1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
85 static inline bool subpool_is_free(struct hugepage_subpool *spool)
89 if (spool->max_hpages != -1)
90 return spool->used_hpages == 0;
91 if (spool->min_hpages != -1)
92 return spool->rsv_hpages == spool->min_hpages;
97 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
99 spin_unlock(&spool->lock);
101 /* If no pages are used, and no other handles to the subpool
102 * remain, give up any reservations based on minimum size and
103 * free the subpool */
104 if (subpool_is_free(spool)) {
105 if (spool->min_hpages != -1)
106 hugetlb_acct_memory(spool->hstate,
112 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
115 struct hugepage_subpool *spool;
117 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
121 spin_lock_init(&spool->lock);
123 spool->max_hpages = max_hpages;
125 spool->min_hpages = min_hpages;
127 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
131 spool->rsv_hpages = min_hpages;
136 void hugepage_put_subpool(struct hugepage_subpool *spool)
138 spin_lock(&spool->lock);
139 BUG_ON(!spool->count);
141 unlock_or_release_subpool(spool);
145 * Subpool accounting for allocating and reserving pages.
146 * Return -ENOMEM if there are not enough resources to satisfy the
147 * request. Otherwise, return the number of pages by which the
148 * global pools must be adjusted (upward). The returned value may
149 * only be different than the passed value (delta) in the case where
150 * a subpool minimum size must be maintained.
152 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
160 spin_lock(&spool->lock);
162 if (spool->max_hpages != -1) { /* maximum size accounting */
163 if ((spool->used_hpages + delta) <= spool->max_hpages)
164 spool->used_hpages += delta;
171 /* minimum size accounting */
172 if (spool->min_hpages != -1 && spool->rsv_hpages) {
173 if (delta > spool->rsv_hpages) {
175 * Asking for more reserves than those already taken on
176 * behalf of subpool. Return difference.
178 ret = delta - spool->rsv_hpages;
179 spool->rsv_hpages = 0;
181 ret = 0; /* reserves already accounted for */
182 spool->rsv_hpages -= delta;
187 spin_unlock(&spool->lock);
192 * Subpool accounting for freeing and unreserving pages.
193 * Return the number of global page reservations that must be dropped.
194 * The return value may only be different than the passed value (delta)
195 * in the case where a subpool minimum size must be maintained.
197 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
205 spin_lock(&spool->lock);
207 if (spool->max_hpages != -1) /* maximum size accounting */
208 spool->used_hpages -= delta;
210 /* minimum size accounting */
211 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
212 if (spool->rsv_hpages + delta <= spool->min_hpages)
215 ret = spool->rsv_hpages + delta - spool->min_hpages;
217 spool->rsv_hpages += delta;
218 if (spool->rsv_hpages > spool->min_hpages)
219 spool->rsv_hpages = spool->min_hpages;
223 * If hugetlbfs_put_super couldn't free spool due to an outstanding
224 * quota reference, free it now.
226 unlock_or_release_subpool(spool);
231 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
233 return HUGETLBFS_SB(inode->i_sb)->spool;
236 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
238 return subpool_inode(file_inode(vma->vm_file));
241 /* Helper that removes a struct file_region from the resv_map cache and returns
244 static struct file_region *
245 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
247 struct file_region *nrg = NULL;
249 VM_BUG_ON(resv->region_cache_count <= 0);
251 resv->region_cache_count--;
252 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
253 list_del(&nrg->link);
261 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
262 struct file_region *rg)
264 #ifdef CONFIG_CGROUP_HUGETLB
265 nrg->reservation_counter = rg->reservation_counter;
272 /* Helper that records hugetlb_cgroup uncharge info. */
273 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
275 struct resv_map *resv,
276 struct file_region *nrg)
278 #ifdef CONFIG_CGROUP_HUGETLB
280 nrg->reservation_counter =
281 &h_cg->rsvd_hugepage[hstate_index(h)];
282 nrg->css = &h_cg->css;
284 * The caller will hold exactly one h_cg->css reference for the
285 * whole contiguous reservation region. But this area might be
286 * scattered when there are already some file_regions reside in
287 * it. As a result, many file_regions may share only one css
288 * reference. In order to ensure that one file_region must hold
289 * exactly one h_cg->css reference, we should do css_get for
290 * each file_region and leave the reference held by caller
294 if (!resv->pages_per_hpage)
295 resv->pages_per_hpage = pages_per_huge_page(h);
296 /* pages_per_hpage should be the same for all entries in
299 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
301 nrg->reservation_counter = NULL;
307 static void put_uncharge_info(struct file_region *rg)
309 #ifdef CONFIG_CGROUP_HUGETLB
315 static bool has_same_uncharge_info(struct file_region *rg,
316 struct file_region *org)
318 #ifdef CONFIG_CGROUP_HUGETLB
320 rg->reservation_counter == org->reservation_counter &&
328 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
330 struct file_region *nrg = NULL, *prg = NULL;
332 prg = list_prev_entry(rg, link);
333 if (&prg->link != &resv->regions && prg->to == rg->from &&
334 has_same_uncharge_info(prg, rg)) {
338 put_uncharge_info(rg);
344 nrg = list_next_entry(rg, link);
345 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
346 has_same_uncharge_info(nrg, rg)) {
347 nrg->from = rg->from;
350 put_uncharge_info(rg);
356 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
357 long to, struct hstate *h, struct hugetlb_cgroup *cg,
358 long *regions_needed)
360 struct file_region *nrg;
362 if (!regions_needed) {
363 nrg = get_file_region_entry_from_cache(map, from, to);
364 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
365 list_add(&nrg->link, rg->link.prev);
366 coalesce_file_region(map, nrg);
368 *regions_needed += 1;
374 * Must be called with resv->lock held.
376 * Calling this with regions_needed != NULL will count the number of pages
377 * to be added but will not modify the linked list. And regions_needed will
378 * indicate the number of file_regions needed in the cache to carry out to add
379 * the regions for this range.
381 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
382 struct hugetlb_cgroup *h_cg,
383 struct hstate *h, long *regions_needed)
386 struct list_head *head = &resv->regions;
387 long last_accounted_offset = f;
388 struct file_region *rg = NULL, *trg = NULL;
393 /* In this loop, we essentially handle an entry for the range
394 * [last_accounted_offset, rg->from), at every iteration, with some
397 list_for_each_entry_safe(rg, trg, head, link) {
398 /* Skip irrelevant regions that start before our range. */
400 /* If this region ends after the last accounted offset,
401 * then we need to update last_accounted_offset.
403 if (rg->to > last_accounted_offset)
404 last_accounted_offset = rg->to;
408 /* When we find a region that starts beyond our range, we've
414 /* Add an entry for last_accounted_offset -> rg->from, and
415 * update last_accounted_offset.
417 if (rg->from > last_accounted_offset)
418 add += hugetlb_resv_map_add(resv, rg,
419 last_accounted_offset,
423 last_accounted_offset = rg->to;
426 /* Handle the case where our range extends beyond
427 * last_accounted_offset.
429 if (last_accounted_offset < t)
430 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
431 t, h, h_cg, regions_needed);
437 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
439 static int allocate_file_region_entries(struct resv_map *resv,
441 __must_hold(&resv->lock)
443 struct list_head allocated_regions;
444 int to_allocate = 0, i = 0;
445 struct file_region *trg = NULL, *rg = NULL;
447 VM_BUG_ON(regions_needed < 0);
449 INIT_LIST_HEAD(&allocated_regions);
452 * Check for sufficient descriptors in the cache to accommodate
453 * the number of in progress add operations plus regions_needed.
455 * This is a while loop because when we drop the lock, some other call
456 * to region_add or region_del may have consumed some region_entries,
457 * so we keep looping here until we finally have enough entries for
458 * (adds_in_progress + regions_needed).
460 while (resv->region_cache_count <
461 (resv->adds_in_progress + regions_needed)) {
462 to_allocate = resv->adds_in_progress + regions_needed -
463 resv->region_cache_count;
465 /* At this point, we should have enough entries in the cache
466 * for all the existings adds_in_progress. We should only be
467 * needing to allocate for regions_needed.
469 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
471 spin_unlock(&resv->lock);
472 for (i = 0; i < to_allocate; i++) {
473 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
476 list_add(&trg->link, &allocated_regions);
479 spin_lock(&resv->lock);
481 list_splice(&allocated_regions, &resv->region_cache);
482 resv->region_cache_count += to_allocate;
488 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
496 * Add the huge page range represented by [f, t) to the reserve
497 * map. Regions will be taken from the cache to fill in this range.
498 * Sufficient regions should exist in the cache due to the previous
499 * call to region_chg with the same range, but in some cases the cache will not
500 * have sufficient entries due to races with other code doing region_add or
501 * region_del. The extra needed entries will be allocated.
503 * regions_needed is the out value provided by a previous call to region_chg.
505 * Return the number of new huge pages added to the map. This number is greater
506 * than or equal to zero. If file_region entries needed to be allocated for
507 * this operation and we were not able to allocate, it returns -ENOMEM.
508 * region_add of regions of length 1 never allocate file_regions and cannot
509 * fail; region_chg will always allocate at least 1 entry and a region_add for
510 * 1 page will only require at most 1 entry.
512 static long region_add(struct resv_map *resv, long f, long t,
513 long in_regions_needed, struct hstate *h,
514 struct hugetlb_cgroup *h_cg)
516 long add = 0, actual_regions_needed = 0;
518 spin_lock(&resv->lock);
521 /* Count how many regions are actually needed to execute this add. */
522 add_reservation_in_range(resv, f, t, NULL, NULL,
523 &actual_regions_needed);
526 * Check for sufficient descriptors in the cache to accommodate
527 * this add operation. Note that actual_regions_needed may be greater
528 * than in_regions_needed, as the resv_map may have been modified since
529 * the region_chg call. In this case, we need to make sure that we
530 * allocate extra entries, such that we have enough for all the
531 * existing adds_in_progress, plus the excess needed for this
534 if (actual_regions_needed > in_regions_needed &&
535 resv->region_cache_count <
536 resv->adds_in_progress +
537 (actual_regions_needed - in_regions_needed)) {
538 /* region_add operation of range 1 should never need to
539 * allocate file_region entries.
541 VM_BUG_ON(t - f <= 1);
543 if (allocate_file_region_entries(
544 resv, actual_regions_needed - in_regions_needed)) {
551 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
553 resv->adds_in_progress -= in_regions_needed;
555 spin_unlock(&resv->lock);
560 * Examine the existing reserve map and determine how many
561 * huge pages in the specified range [f, t) are NOT currently
562 * represented. This routine is called before a subsequent
563 * call to region_add that will actually modify the reserve
564 * map to add the specified range [f, t). region_chg does
565 * not change the number of huge pages represented by the
566 * map. A number of new file_region structures is added to the cache as a
567 * placeholder, for the subsequent region_add call to use. At least 1
568 * file_region structure is added.
570 * out_regions_needed is the number of regions added to the
571 * resv->adds_in_progress. This value needs to be provided to a follow up call
572 * to region_add or region_abort for proper accounting.
574 * Returns the number of huge pages that need to be added to the existing
575 * reservation map for the range [f, t). This number is greater or equal to
576 * zero. -ENOMEM is returned if a new file_region structure or cache entry
577 * is needed and can not be allocated.
579 static long region_chg(struct resv_map *resv, long f, long t,
580 long *out_regions_needed)
584 spin_lock(&resv->lock);
586 /* Count how many hugepages in this range are NOT represented. */
587 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
590 if (*out_regions_needed == 0)
591 *out_regions_needed = 1;
593 if (allocate_file_region_entries(resv, *out_regions_needed))
596 resv->adds_in_progress += *out_regions_needed;
598 spin_unlock(&resv->lock);
603 * Abort the in progress add operation. The adds_in_progress field
604 * of the resv_map keeps track of the operations in progress between
605 * calls to region_chg and region_add. Operations are sometimes
606 * aborted after the call to region_chg. In such cases, region_abort
607 * is called to decrement the adds_in_progress counter. regions_needed
608 * is the value returned by the region_chg call, it is used to decrement
609 * the adds_in_progress counter.
611 * NOTE: The range arguments [f, t) are not needed or used in this
612 * routine. They are kept to make reading the calling code easier as
613 * arguments will match the associated region_chg call.
615 static void region_abort(struct resv_map *resv, long f, long t,
618 spin_lock(&resv->lock);
619 VM_BUG_ON(!resv->region_cache_count);
620 resv->adds_in_progress -= regions_needed;
621 spin_unlock(&resv->lock);
625 * Delete the specified range [f, t) from the reserve map. If the
626 * t parameter is LONG_MAX, this indicates that ALL regions after f
627 * should be deleted. Locate the regions which intersect [f, t)
628 * and either trim, delete or split the existing regions.
630 * Returns the number of huge pages deleted from the reserve map.
631 * In the normal case, the return value is zero or more. In the
632 * case where a region must be split, a new region descriptor must
633 * be allocated. If the allocation fails, -ENOMEM will be returned.
634 * NOTE: If the parameter t == LONG_MAX, then we will never split
635 * a region and possibly return -ENOMEM. Callers specifying
636 * t == LONG_MAX do not need to check for -ENOMEM error.
638 static long region_del(struct resv_map *resv, long f, long t)
640 struct list_head *head = &resv->regions;
641 struct file_region *rg, *trg;
642 struct file_region *nrg = NULL;
646 spin_lock(&resv->lock);
647 list_for_each_entry_safe(rg, trg, head, link) {
649 * Skip regions before the range to be deleted. file_region
650 * ranges are normally of the form [from, to). However, there
651 * may be a "placeholder" entry in the map which is of the form
652 * (from, to) with from == to. Check for placeholder entries
653 * at the beginning of the range to be deleted.
655 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
661 if (f > rg->from && t < rg->to) { /* Must split region */
663 * Check for an entry in the cache before dropping
664 * lock and attempting allocation.
667 resv->region_cache_count > resv->adds_in_progress) {
668 nrg = list_first_entry(&resv->region_cache,
671 list_del(&nrg->link);
672 resv->region_cache_count--;
676 spin_unlock(&resv->lock);
677 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
684 hugetlb_cgroup_uncharge_file_region(
685 resv, rg, t - f, false);
687 /* New entry for end of split region */
691 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
693 INIT_LIST_HEAD(&nrg->link);
695 /* Original entry is trimmed */
698 list_add(&nrg->link, &rg->link);
703 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
704 del += rg->to - rg->from;
705 hugetlb_cgroup_uncharge_file_region(resv, rg,
706 rg->to - rg->from, true);
712 if (f <= rg->from) { /* Trim beginning of region */
713 hugetlb_cgroup_uncharge_file_region(resv, rg,
714 t - rg->from, false);
718 } else { /* Trim end of region */
719 hugetlb_cgroup_uncharge_file_region(resv, rg,
727 spin_unlock(&resv->lock);
733 * A rare out of memory error was encountered which prevented removal of
734 * the reserve map region for a page. The huge page itself was free'ed
735 * and removed from the page cache. This routine will adjust the subpool
736 * usage count, and the global reserve count if needed. By incrementing
737 * these counts, the reserve map entry which could not be deleted will
738 * appear as a "reserved" entry instead of simply dangling with incorrect
741 void hugetlb_fix_reserve_counts(struct inode *inode)
743 struct hugepage_subpool *spool = subpool_inode(inode);
746 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
748 struct hstate *h = hstate_inode(inode);
750 hugetlb_acct_memory(h, 1);
755 * Count and return the number of huge pages in the reserve map
756 * that intersect with the range [f, t).
758 static long region_count(struct resv_map *resv, long f, long t)
760 struct list_head *head = &resv->regions;
761 struct file_region *rg;
764 spin_lock(&resv->lock);
765 /* Locate each segment we overlap with, and count that overlap. */
766 list_for_each_entry(rg, head, link) {
775 seg_from = max(rg->from, f);
776 seg_to = min(rg->to, t);
778 chg += seg_to - seg_from;
780 spin_unlock(&resv->lock);
786 * Convert the address within this vma to the page offset within
787 * the mapping, in pagecache page units; huge pages here.
789 static pgoff_t vma_hugecache_offset(struct hstate *h,
790 struct vm_area_struct *vma, unsigned long address)
792 return ((address - vma->vm_start) >> huge_page_shift(h)) +
793 (vma->vm_pgoff >> huge_page_order(h));
796 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
797 unsigned long address)
799 return vma_hugecache_offset(hstate_vma(vma), vma, address);
801 EXPORT_SYMBOL_GPL(linear_hugepage_index);
804 * Return the size of the pages allocated when backing a VMA. In the majority
805 * cases this will be same size as used by the page table entries.
807 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
809 if (vma->vm_ops && vma->vm_ops->pagesize)
810 return vma->vm_ops->pagesize(vma);
813 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
816 * Return the page size being used by the MMU to back a VMA. In the majority
817 * of cases, the page size used by the kernel matches the MMU size. On
818 * architectures where it differs, an architecture-specific 'strong'
819 * version of this symbol is required.
821 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
823 return vma_kernel_pagesize(vma);
827 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
828 * bits of the reservation map pointer, which are always clear due to
831 #define HPAGE_RESV_OWNER (1UL << 0)
832 #define HPAGE_RESV_UNMAPPED (1UL << 1)
833 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
836 * These helpers are used to track how many pages are reserved for
837 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
838 * is guaranteed to have their future faults succeed.
840 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
841 * the reserve counters are updated with the hugetlb_lock held. It is safe
842 * to reset the VMA at fork() time as it is not in use yet and there is no
843 * chance of the global counters getting corrupted as a result of the values.
845 * The private mapping reservation is represented in a subtly different
846 * manner to a shared mapping. A shared mapping has a region map associated
847 * with the underlying file, this region map represents the backing file
848 * pages which have ever had a reservation assigned which this persists even
849 * after the page is instantiated. A private mapping has a region map
850 * associated with the original mmap which is attached to all VMAs which
851 * reference it, this region map represents those offsets which have consumed
852 * reservation ie. where pages have been instantiated.
854 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
856 return (unsigned long)vma->vm_private_data;
859 static void set_vma_private_data(struct vm_area_struct *vma,
862 vma->vm_private_data = (void *)value;
866 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
867 struct hugetlb_cgroup *h_cg,
870 #ifdef CONFIG_CGROUP_HUGETLB
872 resv_map->reservation_counter = NULL;
873 resv_map->pages_per_hpage = 0;
874 resv_map->css = NULL;
876 resv_map->reservation_counter =
877 &h_cg->rsvd_hugepage[hstate_index(h)];
878 resv_map->pages_per_hpage = pages_per_huge_page(h);
879 resv_map->css = &h_cg->css;
884 struct resv_map *resv_map_alloc(void)
886 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
887 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
889 if (!resv_map || !rg) {
895 kref_init(&resv_map->refs);
896 spin_lock_init(&resv_map->lock);
897 INIT_LIST_HEAD(&resv_map->regions);
899 resv_map->adds_in_progress = 0;
901 * Initialize these to 0. On shared mappings, 0's here indicate these
902 * fields don't do cgroup accounting. On private mappings, these will be
903 * re-initialized to the proper values, to indicate that hugetlb cgroup
904 * reservations are to be un-charged from here.
906 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
908 INIT_LIST_HEAD(&resv_map->region_cache);
909 list_add(&rg->link, &resv_map->region_cache);
910 resv_map->region_cache_count = 1;
915 void resv_map_release(struct kref *ref)
917 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
918 struct list_head *head = &resv_map->region_cache;
919 struct file_region *rg, *trg;
921 /* Clear out any active regions before we release the map. */
922 region_del(resv_map, 0, LONG_MAX);
924 /* ... and any entries left in the cache */
925 list_for_each_entry_safe(rg, trg, head, link) {
930 VM_BUG_ON(resv_map->adds_in_progress);
935 static inline struct resv_map *inode_resv_map(struct inode *inode)
938 * At inode evict time, i_mapping may not point to the original
939 * address space within the inode. This original address space
940 * contains the pointer to the resv_map. So, always use the
941 * address space embedded within the inode.
942 * The VERY common case is inode->mapping == &inode->i_data but,
943 * this may not be true for device special inodes.
945 return (struct resv_map *)(&inode->i_data)->private_data;
948 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
950 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
951 if (vma->vm_flags & VM_MAYSHARE) {
952 struct address_space *mapping = vma->vm_file->f_mapping;
953 struct inode *inode = mapping->host;
955 return inode_resv_map(inode);
958 return (struct resv_map *)(get_vma_private_data(vma) &
963 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
965 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
966 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
968 set_vma_private_data(vma, (get_vma_private_data(vma) &
969 HPAGE_RESV_MASK) | (unsigned long)map);
972 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
974 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
975 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
977 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
980 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
982 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
984 return (get_vma_private_data(vma) & flag) != 0;
987 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
988 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
990 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
991 if (!(vma->vm_flags & VM_MAYSHARE))
992 vma->vm_private_data = (void *)0;
995 /* Returns true if the VMA has associated reserve pages */
996 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
998 if (vma->vm_flags & VM_NORESERVE) {
1000 * This address is already reserved by other process(chg == 0),
1001 * so, we should decrement reserved count. Without decrementing,
1002 * reserve count remains after releasing inode, because this
1003 * allocated page will go into page cache and is regarded as
1004 * coming from reserved pool in releasing step. Currently, we
1005 * don't have any other solution to deal with this situation
1006 * properly, so add work-around here.
1008 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1014 /* Shared mappings always use reserves */
1015 if (vma->vm_flags & VM_MAYSHARE) {
1017 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1018 * be a region map for all pages. The only situation where
1019 * there is no region map is if a hole was punched via
1020 * fallocate. In this case, there really are no reserves to
1021 * use. This situation is indicated if chg != 0.
1030 * Only the process that called mmap() has reserves for
1033 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1035 * Like the shared case above, a hole punch or truncate
1036 * could have been performed on the private mapping.
1037 * Examine the value of chg to determine if reserves
1038 * actually exist or were previously consumed.
1039 * Very Subtle - The value of chg comes from a previous
1040 * call to vma_needs_reserves(). The reserve map for
1041 * private mappings has different (opposite) semantics
1042 * than that of shared mappings. vma_needs_reserves()
1043 * has already taken this difference in semantics into
1044 * account. Therefore, the meaning of chg is the same
1045 * as in the shared case above. Code could easily be
1046 * combined, but keeping it separate draws attention to
1047 * subtle differences.
1058 static void enqueue_huge_page(struct hstate *h, struct page *page)
1060 int nid = page_to_nid(page);
1061 list_move(&page->lru, &h->hugepage_freelists[nid]);
1062 h->free_huge_pages++;
1063 h->free_huge_pages_node[nid]++;
1064 SetHPageFreed(page);
1067 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1070 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1072 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1073 if (nocma && is_migrate_cma_page(page))
1076 if (PageHWPoison(page))
1079 list_move(&page->lru, &h->hugepage_activelist);
1080 set_page_refcounted(page);
1081 ClearHPageFreed(page);
1082 h->free_huge_pages--;
1083 h->free_huge_pages_node[nid]--;
1090 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1093 unsigned int cpuset_mems_cookie;
1094 struct zonelist *zonelist;
1097 int node = NUMA_NO_NODE;
1099 zonelist = node_zonelist(nid, gfp_mask);
1102 cpuset_mems_cookie = read_mems_allowed_begin();
1103 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1106 if (!cpuset_zone_allowed(zone, gfp_mask))
1109 * no need to ask again on the same node. Pool is node rather than
1112 if (zone_to_nid(zone) == node)
1114 node = zone_to_nid(zone);
1116 page = dequeue_huge_page_node_exact(h, node);
1120 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1126 static struct page *dequeue_huge_page_vma(struct hstate *h,
1127 struct vm_area_struct *vma,
1128 unsigned long address, int avoid_reserve,
1132 struct mempolicy *mpol;
1134 nodemask_t *nodemask;
1138 * A child process with MAP_PRIVATE mappings created by their parent
1139 * have no page reserves. This check ensures that reservations are
1140 * not "stolen". The child may still get SIGKILLed
1142 if (!vma_has_reserves(vma, chg) &&
1143 h->free_huge_pages - h->resv_huge_pages == 0)
1146 /* If reserves cannot be used, ensure enough pages are in the pool */
1147 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1150 gfp_mask = htlb_alloc_mask(h);
1151 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1152 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1153 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1154 SetHPageRestoreReserve(page);
1155 h->resv_huge_pages--;
1158 mpol_cond_put(mpol);
1166 * common helper functions for hstate_next_node_to_{alloc|free}.
1167 * We may have allocated or freed a huge page based on a different
1168 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1169 * be outside of *nodes_allowed. Ensure that we use an allowed
1170 * node for alloc or free.
1172 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1174 nid = next_node_in(nid, *nodes_allowed);
1175 VM_BUG_ON(nid >= MAX_NUMNODES);
1180 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1182 if (!node_isset(nid, *nodes_allowed))
1183 nid = next_node_allowed(nid, nodes_allowed);
1188 * returns the previously saved node ["this node"] from which to
1189 * allocate a persistent huge page for the pool and advance the
1190 * next node from which to allocate, handling wrap at end of node
1193 static int hstate_next_node_to_alloc(struct hstate *h,
1194 nodemask_t *nodes_allowed)
1198 VM_BUG_ON(!nodes_allowed);
1200 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1201 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1207 * helper for free_pool_huge_page() - return the previously saved
1208 * node ["this node"] from which to free a huge page. Advance the
1209 * next node id whether or not we find a free huge page to free so
1210 * that the next attempt to free addresses the next node.
1212 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1216 VM_BUG_ON(!nodes_allowed);
1218 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1219 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1224 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1225 for (nr_nodes = nodes_weight(*mask); \
1227 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1230 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1231 for (nr_nodes = nodes_weight(*mask); \
1233 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1236 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1237 static void destroy_compound_gigantic_page(struct page *page,
1241 int nr_pages = 1 << order;
1242 struct page *p = page + 1;
1244 atomic_set(compound_mapcount_ptr(page), 0);
1245 atomic_set(compound_pincount_ptr(page), 0);
1247 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1248 clear_compound_head(p);
1249 set_page_refcounted(p);
1252 set_compound_order(page, 0);
1253 page[1].compound_nr = 0;
1254 __ClearPageHead(page);
1257 static void free_gigantic_page(struct page *page, unsigned int order)
1260 * If the page isn't allocated using the cma allocator,
1261 * cma_release() returns false.
1264 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1268 free_contig_range(page_to_pfn(page), 1 << order);
1271 #ifdef CONFIG_CONTIG_ALLOC
1272 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1273 int nid, nodemask_t *nodemask)
1275 unsigned long nr_pages = pages_per_huge_page(h);
1276 if (nid == NUMA_NO_NODE)
1277 nid = numa_mem_id();
1284 if (hugetlb_cma[nid]) {
1285 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1286 huge_page_order(h), true);
1291 if (!(gfp_mask & __GFP_THISNODE)) {
1292 for_each_node_mask(node, *nodemask) {
1293 if (node == nid || !hugetlb_cma[node])
1296 page = cma_alloc(hugetlb_cma[node], nr_pages,
1297 huge_page_order(h), true);
1305 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1308 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1309 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1310 #else /* !CONFIG_CONTIG_ALLOC */
1311 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1312 int nid, nodemask_t *nodemask)
1316 #endif /* CONFIG_CONTIG_ALLOC */
1318 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1319 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1320 int nid, nodemask_t *nodemask)
1324 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1325 static inline void destroy_compound_gigantic_page(struct page *page,
1326 unsigned int order) { }
1329 static void update_and_free_page(struct hstate *h, struct page *page)
1332 struct page *subpage = page;
1334 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1338 h->nr_huge_pages_node[page_to_nid(page)]--;
1339 for (i = 0; i < pages_per_huge_page(h);
1340 i++, subpage = mem_map_next(subpage, page, i)) {
1341 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1342 1 << PG_referenced | 1 << PG_dirty |
1343 1 << PG_active | 1 << PG_private |
1346 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1347 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1348 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1349 set_page_refcounted(page);
1350 if (hstate_is_gigantic(h)) {
1352 * Temporarily drop the hugetlb_lock, because
1353 * we might block in free_gigantic_page().
1355 spin_unlock(&hugetlb_lock);
1356 destroy_compound_gigantic_page(page, huge_page_order(h));
1357 free_gigantic_page(page, huge_page_order(h));
1358 spin_lock(&hugetlb_lock);
1360 __free_pages(page, huge_page_order(h));
1364 struct hstate *size_to_hstate(unsigned long size)
1368 for_each_hstate(h) {
1369 if (huge_page_size(h) == size)
1375 static void __free_huge_page(struct page *page)
1378 * Can't pass hstate in here because it is called from the
1379 * compound page destructor.
1381 struct hstate *h = page_hstate(page);
1382 int nid = page_to_nid(page);
1383 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1384 bool restore_reserve;
1386 VM_BUG_ON_PAGE(page_count(page), page);
1387 VM_BUG_ON_PAGE(page_mapcount(page), page);
1389 hugetlb_set_page_subpool(page, NULL);
1390 page->mapping = NULL;
1391 restore_reserve = HPageRestoreReserve(page);
1392 ClearHPageRestoreReserve(page);
1395 * If HPageRestoreReserve was set on page, page allocation consumed a
1396 * reservation. If the page was associated with a subpool, there
1397 * would have been a page reserved in the subpool before allocation
1398 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1399 * reservation, do not call hugepage_subpool_put_pages() as this will
1400 * remove the reserved page from the subpool.
1402 if (!restore_reserve) {
1404 * A return code of zero implies that the subpool will be
1405 * under its minimum size if the reservation is not restored
1406 * after page is free. Therefore, force restore_reserve
1409 if (hugepage_subpool_put_pages(spool, 1) == 0)
1410 restore_reserve = true;
1413 spin_lock(&hugetlb_lock);
1414 ClearHPageMigratable(page);
1415 hugetlb_cgroup_uncharge_page(hstate_index(h),
1416 pages_per_huge_page(h), page);
1417 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1418 pages_per_huge_page(h), page);
1419 if (restore_reserve)
1420 h->resv_huge_pages++;
1422 if (HPageTemporary(page)) {
1423 list_del(&page->lru);
1424 ClearHPageTemporary(page);
1425 update_and_free_page(h, page);
1426 } else if (h->surplus_huge_pages_node[nid]) {
1427 /* remove the page from active list */
1428 list_del(&page->lru);
1429 update_and_free_page(h, page);
1430 h->surplus_huge_pages--;
1431 h->surplus_huge_pages_node[nid]--;
1433 arch_clear_hugepage_flags(page);
1434 enqueue_huge_page(h, page);
1436 spin_unlock(&hugetlb_lock);
1440 * As free_huge_page() can be called from a non-task context, we have
1441 * to defer the actual freeing in a workqueue to prevent potential
1442 * hugetlb_lock deadlock.
1444 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1445 * be freed and frees them one-by-one. As the page->mapping pointer is
1446 * going to be cleared in __free_huge_page() anyway, it is reused as the
1447 * llist_node structure of a lockless linked list of huge pages to be freed.
1449 static LLIST_HEAD(hpage_freelist);
1451 static void free_hpage_workfn(struct work_struct *work)
1453 struct llist_node *node;
1456 node = llist_del_all(&hpage_freelist);
1459 page = container_of((struct address_space **)node,
1460 struct page, mapping);
1462 __free_huge_page(page);
1465 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1467 void free_huge_page(struct page *page)
1470 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1474 * Only call schedule_work() if hpage_freelist is previously
1475 * empty. Otherwise, schedule_work() had been called but the
1476 * workfn hasn't retrieved the list yet.
1478 if (llist_add((struct llist_node *)&page->mapping,
1480 schedule_work(&free_hpage_work);
1484 __free_huge_page(page);
1487 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1489 INIT_LIST_HEAD(&page->lru);
1490 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1491 hugetlb_set_page_subpool(page, NULL);
1492 set_hugetlb_cgroup(page, NULL);
1493 set_hugetlb_cgroup_rsvd(page, NULL);
1494 spin_lock(&hugetlb_lock);
1496 h->nr_huge_pages_node[nid]++;
1497 ClearHPageFreed(page);
1498 spin_unlock(&hugetlb_lock);
1501 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1504 int nr_pages = 1 << order;
1505 struct page *p = page + 1;
1507 /* we rely on prep_new_huge_page to set the destructor */
1508 set_compound_order(page, order);
1509 __ClearPageReserved(page);
1510 __SetPageHead(page);
1511 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1513 * For gigantic hugepages allocated through bootmem at
1514 * boot, it's safer to be consistent with the not-gigantic
1515 * hugepages and clear the PG_reserved bit from all tail pages
1516 * too. Otherwise drivers using get_user_pages() to access tail
1517 * pages may get the reference counting wrong if they see
1518 * PG_reserved set on a tail page (despite the head page not
1519 * having PG_reserved set). Enforcing this consistency between
1520 * head and tail pages allows drivers to optimize away a check
1521 * on the head page when they need know if put_page() is needed
1522 * after get_user_pages().
1524 __ClearPageReserved(p);
1525 set_page_count(p, 0);
1526 set_compound_head(p, page);
1528 atomic_set(compound_mapcount_ptr(page), -1);
1529 atomic_set(compound_pincount_ptr(page), 0);
1533 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1534 * transparent huge pages. See the PageTransHuge() documentation for more
1537 int PageHuge(struct page *page)
1539 if (!PageCompound(page))
1542 page = compound_head(page);
1543 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1545 EXPORT_SYMBOL_GPL(PageHuge);
1548 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1549 * normal or transparent huge pages.
1551 int PageHeadHuge(struct page *page_head)
1553 if (!PageHead(page_head))
1556 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1560 * Find and lock address space (mapping) in write mode.
1562 * Upon entry, the page is locked which means that page_mapping() is
1563 * stable. Due to locking order, we can only trylock_write. If we can
1564 * not get the lock, simply return NULL to caller.
1566 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1568 struct address_space *mapping = page_mapping(hpage);
1573 if (i_mmap_trylock_write(mapping))
1579 pgoff_t __basepage_index(struct page *page)
1581 struct page *page_head = compound_head(page);
1582 pgoff_t index = page_index(page_head);
1583 unsigned long compound_idx;
1585 if (!PageHuge(page_head))
1586 return page_index(page);
1588 if (compound_order(page_head) >= MAX_ORDER)
1589 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1591 compound_idx = page - page_head;
1593 return (index << compound_order(page_head)) + compound_idx;
1596 static struct page *alloc_buddy_huge_page(struct hstate *h,
1597 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1598 nodemask_t *node_alloc_noretry)
1600 int order = huge_page_order(h);
1602 bool alloc_try_hard = true;
1605 * By default we always try hard to allocate the page with
1606 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1607 * a loop (to adjust global huge page counts) and previous allocation
1608 * failed, do not continue to try hard on the same node. Use the
1609 * node_alloc_noretry bitmap to manage this state information.
1611 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1612 alloc_try_hard = false;
1613 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1615 gfp_mask |= __GFP_RETRY_MAYFAIL;
1616 if (nid == NUMA_NO_NODE)
1617 nid = numa_mem_id();
1618 page = __alloc_pages(gfp_mask, order, nid, nmask);
1620 __count_vm_event(HTLB_BUDDY_PGALLOC);
1622 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1625 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1626 * indicates an overall state change. Clear bit so that we resume
1627 * normal 'try hard' allocations.
1629 if (node_alloc_noretry && page && !alloc_try_hard)
1630 node_clear(nid, *node_alloc_noretry);
1633 * If we tried hard to get a page but failed, set bit so that
1634 * subsequent attempts will not try as hard until there is an
1635 * overall state change.
1637 if (node_alloc_noretry && !page && alloc_try_hard)
1638 node_set(nid, *node_alloc_noretry);
1644 * Common helper to allocate a fresh hugetlb page. All specific allocators
1645 * should use this function to get new hugetlb pages
1647 static struct page *alloc_fresh_huge_page(struct hstate *h,
1648 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1649 nodemask_t *node_alloc_noretry)
1653 if (hstate_is_gigantic(h))
1654 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1656 page = alloc_buddy_huge_page(h, gfp_mask,
1657 nid, nmask, node_alloc_noretry);
1661 if (hstate_is_gigantic(h))
1662 prep_compound_gigantic_page(page, huge_page_order(h));
1663 prep_new_huge_page(h, page, page_to_nid(page));
1669 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1672 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1673 nodemask_t *node_alloc_noretry)
1677 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1679 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1680 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1681 node_alloc_noretry);
1689 put_page(page); /* free it into the hugepage allocator */
1695 * Free huge page from pool from next node to free.
1696 * Attempt to keep persistent huge pages more or less
1697 * balanced over allowed nodes.
1698 * Called with hugetlb_lock locked.
1700 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1706 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1708 * If we're returning unused surplus pages, only examine
1709 * nodes with surplus pages.
1711 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1712 !list_empty(&h->hugepage_freelists[node])) {
1714 list_entry(h->hugepage_freelists[node].next,
1716 list_del(&page->lru);
1717 h->free_huge_pages--;
1718 h->free_huge_pages_node[node]--;
1720 h->surplus_huge_pages--;
1721 h->surplus_huge_pages_node[node]--;
1723 update_and_free_page(h, page);
1733 * Dissolve a given free hugepage into free buddy pages. This function does
1734 * nothing for in-use hugepages and non-hugepages.
1735 * This function returns values like below:
1737 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1738 * (allocated or reserved.)
1739 * 0: successfully dissolved free hugepages or the page is not a
1740 * hugepage (considered as already dissolved)
1742 int dissolve_free_huge_page(struct page *page)
1747 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1748 if (!PageHuge(page))
1751 spin_lock(&hugetlb_lock);
1752 if (!PageHuge(page)) {
1757 if (!page_count(page)) {
1758 struct page *head = compound_head(page);
1759 struct hstate *h = page_hstate(head);
1760 int nid = page_to_nid(head);
1761 if (h->free_huge_pages - h->resv_huge_pages == 0)
1765 * We should make sure that the page is already on the free list
1766 * when it is dissolved.
1768 if (unlikely(!HPageFreed(head))) {
1769 spin_unlock(&hugetlb_lock);
1773 * Theoretically, we should return -EBUSY when we
1774 * encounter this race. In fact, we have a chance
1775 * to successfully dissolve the page if we do a
1776 * retry. Because the race window is quite small.
1777 * If we seize this opportunity, it is an optimization
1778 * for increasing the success rate of dissolving page.
1784 * Move PageHWPoison flag from head page to the raw error page,
1785 * which makes any subpages rather than the error page reusable.
1787 if (PageHWPoison(head) && page != head) {
1788 SetPageHWPoison(page);
1789 ClearPageHWPoison(head);
1791 list_del(&head->lru);
1792 h->free_huge_pages--;
1793 h->free_huge_pages_node[nid]--;
1794 h->max_huge_pages--;
1795 update_and_free_page(h, head);
1799 spin_unlock(&hugetlb_lock);
1804 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1805 * make specified memory blocks removable from the system.
1806 * Note that this will dissolve a free gigantic hugepage completely, if any
1807 * part of it lies within the given range.
1808 * Also note that if dissolve_free_huge_page() returns with an error, all
1809 * free hugepages that were dissolved before that error are lost.
1811 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1817 if (!hugepages_supported())
1820 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1821 page = pfn_to_page(pfn);
1822 rc = dissolve_free_huge_page(page);
1831 * Allocates a fresh surplus page from the page allocator.
1833 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1834 int nid, nodemask_t *nmask)
1836 struct page *page = NULL;
1838 if (hstate_is_gigantic(h))
1841 spin_lock(&hugetlb_lock);
1842 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1844 spin_unlock(&hugetlb_lock);
1846 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1850 spin_lock(&hugetlb_lock);
1852 * We could have raced with the pool size change.
1853 * Double check that and simply deallocate the new page
1854 * if we would end up overcommiting the surpluses. Abuse
1855 * temporary page to workaround the nasty free_huge_page
1858 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1859 SetHPageTemporary(page);
1860 spin_unlock(&hugetlb_lock);
1864 h->surplus_huge_pages++;
1865 h->surplus_huge_pages_node[page_to_nid(page)]++;
1869 spin_unlock(&hugetlb_lock);
1874 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1875 int nid, nodemask_t *nmask)
1879 if (hstate_is_gigantic(h))
1882 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1887 * We do not account these pages as surplus because they are only
1888 * temporary and will be released properly on the last reference
1890 SetHPageTemporary(page);
1896 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1899 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1900 struct vm_area_struct *vma, unsigned long addr)
1903 struct mempolicy *mpol;
1904 gfp_t gfp_mask = htlb_alloc_mask(h);
1906 nodemask_t *nodemask;
1908 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1909 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1910 mpol_cond_put(mpol);
1915 /* page migration callback function */
1916 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1917 nodemask_t *nmask, gfp_t gfp_mask)
1919 spin_lock(&hugetlb_lock);
1920 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1923 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1925 spin_unlock(&hugetlb_lock);
1929 spin_unlock(&hugetlb_lock);
1931 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1934 /* mempolicy aware migration callback */
1935 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1936 unsigned long address)
1938 struct mempolicy *mpol;
1939 nodemask_t *nodemask;
1944 gfp_mask = htlb_alloc_mask(h);
1945 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1946 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1947 mpol_cond_put(mpol);
1953 * Increase the hugetlb pool such that it can accommodate a reservation
1956 static int gather_surplus_pages(struct hstate *h, long delta)
1957 __must_hold(&hugetlb_lock)
1959 struct list_head surplus_list;
1960 struct page *page, *tmp;
1963 long needed, allocated;
1964 bool alloc_ok = true;
1966 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1968 h->resv_huge_pages += delta;
1973 INIT_LIST_HEAD(&surplus_list);
1977 spin_unlock(&hugetlb_lock);
1978 for (i = 0; i < needed; i++) {
1979 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1980 NUMA_NO_NODE, NULL);
1985 list_add(&page->lru, &surplus_list);
1991 * After retaking hugetlb_lock, we need to recalculate 'needed'
1992 * because either resv_huge_pages or free_huge_pages may have changed.
1994 spin_lock(&hugetlb_lock);
1995 needed = (h->resv_huge_pages + delta) -
1996 (h->free_huge_pages + allocated);
2001 * We were not able to allocate enough pages to
2002 * satisfy the entire reservation so we free what
2003 * we've allocated so far.
2008 * The surplus_list now contains _at_least_ the number of extra pages
2009 * needed to accommodate the reservation. Add the appropriate number
2010 * of pages to the hugetlb pool and free the extras back to the buddy
2011 * allocator. Commit the entire reservation here to prevent another
2012 * process from stealing the pages as they are added to the pool but
2013 * before they are reserved.
2015 needed += allocated;
2016 h->resv_huge_pages += delta;
2019 /* Free the needed pages to the hugetlb pool */
2020 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2026 * This page is now managed by the hugetlb allocator and has
2027 * no users -- drop the buddy allocator's reference.
2029 zeroed = put_page_testzero(page);
2030 VM_BUG_ON_PAGE(!zeroed, page);
2031 enqueue_huge_page(h, page);
2034 spin_unlock(&hugetlb_lock);
2036 /* Free unnecessary surplus pages to the buddy allocator */
2037 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2039 spin_lock(&hugetlb_lock);
2045 * This routine has two main purposes:
2046 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2047 * in unused_resv_pages. This corresponds to the prior adjustments made
2048 * to the associated reservation map.
2049 * 2) Free any unused surplus pages that may have been allocated to satisfy
2050 * the reservation. As many as unused_resv_pages may be freed.
2052 * Called with hugetlb_lock held. However, the lock could be dropped (and
2053 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2054 * we must make sure nobody else can claim pages we are in the process of
2055 * freeing. Do this by ensuring resv_huge_page always is greater than the
2056 * number of huge pages we plan to free when dropping the lock.
2058 static void return_unused_surplus_pages(struct hstate *h,
2059 unsigned long unused_resv_pages)
2061 unsigned long nr_pages;
2063 /* Cannot return gigantic pages currently */
2064 if (hstate_is_gigantic(h))
2068 * Part (or even all) of the reservation could have been backed
2069 * by pre-allocated pages. Only free surplus pages.
2071 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2074 * We want to release as many surplus pages as possible, spread
2075 * evenly across all nodes with memory. Iterate across these nodes
2076 * until we can no longer free unreserved surplus pages. This occurs
2077 * when the nodes with surplus pages have no free pages.
2078 * free_pool_huge_page() will balance the freed pages across the
2079 * on-line nodes with memory and will handle the hstate accounting.
2081 * Note that we decrement resv_huge_pages as we free the pages. If
2082 * we drop the lock, resv_huge_pages will still be sufficiently large
2083 * to cover subsequent pages we may free.
2085 while (nr_pages--) {
2086 h->resv_huge_pages--;
2087 unused_resv_pages--;
2088 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2090 cond_resched_lock(&hugetlb_lock);
2094 /* Fully uncommit the reservation */
2095 h->resv_huge_pages -= unused_resv_pages;
2100 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2101 * are used by the huge page allocation routines to manage reservations.
2103 * vma_needs_reservation is called to determine if the huge page at addr
2104 * within the vma has an associated reservation. If a reservation is
2105 * needed, the value 1 is returned. The caller is then responsible for
2106 * managing the global reservation and subpool usage counts. After
2107 * the huge page has been allocated, vma_commit_reservation is called
2108 * to add the page to the reservation map. If the page allocation fails,
2109 * the reservation must be ended instead of committed. vma_end_reservation
2110 * is called in such cases.
2112 * In the normal case, vma_commit_reservation returns the same value
2113 * as the preceding vma_needs_reservation call. The only time this
2114 * is not the case is if a reserve map was changed between calls. It
2115 * is the responsibility of the caller to notice the difference and
2116 * take appropriate action.
2118 * vma_add_reservation is used in error paths where a reservation must
2119 * be restored when a newly allocated huge page must be freed. It is
2120 * to be called after calling vma_needs_reservation to determine if a
2121 * reservation exists.
2123 enum vma_resv_mode {
2129 static long __vma_reservation_common(struct hstate *h,
2130 struct vm_area_struct *vma, unsigned long addr,
2131 enum vma_resv_mode mode)
2133 struct resv_map *resv;
2136 long dummy_out_regions_needed;
2138 resv = vma_resv_map(vma);
2142 idx = vma_hugecache_offset(h, vma, addr);
2144 case VMA_NEEDS_RESV:
2145 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2146 /* We assume that vma_reservation_* routines always operate on
2147 * 1 page, and that adding to resv map a 1 page entry can only
2148 * ever require 1 region.
2150 VM_BUG_ON(dummy_out_regions_needed != 1);
2152 case VMA_COMMIT_RESV:
2153 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2154 /* region_add calls of range 1 should never fail. */
2158 region_abort(resv, idx, idx + 1, 1);
2162 if (vma->vm_flags & VM_MAYSHARE) {
2163 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2164 /* region_add calls of range 1 should never fail. */
2167 region_abort(resv, idx, idx + 1, 1);
2168 ret = region_del(resv, idx, idx + 1);
2175 if (vma->vm_flags & VM_MAYSHARE)
2177 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2179 * In most cases, reserves always exist for private mappings.
2180 * However, a file associated with mapping could have been
2181 * hole punched or truncated after reserves were consumed.
2182 * As subsequent fault on such a range will not use reserves.
2183 * Subtle - The reserve map for private mappings has the
2184 * opposite meaning than that of shared mappings. If NO
2185 * entry is in the reserve map, it means a reservation exists.
2186 * If an entry exists in the reserve map, it means the
2187 * reservation has already been consumed. As a result, the
2188 * return value of this routine is the opposite of the
2189 * value returned from reserve map manipulation routines above.
2197 return ret < 0 ? ret : 0;
2200 static long vma_needs_reservation(struct hstate *h,
2201 struct vm_area_struct *vma, unsigned long addr)
2203 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2206 static long vma_commit_reservation(struct hstate *h,
2207 struct vm_area_struct *vma, unsigned long addr)
2209 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2212 static void vma_end_reservation(struct hstate *h,
2213 struct vm_area_struct *vma, unsigned long addr)
2215 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2218 static long vma_add_reservation(struct hstate *h,
2219 struct vm_area_struct *vma, unsigned long addr)
2221 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2225 * This routine is called to restore a reservation on error paths. In the
2226 * specific error paths, a huge page was allocated (via alloc_huge_page)
2227 * and is about to be freed. If a reservation for the page existed,
2228 * alloc_huge_page would have consumed the reservation and set
2229 * HPageRestoreReserve in the newly allocated page. When the page is freed
2230 * via free_huge_page, the global reservation count will be incremented if
2231 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2232 * reserve map. Adjust the reserve map here to be consistent with global
2233 * reserve count adjustments to be made by free_huge_page.
2235 static void restore_reserve_on_error(struct hstate *h,
2236 struct vm_area_struct *vma, unsigned long address,
2239 if (unlikely(HPageRestoreReserve(page))) {
2240 long rc = vma_needs_reservation(h, vma, address);
2242 if (unlikely(rc < 0)) {
2244 * Rare out of memory condition in reserve map
2245 * manipulation. Clear HPageRestoreReserve so that
2246 * global reserve count will not be incremented
2247 * by free_huge_page. This will make it appear
2248 * as though the reservation for this page was
2249 * consumed. This may prevent the task from
2250 * faulting in the page at a later time. This
2251 * is better than inconsistent global huge page
2252 * accounting of reserve counts.
2254 ClearHPageRestoreReserve(page);
2256 rc = vma_add_reservation(h, vma, address);
2257 if (unlikely(rc < 0))
2259 * See above comment about rare out of
2262 ClearHPageRestoreReserve(page);
2264 vma_end_reservation(h, vma, address);
2268 struct page *alloc_huge_page(struct vm_area_struct *vma,
2269 unsigned long addr, int avoid_reserve)
2271 struct hugepage_subpool *spool = subpool_vma(vma);
2272 struct hstate *h = hstate_vma(vma);
2274 long map_chg, map_commit;
2277 struct hugetlb_cgroup *h_cg;
2278 bool deferred_reserve;
2280 idx = hstate_index(h);
2282 * Examine the region/reserve map to determine if the process
2283 * has a reservation for the page to be allocated. A return
2284 * code of zero indicates a reservation exists (no change).
2286 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2288 return ERR_PTR(-ENOMEM);
2291 * Processes that did not create the mapping will have no
2292 * reserves as indicated by the region/reserve map. Check
2293 * that the allocation will not exceed the subpool limit.
2294 * Allocations for MAP_NORESERVE mappings also need to be
2295 * checked against any subpool limit.
2297 if (map_chg || avoid_reserve) {
2298 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2300 vma_end_reservation(h, vma, addr);
2301 return ERR_PTR(-ENOSPC);
2305 * Even though there was no reservation in the region/reserve
2306 * map, there could be reservations associated with the
2307 * subpool that can be used. This would be indicated if the
2308 * return value of hugepage_subpool_get_pages() is zero.
2309 * However, if avoid_reserve is specified we still avoid even
2310 * the subpool reservations.
2316 /* If this allocation is not consuming a reservation, charge it now.
2318 deferred_reserve = map_chg || avoid_reserve;
2319 if (deferred_reserve) {
2320 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2321 idx, pages_per_huge_page(h), &h_cg);
2323 goto out_subpool_put;
2326 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2328 goto out_uncharge_cgroup_reservation;
2330 spin_lock(&hugetlb_lock);
2332 * glb_chg is passed to indicate whether or not a page must be taken
2333 * from the global free pool (global change). gbl_chg == 0 indicates
2334 * a reservation exists for the allocation.
2336 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2338 spin_unlock(&hugetlb_lock);
2339 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2341 goto out_uncharge_cgroup;
2342 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2343 SetHPageRestoreReserve(page);
2344 h->resv_huge_pages--;
2346 spin_lock(&hugetlb_lock);
2347 list_add(&page->lru, &h->hugepage_activelist);
2350 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2351 /* If allocation is not consuming a reservation, also store the
2352 * hugetlb_cgroup pointer on the page.
2354 if (deferred_reserve) {
2355 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2359 spin_unlock(&hugetlb_lock);
2361 hugetlb_set_page_subpool(page, spool);
2363 map_commit = vma_commit_reservation(h, vma, addr);
2364 if (unlikely(map_chg > map_commit)) {
2366 * The page was added to the reservation map between
2367 * vma_needs_reservation and vma_commit_reservation.
2368 * This indicates a race with hugetlb_reserve_pages.
2369 * Adjust for the subpool count incremented above AND
2370 * in hugetlb_reserve_pages for the same page. Also,
2371 * the reservation count added in hugetlb_reserve_pages
2372 * no longer applies.
2376 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2377 hugetlb_acct_memory(h, -rsv_adjust);
2378 if (deferred_reserve)
2379 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2380 pages_per_huge_page(h), page);
2384 out_uncharge_cgroup:
2385 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2386 out_uncharge_cgroup_reservation:
2387 if (deferred_reserve)
2388 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2391 if (map_chg || avoid_reserve)
2392 hugepage_subpool_put_pages(spool, 1);
2393 vma_end_reservation(h, vma, addr);
2394 return ERR_PTR(-ENOSPC);
2397 int alloc_bootmem_huge_page(struct hstate *h)
2398 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2399 int __alloc_bootmem_huge_page(struct hstate *h)
2401 struct huge_bootmem_page *m;
2404 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2407 addr = memblock_alloc_try_nid_raw(
2408 huge_page_size(h), huge_page_size(h),
2409 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2412 * Use the beginning of the huge page to store the
2413 * huge_bootmem_page struct (until gather_bootmem
2414 * puts them into the mem_map).
2423 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2424 /* Put them into a private list first because mem_map is not up yet */
2425 INIT_LIST_HEAD(&m->list);
2426 list_add(&m->list, &huge_boot_pages);
2431 static void __init prep_compound_huge_page(struct page *page,
2434 if (unlikely(order > (MAX_ORDER - 1)))
2435 prep_compound_gigantic_page(page, order);
2437 prep_compound_page(page, order);
2440 /* Put bootmem huge pages into the standard lists after mem_map is up */
2441 static void __init gather_bootmem_prealloc(void)
2443 struct huge_bootmem_page *m;
2445 list_for_each_entry(m, &huge_boot_pages, list) {
2446 struct page *page = virt_to_page(m);
2447 struct hstate *h = m->hstate;
2449 WARN_ON(page_count(page) != 1);
2450 prep_compound_huge_page(page, huge_page_order(h));
2451 WARN_ON(PageReserved(page));
2452 prep_new_huge_page(h, page, page_to_nid(page));
2453 put_page(page); /* free it into the hugepage allocator */
2456 * If we had gigantic hugepages allocated at boot time, we need
2457 * to restore the 'stolen' pages to totalram_pages in order to
2458 * fix confusing memory reports from free(1) and another
2459 * side-effects, like CommitLimit going negative.
2461 if (hstate_is_gigantic(h))
2462 adjust_managed_page_count(page, pages_per_huge_page(h));
2467 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2470 nodemask_t *node_alloc_noretry;
2472 if (!hstate_is_gigantic(h)) {
2474 * Bit mask controlling how hard we retry per-node allocations.
2475 * Ignore errors as lower level routines can deal with
2476 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2477 * time, we are likely in bigger trouble.
2479 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2482 /* allocations done at boot time */
2483 node_alloc_noretry = NULL;
2486 /* bit mask controlling how hard we retry per-node allocations */
2487 if (node_alloc_noretry)
2488 nodes_clear(*node_alloc_noretry);
2490 for (i = 0; i < h->max_huge_pages; ++i) {
2491 if (hstate_is_gigantic(h)) {
2492 if (hugetlb_cma_size) {
2493 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2496 if (!alloc_bootmem_huge_page(h))
2498 } else if (!alloc_pool_huge_page(h,
2499 &node_states[N_MEMORY],
2500 node_alloc_noretry))
2504 if (i < h->max_huge_pages) {
2507 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2508 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2509 h->max_huge_pages, buf, i);
2510 h->max_huge_pages = i;
2513 kfree(node_alloc_noretry);
2516 static void __init hugetlb_init_hstates(void)
2520 for_each_hstate(h) {
2521 if (minimum_order > huge_page_order(h))
2522 minimum_order = huge_page_order(h);
2524 /* oversize hugepages were init'ed in early boot */
2525 if (!hstate_is_gigantic(h))
2526 hugetlb_hstate_alloc_pages(h);
2528 VM_BUG_ON(minimum_order == UINT_MAX);
2531 static void __init report_hugepages(void)
2535 for_each_hstate(h) {
2538 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2539 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2540 buf, h->free_huge_pages);
2544 #ifdef CONFIG_HIGHMEM
2545 static void try_to_free_low(struct hstate *h, unsigned long count,
2546 nodemask_t *nodes_allowed)
2550 if (hstate_is_gigantic(h))
2553 for_each_node_mask(i, *nodes_allowed) {
2554 struct page *page, *next;
2555 struct list_head *freel = &h->hugepage_freelists[i];
2556 list_for_each_entry_safe(page, next, freel, lru) {
2557 if (count >= h->nr_huge_pages)
2559 if (PageHighMem(page))
2561 list_del(&page->lru);
2562 update_and_free_page(h, page);
2563 h->free_huge_pages--;
2564 h->free_huge_pages_node[page_to_nid(page)]--;
2569 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2570 nodemask_t *nodes_allowed)
2576 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2577 * balanced by operating on them in a round-robin fashion.
2578 * Returns 1 if an adjustment was made.
2580 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2585 VM_BUG_ON(delta != -1 && delta != 1);
2588 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2589 if (h->surplus_huge_pages_node[node])
2593 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2594 if (h->surplus_huge_pages_node[node] <
2595 h->nr_huge_pages_node[node])
2602 h->surplus_huge_pages += delta;
2603 h->surplus_huge_pages_node[node] += delta;
2607 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2608 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2609 nodemask_t *nodes_allowed)
2611 unsigned long min_count, ret;
2612 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2615 * Bit mask controlling how hard we retry per-node allocations.
2616 * If we can not allocate the bit mask, do not attempt to allocate
2617 * the requested huge pages.
2619 if (node_alloc_noretry)
2620 nodes_clear(*node_alloc_noretry);
2624 spin_lock(&hugetlb_lock);
2627 * Check for a node specific request.
2628 * Changing node specific huge page count may require a corresponding
2629 * change to the global count. In any case, the passed node mask
2630 * (nodes_allowed) will restrict alloc/free to the specified node.
2632 if (nid != NUMA_NO_NODE) {
2633 unsigned long old_count = count;
2635 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2637 * User may have specified a large count value which caused the
2638 * above calculation to overflow. In this case, they wanted
2639 * to allocate as many huge pages as possible. Set count to
2640 * largest possible value to align with their intention.
2642 if (count < old_count)
2647 * Gigantic pages runtime allocation depend on the capability for large
2648 * page range allocation.
2649 * If the system does not provide this feature, return an error when
2650 * the user tries to allocate gigantic pages but let the user free the
2651 * boottime allocated gigantic pages.
2653 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2654 if (count > persistent_huge_pages(h)) {
2655 spin_unlock(&hugetlb_lock);
2656 NODEMASK_FREE(node_alloc_noretry);
2659 /* Fall through to decrease pool */
2663 * Increase the pool size
2664 * First take pages out of surplus state. Then make up the
2665 * remaining difference by allocating fresh huge pages.
2667 * We might race with alloc_surplus_huge_page() here and be unable
2668 * to convert a surplus huge page to a normal huge page. That is
2669 * not critical, though, it just means the overall size of the
2670 * pool might be one hugepage larger than it needs to be, but
2671 * within all the constraints specified by the sysctls.
2673 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2674 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2678 while (count > persistent_huge_pages(h)) {
2680 * If this allocation races such that we no longer need the
2681 * page, free_huge_page will handle it by freeing the page
2682 * and reducing the surplus.
2684 spin_unlock(&hugetlb_lock);
2686 /* yield cpu to avoid soft lockup */
2689 ret = alloc_pool_huge_page(h, nodes_allowed,
2690 node_alloc_noretry);
2691 spin_lock(&hugetlb_lock);
2695 /* Bail for signals. Probably ctrl-c from user */
2696 if (signal_pending(current))
2701 * Decrease the pool size
2702 * First return free pages to the buddy allocator (being careful
2703 * to keep enough around to satisfy reservations). Then place
2704 * pages into surplus state as needed so the pool will shrink
2705 * to the desired size as pages become free.
2707 * By placing pages into the surplus state independent of the
2708 * overcommit value, we are allowing the surplus pool size to
2709 * exceed overcommit. There are few sane options here. Since
2710 * alloc_surplus_huge_page() is checking the global counter,
2711 * though, we'll note that we're not allowed to exceed surplus
2712 * and won't grow the pool anywhere else. Not until one of the
2713 * sysctls are changed, or the surplus pages go out of use.
2715 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2716 min_count = max(count, min_count);
2717 try_to_free_low(h, min_count, nodes_allowed);
2718 while (min_count < persistent_huge_pages(h)) {
2719 if (!free_pool_huge_page(h, nodes_allowed, 0))
2721 cond_resched_lock(&hugetlb_lock);
2723 while (count < persistent_huge_pages(h)) {
2724 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2728 h->max_huge_pages = persistent_huge_pages(h);
2729 spin_unlock(&hugetlb_lock);
2731 NODEMASK_FREE(node_alloc_noretry);
2736 #define HSTATE_ATTR_RO(_name) \
2737 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2739 #define HSTATE_ATTR(_name) \
2740 static struct kobj_attribute _name##_attr = \
2741 __ATTR(_name, 0644, _name##_show, _name##_store)
2743 static struct kobject *hugepages_kobj;
2744 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2746 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2748 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2752 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2753 if (hstate_kobjs[i] == kobj) {
2755 *nidp = NUMA_NO_NODE;
2759 return kobj_to_node_hstate(kobj, nidp);
2762 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2763 struct kobj_attribute *attr, char *buf)
2766 unsigned long nr_huge_pages;
2769 h = kobj_to_hstate(kobj, &nid);
2770 if (nid == NUMA_NO_NODE)
2771 nr_huge_pages = h->nr_huge_pages;
2773 nr_huge_pages = h->nr_huge_pages_node[nid];
2775 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2778 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2779 struct hstate *h, int nid,
2780 unsigned long count, size_t len)
2783 nodemask_t nodes_allowed, *n_mask;
2785 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2788 if (nid == NUMA_NO_NODE) {
2790 * global hstate attribute
2792 if (!(obey_mempolicy &&
2793 init_nodemask_of_mempolicy(&nodes_allowed)))
2794 n_mask = &node_states[N_MEMORY];
2796 n_mask = &nodes_allowed;
2799 * Node specific request. count adjustment happens in
2800 * set_max_huge_pages() after acquiring hugetlb_lock.
2802 init_nodemask_of_node(&nodes_allowed, nid);
2803 n_mask = &nodes_allowed;
2806 err = set_max_huge_pages(h, count, nid, n_mask);
2808 return err ? err : len;
2811 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2812 struct kobject *kobj, const char *buf,
2816 unsigned long count;
2820 err = kstrtoul(buf, 10, &count);
2824 h = kobj_to_hstate(kobj, &nid);
2825 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2828 static ssize_t nr_hugepages_show(struct kobject *kobj,
2829 struct kobj_attribute *attr, char *buf)
2831 return nr_hugepages_show_common(kobj, attr, buf);
2834 static ssize_t nr_hugepages_store(struct kobject *kobj,
2835 struct kobj_attribute *attr, const char *buf, size_t len)
2837 return nr_hugepages_store_common(false, kobj, buf, len);
2839 HSTATE_ATTR(nr_hugepages);
2844 * hstate attribute for optionally mempolicy-based constraint on persistent
2845 * huge page alloc/free.
2847 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2848 struct kobj_attribute *attr,
2851 return nr_hugepages_show_common(kobj, attr, buf);
2854 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2855 struct kobj_attribute *attr, const char *buf, size_t len)
2857 return nr_hugepages_store_common(true, kobj, buf, len);
2859 HSTATE_ATTR(nr_hugepages_mempolicy);
2863 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2864 struct kobj_attribute *attr, char *buf)
2866 struct hstate *h = kobj_to_hstate(kobj, NULL);
2867 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2870 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2871 struct kobj_attribute *attr, const char *buf, size_t count)
2874 unsigned long input;
2875 struct hstate *h = kobj_to_hstate(kobj, NULL);
2877 if (hstate_is_gigantic(h))
2880 err = kstrtoul(buf, 10, &input);
2884 spin_lock(&hugetlb_lock);
2885 h->nr_overcommit_huge_pages = input;
2886 spin_unlock(&hugetlb_lock);
2890 HSTATE_ATTR(nr_overcommit_hugepages);
2892 static ssize_t free_hugepages_show(struct kobject *kobj,
2893 struct kobj_attribute *attr, char *buf)
2896 unsigned long free_huge_pages;
2899 h = kobj_to_hstate(kobj, &nid);
2900 if (nid == NUMA_NO_NODE)
2901 free_huge_pages = h->free_huge_pages;
2903 free_huge_pages = h->free_huge_pages_node[nid];
2905 return sysfs_emit(buf, "%lu\n", free_huge_pages);
2907 HSTATE_ATTR_RO(free_hugepages);
2909 static ssize_t resv_hugepages_show(struct kobject *kobj,
2910 struct kobj_attribute *attr, char *buf)
2912 struct hstate *h = kobj_to_hstate(kobj, NULL);
2913 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2915 HSTATE_ATTR_RO(resv_hugepages);
2917 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2918 struct kobj_attribute *attr, char *buf)
2921 unsigned long surplus_huge_pages;
2924 h = kobj_to_hstate(kobj, &nid);
2925 if (nid == NUMA_NO_NODE)
2926 surplus_huge_pages = h->surplus_huge_pages;
2928 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2930 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2932 HSTATE_ATTR_RO(surplus_hugepages);
2934 static struct attribute *hstate_attrs[] = {
2935 &nr_hugepages_attr.attr,
2936 &nr_overcommit_hugepages_attr.attr,
2937 &free_hugepages_attr.attr,
2938 &resv_hugepages_attr.attr,
2939 &surplus_hugepages_attr.attr,
2941 &nr_hugepages_mempolicy_attr.attr,
2946 static const struct attribute_group hstate_attr_group = {
2947 .attrs = hstate_attrs,
2950 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2951 struct kobject **hstate_kobjs,
2952 const struct attribute_group *hstate_attr_group)
2955 int hi = hstate_index(h);
2957 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2958 if (!hstate_kobjs[hi])
2961 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2963 kobject_put(hstate_kobjs[hi]);
2964 hstate_kobjs[hi] = NULL;
2970 static void __init hugetlb_sysfs_init(void)
2975 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2976 if (!hugepages_kobj)
2979 for_each_hstate(h) {
2980 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2981 hstate_kobjs, &hstate_attr_group);
2983 pr_err("HugeTLB: Unable to add hstate %s", h->name);
2990 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2991 * with node devices in node_devices[] using a parallel array. The array
2992 * index of a node device or _hstate == node id.
2993 * This is here to avoid any static dependency of the node device driver, in
2994 * the base kernel, on the hugetlb module.
2996 struct node_hstate {
2997 struct kobject *hugepages_kobj;
2998 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3000 static struct node_hstate node_hstates[MAX_NUMNODES];
3003 * A subset of global hstate attributes for node devices
3005 static struct attribute *per_node_hstate_attrs[] = {
3006 &nr_hugepages_attr.attr,
3007 &free_hugepages_attr.attr,
3008 &surplus_hugepages_attr.attr,
3012 static const struct attribute_group per_node_hstate_attr_group = {
3013 .attrs = per_node_hstate_attrs,
3017 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3018 * Returns node id via non-NULL nidp.
3020 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3024 for (nid = 0; nid < nr_node_ids; nid++) {
3025 struct node_hstate *nhs = &node_hstates[nid];
3027 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3028 if (nhs->hstate_kobjs[i] == kobj) {
3040 * Unregister hstate attributes from a single node device.
3041 * No-op if no hstate attributes attached.
3043 static void hugetlb_unregister_node(struct node *node)
3046 struct node_hstate *nhs = &node_hstates[node->dev.id];
3048 if (!nhs->hugepages_kobj)
3049 return; /* no hstate attributes */
3051 for_each_hstate(h) {
3052 int idx = hstate_index(h);
3053 if (nhs->hstate_kobjs[idx]) {
3054 kobject_put(nhs->hstate_kobjs[idx]);
3055 nhs->hstate_kobjs[idx] = NULL;
3059 kobject_put(nhs->hugepages_kobj);
3060 nhs->hugepages_kobj = NULL;
3065 * Register hstate attributes for a single node device.
3066 * No-op if attributes already registered.
3068 static void hugetlb_register_node(struct node *node)
3071 struct node_hstate *nhs = &node_hstates[node->dev.id];
3074 if (nhs->hugepages_kobj)
3075 return; /* already allocated */
3077 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3079 if (!nhs->hugepages_kobj)
3082 for_each_hstate(h) {
3083 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3085 &per_node_hstate_attr_group);
3087 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3088 h->name, node->dev.id);
3089 hugetlb_unregister_node(node);
3096 * hugetlb init time: register hstate attributes for all registered node
3097 * devices of nodes that have memory. All on-line nodes should have
3098 * registered their associated device by this time.
3100 static void __init hugetlb_register_all_nodes(void)
3104 for_each_node_state(nid, N_MEMORY) {
3105 struct node *node = node_devices[nid];
3106 if (node->dev.id == nid)
3107 hugetlb_register_node(node);
3111 * Let the node device driver know we're here so it can
3112 * [un]register hstate attributes on node hotplug.
3114 register_hugetlbfs_with_node(hugetlb_register_node,
3115 hugetlb_unregister_node);
3117 #else /* !CONFIG_NUMA */
3119 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3127 static void hugetlb_register_all_nodes(void) { }
3131 static int __init hugetlb_init(void)
3135 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3138 if (!hugepages_supported()) {
3139 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3140 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3145 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3146 * architectures depend on setup being done here.
3148 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3149 if (!parsed_default_hugepagesz) {
3151 * If we did not parse a default huge page size, set
3152 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3153 * number of huge pages for this default size was implicitly
3154 * specified, set that here as well.
3155 * Note that the implicit setting will overwrite an explicit
3156 * setting. A warning will be printed in this case.
3158 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3159 if (default_hstate_max_huge_pages) {
3160 if (default_hstate.max_huge_pages) {
3163 string_get_size(huge_page_size(&default_hstate),
3164 1, STRING_UNITS_2, buf, 32);
3165 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3166 default_hstate.max_huge_pages, buf);
3167 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3168 default_hstate_max_huge_pages);
3170 default_hstate.max_huge_pages =
3171 default_hstate_max_huge_pages;
3175 hugetlb_cma_check();
3176 hugetlb_init_hstates();
3177 gather_bootmem_prealloc();
3180 hugetlb_sysfs_init();
3181 hugetlb_register_all_nodes();
3182 hugetlb_cgroup_file_init();
3185 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3187 num_fault_mutexes = 1;
3189 hugetlb_fault_mutex_table =
3190 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3192 BUG_ON(!hugetlb_fault_mutex_table);
3194 for (i = 0; i < num_fault_mutexes; i++)
3195 mutex_init(&hugetlb_fault_mutex_table[i]);
3198 subsys_initcall(hugetlb_init);
3200 /* Overwritten by architectures with more huge page sizes */
3201 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3203 return size == HPAGE_SIZE;
3206 void __init hugetlb_add_hstate(unsigned int order)
3211 if (size_to_hstate(PAGE_SIZE << order)) {
3214 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3216 h = &hstates[hugetlb_max_hstate++];
3218 h->mask = ~(huge_page_size(h) - 1);
3219 for (i = 0; i < MAX_NUMNODES; ++i)
3220 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3221 INIT_LIST_HEAD(&h->hugepage_activelist);
3222 h->next_nid_to_alloc = first_memory_node;
3223 h->next_nid_to_free = first_memory_node;
3224 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3225 huge_page_size(h)/1024);
3231 * hugepages command line processing
3232 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3233 * specification. If not, ignore the hugepages value. hugepages can also
3234 * be the first huge page command line option in which case it implicitly
3235 * specifies the number of huge pages for the default size.
3237 static int __init hugepages_setup(char *s)
3240 static unsigned long *last_mhp;
3242 if (!parsed_valid_hugepagesz) {
3243 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3244 parsed_valid_hugepagesz = true;
3249 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3250 * yet, so this hugepages= parameter goes to the "default hstate".
3251 * Otherwise, it goes with the previously parsed hugepagesz or
3252 * default_hugepagesz.
3254 else if (!hugetlb_max_hstate)
3255 mhp = &default_hstate_max_huge_pages;
3257 mhp = &parsed_hstate->max_huge_pages;
3259 if (mhp == last_mhp) {
3260 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3264 if (sscanf(s, "%lu", mhp) <= 0)
3268 * Global state is always initialized later in hugetlb_init.
3269 * But we need to allocate gigantic hstates here early to still
3270 * use the bootmem allocator.
3272 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3273 hugetlb_hstate_alloc_pages(parsed_hstate);
3279 __setup("hugepages=", hugepages_setup);
3282 * hugepagesz command line processing
3283 * A specific huge page size can only be specified once with hugepagesz.
3284 * hugepagesz is followed by hugepages on the command line. The global
3285 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3286 * hugepagesz argument was valid.
3288 static int __init hugepagesz_setup(char *s)
3293 parsed_valid_hugepagesz = false;
3294 size = (unsigned long)memparse(s, NULL);
3296 if (!arch_hugetlb_valid_size(size)) {
3297 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3301 h = size_to_hstate(size);
3304 * hstate for this size already exists. This is normally
3305 * an error, but is allowed if the existing hstate is the
3306 * default hstate. More specifically, it is only allowed if
3307 * the number of huge pages for the default hstate was not
3308 * previously specified.
3310 if (!parsed_default_hugepagesz || h != &default_hstate ||
3311 default_hstate.max_huge_pages) {
3312 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3317 * No need to call hugetlb_add_hstate() as hstate already
3318 * exists. But, do set parsed_hstate so that a following
3319 * hugepages= parameter will be applied to this hstate.
3322 parsed_valid_hugepagesz = true;
3326 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3327 parsed_valid_hugepagesz = true;
3330 __setup("hugepagesz=", hugepagesz_setup);
3333 * default_hugepagesz command line input
3334 * Only one instance of default_hugepagesz allowed on command line.
3336 static int __init default_hugepagesz_setup(char *s)
3340 parsed_valid_hugepagesz = false;
3341 if (parsed_default_hugepagesz) {
3342 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3346 size = (unsigned long)memparse(s, NULL);
3348 if (!arch_hugetlb_valid_size(size)) {
3349 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3353 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3354 parsed_valid_hugepagesz = true;
3355 parsed_default_hugepagesz = true;
3356 default_hstate_idx = hstate_index(size_to_hstate(size));
3359 * The number of default huge pages (for this size) could have been
3360 * specified as the first hugetlb parameter: hugepages=X. If so,
3361 * then default_hstate_max_huge_pages is set. If the default huge
3362 * page size is gigantic (>= MAX_ORDER), then the pages must be
3363 * allocated here from bootmem allocator.
3365 if (default_hstate_max_huge_pages) {
3366 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3367 if (hstate_is_gigantic(&default_hstate))
3368 hugetlb_hstate_alloc_pages(&default_hstate);
3369 default_hstate_max_huge_pages = 0;
3374 __setup("default_hugepagesz=", default_hugepagesz_setup);
3376 static unsigned int allowed_mems_nr(struct hstate *h)
3379 unsigned int nr = 0;
3380 nodemask_t *mpol_allowed;
3381 unsigned int *array = h->free_huge_pages_node;
3382 gfp_t gfp_mask = htlb_alloc_mask(h);
3384 mpol_allowed = policy_nodemask_current(gfp_mask);
3386 for_each_node_mask(node, cpuset_current_mems_allowed) {
3387 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3394 #ifdef CONFIG_SYSCTL
3395 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3396 void *buffer, size_t *length,
3397 loff_t *ppos, unsigned long *out)
3399 struct ctl_table dup_table;
3402 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3403 * can duplicate the @table and alter the duplicate of it.
3406 dup_table.data = out;
3408 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3411 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3412 struct ctl_table *table, int write,
3413 void *buffer, size_t *length, loff_t *ppos)
3415 struct hstate *h = &default_hstate;
3416 unsigned long tmp = h->max_huge_pages;
3419 if (!hugepages_supported())
3422 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3428 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3429 NUMA_NO_NODE, tmp, *length);
3434 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3435 void *buffer, size_t *length, loff_t *ppos)
3438 return hugetlb_sysctl_handler_common(false, table, write,
3439 buffer, length, ppos);
3443 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3444 void *buffer, size_t *length, loff_t *ppos)
3446 return hugetlb_sysctl_handler_common(true, table, write,
3447 buffer, length, ppos);
3449 #endif /* CONFIG_NUMA */
3451 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3452 void *buffer, size_t *length, loff_t *ppos)
3454 struct hstate *h = &default_hstate;
3458 if (!hugepages_supported())
3461 tmp = h->nr_overcommit_huge_pages;
3463 if (write && hstate_is_gigantic(h))
3466 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3472 spin_lock(&hugetlb_lock);
3473 h->nr_overcommit_huge_pages = tmp;
3474 spin_unlock(&hugetlb_lock);
3480 #endif /* CONFIG_SYSCTL */
3482 void hugetlb_report_meminfo(struct seq_file *m)
3485 unsigned long total = 0;
3487 if (!hugepages_supported())
3490 for_each_hstate(h) {
3491 unsigned long count = h->nr_huge_pages;
3493 total += huge_page_size(h) * count;
3495 if (h == &default_hstate)
3497 "HugePages_Total: %5lu\n"
3498 "HugePages_Free: %5lu\n"
3499 "HugePages_Rsvd: %5lu\n"
3500 "HugePages_Surp: %5lu\n"
3501 "Hugepagesize: %8lu kB\n",
3505 h->surplus_huge_pages,
3506 huge_page_size(h) / SZ_1K);
3509 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3512 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3514 struct hstate *h = &default_hstate;
3516 if (!hugepages_supported())
3519 return sysfs_emit_at(buf, len,
3520 "Node %d HugePages_Total: %5u\n"
3521 "Node %d HugePages_Free: %5u\n"
3522 "Node %d HugePages_Surp: %5u\n",
3523 nid, h->nr_huge_pages_node[nid],
3524 nid, h->free_huge_pages_node[nid],
3525 nid, h->surplus_huge_pages_node[nid]);
3528 void hugetlb_show_meminfo(void)
3533 if (!hugepages_supported())
3536 for_each_node_state(nid, N_MEMORY)
3538 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3540 h->nr_huge_pages_node[nid],
3541 h->free_huge_pages_node[nid],
3542 h->surplus_huge_pages_node[nid],
3543 huge_page_size(h) / SZ_1K);
3546 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3548 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3549 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3552 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3553 unsigned long hugetlb_total_pages(void)
3556 unsigned long nr_total_pages = 0;
3559 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3560 return nr_total_pages;
3563 static int hugetlb_acct_memory(struct hstate *h, long delta)
3570 spin_lock(&hugetlb_lock);
3572 * When cpuset is configured, it breaks the strict hugetlb page
3573 * reservation as the accounting is done on a global variable. Such
3574 * reservation is completely rubbish in the presence of cpuset because
3575 * the reservation is not checked against page availability for the
3576 * current cpuset. Application can still potentially OOM'ed by kernel
3577 * with lack of free htlb page in cpuset that the task is in.
3578 * Attempt to enforce strict accounting with cpuset is almost
3579 * impossible (or too ugly) because cpuset is too fluid that
3580 * task or memory node can be dynamically moved between cpusets.
3582 * The change of semantics for shared hugetlb mapping with cpuset is
3583 * undesirable. However, in order to preserve some of the semantics,
3584 * we fall back to check against current free page availability as
3585 * a best attempt and hopefully to minimize the impact of changing
3586 * semantics that cpuset has.
3588 * Apart from cpuset, we also have memory policy mechanism that
3589 * also determines from which node the kernel will allocate memory
3590 * in a NUMA system. So similar to cpuset, we also should consider
3591 * the memory policy of the current task. Similar to the description
3595 if (gather_surplus_pages(h, delta) < 0)
3598 if (delta > allowed_mems_nr(h)) {
3599 return_unused_surplus_pages(h, delta);
3606 return_unused_surplus_pages(h, (unsigned long) -delta);
3609 spin_unlock(&hugetlb_lock);
3613 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3615 struct resv_map *resv = vma_resv_map(vma);
3618 * This new VMA should share its siblings reservation map if present.
3619 * The VMA will only ever have a valid reservation map pointer where
3620 * it is being copied for another still existing VMA. As that VMA
3621 * has a reference to the reservation map it cannot disappear until
3622 * after this open call completes. It is therefore safe to take a
3623 * new reference here without additional locking.
3625 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3626 kref_get(&resv->refs);
3629 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3631 struct hstate *h = hstate_vma(vma);
3632 struct resv_map *resv = vma_resv_map(vma);
3633 struct hugepage_subpool *spool = subpool_vma(vma);
3634 unsigned long reserve, start, end;
3637 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3640 start = vma_hugecache_offset(h, vma, vma->vm_start);
3641 end = vma_hugecache_offset(h, vma, vma->vm_end);
3643 reserve = (end - start) - region_count(resv, start, end);
3644 hugetlb_cgroup_uncharge_counter(resv, start, end);
3647 * Decrement reserve counts. The global reserve count may be
3648 * adjusted if the subpool has a minimum size.
3650 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3651 hugetlb_acct_memory(h, -gbl_reserve);
3654 kref_put(&resv->refs, resv_map_release);
3657 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3659 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3664 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3666 return huge_page_size(hstate_vma(vma));
3670 * We cannot handle pagefaults against hugetlb pages at all. They cause
3671 * handle_mm_fault() to try to instantiate regular-sized pages in the
3672 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3675 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3682 * When a new function is introduced to vm_operations_struct and added
3683 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3684 * This is because under System V memory model, mappings created via
3685 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3686 * their original vm_ops are overwritten with shm_vm_ops.
3688 const struct vm_operations_struct hugetlb_vm_ops = {
3689 .fault = hugetlb_vm_op_fault,
3690 .open = hugetlb_vm_op_open,
3691 .close = hugetlb_vm_op_close,
3692 .may_split = hugetlb_vm_op_split,
3693 .pagesize = hugetlb_vm_op_pagesize,
3696 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3702 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3703 vma->vm_page_prot)));
3705 entry = huge_pte_wrprotect(mk_huge_pte(page,
3706 vma->vm_page_prot));
3708 entry = pte_mkyoung(entry);
3709 entry = pte_mkhuge(entry);
3710 entry = arch_make_huge_pte(entry, vma, page, writable);
3715 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3716 unsigned long address, pte_t *ptep)
3720 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3721 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3722 update_mmu_cache(vma, address, ptep);
3725 bool is_hugetlb_entry_migration(pte_t pte)
3729 if (huge_pte_none(pte) || pte_present(pte))
3731 swp = pte_to_swp_entry(pte);
3732 if (is_migration_entry(swp))
3738 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3742 if (huge_pte_none(pte) || pte_present(pte))
3744 swp = pte_to_swp_entry(pte);
3745 if (is_hwpoison_entry(swp))
3752 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3753 struct page *new_page)
3755 __SetPageUptodate(new_page);
3756 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3757 hugepage_add_new_anon_rmap(new_page, vma, addr);
3758 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3759 ClearHPageRestoreReserve(new_page);
3760 SetHPageMigratable(new_page);
3763 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3764 struct vm_area_struct *vma)
3766 pte_t *src_pte, *dst_pte, entry, dst_entry;
3767 struct page *ptepage;
3769 bool cow = is_cow_mapping(vma->vm_flags);
3770 struct hstate *h = hstate_vma(vma);
3771 unsigned long sz = huge_page_size(h);
3772 unsigned long npages = pages_per_huge_page(h);
3773 struct address_space *mapping = vma->vm_file->f_mapping;
3774 struct mmu_notifier_range range;
3778 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3781 mmu_notifier_invalidate_range_start(&range);
3784 * For shared mappings i_mmap_rwsem must be held to call
3785 * huge_pte_alloc, otherwise the returned ptep could go
3786 * away if part of a shared pmd and another thread calls
3789 i_mmap_lock_read(mapping);
3792 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3793 spinlock_t *src_ptl, *dst_ptl;
3794 src_pte = huge_pte_offset(src, addr, sz);
3797 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
3804 * If the pagetables are shared don't copy or take references.
3805 * dst_pte == src_pte is the common case of src/dest sharing.
3807 * However, src could have 'unshared' and dst shares with
3808 * another vma. If dst_pte !none, this implies sharing.
3809 * Check here before taking page table lock, and once again
3810 * after taking the lock below.
3812 dst_entry = huge_ptep_get(dst_pte);
3813 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3816 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3817 src_ptl = huge_pte_lockptr(h, src, src_pte);
3818 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3819 entry = huge_ptep_get(src_pte);
3820 dst_entry = huge_ptep_get(dst_pte);
3822 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3824 * Skip if src entry none. Also, skip in the
3825 * unlikely case dst entry !none as this implies
3826 * sharing with another vma.
3829 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3830 is_hugetlb_entry_hwpoisoned(entry))) {
3831 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3833 if (is_write_migration_entry(swp_entry) && cow) {
3835 * COW mappings require pages in both
3836 * parent and child to be set to read.
3838 make_migration_entry_read(&swp_entry);
3839 entry = swp_entry_to_pte(swp_entry);
3840 set_huge_swap_pte_at(src, addr, src_pte,
3843 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3845 entry = huge_ptep_get(src_pte);
3846 ptepage = pte_page(entry);
3850 * This is a rare case where we see pinned hugetlb
3851 * pages while they're prone to COW. We need to do the
3852 * COW earlier during fork.
3854 * When pre-allocating the page or copying data, we
3855 * need to be without the pgtable locks since we could
3856 * sleep during the process.
3858 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3859 pte_t src_pte_old = entry;
3862 spin_unlock(src_ptl);
3863 spin_unlock(dst_ptl);
3864 /* Do not use reserve as it's private owned */
3865 new = alloc_huge_page(vma, addr, 1);
3871 copy_user_huge_page(new, ptepage, addr, vma,
3875 /* Install the new huge page if src pte stable */
3876 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3877 src_ptl = huge_pte_lockptr(h, src, src_pte);
3878 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3879 entry = huge_ptep_get(src_pte);
3880 if (!pte_same(src_pte_old, entry)) {
3882 /* dst_entry won't change as in child */
3885 hugetlb_install_page(vma, dst_pte, addr, new);
3886 spin_unlock(src_ptl);
3887 spin_unlock(dst_ptl);
3893 * No need to notify as we are downgrading page
3894 * table protection not changing it to point
3897 * See Documentation/vm/mmu_notifier.rst
3899 huge_ptep_set_wrprotect(src, addr, src_pte);
3902 page_dup_rmap(ptepage, true);
3903 set_huge_pte_at(dst, addr, dst_pte, entry);
3904 hugetlb_count_add(npages, dst);
3906 spin_unlock(src_ptl);
3907 spin_unlock(dst_ptl);
3911 mmu_notifier_invalidate_range_end(&range);
3913 i_mmap_unlock_read(mapping);
3918 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3919 unsigned long start, unsigned long end,
3920 struct page *ref_page)
3922 struct mm_struct *mm = vma->vm_mm;
3923 unsigned long address;
3928 struct hstate *h = hstate_vma(vma);
3929 unsigned long sz = huge_page_size(h);
3930 struct mmu_notifier_range range;
3932 WARN_ON(!is_vm_hugetlb_page(vma));
3933 BUG_ON(start & ~huge_page_mask(h));
3934 BUG_ON(end & ~huge_page_mask(h));
3937 * This is a hugetlb vma, all the pte entries should point
3940 tlb_change_page_size(tlb, sz);
3941 tlb_start_vma(tlb, vma);
3944 * If sharing possible, alert mmu notifiers of worst case.
3946 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3948 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3949 mmu_notifier_invalidate_range_start(&range);
3951 for (; address < end; address += sz) {
3952 ptep = huge_pte_offset(mm, address, sz);
3956 ptl = huge_pte_lock(h, mm, ptep);
3957 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3960 * We just unmapped a page of PMDs by clearing a PUD.
3961 * The caller's TLB flush range should cover this area.
3966 pte = huge_ptep_get(ptep);
3967 if (huge_pte_none(pte)) {
3973 * Migrating hugepage or HWPoisoned hugepage is already
3974 * unmapped and its refcount is dropped, so just clear pte here.
3976 if (unlikely(!pte_present(pte))) {
3977 huge_pte_clear(mm, address, ptep, sz);
3982 page = pte_page(pte);
3984 * If a reference page is supplied, it is because a specific
3985 * page is being unmapped, not a range. Ensure the page we
3986 * are about to unmap is the actual page of interest.
3989 if (page != ref_page) {
3994 * Mark the VMA as having unmapped its page so that
3995 * future faults in this VMA will fail rather than
3996 * looking like data was lost
3998 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4001 pte = huge_ptep_get_and_clear(mm, address, ptep);
4002 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4003 if (huge_pte_dirty(pte))
4004 set_page_dirty(page);
4006 hugetlb_count_sub(pages_per_huge_page(h), mm);
4007 page_remove_rmap(page, true);
4010 tlb_remove_page_size(tlb, page, huge_page_size(h));
4012 * Bail out after unmapping reference page if supplied
4017 mmu_notifier_invalidate_range_end(&range);
4018 tlb_end_vma(tlb, vma);
4021 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4022 struct vm_area_struct *vma, unsigned long start,
4023 unsigned long end, struct page *ref_page)
4025 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4028 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4029 * test will fail on a vma being torn down, and not grab a page table
4030 * on its way out. We're lucky that the flag has such an appropriate
4031 * name, and can in fact be safely cleared here. We could clear it
4032 * before the __unmap_hugepage_range above, but all that's necessary
4033 * is to clear it before releasing the i_mmap_rwsem. This works
4034 * because in the context this is called, the VMA is about to be
4035 * destroyed and the i_mmap_rwsem is held.
4037 vma->vm_flags &= ~VM_MAYSHARE;
4040 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4041 unsigned long end, struct page *ref_page)
4043 struct mmu_gather tlb;
4045 tlb_gather_mmu(&tlb, vma->vm_mm);
4046 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4047 tlb_finish_mmu(&tlb);
4051 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4052 * mapping it owns the reserve page for. The intention is to unmap the page
4053 * from other VMAs and let the children be SIGKILLed if they are faulting the
4056 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4057 struct page *page, unsigned long address)
4059 struct hstate *h = hstate_vma(vma);
4060 struct vm_area_struct *iter_vma;
4061 struct address_space *mapping;
4065 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4066 * from page cache lookup which is in HPAGE_SIZE units.
4068 address = address & huge_page_mask(h);
4069 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4071 mapping = vma->vm_file->f_mapping;
4074 * Take the mapping lock for the duration of the table walk. As
4075 * this mapping should be shared between all the VMAs,
4076 * __unmap_hugepage_range() is called as the lock is already held
4078 i_mmap_lock_write(mapping);
4079 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4080 /* Do not unmap the current VMA */
4081 if (iter_vma == vma)
4085 * Shared VMAs have their own reserves and do not affect
4086 * MAP_PRIVATE accounting but it is possible that a shared
4087 * VMA is using the same page so check and skip such VMAs.
4089 if (iter_vma->vm_flags & VM_MAYSHARE)
4093 * Unmap the page from other VMAs without their own reserves.
4094 * They get marked to be SIGKILLed if they fault in these
4095 * areas. This is because a future no-page fault on this VMA
4096 * could insert a zeroed page instead of the data existing
4097 * from the time of fork. This would look like data corruption
4099 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4100 unmap_hugepage_range(iter_vma, address,
4101 address + huge_page_size(h), page);
4103 i_mmap_unlock_write(mapping);
4107 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4108 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4109 * cannot race with other handlers or page migration.
4110 * Keep the pte_same checks anyway to make transition from the mutex easier.
4112 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4113 unsigned long address, pte_t *ptep,
4114 struct page *pagecache_page, spinlock_t *ptl)
4117 struct hstate *h = hstate_vma(vma);
4118 struct page *old_page, *new_page;
4119 int outside_reserve = 0;
4121 unsigned long haddr = address & huge_page_mask(h);
4122 struct mmu_notifier_range range;
4124 pte = huge_ptep_get(ptep);
4125 old_page = pte_page(pte);
4128 /* If no-one else is actually using this page, avoid the copy
4129 * and just make the page writable */
4130 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4131 page_move_anon_rmap(old_page, vma);
4132 set_huge_ptep_writable(vma, haddr, ptep);
4137 * If the process that created a MAP_PRIVATE mapping is about to
4138 * perform a COW due to a shared page count, attempt to satisfy
4139 * the allocation without using the existing reserves. The pagecache
4140 * page is used to determine if the reserve at this address was
4141 * consumed or not. If reserves were used, a partial faulted mapping
4142 * at the time of fork() could consume its reserves on COW instead
4143 * of the full address range.
4145 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4146 old_page != pagecache_page)
4147 outside_reserve = 1;
4152 * Drop page table lock as buddy allocator may be called. It will
4153 * be acquired again before returning to the caller, as expected.
4156 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4158 if (IS_ERR(new_page)) {
4160 * If a process owning a MAP_PRIVATE mapping fails to COW,
4161 * it is due to references held by a child and an insufficient
4162 * huge page pool. To guarantee the original mappers
4163 * reliability, unmap the page from child processes. The child
4164 * may get SIGKILLed if it later faults.
4166 if (outside_reserve) {
4167 struct address_space *mapping = vma->vm_file->f_mapping;
4172 BUG_ON(huge_pte_none(pte));
4174 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4175 * unmapping. unmapping needs to hold i_mmap_rwsem
4176 * in write mode. Dropping i_mmap_rwsem in read mode
4177 * here is OK as COW mappings do not interact with
4180 * Reacquire both after unmap operation.
4182 idx = vma_hugecache_offset(h, vma, haddr);
4183 hash = hugetlb_fault_mutex_hash(mapping, idx);
4184 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4185 i_mmap_unlock_read(mapping);
4187 unmap_ref_private(mm, vma, old_page, haddr);
4189 i_mmap_lock_read(mapping);
4190 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4192 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4194 pte_same(huge_ptep_get(ptep), pte)))
4195 goto retry_avoidcopy;
4197 * race occurs while re-acquiring page table
4198 * lock, and our job is done.
4203 ret = vmf_error(PTR_ERR(new_page));
4204 goto out_release_old;
4208 * When the original hugepage is shared one, it does not have
4209 * anon_vma prepared.
4211 if (unlikely(anon_vma_prepare(vma))) {
4213 goto out_release_all;
4216 copy_user_huge_page(new_page, old_page, address, vma,
4217 pages_per_huge_page(h));
4218 __SetPageUptodate(new_page);
4220 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4221 haddr + huge_page_size(h));
4222 mmu_notifier_invalidate_range_start(&range);
4225 * Retake the page table lock to check for racing updates
4226 * before the page tables are altered
4229 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4230 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4231 ClearHPageRestoreReserve(new_page);
4234 huge_ptep_clear_flush(vma, haddr, ptep);
4235 mmu_notifier_invalidate_range(mm, range.start, range.end);
4236 set_huge_pte_at(mm, haddr, ptep,
4237 make_huge_pte(vma, new_page, 1));
4238 page_remove_rmap(old_page, true);
4239 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4240 SetHPageMigratable(new_page);
4241 /* Make the old page be freed below */
4242 new_page = old_page;
4245 mmu_notifier_invalidate_range_end(&range);
4247 restore_reserve_on_error(h, vma, haddr, new_page);
4252 spin_lock(ptl); /* Caller expects lock to be held */
4256 /* Return the pagecache page at a given address within a VMA */
4257 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4258 struct vm_area_struct *vma, unsigned long address)
4260 struct address_space *mapping;
4263 mapping = vma->vm_file->f_mapping;
4264 idx = vma_hugecache_offset(h, vma, address);
4266 return find_lock_page(mapping, idx);
4270 * Return whether there is a pagecache page to back given address within VMA.
4271 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4273 static bool hugetlbfs_pagecache_present(struct hstate *h,
4274 struct vm_area_struct *vma, unsigned long address)
4276 struct address_space *mapping;
4280 mapping = vma->vm_file->f_mapping;
4281 idx = vma_hugecache_offset(h, vma, address);
4283 page = find_get_page(mapping, idx);
4286 return page != NULL;
4289 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4292 struct inode *inode = mapping->host;
4293 struct hstate *h = hstate_inode(inode);
4294 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4298 ClearHPageRestoreReserve(page);
4301 * set page dirty so that it will not be removed from cache/file
4302 * by non-hugetlbfs specific code paths.
4304 set_page_dirty(page);
4306 spin_lock(&inode->i_lock);
4307 inode->i_blocks += blocks_per_huge_page(h);
4308 spin_unlock(&inode->i_lock);
4312 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4313 struct vm_area_struct *vma,
4314 struct address_space *mapping, pgoff_t idx,
4315 unsigned long address, pte_t *ptep, unsigned int flags)
4317 struct hstate *h = hstate_vma(vma);
4318 vm_fault_t ret = VM_FAULT_SIGBUS;
4324 unsigned long haddr = address & huge_page_mask(h);
4325 bool new_page = false;
4328 * Currently, we are forced to kill the process in the event the
4329 * original mapper has unmapped pages from the child due to a failed
4330 * COW. Warn that such a situation has occurred as it may not be obvious
4332 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4333 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4339 * We can not race with truncation due to holding i_mmap_rwsem.
4340 * i_size is modified when holding i_mmap_rwsem, so check here
4341 * once for faults beyond end of file.
4343 size = i_size_read(mapping->host) >> huge_page_shift(h);
4348 page = find_lock_page(mapping, idx);
4351 * Check for page in userfault range
4353 if (userfaultfd_missing(vma)) {
4355 struct vm_fault vmf = {
4360 * Hard to debug if it ends up being
4361 * used by a callee that assumes
4362 * something about the other
4363 * uninitialized fields... same as in
4369 * hugetlb_fault_mutex and i_mmap_rwsem must be
4370 * dropped before handling userfault. Reacquire
4371 * after handling fault to make calling code simpler.
4373 hash = hugetlb_fault_mutex_hash(mapping, idx);
4374 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4375 i_mmap_unlock_read(mapping);
4376 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4377 i_mmap_lock_read(mapping);
4378 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4382 page = alloc_huge_page(vma, haddr, 0);
4385 * Returning error will result in faulting task being
4386 * sent SIGBUS. The hugetlb fault mutex prevents two
4387 * tasks from racing to fault in the same page which
4388 * could result in false unable to allocate errors.
4389 * Page migration does not take the fault mutex, but
4390 * does a clear then write of pte's under page table
4391 * lock. Page fault code could race with migration,
4392 * notice the clear pte and try to allocate a page
4393 * here. Before returning error, get ptl and make
4394 * sure there really is no pte entry.
4396 ptl = huge_pte_lock(h, mm, ptep);
4398 if (huge_pte_none(huge_ptep_get(ptep)))
4399 ret = vmf_error(PTR_ERR(page));
4403 clear_huge_page(page, address, pages_per_huge_page(h));
4404 __SetPageUptodate(page);
4407 if (vma->vm_flags & VM_MAYSHARE) {
4408 int err = huge_add_to_page_cache(page, mapping, idx);
4417 if (unlikely(anon_vma_prepare(vma))) {
4419 goto backout_unlocked;
4425 * If memory error occurs between mmap() and fault, some process
4426 * don't have hwpoisoned swap entry for errored virtual address.
4427 * So we need to block hugepage fault by PG_hwpoison bit check.
4429 if (unlikely(PageHWPoison(page))) {
4430 ret = VM_FAULT_HWPOISON_LARGE |
4431 VM_FAULT_SET_HINDEX(hstate_index(h));
4432 goto backout_unlocked;
4437 * If we are going to COW a private mapping later, we examine the
4438 * pending reservations for this page now. This will ensure that
4439 * any allocations necessary to record that reservation occur outside
4442 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4443 if (vma_needs_reservation(h, vma, haddr) < 0) {
4445 goto backout_unlocked;
4447 /* Just decrements count, does not deallocate */
4448 vma_end_reservation(h, vma, haddr);
4451 ptl = huge_pte_lock(h, mm, ptep);
4453 if (!huge_pte_none(huge_ptep_get(ptep)))
4457 ClearHPageRestoreReserve(page);
4458 hugepage_add_new_anon_rmap(page, vma, haddr);
4460 page_dup_rmap(page, true);
4461 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4462 && (vma->vm_flags & VM_SHARED)));
4463 set_huge_pte_at(mm, haddr, ptep, new_pte);
4465 hugetlb_count_add(pages_per_huge_page(h), mm);
4466 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4467 /* Optimization, do the COW without a second fault */
4468 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4474 * Only set HPageMigratable in newly allocated pages. Existing pages
4475 * found in the pagecache may not have HPageMigratableset if they have
4476 * been isolated for migration.
4479 SetHPageMigratable(page);
4489 restore_reserve_on_error(h, vma, haddr, page);
4495 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4497 unsigned long key[2];
4500 key[0] = (unsigned long) mapping;
4503 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4505 return hash & (num_fault_mutexes - 1);
4509 * For uniprocessor systems we always use a single mutex, so just
4510 * return 0 and avoid the hashing overhead.
4512 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4518 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4519 unsigned long address, unsigned int flags)
4526 struct page *page = NULL;
4527 struct page *pagecache_page = NULL;
4528 struct hstate *h = hstate_vma(vma);
4529 struct address_space *mapping;
4530 int need_wait_lock = 0;
4531 unsigned long haddr = address & huge_page_mask(h);
4533 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4536 * Since we hold no locks, ptep could be stale. That is
4537 * OK as we are only making decisions based on content and
4538 * not actually modifying content here.
4540 entry = huge_ptep_get(ptep);
4541 if (unlikely(is_hugetlb_entry_migration(entry))) {
4542 migration_entry_wait_huge(vma, mm, ptep);
4544 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4545 return VM_FAULT_HWPOISON_LARGE |
4546 VM_FAULT_SET_HINDEX(hstate_index(h));
4550 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4551 * until finished with ptep. This serves two purposes:
4552 * 1) It prevents huge_pmd_unshare from being called elsewhere
4553 * and making the ptep no longer valid.
4554 * 2) It synchronizes us with i_size modifications during truncation.
4556 * ptep could have already be assigned via huge_pte_offset. That
4557 * is OK, as huge_pte_alloc will return the same value unless
4558 * something has changed.
4560 mapping = vma->vm_file->f_mapping;
4561 i_mmap_lock_read(mapping);
4562 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4564 i_mmap_unlock_read(mapping);
4565 return VM_FAULT_OOM;
4569 * Serialize hugepage allocation and instantiation, so that we don't
4570 * get spurious allocation failures if two CPUs race to instantiate
4571 * the same page in the page cache.
4573 idx = vma_hugecache_offset(h, vma, haddr);
4574 hash = hugetlb_fault_mutex_hash(mapping, idx);
4575 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4577 entry = huge_ptep_get(ptep);
4578 if (huge_pte_none(entry)) {
4579 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4586 * entry could be a migration/hwpoison entry at this point, so this
4587 * check prevents the kernel from going below assuming that we have
4588 * an active hugepage in pagecache. This goto expects the 2nd page
4589 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4590 * properly handle it.
4592 if (!pte_present(entry))
4596 * If we are going to COW the mapping later, we examine the pending
4597 * reservations for this page now. This will ensure that any
4598 * allocations necessary to record that reservation occur outside the
4599 * spinlock. For private mappings, we also lookup the pagecache
4600 * page now as it is used to determine if a reservation has been
4603 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4604 if (vma_needs_reservation(h, vma, haddr) < 0) {
4608 /* Just decrements count, does not deallocate */
4609 vma_end_reservation(h, vma, haddr);
4611 if (!(vma->vm_flags & VM_MAYSHARE))
4612 pagecache_page = hugetlbfs_pagecache_page(h,
4616 ptl = huge_pte_lock(h, mm, ptep);
4618 /* Check for a racing update before calling hugetlb_cow */
4619 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4623 * hugetlb_cow() requires page locks of pte_page(entry) and
4624 * pagecache_page, so here we need take the former one
4625 * when page != pagecache_page or !pagecache_page.
4627 page = pte_page(entry);
4628 if (page != pagecache_page)
4629 if (!trylock_page(page)) {
4636 if (flags & FAULT_FLAG_WRITE) {
4637 if (!huge_pte_write(entry)) {
4638 ret = hugetlb_cow(mm, vma, address, ptep,
4639 pagecache_page, ptl);
4642 entry = huge_pte_mkdirty(entry);
4644 entry = pte_mkyoung(entry);
4645 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4646 flags & FAULT_FLAG_WRITE))
4647 update_mmu_cache(vma, haddr, ptep);
4649 if (page != pagecache_page)
4655 if (pagecache_page) {
4656 unlock_page(pagecache_page);
4657 put_page(pagecache_page);
4660 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4661 i_mmap_unlock_read(mapping);
4663 * Generally it's safe to hold refcount during waiting page lock. But
4664 * here we just wait to defer the next page fault to avoid busy loop and
4665 * the page is not used after unlocked before returning from the current
4666 * page fault. So we are safe from accessing freed page, even if we wait
4667 * here without taking refcount.
4670 wait_on_page_locked(page);
4675 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4676 * modifications for huge pages.
4678 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4680 struct vm_area_struct *dst_vma,
4681 unsigned long dst_addr,
4682 unsigned long src_addr,
4683 struct page **pagep)
4685 struct address_space *mapping;
4688 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4689 struct hstate *h = hstate_vma(dst_vma);
4697 page = alloc_huge_page(dst_vma, dst_addr, 0);
4701 ret = copy_huge_page_from_user(page,
4702 (const void __user *) src_addr,
4703 pages_per_huge_page(h), false);
4705 /* fallback to copy_from_user outside mmap_lock */
4706 if (unlikely(ret)) {
4709 /* don't free the page */
4718 * The memory barrier inside __SetPageUptodate makes sure that
4719 * preceding stores to the page contents become visible before
4720 * the set_pte_at() write.
4722 __SetPageUptodate(page);
4724 mapping = dst_vma->vm_file->f_mapping;
4725 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4728 * If shared, add to page cache
4731 size = i_size_read(mapping->host) >> huge_page_shift(h);
4734 goto out_release_nounlock;
4737 * Serialization between remove_inode_hugepages() and
4738 * huge_add_to_page_cache() below happens through the
4739 * hugetlb_fault_mutex_table that here must be hold by
4742 ret = huge_add_to_page_cache(page, mapping, idx);
4744 goto out_release_nounlock;
4747 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4751 * Recheck the i_size after holding PT lock to make sure not
4752 * to leave any page mapped (as page_mapped()) beyond the end
4753 * of the i_size (remove_inode_hugepages() is strict about
4754 * enforcing that). If we bail out here, we'll also leave a
4755 * page in the radix tree in the vm_shared case beyond the end
4756 * of the i_size, but remove_inode_hugepages() will take care
4757 * of it as soon as we drop the hugetlb_fault_mutex_table.
4759 size = i_size_read(mapping->host) >> huge_page_shift(h);
4762 goto out_release_unlock;
4765 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4766 goto out_release_unlock;
4769 page_dup_rmap(page, true);
4771 ClearHPageRestoreReserve(page);
4772 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4775 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4776 if (dst_vma->vm_flags & VM_WRITE)
4777 _dst_pte = huge_pte_mkdirty(_dst_pte);
4778 _dst_pte = pte_mkyoung(_dst_pte);
4780 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4782 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4783 dst_vma->vm_flags & VM_WRITE);
4784 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4786 /* No need to invalidate - it was non-present before */
4787 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4790 SetHPageMigratable(page);
4800 out_release_nounlock:
4805 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4806 int refs, struct page **pages,
4807 struct vm_area_struct **vmas)
4811 for (nr = 0; nr < refs; nr++) {
4813 pages[nr] = mem_map_offset(page, nr);
4819 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4820 struct page **pages, struct vm_area_struct **vmas,
4821 unsigned long *position, unsigned long *nr_pages,
4822 long i, unsigned int flags, int *locked)
4824 unsigned long pfn_offset;
4825 unsigned long vaddr = *position;
4826 unsigned long remainder = *nr_pages;
4827 struct hstate *h = hstate_vma(vma);
4828 int err = -EFAULT, refs;
4830 while (vaddr < vma->vm_end && remainder) {
4832 spinlock_t *ptl = NULL;
4837 * If we have a pending SIGKILL, don't keep faulting pages and
4838 * potentially allocating memory.
4840 if (fatal_signal_pending(current)) {
4846 * Some archs (sparc64, sh*) have multiple pte_ts to
4847 * each hugepage. We have to make sure we get the
4848 * first, for the page indexing below to work.
4850 * Note that page table lock is not held when pte is null.
4852 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4855 ptl = huge_pte_lock(h, mm, pte);
4856 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4859 * When coredumping, it suits get_dump_page if we just return
4860 * an error where there's an empty slot with no huge pagecache
4861 * to back it. This way, we avoid allocating a hugepage, and
4862 * the sparse dumpfile avoids allocating disk blocks, but its
4863 * huge holes still show up with zeroes where they need to be.
4865 if (absent && (flags & FOLL_DUMP) &&
4866 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4874 * We need call hugetlb_fault for both hugepages under migration
4875 * (in which case hugetlb_fault waits for the migration,) and
4876 * hwpoisoned hugepages (in which case we need to prevent the
4877 * caller from accessing to them.) In order to do this, we use
4878 * here is_swap_pte instead of is_hugetlb_entry_migration and
4879 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4880 * both cases, and because we can't follow correct pages
4881 * directly from any kind of swap entries.
4883 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4884 ((flags & FOLL_WRITE) &&
4885 !huge_pte_write(huge_ptep_get(pte)))) {
4887 unsigned int fault_flags = 0;
4891 if (flags & FOLL_WRITE)
4892 fault_flags |= FAULT_FLAG_WRITE;
4894 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4895 FAULT_FLAG_KILLABLE;
4896 if (flags & FOLL_NOWAIT)
4897 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4898 FAULT_FLAG_RETRY_NOWAIT;
4899 if (flags & FOLL_TRIED) {
4901 * Note: FAULT_FLAG_ALLOW_RETRY and
4902 * FAULT_FLAG_TRIED can co-exist
4904 fault_flags |= FAULT_FLAG_TRIED;
4906 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4907 if (ret & VM_FAULT_ERROR) {
4908 err = vm_fault_to_errno(ret, flags);
4912 if (ret & VM_FAULT_RETRY) {
4914 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4918 * VM_FAULT_RETRY must not return an
4919 * error, it will return zero
4922 * No need to update "position" as the
4923 * caller will not check it after
4924 * *nr_pages is set to 0.
4931 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4932 page = pte_page(huge_ptep_get(pte));
4935 * If subpage information not requested, update counters
4936 * and skip the same_page loop below.
4938 if (!pages && !vmas && !pfn_offset &&
4939 (vaddr + huge_page_size(h) < vma->vm_end) &&
4940 (remainder >= pages_per_huge_page(h))) {
4941 vaddr += huge_page_size(h);
4942 remainder -= pages_per_huge_page(h);
4943 i += pages_per_huge_page(h);
4948 refs = min3(pages_per_huge_page(h) - pfn_offset,
4949 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4952 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4954 likely(pages) ? pages + i : NULL,
4955 vmas ? vmas + i : NULL);
4959 * try_grab_compound_head() should always succeed here,
4960 * because: a) we hold the ptl lock, and b) we've just
4961 * checked that the huge page is present in the page
4962 * tables. If the huge page is present, then the tail
4963 * pages must also be present. The ptl prevents the
4964 * head page and tail pages from being rearranged in
4965 * any way. So this page must be available at this
4966 * point, unless the page refcount overflowed:
4968 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
4978 vaddr += (refs << PAGE_SHIFT);
4984 *nr_pages = remainder;
4986 * setting position is actually required only if remainder is
4987 * not zero but it's faster not to add a "if (remainder)"
4995 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4996 unsigned long address, unsigned long end, pgprot_t newprot)
4998 struct mm_struct *mm = vma->vm_mm;
4999 unsigned long start = address;
5002 struct hstate *h = hstate_vma(vma);
5003 unsigned long pages = 0;
5004 bool shared_pmd = false;
5005 struct mmu_notifier_range range;
5008 * In the case of shared PMDs, the area to flush could be beyond
5009 * start/end. Set range.start/range.end to cover the maximum possible
5010 * range if PMD sharing is possible.
5012 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5013 0, vma, mm, start, end);
5014 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5016 BUG_ON(address >= end);
5017 flush_cache_range(vma, range.start, range.end);
5019 mmu_notifier_invalidate_range_start(&range);
5020 i_mmap_lock_write(vma->vm_file->f_mapping);
5021 for (; address < end; address += huge_page_size(h)) {
5023 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5026 ptl = huge_pte_lock(h, mm, ptep);
5027 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5033 pte = huge_ptep_get(ptep);
5034 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5038 if (unlikely(is_hugetlb_entry_migration(pte))) {
5039 swp_entry_t entry = pte_to_swp_entry(pte);
5041 if (is_write_migration_entry(entry)) {
5044 make_migration_entry_read(&entry);
5045 newpte = swp_entry_to_pte(entry);
5046 set_huge_swap_pte_at(mm, address, ptep,
5047 newpte, huge_page_size(h));
5053 if (!huge_pte_none(pte)) {
5056 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5057 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5058 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5059 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5065 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5066 * may have cleared our pud entry and done put_page on the page table:
5067 * once we release i_mmap_rwsem, another task can do the final put_page
5068 * and that page table be reused and filled with junk. If we actually
5069 * did unshare a page of pmds, flush the range corresponding to the pud.
5072 flush_hugetlb_tlb_range(vma, range.start, range.end);
5074 flush_hugetlb_tlb_range(vma, start, end);
5076 * No need to call mmu_notifier_invalidate_range() we are downgrading
5077 * page table protection not changing it to point to a new page.
5079 * See Documentation/vm/mmu_notifier.rst
5081 i_mmap_unlock_write(vma->vm_file->f_mapping);
5082 mmu_notifier_invalidate_range_end(&range);
5084 return pages << h->order;
5087 /* Return true if reservation was successful, false otherwise. */
5088 bool hugetlb_reserve_pages(struct inode *inode,
5090 struct vm_area_struct *vma,
5091 vm_flags_t vm_flags)
5094 struct hstate *h = hstate_inode(inode);
5095 struct hugepage_subpool *spool = subpool_inode(inode);
5096 struct resv_map *resv_map;
5097 struct hugetlb_cgroup *h_cg = NULL;
5098 long gbl_reserve, regions_needed = 0;
5100 /* This should never happen */
5102 VM_WARN(1, "%s called with a negative range\n", __func__);
5107 * Only apply hugepage reservation if asked. At fault time, an
5108 * attempt will be made for VM_NORESERVE to allocate a page
5109 * without using reserves
5111 if (vm_flags & VM_NORESERVE)
5115 * Shared mappings base their reservation on the number of pages that
5116 * are already allocated on behalf of the file. Private mappings need
5117 * to reserve the full area even if read-only as mprotect() may be
5118 * called to make the mapping read-write. Assume !vma is a shm mapping
5120 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5122 * resv_map can not be NULL as hugetlb_reserve_pages is only
5123 * called for inodes for which resv_maps were created (see
5124 * hugetlbfs_get_inode).
5126 resv_map = inode_resv_map(inode);
5128 chg = region_chg(resv_map, from, to, ®ions_needed);
5131 /* Private mapping. */
5132 resv_map = resv_map_alloc();
5138 set_vma_resv_map(vma, resv_map);
5139 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5145 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5146 chg * pages_per_huge_page(h), &h_cg) < 0)
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)
5163 goto out_uncharge_cgroup;
5166 * Check enough hugepages are available for the reservation.
5167 * Hand the pages back to the subpool if there are not
5169 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5173 * Account for the reservations made. Shared mappings record regions
5174 * that have reservations as they are shared by multiple VMAs.
5175 * When the last VMA disappears, the region map says how much
5176 * the reservation was and the page cache tells how much of
5177 * the reservation was consumed. Private mappings are per-VMA and
5178 * only the consumed reservations are tracked. When the VMA
5179 * disappears, the original reservation is the VMA size and the
5180 * consumed reservations are stored in the map. Hence, nothing
5181 * else has to be done for private mappings here
5183 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5184 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5186 if (unlikely(add < 0)) {
5187 hugetlb_acct_memory(h, -gbl_reserve);
5189 } else if (unlikely(chg > add)) {
5191 * pages in this range were added to the reserve
5192 * map between region_chg and region_add. This
5193 * indicates a race with alloc_huge_page. Adjust
5194 * the subpool and reserve counts modified above
5195 * based on the difference.
5200 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5201 * reference to h_cg->css. See comment below for detail.
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);
5212 * The file_regions will hold their own reference to
5213 * h_cg->css. So we should release the reference held
5214 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5217 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5223 /* put back original number of pages, chg */
5224 (void)hugepage_subpool_put_pages(spool, chg);
5225 out_uncharge_cgroup:
5226 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5227 chg * pages_per_huge_page(h), h_cg);
5229 if (!vma || vma->vm_flags & VM_MAYSHARE)
5230 /* Only call region_abort if the region_chg succeeded but the
5231 * region_add failed or didn't run.
5233 if (chg >= 0 && add < 0)
5234 region_abort(resv_map, from, to, regions_needed);
5235 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5236 kref_put(&resv_map->refs, resv_map_release);
5240 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5243 struct hstate *h = hstate_inode(inode);
5244 struct resv_map *resv_map = inode_resv_map(inode);
5246 struct hugepage_subpool *spool = subpool_inode(inode);
5250 * Since this routine can be called in the evict inode path for all
5251 * hugetlbfs inodes, resv_map could be NULL.
5254 chg = region_del(resv_map, start, end);
5256 * region_del() can fail in the rare case where a region
5257 * must be split and another region descriptor can not be
5258 * allocated. If end == LONG_MAX, it will not fail.
5264 spin_lock(&inode->i_lock);
5265 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5266 spin_unlock(&inode->i_lock);
5269 * If the subpool has a minimum size, the number of global
5270 * reservations to be released may be adjusted.
5272 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5273 hugetlb_acct_memory(h, -gbl_reserve);
5278 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5279 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5280 struct vm_area_struct *vma,
5281 unsigned long addr, pgoff_t idx)
5283 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5285 unsigned long sbase = saddr & PUD_MASK;
5286 unsigned long s_end = sbase + PUD_SIZE;
5288 /* Allow segments to share if only one is marked locked */
5289 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5290 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5293 * match the virtual addresses, permission and the alignment of the
5296 if (pmd_index(addr) != pmd_index(saddr) ||
5297 vm_flags != svm_flags ||
5298 !range_in_vma(svma, sbase, s_end))
5304 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5306 unsigned long base = addr & PUD_MASK;
5307 unsigned long end = base + PUD_SIZE;
5310 * check on proper vm_flags and page table alignment
5312 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5317 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5319 #ifdef CONFIG_USERFAULTFD
5320 if (uffd_disable_huge_pmd_share(vma))
5323 return vma_shareable(vma, addr);
5327 * Determine if start,end range within vma could be mapped by shared pmd.
5328 * If yes, adjust start and end to cover range associated with possible
5329 * shared pmd mappings.
5331 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5332 unsigned long *start, unsigned long *end)
5334 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5335 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5338 * vma need span at least one aligned PUD size and the start,end range
5339 * must at least partialy within it.
5341 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5342 (*end <= v_start) || (*start >= v_end))
5345 /* Extend the range to be PUD aligned for a worst case scenario */
5346 if (*start > v_start)
5347 *start = ALIGN_DOWN(*start, PUD_SIZE);
5350 *end = ALIGN(*end, PUD_SIZE);
5354 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5355 * and returns the corresponding pte. While this is not necessary for the
5356 * !shared pmd case because we can allocate the pmd later as well, it makes the
5357 * code much cleaner.
5359 * This routine must be called with i_mmap_rwsem held in at least read mode if
5360 * sharing is possible. For hugetlbfs, this prevents removal of any page
5361 * table entries associated with the address space. This is important as we
5362 * are setting up sharing based on existing page table entries (mappings).
5364 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5365 * huge_pte_alloc know that sharing is not possible and do not take
5366 * i_mmap_rwsem as a performance optimization. This is handled by the
5367 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5368 * only required for subsequent processing.
5370 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5371 unsigned long addr, pud_t *pud)
5373 struct address_space *mapping = vma->vm_file->f_mapping;
5374 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5376 struct vm_area_struct *svma;
5377 unsigned long saddr;
5382 i_mmap_assert_locked(mapping);
5383 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5387 saddr = page_table_shareable(svma, vma, addr, idx);
5389 spte = huge_pte_offset(svma->vm_mm, saddr,
5390 vma_mmu_pagesize(svma));
5392 get_page(virt_to_page(spte));
5401 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5402 if (pud_none(*pud)) {
5403 pud_populate(mm, pud,
5404 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5407 put_page(virt_to_page(spte));
5411 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5416 * unmap huge page backed by shared pte.
5418 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5419 * indicated by page_count > 1, unmap is achieved by clearing pud and
5420 * decrementing the ref count. If count == 1, the pte page is not shared.
5422 * Called with page table lock held and i_mmap_rwsem held in write mode.
5424 * returns: 1 successfully unmapped a shared pte page
5425 * 0 the underlying pte page is not shared, or it is the last user
5427 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5428 unsigned long *addr, pte_t *ptep)
5430 pgd_t *pgd = pgd_offset(mm, *addr);
5431 p4d_t *p4d = p4d_offset(pgd, *addr);
5432 pud_t *pud = pud_offset(p4d, *addr);
5434 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5435 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5436 if (page_count(virt_to_page(ptep)) == 1)
5440 put_page(virt_to_page(ptep));
5442 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5446 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5447 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5448 unsigned long addr, pud_t *pud)
5453 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5454 unsigned long *addr, pte_t *ptep)
5459 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5460 unsigned long *start, unsigned long *end)
5464 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5468 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5470 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5471 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5472 unsigned long addr, unsigned long sz)
5479 pgd = pgd_offset(mm, addr);
5480 p4d = p4d_alloc(mm, pgd, addr);
5483 pud = pud_alloc(mm, p4d, addr);
5485 if (sz == PUD_SIZE) {
5488 BUG_ON(sz != PMD_SIZE);
5489 if (want_pmd_share(vma, addr) && pud_none(*pud))
5490 pte = huge_pmd_share(mm, vma, addr, pud);
5492 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5495 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5501 * huge_pte_offset() - Walk the page table to resolve the hugepage
5502 * entry at address @addr
5504 * Return: Pointer to page table entry (PUD or PMD) for
5505 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5506 * size @sz doesn't match the hugepage size at this level of the page
5509 pte_t *huge_pte_offset(struct mm_struct *mm,
5510 unsigned long addr, unsigned long sz)
5517 pgd = pgd_offset(mm, addr);
5518 if (!pgd_present(*pgd))
5520 p4d = p4d_offset(pgd, addr);
5521 if (!p4d_present(*p4d))
5524 pud = pud_offset(p4d, addr);
5526 /* must be pud huge, non-present or none */
5527 return (pte_t *)pud;
5528 if (!pud_present(*pud))
5530 /* must have a valid entry and size to go further */
5532 pmd = pmd_offset(pud, addr);
5533 /* must be pmd huge, non-present or none */
5534 return (pte_t *)pmd;
5537 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5540 * These functions are overwritable if your architecture needs its own
5543 struct page * __weak
5544 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5547 return ERR_PTR(-EINVAL);
5550 struct page * __weak
5551 follow_huge_pd(struct vm_area_struct *vma,
5552 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5554 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5558 struct page * __weak
5559 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5560 pmd_t *pmd, int flags)
5562 struct page *page = NULL;
5566 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5567 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5568 (FOLL_PIN | FOLL_GET)))
5572 ptl = pmd_lockptr(mm, pmd);
5575 * make sure that the address range covered by this pmd is not
5576 * unmapped from other threads.
5578 if (!pmd_huge(*pmd))
5580 pte = huge_ptep_get((pte_t *)pmd);
5581 if (pte_present(pte)) {
5582 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5584 * try_grab_page() should always succeed here, because: a) we
5585 * hold the pmd (ptl) lock, and b) we've just checked that the
5586 * huge pmd (head) page is present in the page tables. The ptl
5587 * prevents the head page and tail pages from being rearranged
5588 * in any way. So this page must be available at this point,
5589 * unless the page refcount overflowed:
5591 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5596 if (is_hugetlb_entry_migration(pte)) {
5598 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5602 * hwpoisoned entry is treated as no_page_table in
5603 * follow_page_mask().
5611 struct page * __weak
5612 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5613 pud_t *pud, int flags)
5615 if (flags & (FOLL_GET | FOLL_PIN))
5618 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5621 struct page * __weak
5622 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5624 if (flags & (FOLL_GET | FOLL_PIN))
5627 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5630 bool isolate_huge_page(struct page *page, struct list_head *list)
5634 spin_lock(&hugetlb_lock);
5635 if (!PageHeadHuge(page) ||
5636 !HPageMigratable(page) ||
5637 !get_page_unless_zero(page)) {
5641 ClearHPageMigratable(page);
5642 list_move_tail(&page->lru, list);
5644 spin_unlock(&hugetlb_lock);
5648 void putback_active_hugepage(struct page *page)
5650 spin_lock(&hugetlb_lock);
5651 SetHPageMigratable(page);
5652 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5653 spin_unlock(&hugetlb_lock);
5657 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5659 struct hstate *h = page_hstate(oldpage);
5661 hugetlb_cgroup_migrate(oldpage, newpage);
5662 set_page_owner_migrate_reason(newpage, reason);
5665 * transfer temporary state of the new huge page. This is
5666 * reverse to other transitions because the newpage is going to
5667 * be final while the old one will be freed so it takes over
5668 * the temporary status.
5670 * Also note that we have to transfer the per-node surplus state
5671 * here as well otherwise the global surplus count will not match
5674 if (HPageTemporary(newpage)) {
5675 int old_nid = page_to_nid(oldpage);
5676 int new_nid = page_to_nid(newpage);
5678 SetHPageTemporary(oldpage);
5679 ClearHPageTemporary(newpage);
5682 * There is no need to transfer the per-node surplus state
5683 * when we do not cross the node.
5685 if (new_nid == old_nid)
5687 spin_lock(&hugetlb_lock);
5688 if (h->surplus_huge_pages_node[old_nid]) {
5689 h->surplus_huge_pages_node[old_nid]--;
5690 h->surplus_huge_pages_node[new_nid]++;
5692 spin_unlock(&hugetlb_lock);
5697 * This function will unconditionally remove all the shared pmd pgtable entries
5698 * within the specific vma for a hugetlbfs memory range.
5700 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5702 struct hstate *h = hstate_vma(vma);
5703 unsigned long sz = huge_page_size(h);
5704 struct mm_struct *mm = vma->vm_mm;
5705 struct mmu_notifier_range range;
5706 unsigned long address, start, end;
5710 if (!(vma->vm_flags & VM_MAYSHARE))
5713 start = ALIGN(vma->vm_start, PUD_SIZE);
5714 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5720 * No need to call adjust_range_if_pmd_sharing_possible(), because
5721 * we have already done the PUD_SIZE alignment.
5723 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5725 mmu_notifier_invalidate_range_start(&range);
5726 i_mmap_lock_write(vma->vm_file->f_mapping);
5727 for (address = start; address < end; address += PUD_SIZE) {
5728 unsigned long tmp = address;
5730 ptep = huge_pte_offset(mm, address, sz);
5733 ptl = huge_pte_lock(h, mm, ptep);
5734 /* We don't want 'address' to be changed */
5735 huge_pmd_unshare(mm, vma, &tmp, ptep);
5738 flush_hugetlb_tlb_range(vma, start, end);
5739 i_mmap_unlock_write(vma->vm_file->f_mapping);
5741 * No need to call mmu_notifier_invalidate_range(), see
5742 * Documentation/vm/mmu_notifier.rst.
5744 mmu_notifier_invalidate_range_end(&range);
5748 static bool cma_reserve_called __initdata;
5750 static int __init cmdline_parse_hugetlb_cma(char *p)
5752 hugetlb_cma_size = memparse(p, &p);
5756 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5758 void __init hugetlb_cma_reserve(int order)
5760 unsigned long size, reserved, per_node;
5763 cma_reserve_called = true;
5765 if (!hugetlb_cma_size)
5768 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5769 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5770 (PAGE_SIZE << order) / SZ_1M);
5775 * If 3 GB area is requested on a machine with 4 numa nodes,
5776 * let's allocate 1 GB on first three nodes and ignore the last one.
5778 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5779 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5780 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5783 for_each_node_state(nid, N_ONLINE) {
5785 char name[CMA_MAX_NAME];
5787 size = min(per_node, hugetlb_cma_size - reserved);
5788 size = round_up(size, PAGE_SIZE << order);
5790 snprintf(name, sizeof(name), "hugetlb%d", nid);
5791 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5793 &hugetlb_cma[nid], nid);
5795 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5801 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5804 if (reserved >= hugetlb_cma_size)
5809 void __init hugetlb_cma_check(void)
5811 if (!hugetlb_cma_size || cma_reserve_called)
5814 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5817 #endif /* CONFIG_CMA */