hugetlb_register_node(node);
3196 * Let the node device driver know we're here so it can
3197 * [un]register hstate attributes on node hotplug.
3199 register_hugetlbfs_with_node(hugetlb_register_node,
3200 hugetlb_unregister_node);
3202 #else /* !CONFIG_NUMA */
3204 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3212 static void hugetlb_register_all_nodes(void) { }
3216 static int __init hugetlb_init(void)
3220 if (!hugepages_supported()) {
3221 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3222 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3227 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3228 * architectures depend on setup being done here.
3230 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3231 if (!parsed_default_hugepagesz) {
3233 * If we did not parse a default huge page size, set
3234 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3235 * number of huge pages for this default size was implicitly
3236 * specified, set that here as well.
3237 * Note that the implicit setting will overwrite an explicit
3238 * setting. A warning will be printed in this case.
3240 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3241 if (default_hstate_max_huge_pages) {
3242 if (default_hstate.max_huge_pages) {
3245 string_get_size(huge_page_size(&default_hstate),
3246 1, STRING_UNITS_2, buf, 32);
3247 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3248 default_hstate.max_huge_pages, buf);
3249 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3250 default_hstate_max_huge_pages);
3252 default_hstate.max_huge_pages =
3253 default_hstate_max_huge_pages;
3257 hugetlb_cma_check();
3258 hugetlb_init_hstates();
3259 gather_bootmem_prealloc();
3262 hugetlb_sysfs_init();
3263 hugetlb_register_all_nodes();
3264 hugetlb_cgroup_file_init();
3267 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3269 num_fault_mutexes = 1;
3271 hugetlb_fault_mutex_table =
3272 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3274 BUG_ON(!hugetlb_fault_mutex_table);
3276 for (i = 0; i < num_fault_mutexes; i++)
3277 mutex_init(&hugetlb_fault_mutex_table[i]);
3280 subsys_initcall(hugetlb_init);
3282 /* Overwritten by architectures with more huge page sizes */
3283 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3285 return size == HPAGE_SIZE;
3288 void __init hugetlb_add_hstate(unsigned int order)
3293 if (size_to_hstate(PAGE_SIZE << order)) {
3296 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3298 h = &hstates[hugetlb_max_hstate++];
3300 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3301 h->nr_huge_pages = 0;
3302 h->free_huge_pages = 0;
3303 for (i = 0; i < MAX_NUMNODES; ++i)
3304 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3305 INIT_LIST_HEAD(&h->hugepage_activelist);
3306 h->next_nid_to_alloc = first_memory_node;
3307 h->next_nid_to_free = first_memory_node;
3308 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3309 huge_page_size(h)/1024);
3315 * hugepages command line processing
3316 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3317 * specification. If not, ignore the hugepages value. hugepages can also
3318 * be the first huge page command line option in which case it implicitly
3319 * specifies the number of huge pages for the default size.
3321 static int __init hugepages_setup(char *s)
3324 static unsigned long *last_mhp;
3326 if (!parsed_valid_hugepagesz) {
3327 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3328 parsed_valid_hugepagesz = true;
3333 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3334 * yet, so this hugepages= parameter goes to the "default hstate".
3335 * Otherwise, it goes with the previously parsed hugepagesz or
3336 * default_hugepagesz.
3338 else if (!hugetlb_max_hstate)
3339 mhp = &default_hstate_max_huge_pages;
3341 mhp = &parsed_hstate->max_huge_pages;
3343 if (mhp == last_mhp) {
3344 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3348 if (sscanf(s, "%lu", mhp) <= 0)
3352 * Global state is always initialized later in hugetlb_init.
3353 * But we need to allocate >= MAX_ORDER hstates here early to still
3354 * use the bootmem allocator.
3356 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3357 hugetlb_hstate_alloc_pages(parsed_hstate);
3363 __setup("hugepages=", hugepages_setup);
3366 * hugepagesz command line processing
3367 * A specific huge page size can only be specified once with hugepagesz.
3368 * hugepagesz is followed by hugepages on the command line. The global
3369 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3370 * hugepagesz argument was valid.
3372 static int __init hugepagesz_setup(char *s)
3377 parsed_valid_hugepagesz = false;
3378 size = (unsigned long)memparse(s, NULL);
3380 if (!arch_hugetlb_valid_size(size)) {
3381 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3385 h = size_to_hstate(size);
3388 * hstate for this size already exists. This is normally
3389 * an error, but is allowed if the existing hstate is the
3390 * default hstate. More specifically, it is only allowed if
3391 * the number of huge pages for the default hstate was not
3392 * previously specified.
3394 if (!parsed_default_hugepagesz || h != &default_hstate ||
3395 default_hstate.max_huge_pages) {
3396 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3401 * No need to call hugetlb_add_hstate() as hstate already
3402 * exists. But, do set parsed_hstate so that a following
3403 * hugepages= parameter will be applied to this hstate.
3406 parsed_valid_hugepagesz = true;
3410 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3411 parsed_valid_hugepagesz = true;
3414 __setup("hugepagesz=", hugepagesz_setup);
3417 * default_hugepagesz command line input
3418 * Only one instance of default_hugepagesz allowed on command line.
3420 static int __init default_hugepagesz_setup(char *s)
3424 parsed_valid_hugepagesz = false;
3425 if (parsed_default_hugepagesz) {
3426 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3430 size = (unsigned long)memparse(s, NULL);
3432 if (!arch_hugetlb_valid_size(size)) {
3433 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3437 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3438 parsed_valid_hugepagesz = true;
3439 parsed_default_hugepagesz = true;
3440 default_hstate_idx = hstate_index(size_to_hstate(size));
3443 * The number of default huge pages (for this size) could have been
3444 * specified as the first hugetlb parameter: hugepages=X. If so,
3445 * then default_hstate_max_huge_pages is set. If the default huge
3446 * page size is gigantic (>= MAX_ORDER), then the pages must be
3447 * allocated here from bootmem allocator.
3449 if (default_hstate_max_huge_pages) {
3450 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3451 if (hstate_is_gigantic(&default_hstate))
3452 hugetlb_hstate_alloc_pages(&default_hstate);
3453 default_hstate_max_huge_pages = 0;
3458 __setup("default_hugepagesz=", default_hugepagesz_setup);
3460 static unsigned int cpuset_mems_nr(unsigned int *array)
3463 unsigned int nr = 0;
3465 for_each_node_mask(node, cpuset_current_mems_allowed)
3471 #ifdef CONFIG_SYSCTL
3472 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3473 struct ctl_table *table, int write,
3474 void *buffer, size_t *length, loff_t *ppos)
3476 struct hstate *h = &default_hstate;
3477 unsigned long tmp = h->max_huge_pages;
3480 if (!hugepages_supported())
3484 table->maxlen = sizeof(unsigned long);
3485 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3490 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3491 NUMA_NO_NODE, tmp, *length);
3496 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3497 void *buffer, size_t *length, loff_t *ppos)
3500 return hugetlb_sysctl_handler_common(false, table, write,
3501 buffer, length, ppos);
3505 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3506 void *buffer, size_t *length, loff_t *ppos)
3508 return hugetlb_sysctl_handler_common(true, table, write,
3509 buffer, length, ppos);
3511 #endif /* CONFIG_NUMA */
3513 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3514 void *buffer, size_t *length, loff_t *ppos)
3516 struct hstate *h = &default_hstate;
3520 if (!hugepages_supported())
3523 tmp = h->nr_overcommit_huge_pages;
3525 if (write && hstate_is_gigantic(h))
3529 table->maxlen = sizeof(unsigned long);
3530 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3535 spin_lock(&hugetlb_lock);
3536 h->nr_overcommit_huge_pages = tmp;
3537 spin_unlock(&hugetlb_lock);
3543 #endif /* CONFIG_SYSCTL */
3545 void hugetlb_report_meminfo(struct seq_file *m)
3548 unsigned long total = 0;
3550 if (!hugepages_supported())
3553 for_each_hstate(h) {
3554 unsigned long count = h->nr_huge_pages;
3556 total += (PAGE_SIZE << huge_page_order(h)) * count;
3558 if (h == &default_hstate)
3560 "HugePages_Total: %5lu\n"
3561 "HugePages_Free: %5lu\n"
3562 "HugePages_Rsvd: %5lu\n"
3563 "HugePages_Surp: %5lu\n"
3564 "Hugepagesize: %8lu kB\n",
3568 h->surplus_huge_pages,
3569 (PAGE_SIZE << huge_page_order(h)) / 1024);
3572 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3575 int hugetlb_report_node_meminfo(int nid, char *buf)
3577 struct hstate *h = &default_hstate;
3578 if (!hugepages_supported())
3581 "Node %d HugePages_Total: %5u\n"
3582 "Node %d HugePages_Free: %5u\n"
3583 "Node %d HugePages_Surp: %5u\n",
3584 nid, h->nr_huge_pages_node[nid],
3585 nid, h->free_huge_pages_node[nid],
3586 nid, h->surplus_huge_pages_node[nid]);
3589 void hugetlb_show_meminfo(void)
3594 if (!hugepages_supported())
3597 for_each_node_state(nid, N_MEMORY)
3599 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3601 h->nr_huge_pages_node[nid],
3602 h->free_huge_pages_node[nid],
3603 h->surplus_huge_pages_node[nid],
3604 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3607 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3609 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3610 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3613 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3614 unsigned long hugetlb_total_pages(void)
3617 unsigned long nr_total_pages = 0;
3620 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3621 return nr_total_pages;
3624 static int hugetlb_acct_memory(struct hstate *h, long delta)
3628 spin_lock(&hugetlb_lock);
3630 * When cpuset is configured, it breaks the strict hugetlb page
3631 * reservation as the accounting is done on a global variable. Such
3632 * reservation is completely rubbish in the presence of cpuset because
3633 * the reservation is not checked against page availability for the
3634 * current cpuset. Application can still potentially OOM'ed by kernel
3635 * with lack of free htlb page in cpuset that the task is in.
3636 * Attempt to enforce strict accounting with cpuset is almost
3637 * impossible (or too ugly) because cpuset is too fluid that
3638 * task or memory node can be dynamically moved between cpusets.
3640 * The change of semantics for shared hugetlb mapping with cpuset is
3641 * undesirable. However, in order to preserve some of the semantics,
3642 * we fall back to check against current free page availability as
3643 * a best attempt and hopefully to minimize the impact of changing
3644 * semantics that cpuset has.
3647 if (gather_surplus_pages(h, delta) < 0)
3650 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3651 return_unused_surplus_pages(h, delta);
3658 return_unused_surplus_pages(h, (unsigned long) -delta);
3661 spin_unlock(&hugetlb_lock);
3665 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3667 struct resv_map *resv = vma_resv_map(vma);
3670 * This new VMA should share its siblings reservation map if present.
3671 * The VMA will only ever have a valid reservation map pointer where
3672 * it is being copied for another still existing VMA. As that VMA
3673 * has a reference to the reservation map it cannot disappear until
3674 * after this open call completes. It is therefore safe to take a
3675 * new reference here without additional locking.
3677 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3678 kref_get(&resv->refs);
3681 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3683 struct hstate *h = hstate_vma(vma);
3684 struct resv_map *resv = vma_resv_map(vma);
3685 struct hugepage_subpool *spool = subpool_vma(vma);
3686 unsigned long reserve, start, end;
3689 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3692 start = vma_hugecache_offset(h, vma, vma->vm_start);
3693 end = vma_hugecache_offset(h, vma, vma->vm_end);
3695 reserve = (end - start) - region_count(resv, start, end);
3696 hugetlb_cgroup_uncharge_counter(resv, start, end);
3699 * Decrement reserve counts. The global reserve count may be
3700 * adjusted if the subpool has a minimum size.
3702 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3703 hugetlb_acct_memory(h, -gbl_reserve);
3706 kref_put(&resv->refs, resv_map_release);
3709 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3711 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3716 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3718 struct hstate *hstate = hstate_vma(vma);
3720 return 1UL << huge_page_shift(hstate);
3724 * We cannot handle pagefaults against hugetlb pages at all. They cause
3725 * handle_mm_fault() to try to instantiate regular-sized pages in the
3726 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3729 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3736 * When a new function is introduced to vm_operations_struct and added
3737 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3738 * This is because under System V memory model, mappings created via
3739 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3740 * their original vm_ops are overwritten with shm_vm_ops.
3742 const struct vm_operations_struct hugetlb_vm_ops = {
3743 .fault = hugetlb_vm_op_fault,
3744 .open = hugetlb_vm_op_open,
3745 .close = hugetlb_vm_op_close,
3746 .split = hugetlb_vm_op_split,
3747 .pagesize = hugetlb_vm_op_pagesize,
3750 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3756 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3757 vma->vm_page_prot)));
3759 entry = huge_pte_wrprotect(mk_huge_pte(page,
3760 vma->vm_page_prot));
3762 entry = pte_mkyoung(entry);
3763 entry = pte_mkhuge(entry);
3764 entry = arch_make_huge_pte(entry, vma, page, writable);
3769 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3770 unsigned long address, pte_t *ptep)
3774 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3775 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3776 update_mmu_cache(vma, address, ptep);
3779 bool is_hugetlb_entry_migration(pte_t pte)
3783 if (huge_pte_none(pte) || pte_present(pte))
3785 swp = pte_to_swp_entry(pte);
3786 if (non_swap_entry(swp) && is_migration_entry(swp))
3792 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3796 if (huge_pte_none(pte) || pte_present(pte))
3798 swp = pte_to_swp_entry(pte);
3799 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3805 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3806 struct vm_area_struct *vma)
3808 pte_t *src_pte, *dst_pte, entry, dst_entry;
3809 struct page *ptepage;
3812 struct hstate *h = hstate_vma(vma);
3813 unsigned long sz = huge_page_size(h);
3814 struct address_space *mapping = vma->vm_file->f_mapping;
3815 struct mmu_notifier_range range;
3818 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3821 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3824 mmu_notifier_invalidate_range_start(&range);
3827 * For shared mappings i_mmap_rwsem must be held to call
3828 * huge_pte_alloc, otherwise the returned ptep could go
3829 * away if part of a shared pmd and another thread calls
3832 i_mmap_lock_read(mapping);
3835 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3836 spinlock_t *src_ptl, *dst_ptl;
3837 src_pte = huge_pte_offset(src, addr, sz);
3840 dst_pte = huge_pte_alloc(dst, addr, sz);
3847 * If the pagetables are shared don't copy or take references.
3848 * dst_pte == src_pte is the common case of src/dest sharing.
3850 * However, src could have 'unshared' and dst shares with
3851 * another vma. If dst_pte !none, this implies sharing.
3852 * Check here before taking page table lock, and once again
3853 * after taking the lock below.
3855 dst_entry = huge_ptep_get(dst_pte);
3856 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3859 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3860 src_ptl = huge_pte_lockptr(h, src, src_pte);
3861 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3862 entry = huge_ptep_get(src_pte);
3863 dst_entry = huge_ptep_get(dst_pte);
3864 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3866 * Skip if src entry none. Also, skip in the
3867 * unlikely case dst entry !none as this implies
3868 * sharing with another vma.
3871 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3872 is_hugetlb_entry_hwpoisoned(entry))) {
3873 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3875 if (is_write_migration_entry(swp_entry) && cow) {
3877 * COW mappings require pages in both
3878 * parent and child to be set to read.
3880 make_migration_entry_read(&swp_entry);
3881 entry = swp_entry_to_pte(swp_entry);
3882 set_huge_swap_pte_at(src, addr, src_pte,
3885 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3889 * No need to notify as we are downgrading page
3890 * table protection not changing it to point
3893 * See Documentation/vm/mmu_notifier.rst
3895 huge_ptep_set_wrprotect(src, addr, src_pte);
3897 entry = huge_ptep_get(src_pte);
3898 ptepage = pte_page(entry);
3900 page_dup_rmap(ptepage, true);
3901 set_huge_pte_at(dst, addr, dst_pte, entry);
3902 hugetlb_count_add(pages_per_huge_page(h), dst);
3904 spin_unlock(src_ptl);
3905 spin_unlock(dst_ptl);
3909 mmu_notifier_invalidate_range_end(&range);
3911 i_mmap_unlock_read(mapping);
3916 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3917 unsigned long start, unsigned long end,
3918 struct page *ref_page)
3920 struct mm_struct *mm = vma->vm_mm;
3921 unsigned long address;
3926 struct hstate *h = hstate_vma(vma);
3927 unsigned long sz = huge_page_size(h);
3928 struct mmu_notifier_range range;
3930 WARN_ON(!is_vm_hugetlb_page(vma));
3931 BUG_ON(start & ~huge_page_mask(h));
3932 BUG_ON(end & ~huge_page_mask(h));
3935 * This is a hugetlb vma, all the pte entries should point
3938 tlb_change_page_size(tlb, sz);
3939 tlb_start_vma(tlb, vma);
3942 * If sharing possible, alert mmu notifiers of worst case.
3944 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3946 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3947 mmu_notifier_invalidate_range_start(&range);
3949 for (; address < end; address += sz) {
3950 ptep = huge_pte_offset(mm, address, sz);
3954 ptl = huge_pte_lock(h, mm, ptep);
3955 if (huge_pmd_unshare(mm, &address, ptep)) {
3958 * We just unmapped a page of PMDs by clearing a PUD.
3959 * The caller's TLB flush range should cover this area.
3964 pte = huge_ptep_get(ptep);
3965 if (huge_pte_none(pte)) {
3971 * Migrating hugepage or HWPoisoned hugepage is already
3972 * unmapped and its refcount is dropped, so just clear pte here.
3974 if (unlikely(!pte_present(pte))) {
3975 huge_pte_clear(mm, address, ptep, sz);
3980 page = pte_page(pte);
3982 * If a reference page is supplied, it is because a specific
3983 * page is being unmapped, not a range. Ensure the page we
3984 * are about to unmap is the actual page of interest.
3987 if (page != ref_page) {
3992 * Mark the VMA as having unmapped its page so that
3993 * future faults in this VMA will fail rather than
3994 * looking like data was lost
3996 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3999 pte = huge_ptep_get_and_clear(mm, address, ptep);
4000 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4001 if (huge_pte_dirty(pte))
4002 set_page_dirty(page);
4004 hugetlb_count_sub(pages_per_huge_page(h), mm);
4005 page_remove_rmap(page, true);
4008 tlb_remove_page_size(tlb, page, huge_page_size(h));
4010 * Bail out after unmapping reference page if supplied
4015 mmu_notifier_invalidate_range_end(&range);
4016 tlb_end_vma(tlb, vma);
4019 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4020 struct vm_area_struct *vma, unsigned long start,
4021 unsigned long end, struct page *ref_page)
4023 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4026 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4027 * test will fail on a vma being torn down, and not grab a page table
4028 * on its way out. We're lucky that the flag has such an appropriate
4029 * name, and can in fact be safely cleared here. We could clear it
4030 * before the __unmap_hugepage_range above, but all that's necessary
4031 * is to clear it before releasing the i_mmap_rwsem. This works
4032 * because in the context this is called, the VMA is about to be
4033 * destroyed and the i_mmap_rwsem is held.
4035 vma->vm_flags &= ~VM_MAYSHARE;
4038 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4039 unsigned long end, struct page *ref_page)
4041 struct mm_struct *mm;
4042 struct mmu_gather tlb;
4043 unsigned long tlb_start = start;
4044 unsigned long tlb_end = end;
4047 * If shared PMDs were possibly used within this vma range, adjust
4048 * start/end for worst case tlb flushing.
4049 * Note that we can not be sure if PMDs are shared until we try to
4050 * unmap pages. However, we want to make sure TLB flushing covers
4051 * the largest possible range.
4053 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4057 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4058 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4059 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4063 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4064 * mappping it owns the reserve page for. The intention is to unmap the page
4065 * from other VMAs and let the children be SIGKILLed if they are faulting the
4068 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4069 struct page *page, unsigned long address)
4071 struct hstate *h = hstate_vma(vma);
4072 struct vm_area_struct *iter_vma;
4073 struct address_space *mapping;
4077 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4078 * from page cache lookup which is in HPAGE_SIZE units.
4080 address = address & huge_page_mask(h);
4081 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4083 mapping = vma->vm_file->f_mapping;
4086 * Take the mapping lock for the duration of the table walk. As
4087 * this mapping should be shared between all the VMAs,
4088 * __unmap_hugepage_range() is called as the lock is already held
4090 i_mmap_lock_write(mapping);
4091 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4092 /* Do not unmap the current VMA */
4093 if (iter_vma == vma)
4097 * Shared VMAs have their own reserves and do not affect
4098 * MAP_PRIVATE accounting but it is possible that a shared
4099 * VMA is using the same page so check and skip such VMAs.
4101 if (iter_vma->vm_flags & VM_MAYSHARE)
4105 * Unmap the page from other VMAs without their own reserves.
4106 * They get marked to be SIGKILLed if they fault in these
4107 * areas. This is because a future no-page fault on this VMA
4108 * could insert a zeroed page instead of the data existing
4109 * from the time of fork. This would look like data corruption
4111 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4112 unmap_hugepage_range(iter_vma, address,
4113 address + huge_page_size(h), page);
4115 i_mmap_unlock_write(mapping);
4119 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4120 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4121 * cannot race with other handlers or page migration.
4122 * Keep the pte_same checks anyway to make transition from the mutex easier.
4124 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4125 unsigned long address, pte_t *ptep,
4126 struct page *pagecache_page, spinlock_t *ptl)
4129 struct hstate *h = hstate_vma(vma);
4130 struct page *old_page, *new_page;
4131 int outside_reserve = 0;
4133 unsigned long haddr = address & huge_page_mask(h);
4134 struct mmu_notifier_range range;
4136 pte = huge_ptep_get(ptep);
4137 old_page = pte_page(pte);
4140 /* If no-one else is actually using this page, avoid the copy
4141 * and just make the page writable */
4142 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4143 page_move_anon_rmap(old_page, vma);
4144 set_huge_ptep_writable(vma, haddr, ptep);
4149 * If the process that created a MAP_PRIVATE mapping is about to
4150 * perform a COW due to a shared page count, attempt to satisfy
4151 * the allocation without using the existing reserves. The pagecache
4152 * page is used to determine if the reserve at this address was
4153 * consumed or not. If reserves were used, a partial faulted mapping
4154 * at the time of fork() could consume its reserves on COW instead
4155 * of the full address range.
4157 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4158 old_page != pagecache_page)
4159 outside_reserve = 1;
4164 * Drop page table lock as buddy allocator may be called. It will
4165 * be acquired again before returning to the caller, as expected.
4168 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4170 if (IS_ERR(new_page)) {
4172 * If a process owning a MAP_PRIVATE mapping fails to COW,
4173 * it is due to references held by a child and an insufficient
4174 * huge page pool. To guarantee the original mappers
4175 * reliability, unmap the page from child processes. The child
4176 * may get SIGKILLed if it later faults.
4178 if (outside_reserve) {
4180 BUG_ON(huge_pte_none(pte));
4181 unmap_ref_private(mm, vma, old_page, haddr);
4182 BUG_ON(huge_pte_none(pte));
4184 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4186 pte_same(huge_ptep_get(ptep), pte)))
4187 goto retry_avoidcopy;
4189 * race occurs while re-acquiring page table
4190 * lock, and our job is done.
4195 ret = vmf_error(PTR_ERR(new_page));
4196 goto out_release_old;
4200 * When the original hugepage is shared one, it does not have
4201 * anon_vma prepared.
4203 if (unlikely(anon_vma_prepare(vma))) {
4205 goto out_release_all;
4208 copy_user_huge_page(new_page, old_page, address, vma,
4209 pages_per_huge_page(h));
4210 __SetPageUptodate(new_page);
4212 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4213 haddr + huge_page_size(h));
4214 mmu_notifier_invalidate_range_start(&range);
4217 * Retake the page table lock to check for racing updates
4218 * before the page tables are altered
4221 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4222 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4223 ClearPagePrivate(new_page);
4226 huge_ptep_clear_flush(vma, haddr, ptep);
4227 mmu_notifier_invalidate_range(mm, range.start, range.end);
4228 set_huge_pte_at(mm, haddr, ptep,
4229 make_huge_pte(vma, new_page, 1));
4230 page_remove_rmap(old_page, true);
4231 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4232 set_page_huge_active(new_page);
4233 /* Make the old page be freed below */
4234 new_page = old_page;
4237 mmu_notifier_invalidate_range_end(&range);
4239 restore_reserve_on_error(h, vma, haddr, new_page);
4244 spin_lock(ptl); /* Caller expects lock to be held */
4248 /* Return the pagecache page at a given address within a VMA */
4249 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4250 struct vm_area_struct *vma, unsigned long address)
4252 struct address_space *mapping;
4255 mapping = vma->vm_file->f_mapping;
4256 idx = vma_hugecache_offset(h, vma, address);
4258 return find_lock_page(mapping, idx);
4262 * Return whether there is a pagecache page to back given address within VMA.
4263 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4265 static bool hugetlbfs_pagecache_present(struct hstate *h,
4266 struct vm_area_struct *vma, unsigned long address)
4268 struct address_space *mapping;
4272 mapping = vma->vm_file->f_mapping;
4273 idx = vma_hugecache_offset(h, vma, address);
4275 page = find_get_page(mapping, idx);
4278 return page != NULL;
4281 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4284 struct inode *inode = mapping->host;
4285 struct hstate *h = hstate_inode(inode);
4286 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4290 ClearPagePrivate(page);
4293 * set page dirty so that it will not be removed from cache/file
4294 * by non-hugetlbfs specific code paths.
4296 set_page_dirty(page);
4298 spin_lock(&inode->i_lock);
4299 inode->i_blocks += blocks_per_huge_page(h);
4300 spin_unlock(&inode->i_lock);
4304 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4305 struct vm_area_struct *vma,
4306 struct address_space *mapping, pgoff_t idx,
4307 unsigned long address, pte_t *ptep, unsigned int flags)
4309 struct hstate *h = hstate_vma(vma);
4310 vm_fault_t ret = VM_FAULT_SIGBUS;
4316 unsigned long haddr = address & huge_page_mask(h);
4317 bool new_page = false;
4320 * Currently, we are forced to kill the process in the event the
4321 * original mapper has unmapped pages from the child due to a failed
4322 * COW. Warn that such a situation has occurred as it may not be obvious
4324 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4325 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4331 * We can not race with truncation due to holding i_mmap_rwsem.
4332 * i_size is modified when holding i_mmap_rwsem, so check here
4333 * once for faults beyond end of file.
4335 size = i_size_read(mapping->host) >> huge_page_shift(h);
4340 page = find_lock_page(mapping, idx);
4343 * Check for page in userfault range
4345 if (userfaultfd_missing(vma)) {
4347 struct vm_fault vmf = {
4352 * Hard to debug if it ends up being
4353 * used by a callee that assumes
4354 * something about the other
4355 * uninitialized fields... same as in
4361 * hugetlb_fault_mutex and i_mmap_rwsem must be
4362 * dropped before handling userfault. Reacquire
4363 * after handling fault to make calling code simpler.
4365 hash = hugetlb_fault_mutex_hash(mapping, idx);
4366 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4367 i_mmap_unlock_read(mapping);
4368 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4369 i_mmap_lock_read(mapping);
4370 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4374 page = alloc_huge_page(vma, haddr, 0);
4377 * Returning error will result in faulting task being
4378 * sent SIGBUS. The hugetlb fault mutex prevents two
4379 * tasks from racing to fault in the same page which
4380 * could result in false unable to allocate errors.
4381 * Page migration does not take the fault mutex, but
4382 * does a clear then write of pte's under page table
4383 * lock. Page fault code could race with migration,
4384 * notice the clear pte and try to allocate a page
4385 * here. Before returning error, get ptl and make
4386 * sure there really is no pte entry.
4388 ptl = huge_pte_lock(h, mm, ptep);
4389 if (!huge_pte_none(huge_ptep_get(ptep))) {
4395 ret = vmf_error(PTR_ERR(page));
4398 clear_huge_page(page, address, pages_per_huge_page(h));
4399 __SetPageUptodate(page);
4402 if (vma->vm_flags & VM_MAYSHARE) {
4403 int err = huge_add_to_page_cache(page, mapping, idx);
4412 if (unlikely(anon_vma_prepare(vma))) {
4414 goto backout_unlocked;
4420 * If memory error occurs between mmap() and fault, some process
4421 * don't have hwpoisoned swap entry for errored virtual address.
4422 * So we need to block hugepage fault by PG_hwpoison bit check.
4424 if (unlikely(PageHWPoison(page))) {
4425 ret = VM_FAULT_HWPOISON |
4426 VM_FAULT_SET_HINDEX(hstate_index(h));
4427 goto backout_unlocked;
4432 * If we are going to COW a private mapping later, we examine the
4433 * pending reservations for this page now. This will ensure that
4434 * any allocations necessary to record that reservation occur outside
4437 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4438 if (vma_needs_reservation(h, vma, haddr) < 0) {
4440 goto backout_unlocked;
4442 /* Just decrements count, does not deallocate */
4443 vma_end_reservation(h, vma, haddr);
4446 ptl = huge_pte_lock(h, mm, ptep);
4448 if (!huge_pte_none(huge_ptep_get(ptep)))
4452 ClearPagePrivate(page);
4453 hugepage_add_new_anon_rmap(page, vma, haddr);
4455 page_dup_rmap(page, true);
4456 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4457 && (vma->vm_flags & VM_SHARED)));
4458 set_huge_pte_at(mm, haddr, ptep, new_pte);
4460 hugetlb_count_add(pages_per_huge_page(h), mm);
4461 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4462 /* Optimization, do the COW without a second fault */
4463 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4469 * Only make newly allocated pages active. Existing pages found
4470 * in the pagecache could be !page_huge_active() if they have been
4471 * isolated for migration.
4474 set_page_huge_active(page);
4484 restore_reserve_on_error(h, vma, haddr, page);
4490 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4492 unsigned long key[2];
4495 key[0] = (unsigned long) mapping;
4498 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4500 return hash & (num_fault_mutexes - 1);
4504 * For uniprocesor systems we always use a single mutex, so just
4505 * return 0 and avoid the hashing overhead.
4507 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4513 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4514 unsigned long address, unsigned int flags)
4521 struct page *page = NULL;
4522 struct page *pagecache_page = NULL;
4523 struct hstate *h = hstate_vma(vma);
4524 struct address_space *mapping;
4525 int need_wait_lock = 0;
4526 unsigned long haddr = address & huge_page_mask(h);
4528 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4531 * Since we hold no locks, ptep could be stale. That is
4532 * OK as we are only making decisions based on content and
4533 * not actually modifying content here.
4535 entry = huge_ptep_get(ptep);
4536 if (unlikely(is_hugetlb_entry_migration(entry))) {
4537 migration_entry_wait_huge(vma, mm, ptep);
4539 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4540 return VM_FAULT_HWPOISON_LARGE |
4541 VM_FAULT_SET_HINDEX(hstate_index(h));
4543 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4545 return VM_FAULT_OOM;
4549 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4550 * until finished with ptep. This serves two purposes:
4551 * 1) It prevents huge_pmd_unshare from being called elsewhere
4552 * and making the ptep no longer valid.
4553 * 2) It synchronizes us with i_size modifications during truncation.
4555 * ptep could have already be assigned via huge_pte_offset. That
4556 * is OK, as huge_pte_alloc will return the same value unless
4557 * something has changed.
4559 mapping = vma->vm_file->f_mapping;
4560 i_mmap_lock_read(mapping);
4561 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4563 i_mmap_unlock_read(mapping);
4564 return VM_FAULT_OOM;
4568 * Serialize hugepage allocation and instantiation, so that we don't
4569 * get spurious allocation failures if two CPUs race to instantiate
4570 * the same page in the page cache.
4572 idx = vma_hugecache_offset(h, vma, haddr);
4573 hash = hugetlb_fault_mutex_hash(mapping, idx);
4574 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4576 entry = huge_ptep_get(ptep);
4577 if (huge_pte_none(entry)) {
4578 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4585 * entry could be a migration/hwpoison entry at this point, so this
4586 * check prevents the kernel from going below assuming that we have
4587 * an active hugepage in pagecache. This goto expects the 2nd page
4588 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4589 * properly handle it.
4591 if (!pte_present(entry))
4595 * If we are going to COW the mapping later, we examine the pending
4596 * reservations for this page now. This will ensure that any
4597 * allocations necessary to record that reservation occur outside the
4598 * spinlock. For private mappings, we also lookup the pagecache
4599 * page now as it is used to determine if a reservation has been
4602 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4603 if (vma_needs_reservation(h, vma, haddr) < 0) {
4607 /* Just decrements count, does not deallocate */
4608 vma_end_reservation(h, vma, haddr);
4610 if (!(vma->vm_flags & VM_MAYSHARE))
4611 pagecache_page = hugetlbfs_pagecache_page(h,
4615 ptl = huge_pte_lock(h, mm, ptep);
4617 /* Check for a racing update before calling hugetlb_cow */
4618 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4622 * hugetlb_cow() requires page locks of pte_page(entry) and
4623 * pagecache_page, so here we need take the former one
4624 * when page != pagecache_page or !pagecache_page.
4626 page = pte_page(entry);
4627 if (page != pagecache_page)
4628 if (!trylock_page(page)) {
4635 if (flags & FAULT_FLAG_WRITE) {
4636 if (!huge_pte_write(entry)) {
4637 ret = hugetlb_cow(mm, vma, address, ptep,
4638 pagecache_page, ptl);
4641 entry = huge_pte_mkdirty(entry);
4643 entry = pte_mkyoung(entry);
4644 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4645 flags & FAULT_FLAG_WRITE))
4646 update_mmu_cache(vma, haddr, ptep);
4648 if (page != pagecache_page)
4654 if (pagecache_page) {
4655 unlock_page(pagecache_page);
4656 put_page(pagecache_page);
4659 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4660 i_mmap_unlock_read(mapping);
4662 * Generally it's safe to hold refcount during waiting page lock. But
4663 * here we just wait to defer the next page fault to avoid busy loop and
4664 * the page is not used after unlocked before returning from the current
4665 * page fault. So we are safe from accessing freed page, even if we wait
4666 * here without taking refcount.
4669 wait_on_page_locked(page);
4674 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4675 * modifications for huge pages.
4677 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4679 struct vm_area_struct *dst_vma,
4680 unsigned long dst_addr,
4681 unsigned long src_addr,
4682 struct page **pagep)
4684 struct address_space *mapping;
4687 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4688 struct hstate *h = hstate_vma(dst_vma);
4696 page = alloc_huge_page(dst_vma, dst_addr, 0);
4700 ret = copy_huge_page_from_user(page,
4701 (const void __user *) src_addr,
4702 pages_per_huge_page(h), false);
4704 /* fallback to copy_from_user outside mmap_lock */
4705 if (unlikely(ret)) {
4708 /* don't free the page */
4717 * The memory barrier inside __SetPageUptodate makes sure that
4718 * preceding stores to the page contents become visible before
4719 * the set_pte_at() write.
4721 __SetPageUptodate(page);
4723 mapping = dst_vma->vm_file->f_mapping;
4724 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4727 * If shared, add to page cache
4730 size = i_size_read(mapping->host) >> huge_page_shift(h);
4733 goto out_release_nounlock;
4736 * Serialization between remove_inode_hugepages() and
4737 * huge_add_to_page_cache() below happens through the
4738 * hugetlb_fault_mutex_table that here must be hold by
4741 ret = huge_add_to_page_cache(page, mapping, idx);
4743 goto out_release_nounlock;
4746 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4750 * Recheck the i_size after holding PT lock to make sure not
4751 * to leave any page mapped (as page_mapped()) beyond the end
4752 * of the i_size (remove_inode_hugepages() is strict about
4753 * enforcing that). If we bail out here, we'll also leave a
4754 * page in the radix tree in the vm_shared case beyond the end
4755 * of the i_size, but remove_inode_hugepages() will take care
4756 * of it as soon as we drop the hugetlb_fault_mutex_table.
4758 size = i_size_read(mapping->host) >> huge_page_shift(h);
4761 goto out_release_unlock;
4764 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4765 goto out_release_unlock;
4768 page_dup_rmap(page, true);
4770 ClearPagePrivate(page);
4771 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4774 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4775 if (dst_vma->vm_flags & VM_WRITE)
4776 _dst_pte = huge_pte_mkdirty(_dst_pte);
4777 _dst_pte = pte_mkyoung(_dst_pte);
4779 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4781 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4782 dst_vma->vm_flags & VM_WRITE);
4783 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4785 /* No need to invalidate - it was non-present before */
4786 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4789 set_page_huge_active(page);
4799 out_release_nounlock:
4804 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4805 struct page **pages, struct vm_area_struct **vmas,
4806 unsigned long *position, unsigned long *nr_pages,
4807 long i, unsigned int flags, int *locked)
4809 unsigned long pfn_offset;
4810 unsigned long vaddr = *position;
4811 unsigned long remainder = *nr_pages;
4812 struct hstate *h = hstate_vma(vma);
4815 while (vaddr < vma->vm_end && remainder) {
4817 spinlock_t *ptl = NULL;
4822 * If we have a pending SIGKILL, don't keep faulting pages and
4823 * potentially allocating memory.
4825 if (fatal_signal_pending(current)) {
4831 * Some archs (sparc64, sh*) have multiple pte_ts to
4832 * each hugepage. We have to make sure we get the
4833 * first, for the page indexing below to work.
4835 * Note that page table lock is not held when pte is null.
4837 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4840 ptl = huge_pte_lock(h, mm, pte);
4841 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4844 * When coredumping, it suits get_dump_page if we just return
4845 * an error where there's an empty slot with no huge pagecache
4846 * to back it. This way, we avoid allocating a hugepage, and
4847 * the sparse dumpfile avoids allocating disk blocks, but its
4848 * huge holes still show up with zeroes where they need to be.
4850 if (absent && (flags & FOLL_DUMP) &&
4851 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4859 * We need call hugetlb_fault for both hugepages under migration
4860 * (in which case hugetlb_fault waits for the migration,) and
4861 * hwpoisoned hugepages (in which case we need to prevent the
4862 * caller from accessing to them.) In order to do this, we use
4863 * here is_swap_pte instead of is_hugetlb_entry_migration and
4864 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4865 * both cases, and because we can't follow correct pages
4866 * directly from any kind of swap entries.
4868 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4869 ((flags & FOLL_WRITE) &&
4870 !huge_pte_write(huge_ptep_get(pte)))) {
4872 unsigned int fault_flags = 0;
4876 if (flags & FOLL_WRITE)
4877 fault_flags |= FAULT_FLAG_WRITE;
4879 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4880 FAULT_FLAG_KILLABLE;
4881 if (flags & FOLL_NOWAIT)
4882 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4883 FAULT_FLAG_RETRY_NOWAIT;
4884 if (flags & FOLL_TRIED) {
4886 * Note: FAULT_FLAG_ALLOW_RETRY and
4887 * FAULT_FLAG_TRIED can co-exist
4889 fault_flags |= FAULT_FLAG_TRIED;
4891 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4892 if (ret & VM_FAULT_ERROR) {
4893 err = vm_fault_to_errno(ret, flags);
4897 if (ret & VM_FAULT_RETRY) {
4899 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4903 * VM_FAULT_RETRY must not return an
4904 * error, it will return zero
4907 * No need to update "position" as the
4908 * caller will not check it after
4909 * *nr_pages is set to 0.
4916 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4917 page = pte_page(huge_ptep_get(pte));
4920 * If subpage information not requested, update counters
4921 * and skip the same_page loop below.
4923 if (!pages && !vmas && !pfn_offset &&
4924 (vaddr + huge_page_size(h) < vma->vm_end) &&
4925 (remainder >= pages_per_huge_page(h))) {
4926 vaddr += huge_page_size(h);
4927 remainder -= pages_per_huge_page(h);
4928 i += pages_per_huge_page(h);
4935 pages[i] = mem_map_offset(page, pfn_offset);
4937 * try_grab_page() should always succeed here, because:
4938 * a) we hold the ptl lock, and b) we've just checked
4939 * that the huge page is present in the page tables. If
4940 * the huge page is present, then the tail pages must
4941 * also be present. The ptl prevents the head page and
4942 * tail pages from being rearranged in any way. So this
4943 * page must be available at this point, unless the page
4944 * refcount overflowed:
4946 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4961 if (vaddr < vma->vm_end && remainder &&
4962 pfn_offset < pages_per_huge_page(h)) {
4964 * We use pfn_offset to avoid touching the pageframes
4965 * of this compound page.
4971 *nr_pages = remainder;
4973 * setting position is actually required only if remainder is
4974 * not zero but it's faster not to add a "if (remainder)"
4982 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4984 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4987 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4990 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4991 unsigned long address, unsigned long end, pgprot_t newprot)
4993 struct mm_struct *mm = vma->vm_mm;
4994 unsigned long start = address;
4997 struct hstate *h = hstate_vma(vma);
4998 unsigned long pages = 0;
4999 bool shared_pmd = false;
5000 struct mmu_notifier_range range;
5003 * In the case of shared PMDs, the area to flush could be beyond
5004 * start/end. Set range.start/range.end to cover the maximum possible
5005 * range if PMD sharing is possible.
5007 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5008 0, vma, mm, start, end);
5009 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5011 BUG_ON(address >= end);
5012 flush_cache_range(vma, range.start, range.end);
5014 mmu_notifier_invalidate_range_start(&range);
5015 i_mmap_lock_write(vma->vm_file->f_mapping);
5016 for (; address < end; address += huge_page_size(h)) {
5018 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5021 ptl = huge_pte_lock(h, mm, ptep);
5022 if (huge_pmd_unshare(mm, &address, ptep)) {
5028 pte = huge_ptep_get(ptep);
5029 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5033 if (unlikely(is_hugetlb_entry_migration(pte))) {
5034 swp_entry_t entry = pte_to_swp_entry(pte);
5036 if (is_write_migration_entry(entry)) {
5039 make_migration_entry_read(&entry);
5040 newpte = swp_entry_to_pte(entry);
5041 set_huge_swap_pte_at(mm, address, ptep,
5042 newpte, huge_page_size(h));
5048 if (!huge_pte_none(pte)) {
5051 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5052 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5053 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5054 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5060 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5061 * may have cleared our pud entry and done put_page on the page table:
5062 * once we release i_mmap_rwsem, another task can do the final put_page
5063 * and that page table be reused and filled with junk. If we actually
5064 * did unshare a page of pmds, flush the range corresponding to the pud.
5067 flush_hugetlb_tlb_range(vma, range.start, range.end);
5069 flush_hugetlb_tlb_range(vma, start, end);
5071 * No need to call mmu_notifier_invalidate_range() we are downgrading
5072 * page table protection not changing it to point to a new page.
5074 * See Documentation/vm/mmu_notifier.rst
5076 i_mmap_unlock_write(vma->vm_file->f_mapping);
5077 mmu_notifier_invalidate_range_end(&range);
5079 return pages << h->order;
5082 int hugetlb_reserve_pages(struct inode *inode,
5084 struct vm_area_struct *vma,
5085 vm_flags_t vm_flags)
5087 long ret, chg, add = -1;
5088 struct hstate *h = hstate_inode(inode);
5089 struct hugepage_subpool *spool = subpool_inode(inode);
5090 struct resv_map *resv_map;
5091 struct hugetlb_cgroup *h_cg = NULL;
5092 long gbl_reserve, regions_needed = 0;
5094 /* This should never happen */
5096 VM_WARN(1, "%s called with a negative range\n", __func__);
5101 * Only apply hugepage reservation if asked. At fault time, an
5102 * attempt will be made for VM_NORESERVE to allocate a page
5103 * without using reserves
5105 if (vm_flags & VM_NORESERVE)
5109 * Shared mappings base their reservation on the number of pages that
5110 * are already allocated on behalf of the file. Private mappings need
5111 * to reserve the full area even if read-only as mprotect() may be
5112 * called to make the mapping read-write. Assume !vma is a shm mapping
5114 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5116 * resv_map can not be NULL as hugetlb_reserve_pages is only
5117 * called for inodes for which resv_maps were created (see
5118 * hugetlbfs_get_inode).
5120 resv_map = inode_resv_map(inode);
5122 chg = region_chg(resv_map, from, to, ®ions_needed);
5125 /* Private mapping. */
5126 resv_map = resv_map_alloc();
5132 set_vma_resv_map(vma, resv_map);
5133 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5141 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5142 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5149 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5150 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5153 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5157 * There must be enough pages in the subpool for the mapping. If
5158 * the subpool has a minimum size, there may be some global
5159 * reservations already in place (gbl_reserve).
5161 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5162 if (gbl_reserve < 0) {
5164 goto out_uncharge_cgroup;
5168 * Check enough hugepages are available for the reservation.
5169 * Hand the pages back to the subpool if there are not
5171 ret = hugetlb_acct_memory(h, gbl_reserve);
5177 * Account for the reservations made. Shared mappings record regions
5178 * that have reservations as they are shared by multiple VMAs.
5179 * When the last VMA disappears, the region map says how much
5180 * the reservation was and the page cache tells how much of
5181 * the reservation was consumed. Private mappings are per-VMA and
5182 * only the consumed reservations are tracked. When the VMA
5183 * disappears, the original reservation is the VMA size and the
5184 * consumed reservations are stored in the map. Hence, nothing
5185 * else has to be done for private mappings here
5187 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5188 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5190 if (unlikely(add < 0)) {
5191 hugetlb_acct_memory(h, -gbl_reserve);
5193 } else if (unlikely(chg > add)) {
5195 * pages in this range were added to the reserve
5196 * map between region_chg and region_add. This
5197 * indicates a race with alloc_huge_page. Adjust
5198 * the subpool and reserve counts modified above
5199 * based on the difference.
5203 hugetlb_cgroup_uncharge_cgroup_rsvd(
5205 (chg - add) * pages_per_huge_page(h), h_cg);
5207 rsv_adjust = hugepage_subpool_put_pages(spool,
5209 hugetlb_acct_memory(h, -rsv_adjust);
5214 /* put back original number of pages, chg */
5215 (void)hugepage_subpool_put_pages(spool, chg);
5216 out_uncharge_cgroup:
5217 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5218 chg * pages_per_huge_page(h), h_cg);
5220 if (!vma || vma->vm_flags & VM_MAYSHARE)
5221 /* Only call region_abort if the region_chg succeeded but the
5222 * region_add failed or didn't run.
5224 if (chg >= 0 && add < 0)
5225 region_abort(resv_map, from, to, regions_needed);
5226 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5227 kref_put(&resv_map->refs, resv_map_release);
5231 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5234 struct hstate *h = hstate_inode(inode);
5235 struct resv_map *resv_map = inode_resv_map(inode);
5237 struct hugepage_subpool *spool = subpool_inode(inode);
5241 * Since this routine can be called in the evict inode path for all
5242 * hugetlbfs inodes, resv_map could be NULL.
5245 chg = region_del(resv_map, start, end);
5247 * region_del() can fail in the rare case where a region
5248 * must be split and another region descriptor can not be
5249 * allocated. If end == LONG_MAX, it will not fail.
5255 spin_lock(&inode->i_lock);
5256 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5257 spin_unlock(&inode->i_lock);
5260 * If the subpool has a minimum size, the number of global
5261 * reservations to be released may be adjusted.
5263 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5264 hugetlb_acct_memory(h, -gbl_reserve);
5269 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5270 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5271 struct vm_area_struct *vma,
5272 unsigned long addr, pgoff_t idx)
5274 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5276 unsigned long sbase = saddr & PUD_MASK;
5277 unsigned long s_end = sbase + PUD_SIZE;
5279 /* Allow segments to share if only one is marked locked */
5280 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5281 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5284 * match the virtual addresses, permission and the alignment of the
5287 if (pmd_index(addr) != pmd_index(saddr) ||
5288 vm_flags != svm_flags ||
5289 sbase < svma->vm_start || svma->vm_end < s_end)
5295 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5297 unsigned long base = addr & PUD_MASK;
5298 unsigned long end = base + PUD_SIZE;
5301 * check on proper vm_flags and page table alignment
5303 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5309 * Determine if start,end range within vma could be mapped by shared pmd.
5310 * If yes, adjust start and end to cover range associated with possible
5311 * shared pmd mappings.
5313 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5314 unsigned long *start, unsigned long *end)
5316 unsigned long check_addr;
5318 if (!(vma->vm_flags & VM_MAYSHARE))
5321 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
5322 unsigned long a_start = check_addr & PUD_MASK;
5323 unsigned long a_end = a_start + PUD_SIZE;
5326 * If sharing is possible, adjust start/end if necessary.
5328 if (range_in_vma(vma, a_start, a_end)) {
5329 if (a_start < *start)
5338 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5339 * and returns the corresponding pte. While this is not necessary for the
5340 * !shared pmd case because we can allocate the pmd later as well, it makes the
5341 * code much cleaner.
5343 * This routine must be called with i_mmap_rwsem held in at least read mode.
5344 * For hugetlbfs, this prevents removal of any page table entries associated
5345 * with the address space. This is important as we are setting up sharing
5346 * based on existing page table entries (mappings).
5348 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5350 struct vm_area_struct *vma = find_vma(mm, addr);
5351 struct address_space *mapping = vma->vm_file->f_mapping;
5352 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5354 struct vm_area_struct *svma;
5355 unsigned long saddr;
5360 if (!vma_shareable(vma, addr))
5361 return (pte_t *)pmd_alloc(mm, pud, addr);
5363 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5367 saddr = page_table_shareable(svma, vma, addr, idx);
5369 spte = huge_pte_offset(svma->vm_mm, saddr,
5370 vma_mmu_pagesize(svma));
5372 get_page(virt_to_page(spte));
5381 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5382 if (pud_none(*pud)) {
5383 pud_populate(mm, pud,
5384 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5387 put_page(virt_to_page(spte));
5391 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5396 * unmap huge page backed by shared pte.
5398 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5399 * indicated by page_count > 1, unmap is achieved by clearing pud and
5400 * decrementing the ref count. If count == 1, the pte page is not shared.
5402 * Called with page table lock held and i_mmap_rwsem held in write mode.
5404 * returns: 1 successfully unmapped a shared pte page
5405 * 0 the underlying pte page is not shared, or it is the last user
5407 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5409 pgd_t *pgd = pgd_offset(mm, *addr);
5410 p4d_t *p4d = p4d_offset(pgd, *addr);
5411 pud_t *pud = pud_offset(p4d, *addr);
5413 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5414 if (page_count(virt_to_page(ptep)) == 1)
5418 put_page(virt_to_page(ptep));
5420 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5423 #define want_pmd_share() (1)
5424 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5425 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5430 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5435 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5436 unsigned long *start, unsigned long *end)
5439 #define want_pmd_share() (0)
5440 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5442 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5443 pte_t *huge_pte_alloc(struct mm_struct *mm,
5444 unsigned long addr, unsigned long sz)
5451 pgd = pgd_offset(mm, addr);
5452 p4d = p4d_alloc(mm, pgd, addr);
5455 pud = pud_alloc(mm, p4d, addr);
5457 if (sz == PUD_SIZE) {
5460 BUG_ON(sz != PMD_SIZE);
5461 if (want_pmd_share() && pud_none(*pud))
5462 pte = huge_pmd_share(mm, addr, pud);
5464 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5467 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5473 * huge_pte_offset() - Walk the page table to resolve the hugepage
5474 * entry at address @addr
5476 * Return: Pointer to page table entry (PUD or PMD) for
5477 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5478 * size @sz doesn't match the hugepage size at this level of the page
5481 pte_t *huge_pte_offset(struct mm_struct *mm,
5482 unsigned long addr, unsigned long sz)
5489 pgd = pgd_offset(mm, addr);
5490 if (!pgd_present(*pgd))
5492 p4d = p4d_offset(pgd, addr);
5493 if (!p4d_present(*p4d))
5496 pud = pud_offset(p4d, addr);
5498 /* must be pud huge, non-present or none */
5499 return (pte_t *)pud;
5500 if (!pud_present(*pud))
5502 /* must have a valid entry and size to go further */
5504 pmd = pmd_offset(pud, addr);
5505 /* must be pmd huge, non-present or none */
5506 return (pte_t *)pmd;
5509 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5512 * These functions are overwritable if your architecture needs its own
5515 struct page * __weak
5516 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5519 return ERR_PTR(-EINVAL);
5522 struct page * __weak
5523 follow_huge_pd(struct vm_area_struct *vma,
5524 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5526 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5530 struct page * __weak
5531 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5532 pmd_t *pmd, int flags)
5534 struct page *page = NULL;
5538 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5539 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5540 (FOLL_PIN | FOLL_GET)))
5544 ptl = pmd_lockptr(mm, pmd);
5547 * make sure that the address range covered by this pmd is not
5548 * unmapped from other threads.
5550 if (!pmd_huge(*pmd))
5552 pte = huge_ptep_get((pte_t *)pmd);
5553 if (pte_present(pte)) {
5554 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5556 * try_grab_page() should always succeed here, because: a) we
5557 * hold the pmd (ptl) lock, and b) we've just checked that the
5558 * huge pmd (head) page is present in the page tables. The ptl
5559 * prevents the head page and tail pages from being rearranged
5560 * in any way. So this page must be available at this point,
5561 * unless the page refcount overflowed:
5563 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5568 if (is_hugetlb_entry_migration(pte)) {
5570 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5574 * hwpoisoned entry is treated as no_page_table in
5575 * follow_page_mask().
5583 struct page * __weak
5584 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5585 pud_t *pud, int flags)
5587 if (flags & (FOLL_GET | FOLL_PIN))
5590 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5593 struct page * __weak
5594 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5596 if (flags & (FOLL_GET | FOLL_PIN))
5599 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5602 bool isolate_huge_page(struct page *page, struct list_head *list)
5606 VM_BUG_ON_PAGE(!PageHead(page), page);
5607 spin_lock(&hugetlb_lock);
5608 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5612 clear_page_huge_active(page);
5613 list_move_tail(&page->lru, list);
5615 spin_unlock(&hugetlb_lock);
5619 void putback_active_hugepage(struct page *page)
5621 VM_BUG_ON_PAGE(!PageHead(page), page);
5622 spin_lock(&hugetlb_lock);
5623 set_page_huge_active(page);
5624 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5625 spin_unlock(&hugetlb_lock);
5629 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5631 struct hstate *h = page_hstate(oldpage);
5633 hugetlb_cgroup_migrate(oldpage, newpage);
5634 set_page_owner_migrate_reason(newpage, reason);
5637 * transfer temporary state of the new huge page. This is
5638 * reverse to other transitions because the newpage is going to
5639 * be final while the old one will be freed so it takes over
5640 * the temporary status.
5642 * Also note that we have to transfer the per-node surplus state
5643 * here as well otherwise the global surplus count will not match
5646 if (PageHugeTemporary(newpage)) {
5647 int old_nid = page_to_nid(oldpage);
5648 int new_nid = page_to_nid(newpage);
5650 SetPageHugeTemporary(oldpage);
5651 ClearPageHugeTemporary(newpage);
5653 spin_lock(&hugetlb_lock);
5654 if (h->surplus_huge_pages_node[old_nid]) {
5655 h->surplus_huge_pages_node[old_nid]--;
5656 h->surplus_huge_pages_node[new_nid]++;
5658 spin_unlock(&hugetlb_lock);
5663 static bool cma_reserve_called __initdata;
5665 static int __init cmdline_parse_hugetlb_cma(char *p)
5667 hugetlb_cma_size = memparse(p, &p);
5671 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5673 void __init hugetlb_cma_reserve(int order)
5675 unsigned long size, reserved, per_node;
5678 cma_reserve_called = true;
5680 if (!hugetlb_cma_size)
5683 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5684 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5685 (PAGE_SIZE << order) / SZ_1M);
5690 * If 3 GB area is requested on a machine with 4 numa nodes,
5691 * let's allocate 1 GB on first three nodes and ignore the last one.
5693 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5694 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5695 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5698 for_each_node_state(nid, N_ONLINE) {
5701 size = min(per_node, hugetlb_cma_size - reserved);
5702 size = round_up(size, PAGE_SIZE << order);
5704 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5705 0, false, "hugetlb",
5706 &hugetlb_cma[nid], nid);
5708 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5714 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5717 if (reserved >= hugetlb_cma_size)
5722 void __init hugetlb_cma_check(void)
5724 if (!hugetlb_cma_size || cma_reserve_called)
5727 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5730 #endif /* CONFIG_CMA */