1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
36 #include <asm/pgalloc.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
45 #include "hugetlb_vmemmap.h"
47 int hugetlb_max_hstate __read_mostly;
48 unsigned int default_hstate_idx;
49 struct hstate hstates[HUGE_MAX_HSTATE];
52 static struct cma *hugetlb_cma[MAX_NUMNODES];
54 static unsigned long hugetlb_cma_size __initdata;
57 * Minimum page order among possible hugepage sizes, set to a proper value
60 static unsigned int minimum_order __read_mostly = UINT_MAX;
62 __initdata LIST_HEAD(huge_boot_pages);
64 /* for command line parsing */
65 static struct hstate * __initdata parsed_hstate;
66 static unsigned long __initdata default_hstate_max_huge_pages;
67 static bool __initdata parsed_valid_hugepagesz = true;
68 static bool __initdata parsed_default_hugepagesz;
71 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
74 DEFINE_SPINLOCK(hugetlb_lock);
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
80 static int num_fault_mutexes;
81 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
83 /* Forward declaration */
84 static int hugetlb_acct_memory(struct hstate *h, long delta);
85 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
86 unsigned long start, unsigned long end);
88 static inline bool subpool_is_free(struct hugepage_subpool *spool)
92 if (spool->max_hpages != -1)
93 return spool->used_hpages == 0;
94 if (spool->min_hpages != -1)
95 return spool->rsv_hpages == spool->min_hpages;
100 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
101 unsigned long irq_flags)
103 spin_unlock_irqrestore(&spool->lock, irq_flags);
105 /* If no pages are used, and no other handles to the subpool
106 * remain, give up any reservations based on minimum size and
107 * free the subpool */
108 if (subpool_is_free(spool)) {
109 if (spool->min_hpages != -1)
110 hugetlb_acct_memory(spool->hstate,
116 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
119 struct hugepage_subpool *spool;
121 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
125 spin_lock_init(&spool->lock);
127 spool->max_hpages = max_hpages;
129 spool->min_hpages = min_hpages;
131 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
135 spool->rsv_hpages = min_hpages;
140 void hugepage_put_subpool(struct hugepage_subpool *spool)
144 spin_lock_irqsave(&spool->lock, flags);
145 BUG_ON(!spool->count);
147 unlock_or_release_subpool(spool, flags);
151 * Subpool accounting for allocating and reserving pages.
152 * Return -ENOMEM if there are not enough resources to satisfy the
153 * request. Otherwise, return the number of pages by which the
154 * global pools must be adjusted (upward). The returned value may
155 * only be different than the passed value (delta) in the case where
156 * a subpool minimum size must be maintained.
158 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
166 spin_lock_irq(&spool->lock);
168 if (spool->max_hpages != -1) { /* maximum size accounting */
169 if ((spool->used_hpages + delta) <= spool->max_hpages)
170 spool->used_hpages += delta;
177 /* minimum size accounting */
178 if (spool->min_hpages != -1 && spool->rsv_hpages) {
179 if (delta > spool->rsv_hpages) {
181 * Asking for more reserves than those already taken on
182 * behalf of subpool. Return difference.
184 ret = delta - spool->rsv_hpages;
185 spool->rsv_hpages = 0;
187 ret = 0; /* reserves already accounted for */
188 spool->rsv_hpages -= delta;
193 spin_unlock_irq(&spool->lock);
198 * Subpool accounting for freeing and unreserving pages.
199 * Return the number of global page reservations that must be dropped.
200 * The return value may only be different than the passed value (delta)
201 * in the case where a subpool minimum size must be maintained.
203 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
212 spin_lock_irqsave(&spool->lock, flags);
214 if (spool->max_hpages != -1) /* maximum size accounting */
215 spool->used_hpages -= delta;
217 /* minimum size accounting */
218 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
219 if (spool->rsv_hpages + delta <= spool->min_hpages)
222 ret = spool->rsv_hpages + delta - spool->min_hpages;
224 spool->rsv_hpages += delta;
225 if (spool->rsv_hpages > spool->min_hpages)
226 spool->rsv_hpages = spool->min_hpages;
230 * If hugetlbfs_put_super couldn't free spool due to an outstanding
231 * quota reference, free it now.
233 unlock_or_release_subpool(spool, flags);
238 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
240 return HUGETLBFS_SB(inode->i_sb)->spool;
243 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
245 return subpool_inode(file_inode(vma->vm_file));
248 /* Helper that removes a struct file_region from the resv_map cache and returns
251 static struct file_region *
252 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
254 struct file_region *nrg = NULL;
256 VM_BUG_ON(resv->region_cache_count <= 0);
258 resv->region_cache_count--;
259 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
260 list_del(&nrg->link);
268 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
269 struct file_region *rg)
271 #ifdef CONFIG_CGROUP_HUGETLB
272 nrg->reservation_counter = rg->reservation_counter;
279 /* Helper that records hugetlb_cgroup uncharge info. */
280 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
282 struct resv_map *resv,
283 struct file_region *nrg)
285 #ifdef CONFIG_CGROUP_HUGETLB
287 nrg->reservation_counter =
288 &h_cg->rsvd_hugepage[hstate_index(h)];
289 nrg->css = &h_cg->css;
291 * The caller will hold exactly one h_cg->css reference for the
292 * whole contiguous reservation region. But this area might be
293 * scattered when there are already some file_regions reside in
294 * it. As a result, many file_regions may share only one css
295 * reference. In order to ensure that one file_region must hold
296 * exactly one h_cg->css reference, we should do css_get for
297 * each file_region and leave the reference held by caller
301 if (!resv->pages_per_hpage)
302 resv->pages_per_hpage = pages_per_huge_page(h);
303 /* pages_per_hpage should be the same for all entries in
306 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
308 nrg->reservation_counter = NULL;
314 static void put_uncharge_info(struct file_region *rg)
316 #ifdef CONFIG_CGROUP_HUGETLB
322 static bool has_same_uncharge_info(struct file_region *rg,
323 struct file_region *org)
325 #ifdef CONFIG_CGROUP_HUGETLB
327 rg->reservation_counter == org->reservation_counter &&
335 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
337 struct file_region *nrg = NULL, *prg = NULL;
339 prg = list_prev_entry(rg, link);
340 if (&prg->link != &resv->regions && prg->to == rg->from &&
341 has_same_uncharge_info(prg, rg)) {
345 put_uncharge_info(rg);
351 nrg = list_next_entry(rg, link);
352 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
353 has_same_uncharge_info(nrg, rg)) {
354 nrg->from = rg->from;
357 put_uncharge_info(rg);
363 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
364 long to, struct hstate *h, struct hugetlb_cgroup *cg,
365 long *regions_needed)
367 struct file_region *nrg;
369 if (!regions_needed) {
370 nrg = get_file_region_entry_from_cache(map, from, to);
371 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
372 list_add(&nrg->link, rg->link.prev);
373 coalesce_file_region(map, nrg);
375 *regions_needed += 1;
381 * Must be called with resv->lock held.
383 * Calling this with regions_needed != NULL will count the number of pages
384 * to be added but will not modify the linked list. And regions_needed will
385 * indicate the number of file_regions needed in the cache to carry out to add
386 * the regions for this range.
388 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
389 struct hugetlb_cgroup *h_cg,
390 struct hstate *h, long *regions_needed)
393 struct list_head *head = &resv->regions;
394 long last_accounted_offset = f;
395 struct file_region *rg = NULL, *trg = NULL;
400 /* In this loop, we essentially handle an entry for the range
401 * [last_accounted_offset, rg->from), at every iteration, with some
404 list_for_each_entry_safe(rg, trg, head, link) {
405 /* Skip irrelevant regions that start before our range. */
407 /* If this region ends after the last accounted offset,
408 * then we need to update last_accounted_offset.
410 if (rg->to > last_accounted_offset)
411 last_accounted_offset = rg->to;
415 /* When we find a region that starts beyond our range, we've
421 /* Add an entry for last_accounted_offset -> rg->from, and
422 * update last_accounted_offset.
424 if (rg->from > last_accounted_offset)
425 add += hugetlb_resv_map_add(resv, rg,
426 last_accounted_offset,
430 last_accounted_offset = rg->to;
433 /* Handle the case where our range extends beyond
434 * last_accounted_offset.
436 if (last_accounted_offset < t)
437 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
438 t, h, h_cg, regions_needed);
444 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
446 static int allocate_file_region_entries(struct resv_map *resv,
448 __must_hold(&resv->lock)
450 struct list_head allocated_regions;
451 int to_allocate = 0, i = 0;
452 struct file_region *trg = NULL, *rg = NULL;
454 VM_BUG_ON(regions_needed < 0);
456 INIT_LIST_HEAD(&allocated_regions);
459 * Check for sufficient descriptors in the cache to accommodate
460 * the number of in progress add operations plus regions_needed.
462 * This is a while loop because when we drop the lock, some other call
463 * to region_add or region_del may have consumed some region_entries,
464 * so we keep looping here until we finally have enough entries for
465 * (adds_in_progress + regions_needed).
467 while (resv->region_cache_count <
468 (resv->adds_in_progress + regions_needed)) {
469 to_allocate = resv->adds_in_progress + regions_needed -
470 resv->region_cache_count;
472 /* At this point, we should have enough entries in the cache
473 * for all the existing adds_in_progress. We should only be
474 * needing to allocate for regions_needed.
476 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
478 spin_unlock(&resv->lock);
479 for (i = 0; i < to_allocate; i++) {
480 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
483 list_add(&trg->link, &allocated_regions);
486 spin_lock(&resv->lock);
488 list_splice(&allocated_regions, &resv->region_cache);
489 resv->region_cache_count += to_allocate;
495 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
503 * Add the huge page range represented by [f, t) to the reserve
504 * map. Regions will be taken from the cache to fill in this range.
505 * Sufficient regions should exist in the cache due to the previous
506 * call to region_chg with the same range, but in some cases the cache will not
507 * have sufficient entries due to races with other code doing region_add or
508 * region_del. The extra needed entries will be allocated.
510 * regions_needed is the out value provided by a previous call to region_chg.
512 * Return the number of new huge pages added to the map. This number is greater
513 * than or equal to zero. If file_region entries needed to be allocated for
514 * this operation and we were not able to allocate, it returns -ENOMEM.
515 * region_add of regions of length 1 never allocate file_regions and cannot
516 * fail; region_chg will always allocate at least 1 entry and a region_add for
517 * 1 page will only require at most 1 entry.
519 static long region_add(struct resv_map *resv, long f, long t,
520 long in_regions_needed, struct hstate *h,
521 struct hugetlb_cgroup *h_cg)
523 long add = 0, actual_regions_needed = 0;
525 spin_lock(&resv->lock);
528 /* Count how many regions are actually needed to execute this add. */
529 add_reservation_in_range(resv, f, t, NULL, NULL,
530 &actual_regions_needed);
533 * Check for sufficient descriptors in the cache to accommodate
534 * this add operation. Note that actual_regions_needed may be greater
535 * than in_regions_needed, as the resv_map may have been modified since
536 * the region_chg call. In this case, we need to make sure that we
537 * allocate extra entries, such that we have enough for all the
538 * existing adds_in_progress, plus the excess needed for this
541 if (actual_regions_needed > in_regions_needed &&
542 resv->region_cache_count <
543 resv->adds_in_progress +
544 (actual_regions_needed - in_regions_needed)) {
545 /* region_add operation of range 1 should never need to
546 * allocate file_region entries.
548 VM_BUG_ON(t - f <= 1);
550 if (allocate_file_region_entries(
551 resv, actual_regions_needed - in_regions_needed)) {
558 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
560 resv->adds_in_progress -= in_regions_needed;
562 spin_unlock(&resv->lock);
567 * Examine the existing reserve map and determine how many
568 * huge pages in the specified range [f, t) are NOT currently
569 * represented. This routine is called before a subsequent
570 * call to region_add that will actually modify the reserve
571 * map to add the specified range [f, t). region_chg does
572 * not change the number of huge pages represented by the
573 * map. A number of new file_region structures is added to the cache as a
574 * placeholder, for the subsequent region_add call to use. At least 1
575 * file_region structure is added.
577 * out_regions_needed is the number of regions added to the
578 * resv->adds_in_progress. This value needs to be provided to a follow up call
579 * to region_add or region_abort for proper accounting.
581 * Returns the number of huge pages that need to be added to the existing
582 * reservation map for the range [f, t). This number is greater or equal to
583 * zero. -ENOMEM is returned if a new file_region structure or cache entry
584 * is needed and can not be allocated.
586 static long region_chg(struct resv_map *resv, long f, long t,
587 long *out_regions_needed)
591 spin_lock(&resv->lock);
593 /* Count how many hugepages in this range are NOT represented. */
594 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
597 if (*out_regions_needed == 0)
598 *out_regions_needed = 1;
600 if (allocate_file_region_entries(resv, *out_regions_needed))
603 resv->adds_in_progress += *out_regions_needed;
605 spin_unlock(&resv->lock);
610 * Abort the in progress add operation. The adds_in_progress field
611 * of the resv_map keeps track of the operations in progress between
612 * calls to region_chg and region_add. Operations are sometimes
613 * aborted after the call to region_chg. In such cases, region_abort
614 * is called to decrement the adds_in_progress counter. regions_needed
615 * is the value returned by the region_chg call, it is used to decrement
616 * the adds_in_progress counter.
618 * NOTE: The range arguments [f, t) are not needed or used in this
619 * routine. They are kept to make reading the calling code easier as
620 * arguments will match the associated region_chg call.
622 static void region_abort(struct resv_map *resv, long f, long t,
625 spin_lock(&resv->lock);
626 VM_BUG_ON(!resv->region_cache_count);
627 resv->adds_in_progress -= regions_needed;
628 spin_unlock(&resv->lock);
632 * Delete the specified range [f, t) from the reserve map. If the
633 * t parameter is LONG_MAX, this indicates that ALL regions after f
634 * should be deleted. Locate the regions which intersect [f, t)
635 * and either trim, delete or split the existing regions.
637 * Returns the number of huge pages deleted from the reserve map.
638 * In the normal case, the return value is zero or more. In the
639 * case where a region must be split, a new region descriptor must
640 * be allocated. If the allocation fails, -ENOMEM will be returned.
641 * NOTE: If the parameter t == LONG_MAX, then we will never split
642 * a region and possibly return -ENOMEM. Callers specifying
643 * t == LONG_MAX do not need to check for -ENOMEM error.
645 static long region_del(struct resv_map *resv, long f, long t)
647 struct list_head *head = &resv->regions;
648 struct file_region *rg, *trg;
649 struct file_region *nrg = NULL;
653 spin_lock(&resv->lock);
654 list_for_each_entry_safe(rg, trg, head, link) {
656 * Skip regions before the range to be deleted. file_region
657 * ranges are normally of the form [from, to). However, there
658 * may be a "placeholder" entry in the map which is of the form
659 * (from, to) with from == to. Check for placeholder entries
660 * at the beginning of the range to be deleted.
662 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
668 if (f > rg->from && t < rg->to) { /* Must split region */
670 * Check for an entry in the cache before dropping
671 * lock and attempting allocation.
674 resv->region_cache_count > resv->adds_in_progress) {
675 nrg = list_first_entry(&resv->region_cache,
678 list_del(&nrg->link);
679 resv->region_cache_count--;
683 spin_unlock(&resv->lock);
684 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
691 hugetlb_cgroup_uncharge_file_region(
692 resv, rg, t - f, false);
694 /* New entry for end of split region */
698 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
700 INIT_LIST_HEAD(&nrg->link);
702 /* Original entry is trimmed */
705 list_add(&nrg->link, &rg->link);
710 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
711 del += rg->to - rg->from;
712 hugetlb_cgroup_uncharge_file_region(resv, rg,
713 rg->to - rg->from, true);
719 if (f <= rg->from) { /* Trim beginning of region */
720 hugetlb_cgroup_uncharge_file_region(resv, rg,
721 t - rg->from, false);
725 } else { /* Trim end of region */
726 hugetlb_cgroup_uncharge_file_region(resv, rg,
734 spin_unlock(&resv->lock);
740 * A rare out of memory error was encountered which prevented removal of
741 * the reserve map region for a page. The huge page itself was free'ed
742 * and removed from the page cache. This routine will adjust the subpool
743 * usage count, and the global reserve count if needed. By incrementing
744 * these counts, the reserve map entry which could not be deleted will
745 * appear as a "reserved" entry instead of simply dangling with incorrect
748 void hugetlb_fix_reserve_counts(struct inode *inode)
750 struct hugepage_subpool *spool = subpool_inode(inode);
752 bool reserved = false;
754 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
755 if (rsv_adjust > 0) {
756 struct hstate *h = hstate_inode(inode);
758 if (!hugetlb_acct_memory(h, 1))
760 } else if (!rsv_adjust) {
765 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
769 * Count and return the number of huge pages in the reserve map
770 * that intersect with the range [f, t).
772 static long region_count(struct resv_map *resv, long f, long t)
774 struct list_head *head = &resv->regions;
775 struct file_region *rg;
778 spin_lock(&resv->lock);
779 /* Locate each segment we overlap with, and count that overlap. */
780 list_for_each_entry(rg, head, link) {
789 seg_from = max(rg->from, f);
790 seg_to = min(rg->to, t);
792 chg += seg_to - seg_from;
794 spin_unlock(&resv->lock);
800 * Convert the address within this vma to the page offset within
801 * the mapping, in pagecache page units; huge pages here.
803 static pgoff_t vma_hugecache_offset(struct hstate *h,
804 struct vm_area_struct *vma, unsigned long address)
806 return ((address - vma->vm_start) >> huge_page_shift(h)) +
807 (vma->vm_pgoff >> huge_page_order(h));
810 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
811 unsigned long address)
813 return vma_hugecache_offset(hstate_vma(vma), vma, address);
815 EXPORT_SYMBOL_GPL(linear_hugepage_index);
818 * Return the size of the pages allocated when backing a VMA. In the majority
819 * cases this will be same size as used by the page table entries.
821 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
823 if (vma->vm_ops && vma->vm_ops->pagesize)
824 return vma->vm_ops->pagesize(vma);
827 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
830 * Return the page size being used by the MMU to back a VMA. In the majority
831 * of cases, the page size used by the kernel matches the MMU size. On
832 * architectures where it differs, an architecture-specific 'strong'
833 * version of this symbol is required.
835 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
837 return vma_kernel_pagesize(vma);
841 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
842 * bits of the reservation map pointer, which are always clear due to
845 #define HPAGE_RESV_OWNER (1UL << 0)
846 #define HPAGE_RESV_UNMAPPED (1UL << 1)
847 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
850 * These helpers are used to track how many pages are reserved for
851 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
852 * is guaranteed to have their future faults succeed.
854 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
855 * the reserve counters are updated with the hugetlb_lock held. It is safe
856 * to reset the VMA at fork() time as it is not in use yet and there is no
857 * chance of the global counters getting corrupted as a result of the values.
859 * The private mapping reservation is represented in a subtly different
860 * manner to a shared mapping. A shared mapping has a region map associated
861 * with the underlying file, this region map represents the backing file
862 * pages which have ever had a reservation assigned which this persists even
863 * after the page is instantiated. A private mapping has a region map
864 * associated with the original mmap which is attached to all VMAs which
865 * reference it, this region map represents those offsets which have consumed
866 * reservation ie. where pages have been instantiated.
868 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
870 return (unsigned long)vma->vm_private_data;
873 static void set_vma_private_data(struct vm_area_struct *vma,
876 vma->vm_private_data = (void *)value;
880 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
881 struct hugetlb_cgroup *h_cg,
884 #ifdef CONFIG_CGROUP_HUGETLB
886 resv_map->reservation_counter = NULL;
887 resv_map->pages_per_hpage = 0;
888 resv_map->css = NULL;
890 resv_map->reservation_counter =
891 &h_cg->rsvd_hugepage[hstate_index(h)];
892 resv_map->pages_per_hpage = pages_per_huge_page(h);
893 resv_map->css = &h_cg->css;
898 struct resv_map *resv_map_alloc(void)
900 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
901 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
903 if (!resv_map || !rg) {
909 kref_init(&resv_map->refs);
910 spin_lock_init(&resv_map->lock);
911 INIT_LIST_HEAD(&resv_map->regions);
913 resv_map->adds_in_progress = 0;
915 * Initialize these to 0. On shared mappings, 0's here indicate these
916 * fields don't do cgroup accounting. On private mappings, these will be
917 * re-initialized to the proper values, to indicate that hugetlb cgroup
918 * reservations are to be un-charged from here.
920 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
922 INIT_LIST_HEAD(&resv_map->region_cache);
923 list_add(&rg->link, &resv_map->region_cache);
924 resv_map->region_cache_count = 1;
929 void resv_map_release(struct kref *ref)
931 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
932 struct list_head *head = &resv_map->region_cache;
933 struct file_region *rg, *trg;
935 /* Clear out any active regions before we release the map. */
936 region_del(resv_map, 0, LONG_MAX);
938 /* ... and any entries left in the cache */
939 list_for_each_entry_safe(rg, trg, head, link) {
944 VM_BUG_ON(resv_map->adds_in_progress);
949 static inline struct resv_map *inode_resv_map(struct inode *inode)
952 * At inode evict time, i_mapping may not point to the original
953 * address space within the inode. This original address space
954 * contains the pointer to the resv_map. So, always use the
955 * address space embedded within the inode.
956 * The VERY common case is inode->mapping == &inode->i_data but,
957 * this may not be true for device special inodes.
959 return (struct resv_map *)(&inode->i_data)->private_data;
962 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
964 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
965 if (vma->vm_flags & VM_MAYSHARE) {
966 struct address_space *mapping = vma->vm_file->f_mapping;
967 struct inode *inode = mapping->host;
969 return inode_resv_map(inode);
972 return (struct resv_map *)(get_vma_private_data(vma) &
977 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
979 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
980 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
982 set_vma_private_data(vma, (get_vma_private_data(vma) &
983 HPAGE_RESV_MASK) | (unsigned long)map);
986 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
988 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
989 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
991 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
994 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
996 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
998 return (get_vma_private_data(vma) & flag) != 0;
1001 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1002 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1004 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1005 if (!(vma->vm_flags & VM_MAYSHARE))
1006 vma->vm_private_data = (void *)0;
1009 /* Returns true if the VMA has associated reserve pages */
1010 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1012 if (vma->vm_flags & VM_NORESERVE) {
1014 * This address is already reserved by other process(chg == 0),
1015 * so, we should decrement reserved count. Without decrementing,
1016 * reserve count remains after releasing inode, because this
1017 * allocated page will go into page cache and is regarded as
1018 * coming from reserved pool in releasing step. Currently, we
1019 * don't have any other solution to deal with this situation
1020 * properly, so add work-around here.
1022 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1028 /* Shared mappings always use reserves */
1029 if (vma->vm_flags & VM_MAYSHARE) {
1031 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1032 * be a region map for all pages. The only situation where
1033 * there is no region map is if a hole was punched via
1034 * fallocate. In this case, there really are no reserves to
1035 * use. This situation is indicated if chg != 0.
1044 * Only the process that called mmap() has reserves for
1047 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1049 * Like the shared case above, a hole punch or truncate
1050 * could have been performed on the private mapping.
1051 * Examine the value of chg to determine if reserves
1052 * actually exist or were previously consumed.
1053 * Very Subtle - The value of chg comes from a previous
1054 * call to vma_needs_reserves(). The reserve map for
1055 * private mappings has different (opposite) semantics
1056 * than that of shared mappings. vma_needs_reserves()
1057 * has already taken this difference in semantics into
1058 * account. Therefore, the meaning of chg is the same
1059 * as in the shared case above. Code could easily be
1060 * combined, but keeping it separate draws attention to
1061 * subtle differences.
1072 static void enqueue_huge_page(struct hstate *h, struct page *page)
1074 int nid = page_to_nid(page);
1076 lockdep_assert_held(&hugetlb_lock);
1077 VM_BUG_ON_PAGE(page_count(page), page);
1079 list_move(&page->lru, &h->hugepage_freelists[nid]);
1080 h->free_huge_pages++;
1081 h->free_huge_pages_node[nid]++;
1082 SetHPageFreed(page);
1085 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1088 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1090 lockdep_assert_held(&hugetlb_lock);
1091 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1092 if (pin && !is_pinnable_page(page))
1095 if (PageHWPoison(page))
1098 list_move(&page->lru, &h->hugepage_activelist);
1099 set_page_refcounted(page);
1100 ClearHPageFreed(page);
1101 h->free_huge_pages--;
1102 h->free_huge_pages_node[nid]--;
1109 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1112 unsigned int cpuset_mems_cookie;
1113 struct zonelist *zonelist;
1116 int node = NUMA_NO_NODE;
1118 zonelist = node_zonelist(nid, gfp_mask);
1121 cpuset_mems_cookie = read_mems_allowed_begin();
1122 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1125 if (!cpuset_zone_allowed(zone, gfp_mask))
1128 * no need to ask again on the same node. Pool is node rather than
1131 if (zone_to_nid(zone) == node)
1133 node = zone_to_nid(zone);
1135 page = dequeue_huge_page_node_exact(h, node);
1139 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1145 static struct page *dequeue_huge_page_vma(struct hstate *h,
1146 struct vm_area_struct *vma,
1147 unsigned long address, int avoid_reserve,
1150 struct page *page = NULL;
1151 struct mempolicy *mpol;
1153 nodemask_t *nodemask;
1157 * A child process with MAP_PRIVATE mappings created by their parent
1158 * have no page reserves. This check ensures that reservations are
1159 * not "stolen". The child may still get SIGKILLed
1161 if (!vma_has_reserves(vma, chg) &&
1162 h->free_huge_pages - h->resv_huge_pages == 0)
1165 /* If reserves cannot be used, ensure enough pages are in the pool */
1166 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1169 gfp_mask = htlb_alloc_mask(h);
1170 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1172 if (mpol_is_preferred_many(mpol)) {
1173 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1175 /* Fallback to all nodes if page==NULL */
1180 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1182 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1183 SetHPageRestoreReserve(page);
1184 h->resv_huge_pages--;
1187 mpol_cond_put(mpol);
1195 * common helper functions for hstate_next_node_to_{alloc|free}.
1196 * We may have allocated or freed a huge page based on a different
1197 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1198 * be outside of *nodes_allowed. Ensure that we use an allowed
1199 * node for alloc or free.
1201 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1203 nid = next_node_in(nid, *nodes_allowed);
1204 VM_BUG_ON(nid >= MAX_NUMNODES);
1209 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1211 if (!node_isset(nid, *nodes_allowed))
1212 nid = next_node_allowed(nid, nodes_allowed);
1217 * returns the previously saved node ["this node"] from which to
1218 * allocate a persistent huge page for the pool and advance the
1219 * next node from which to allocate, handling wrap at end of node
1222 static int hstate_next_node_to_alloc(struct hstate *h,
1223 nodemask_t *nodes_allowed)
1227 VM_BUG_ON(!nodes_allowed);
1229 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1230 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1236 * helper for remove_pool_huge_page() - return the previously saved
1237 * node ["this node"] from which to free a huge page. Advance the
1238 * next node id whether or not we find a free huge page to free so
1239 * that the next attempt to free addresses the next node.
1241 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1245 VM_BUG_ON(!nodes_allowed);
1247 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1248 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1253 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1254 for (nr_nodes = nodes_weight(*mask); \
1256 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1259 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1260 for (nr_nodes = nodes_weight(*mask); \
1262 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1265 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1266 static void destroy_compound_gigantic_page(struct page *page,
1270 int nr_pages = 1 << order;
1271 struct page *p = page + 1;
1273 atomic_set(compound_mapcount_ptr(page), 0);
1274 atomic_set(compound_pincount_ptr(page), 0);
1276 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1277 clear_compound_head(p);
1278 set_page_refcounted(p);
1281 set_compound_order(page, 0);
1282 page[1].compound_nr = 0;
1283 __ClearPageHead(page);
1286 static void free_gigantic_page(struct page *page, unsigned int order)
1289 * If the page isn't allocated using the cma allocator,
1290 * cma_release() returns false.
1293 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1297 free_contig_range(page_to_pfn(page), 1 << order);
1300 #ifdef CONFIG_CONTIG_ALLOC
1301 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1302 int nid, nodemask_t *nodemask)
1304 unsigned long nr_pages = pages_per_huge_page(h);
1305 if (nid == NUMA_NO_NODE)
1306 nid = numa_mem_id();
1313 if (hugetlb_cma[nid]) {
1314 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1315 huge_page_order(h), true);
1320 if (!(gfp_mask & __GFP_THISNODE)) {
1321 for_each_node_mask(node, *nodemask) {
1322 if (node == nid || !hugetlb_cma[node])
1325 page = cma_alloc(hugetlb_cma[node], nr_pages,
1326 huge_page_order(h), true);
1334 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1337 #else /* !CONFIG_CONTIG_ALLOC */
1338 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1339 int nid, nodemask_t *nodemask)
1343 #endif /* CONFIG_CONTIG_ALLOC */
1345 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1346 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1347 int nid, nodemask_t *nodemask)
1351 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1352 static inline void destroy_compound_gigantic_page(struct page *page,
1353 unsigned int order) { }
1357 * Remove hugetlb page from lists, and update dtor so that page appears
1358 * as just a compound page. A reference is held on the page.
1360 * Must be called with hugetlb lock held.
1362 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1363 bool adjust_surplus)
1365 int nid = page_to_nid(page);
1367 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1368 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1370 lockdep_assert_held(&hugetlb_lock);
1371 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1374 list_del(&page->lru);
1376 if (HPageFreed(page)) {
1377 h->free_huge_pages--;
1378 h->free_huge_pages_node[nid]--;
1380 if (adjust_surplus) {
1381 h->surplus_huge_pages--;
1382 h->surplus_huge_pages_node[nid]--;
1388 * For non-gigantic pages set the destructor to the normal compound
1389 * page dtor. This is needed in case someone takes an additional
1390 * temporary ref to the page, and freeing is delayed until they drop
1393 * For gigantic pages set the destructor to the null dtor. This
1394 * destructor will never be called. Before freeing the gigantic
1395 * page destroy_compound_gigantic_page will turn the compound page
1396 * into a simple group of pages. After this the destructor does not
1399 * This handles the case where more than one ref is held when and
1400 * after update_and_free_page is called.
1402 set_page_refcounted(page);
1403 if (hstate_is_gigantic(h))
1404 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1406 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1409 h->nr_huge_pages_node[nid]--;
1412 static void add_hugetlb_page(struct hstate *h, struct page *page,
1413 bool adjust_surplus)
1416 int nid = page_to_nid(page);
1418 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1420 lockdep_assert_held(&hugetlb_lock);
1422 INIT_LIST_HEAD(&page->lru);
1424 h->nr_huge_pages_node[nid]++;
1426 if (adjust_surplus) {
1427 h->surplus_huge_pages++;
1428 h->surplus_huge_pages_node[nid]++;
1431 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1432 set_page_private(page, 0);
1433 SetHPageVmemmapOptimized(page);
1436 * This page is about to be managed by the hugetlb allocator and
1437 * should have no users. Drop our reference, and check for others
1440 zeroed = put_page_testzero(page);
1443 * It is VERY unlikely soneone else has taken a ref on
1444 * the page. In this case, we simply return as the
1445 * hugetlb destructor (free_huge_page) will be called
1446 * when this other ref is dropped.
1450 arch_clear_hugepage_flags(page);
1451 enqueue_huge_page(h, page);
1454 static void __update_and_free_page(struct hstate *h, struct page *page)
1457 struct page *subpage = page;
1459 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1462 if (alloc_huge_page_vmemmap(h, page)) {
1463 spin_lock_irq(&hugetlb_lock);
1465 * If we cannot allocate vmemmap pages, just refuse to free the
1466 * page and put the page back on the hugetlb free list and treat
1467 * as a surplus page.
1469 add_hugetlb_page(h, page, true);
1470 spin_unlock_irq(&hugetlb_lock);
1474 for (i = 0; i < pages_per_huge_page(h);
1475 i++, subpage = mem_map_next(subpage, page, i)) {
1476 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1477 1 << PG_referenced | 1 << PG_dirty |
1478 1 << PG_active | 1 << PG_private |
1481 if (hstate_is_gigantic(h)) {
1482 destroy_compound_gigantic_page(page, huge_page_order(h));
1483 free_gigantic_page(page, huge_page_order(h));
1485 __free_pages(page, huge_page_order(h));
1490 * As update_and_free_page() can be called under any context, so we cannot
1491 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1492 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1493 * the vmemmap pages.
1495 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1496 * freed and frees them one-by-one. As the page->mapping pointer is going
1497 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1498 * structure of a lockless linked list of huge pages to be freed.
1500 static LLIST_HEAD(hpage_freelist);
1502 static void free_hpage_workfn(struct work_struct *work)
1504 struct llist_node *node;
1506 node = llist_del_all(&hpage_freelist);
1512 page = container_of((struct address_space **)node,
1513 struct page, mapping);
1515 page->mapping = NULL;
1517 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1518 * is going to trigger because a previous call to
1519 * remove_hugetlb_page() will set_compound_page_dtor(page,
1520 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1522 h = size_to_hstate(page_size(page));
1524 __update_and_free_page(h, page);
1529 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1531 static inline void flush_free_hpage_work(struct hstate *h)
1533 if (free_vmemmap_pages_per_hpage(h))
1534 flush_work(&free_hpage_work);
1537 static void update_and_free_page(struct hstate *h, struct page *page,
1540 if (!HPageVmemmapOptimized(page) || !atomic) {
1541 __update_and_free_page(h, page);
1546 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1548 * Only call schedule_work() if hpage_freelist is previously
1549 * empty. Otherwise, schedule_work() had been called but the workfn
1550 * hasn't retrieved the list yet.
1552 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1553 schedule_work(&free_hpage_work);
1556 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1558 struct page *page, *t_page;
1560 list_for_each_entry_safe(page, t_page, list, lru) {
1561 update_and_free_page(h, page, false);
1566 struct hstate *size_to_hstate(unsigned long size)
1570 for_each_hstate(h) {
1571 if (huge_page_size(h) == size)
1577 void free_huge_page(struct page *page)
1580 * Can't pass hstate in here because it is called from the
1581 * compound page destructor.
1583 struct hstate *h = page_hstate(page);
1584 int nid = page_to_nid(page);
1585 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1586 bool restore_reserve;
1587 unsigned long flags;
1589 VM_BUG_ON_PAGE(page_count(page), page);
1590 VM_BUG_ON_PAGE(page_mapcount(page), page);
1592 hugetlb_set_page_subpool(page, NULL);
1593 page->mapping = NULL;
1594 restore_reserve = HPageRestoreReserve(page);
1595 ClearHPageRestoreReserve(page);
1598 * If HPageRestoreReserve was set on page, page allocation consumed a
1599 * reservation. If the page was associated with a subpool, there
1600 * would have been a page reserved in the subpool before allocation
1601 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1602 * reservation, do not call hugepage_subpool_put_pages() as this will
1603 * remove the reserved page from the subpool.
1605 if (!restore_reserve) {
1607 * A return code of zero implies that the subpool will be
1608 * under its minimum size if the reservation is not restored
1609 * after page is free. Therefore, force restore_reserve
1612 if (hugepage_subpool_put_pages(spool, 1) == 0)
1613 restore_reserve = true;
1616 spin_lock_irqsave(&hugetlb_lock, flags);
1617 ClearHPageMigratable(page);
1618 hugetlb_cgroup_uncharge_page(hstate_index(h),
1619 pages_per_huge_page(h), page);
1620 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1621 pages_per_huge_page(h), page);
1622 if (restore_reserve)
1623 h->resv_huge_pages++;
1625 if (HPageTemporary(page)) {
1626 remove_hugetlb_page(h, page, false);
1627 spin_unlock_irqrestore(&hugetlb_lock, flags);
1628 update_and_free_page(h, page, true);
1629 } else if (h->surplus_huge_pages_node[nid]) {
1630 /* remove the page from active list */
1631 remove_hugetlb_page(h, page, true);
1632 spin_unlock_irqrestore(&hugetlb_lock, flags);
1633 update_and_free_page(h, page, true);
1635 arch_clear_hugepage_flags(page);
1636 enqueue_huge_page(h, page);
1637 spin_unlock_irqrestore(&hugetlb_lock, flags);
1642 * Must be called with the hugetlb lock held
1644 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1646 lockdep_assert_held(&hugetlb_lock);
1648 h->nr_huge_pages_node[nid]++;
1651 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1653 free_huge_page_vmemmap(h, page);
1654 INIT_LIST_HEAD(&page->lru);
1655 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1656 hugetlb_set_page_subpool(page, NULL);
1657 set_hugetlb_cgroup(page, NULL);
1658 set_hugetlb_cgroup_rsvd(page, NULL);
1661 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1663 __prep_new_huge_page(h, page);
1664 spin_lock_irq(&hugetlb_lock);
1665 __prep_account_new_huge_page(h, nid);
1666 spin_unlock_irq(&hugetlb_lock);
1669 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1672 int nr_pages = 1 << order;
1673 struct page *p = page + 1;
1675 /* we rely on prep_new_huge_page to set the destructor */
1676 set_compound_order(page, order);
1677 __ClearPageReserved(page);
1678 __SetPageHead(page);
1679 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1681 * For gigantic hugepages allocated through bootmem at
1682 * boot, it's safer to be consistent with the not-gigantic
1683 * hugepages and clear the PG_reserved bit from all tail pages
1684 * too. Otherwise drivers using get_user_pages() to access tail
1685 * pages may get the reference counting wrong if they see
1686 * PG_reserved set on a tail page (despite the head page not
1687 * having PG_reserved set). Enforcing this consistency between
1688 * head and tail pages allows drivers to optimize away a check
1689 * on the head page when they need know if put_page() is needed
1690 * after get_user_pages().
1692 __ClearPageReserved(p);
1694 * Subtle and very unlikely
1696 * Gigantic 'page allocators' such as memblock or cma will
1697 * return a set of pages with each page ref counted. We need
1698 * to turn this set of pages into a compound page with tail
1699 * page ref counts set to zero. Code such as speculative page
1700 * cache adding could take a ref on a 'to be' tail page.
1701 * We need to respect any increased ref count, and only set
1702 * the ref count to zero if count is currently 1. If count
1703 * is not 1, we return an error. An error return indicates
1704 * the set of pages can not be converted to a gigantic page.
1705 * The caller who allocated the pages should then discard the
1706 * pages using the appropriate free interface.
1708 if (!page_ref_freeze(p, 1)) {
1709 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1712 set_page_count(p, 0);
1713 set_compound_head(p, page);
1715 atomic_set(compound_mapcount_ptr(page), -1);
1716 atomic_set(compound_pincount_ptr(page), 0);
1720 /* undo tail page modifications made above */
1722 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1723 clear_compound_head(p);
1724 set_page_refcounted(p);
1726 /* need to clear PG_reserved on remaining tail pages */
1727 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1728 __ClearPageReserved(p);
1729 set_compound_order(page, 0);
1730 page[1].compound_nr = 0;
1731 __ClearPageHead(page);
1736 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1737 * transparent huge pages. See the PageTransHuge() documentation for more
1740 int PageHuge(struct page *page)
1742 if (!PageCompound(page))
1745 page = compound_head(page);
1746 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1748 EXPORT_SYMBOL_GPL(PageHuge);
1751 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1752 * normal or transparent huge pages.
1754 int PageHeadHuge(struct page *page_head)
1756 if (!PageHead(page_head))
1759 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1763 * Find and lock address space (mapping) in write mode.
1765 * Upon entry, the page is locked which means that page_mapping() is
1766 * stable. Due to locking order, we can only trylock_write. If we can
1767 * not get the lock, simply return NULL to caller.
1769 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1771 struct address_space *mapping = page_mapping(hpage);
1776 if (i_mmap_trylock_write(mapping))
1782 pgoff_t hugetlb_basepage_index(struct page *page)
1784 struct page *page_head = compound_head(page);
1785 pgoff_t index = page_index(page_head);
1786 unsigned long compound_idx;
1788 if (compound_order(page_head) >= MAX_ORDER)
1789 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1791 compound_idx = page - page_head;
1793 return (index << compound_order(page_head)) + compound_idx;
1796 static struct page *alloc_buddy_huge_page(struct hstate *h,
1797 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1798 nodemask_t *node_alloc_noretry)
1800 int order = huge_page_order(h);
1802 bool alloc_try_hard = true;
1805 * By default we always try hard to allocate the page with
1806 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1807 * a loop (to adjust global huge page counts) and previous allocation
1808 * failed, do not continue to try hard on the same node. Use the
1809 * node_alloc_noretry bitmap to manage this state information.
1811 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1812 alloc_try_hard = false;
1813 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1815 gfp_mask |= __GFP_RETRY_MAYFAIL;
1816 if (nid == NUMA_NO_NODE)
1817 nid = numa_mem_id();
1818 page = __alloc_pages(gfp_mask, order, nid, nmask);
1820 __count_vm_event(HTLB_BUDDY_PGALLOC);
1822 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1825 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1826 * indicates an overall state change. Clear bit so that we resume
1827 * normal 'try hard' allocations.
1829 if (node_alloc_noretry && page && !alloc_try_hard)
1830 node_clear(nid, *node_alloc_noretry);
1833 * If we tried hard to get a page but failed, set bit so that
1834 * subsequent attempts will not try as hard until there is an
1835 * overall state change.
1837 if (node_alloc_noretry && !page && alloc_try_hard)
1838 node_set(nid, *node_alloc_noretry);
1844 * Common helper to allocate a fresh hugetlb page. All specific allocators
1845 * should use this function to get new hugetlb pages
1847 static struct page *alloc_fresh_huge_page(struct hstate *h,
1848 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1849 nodemask_t *node_alloc_noretry)
1855 if (hstate_is_gigantic(h))
1856 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1858 page = alloc_buddy_huge_page(h, gfp_mask,
1859 nid, nmask, node_alloc_noretry);
1863 if (hstate_is_gigantic(h)) {
1864 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1866 * Rare failure to convert pages to compound page.
1867 * Free pages and try again - ONCE!
1869 free_gigantic_page(page, huge_page_order(h));
1877 prep_new_huge_page(h, page, page_to_nid(page));
1883 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1886 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1887 nodemask_t *node_alloc_noretry)
1891 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1893 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1894 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1895 node_alloc_noretry);
1903 put_page(page); /* free it into the hugepage allocator */
1909 * Remove huge page from pool from next node to free. Attempt to keep
1910 * persistent huge pages more or less balanced over allowed nodes.
1911 * This routine only 'removes' the hugetlb page. The caller must make
1912 * an additional call to free the page to low level allocators.
1913 * Called with hugetlb_lock locked.
1915 static struct page *remove_pool_huge_page(struct hstate *h,
1916 nodemask_t *nodes_allowed,
1920 struct page *page = NULL;
1922 lockdep_assert_held(&hugetlb_lock);
1923 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1925 * If we're returning unused surplus pages, only examine
1926 * nodes with surplus pages.
1928 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1929 !list_empty(&h->hugepage_freelists[node])) {
1930 page = list_entry(h->hugepage_freelists[node].next,
1932 remove_hugetlb_page(h, page, acct_surplus);
1941 * Dissolve a given free hugepage into free buddy pages. This function does
1942 * nothing for in-use hugepages and non-hugepages.
1943 * This function returns values like below:
1945 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1946 * when the system is under memory pressure and the feature of
1947 * freeing unused vmemmap pages associated with each hugetlb page
1949 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1950 * (allocated or reserved.)
1951 * 0: successfully dissolved free hugepages or the page is not a
1952 * hugepage (considered as already dissolved)
1954 int dissolve_free_huge_page(struct page *page)
1959 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1960 if (!PageHuge(page))
1963 spin_lock_irq(&hugetlb_lock);
1964 if (!PageHuge(page)) {
1969 if (!page_count(page)) {
1970 struct page *head = compound_head(page);
1971 struct hstate *h = page_hstate(head);
1972 if (h->free_huge_pages - h->resv_huge_pages == 0)
1976 * We should make sure that the page is already on the free list
1977 * when it is dissolved.
1979 if (unlikely(!HPageFreed(head))) {
1980 spin_unlock_irq(&hugetlb_lock);
1984 * Theoretically, we should return -EBUSY when we
1985 * encounter this race. In fact, we have a chance
1986 * to successfully dissolve the page if we do a
1987 * retry. Because the race window is quite small.
1988 * If we seize this opportunity, it is an optimization
1989 * for increasing the success rate of dissolving page.
1994 remove_hugetlb_page(h, head, false);
1995 h->max_huge_pages--;
1996 spin_unlock_irq(&hugetlb_lock);
1999 * Normally update_and_free_page will allocate required vmemmmap
2000 * before freeing the page. update_and_free_page will fail to
2001 * free the page if it can not allocate required vmemmap. We
2002 * need to adjust max_huge_pages if the page is not freed.
2003 * Attempt to allocate vmemmmap here so that we can take
2004 * appropriate action on failure.
2006 rc = alloc_huge_page_vmemmap(h, head);
2009 * Move PageHWPoison flag from head page to the raw
2010 * error page, which makes any subpages rather than
2011 * the error page reusable.
2013 if (PageHWPoison(head) && page != head) {
2014 SetPageHWPoison(page);
2015 ClearPageHWPoison(head);
2017 update_and_free_page(h, head, false);
2019 spin_lock_irq(&hugetlb_lock);
2020 add_hugetlb_page(h, head, false);
2021 h->max_huge_pages++;
2022 spin_unlock_irq(&hugetlb_lock);
2028 spin_unlock_irq(&hugetlb_lock);
2033 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2034 * make specified memory blocks removable from the system.
2035 * Note that this will dissolve a free gigantic hugepage completely, if any
2036 * part of it lies within the given range.
2037 * Also note that if dissolve_free_huge_page() returns with an error, all
2038 * free hugepages that were dissolved before that error are lost.
2040 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2046 if (!hugepages_supported())
2049 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2050 page = pfn_to_page(pfn);
2051 rc = dissolve_free_huge_page(page);
2060 * Allocates a fresh surplus page from the page allocator.
2062 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2063 int nid, nodemask_t *nmask, bool zero_ref)
2065 struct page *page = NULL;
2068 if (hstate_is_gigantic(h))
2071 spin_lock_irq(&hugetlb_lock);
2072 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2074 spin_unlock_irq(&hugetlb_lock);
2077 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2081 spin_lock_irq(&hugetlb_lock);
2083 * We could have raced with the pool size change.
2084 * Double check that and simply deallocate the new page
2085 * if we would end up overcommiting the surpluses. Abuse
2086 * temporary page to workaround the nasty free_huge_page
2089 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2090 SetHPageTemporary(page);
2091 spin_unlock_irq(&hugetlb_lock);
2098 * Caller requires a page with zero ref count.
2099 * We will drop ref count here. If someone else is holding
2100 * a ref, the page will be freed when they drop it. Abuse
2101 * temporary page flag to accomplish this.
2103 SetHPageTemporary(page);
2104 if (!put_page_testzero(page)) {
2106 * Unexpected inflated ref count on freshly allocated
2109 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2110 spin_unlock_irq(&hugetlb_lock);
2117 ClearHPageTemporary(page);
2120 h->surplus_huge_pages++;
2121 h->surplus_huge_pages_node[page_to_nid(page)]++;
2124 spin_unlock_irq(&hugetlb_lock);
2129 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2130 int nid, nodemask_t *nmask)
2134 if (hstate_is_gigantic(h))
2137 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2142 * We do not account these pages as surplus because they are only
2143 * temporary and will be released properly on the last reference
2145 SetHPageTemporary(page);
2151 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2154 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2155 struct vm_area_struct *vma, unsigned long addr)
2157 struct page *page = NULL;
2158 struct mempolicy *mpol;
2159 gfp_t gfp_mask = htlb_alloc_mask(h);
2161 nodemask_t *nodemask;
2163 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2164 if (mpol_is_preferred_many(mpol)) {
2165 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2167 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2168 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2170 /* Fallback to all nodes if page==NULL */
2175 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2176 mpol_cond_put(mpol);
2180 /* page migration callback function */
2181 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2182 nodemask_t *nmask, gfp_t gfp_mask)
2184 spin_lock_irq(&hugetlb_lock);
2185 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2188 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2190 spin_unlock_irq(&hugetlb_lock);
2194 spin_unlock_irq(&hugetlb_lock);
2196 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2199 /* mempolicy aware migration callback */
2200 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2201 unsigned long address)
2203 struct mempolicy *mpol;
2204 nodemask_t *nodemask;
2209 gfp_mask = htlb_alloc_mask(h);
2210 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2211 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2212 mpol_cond_put(mpol);
2218 * Increase the hugetlb pool such that it can accommodate a reservation
2221 static int gather_surplus_pages(struct hstate *h, long delta)
2222 __must_hold(&hugetlb_lock)
2224 struct list_head surplus_list;
2225 struct page *page, *tmp;
2228 long needed, allocated;
2229 bool alloc_ok = true;
2231 lockdep_assert_held(&hugetlb_lock);
2232 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2234 h->resv_huge_pages += delta;
2239 INIT_LIST_HEAD(&surplus_list);
2243 spin_unlock_irq(&hugetlb_lock);
2244 for (i = 0; i < needed; i++) {
2245 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2246 NUMA_NO_NODE, NULL, true);
2251 list_add(&page->lru, &surplus_list);
2257 * After retaking hugetlb_lock, we need to recalculate 'needed'
2258 * because either resv_huge_pages or free_huge_pages may have changed.
2260 spin_lock_irq(&hugetlb_lock);
2261 needed = (h->resv_huge_pages + delta) -
2262 (h->free_huge_pages + allocated);
2267 * We were not able to allocate enough pages to
2268 * satisfy the entire reservation so we free what
2269 * we've allocated so far.
2274 * The surplus_list now contains _at_least_ the number of extra pages
2275 * needed to accommodate the reservation. Add the appropriate number
2276 * of pages to the hugetlb pool and free the extras back to the buddy
2277 * allocator. Commit the entire reservation here to prevent another
2278 * process from stealing the pages as they are added to the pool but
2279 * before they are reserved.
2281 needed += allocated;
2282 h->resv_huge_pages += delta;
2285 /* Free the needed pages to the hugetlb pool */
2286 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2289 /* Add the page to the hugetlb allocator */
2290 enqueue_huge_page(h, page);
2293 spin_unlock_irq(&hugetlb_lock);
2296 * Free unnecessary surplus pages to the buddy allocator.
2297 * Pages have no ref count, call free_huge_page directly.
2299 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2300 free_huge_page(page);
2301 spin_lock_irq(&hugetlb_lock);
2307 * This routine has two main purposes:
2308 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2309 * in unused_resv_pages. This corresponds to the prior adjustments made
2310 * to the associated reservation map.
2311 * 2) Free any unused surplus pages that may have been allocated to satisfy
2312 * the reservation. As many as unused_resv_pages may be freed.
2314 static void return_unused_surplus_pages(struct hstate *h,
2315 unsigned long unused_resv_pages)
2317 unsigned long nr_pages;
2319 LIST_HEAD(page_list);
2321 lockdep_assert_held(&hugetlb_lock);
2322 /* Uncommit the reservation */
2323 h->resv_huge_pages -= unused_resv_pages;
2325 /* Cannot return gigantic pages currently */
2326 if (hstate_is_gigantic(h))
2330 * Part (or even all) of the reservation could have been backed
2331 * by pre-allocated pages. Only free surplus pages.
2333 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2336 * We want to release as many surplus pages as possible, spread
2337 * evenly across all nodes with memory. Iterate across these nodes
2338 * until we can no longer free unreserved surplus pages. This occurs
2339 * when the nodes with surplus pages have no free pages.
2340 * remove_pool_huge_page() will balance the freed pages across the
2341 * on-line nodes with memory and will handle the hstate accounting.
2343 while (nr_pages--) {
2344 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2348 list_add(&page->lru, &page_list);
2352 spin_unlock_irq(&hugetlb_lock);
2353 update_and_free_pages_bulk(h, &page_list);
2354 spin_lock_irq(&hugetlb_lock);
2359 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2360 * are used by the huge page allocation routines to manage reservations.
2362 * vma_needs_reservation is called to determine if the huge page at addr
2363 * within the vma has an associated reservation. If a reservation is
2364 * needed, the value 1 is returned. The caller is then responsible for
2365 * managing the global reservation and subpool usage counts. After
2366 * the huge page has been allocated, vma_commit_reservation is called
2367 * to add the page to the reservation map. If the page allocation fails,
2368 * the reservation must be ended instead of committed. vma_end_reservation
2369 * is called in such cases.
2371 * In the normal case, vma_commit_reservation returns the same value
2372 * as the preceding vma_needs_reservation call. The only time this
2373 * is not the case is if a reserve map was changed between calls. It
2374 * is the responsibility of the caller to notice the difference and
2375 * take appropriate action.
2377 * vma_add_reservation is used in error paths where a reservation must
2378 * be restored when a newly allocated huge page must be freed. It is
2379 * to be called after calling vma_needs_reservation to determine if a
2380 * reservation exists.
2382 * vma_del_reservation is used in error paths where an entry in the reserve
2383 * map was created during huge page allocation and must be removed. It is to
2384 * be called after calling vma_needs_reservation to determine if a reservation
2387 enum vma_resv_mode {
2394 static long __vma_reservation_common(struct hstate *h,
2395 struct vm_area_struct *vma, unsigned long addr,
2396 enum vma_resv_mode mode)
2398 struct resv_map *resv;
2401 long dummy_out_regions_needed;
2403 resv = vma_resv_map(vma);
2407 idx = vma_hugecache_offset(h, vma, addr);
2409 case VMA_NEEDS_RESV:
2410 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2411 /* We assume that vma_reservation_* routines always operate on
2412 * 1 page, and that adding to resv map a 1 page entry can only
2413 * ever require 1 region.
2415 VM_BUG_ON(dummy_out_regions_needed != 1);
2417 case VMA_COMMIT_RESV:
2418 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2419 /* region_add calls of range 1 should never fail. */
2423 region_abort(resv, idx, idx + 1, 1);
2427 if (vma->vm_flags & VM_MAYSHARE) {
2428 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2429 /* region_add calls of range 1 should never fail. */
2432 region_abort(resv, idx, idx + 1, 1);
2433 ret = region_del(resv, idx, idx + 1);
2437 if (vma->vm_flags & VM_MAYSHARE) {
2438 region_abort(resv, idx, idx + 1, 1);
2439 ret = region_del(resv, idx, idx + 1);
2441 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2442 /* region_add calls of range 1 should never fail. */
2450 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2453 * We know private mapping must have HPAGE_RESV_OWNER set.
2455 * In most cases, reserves always exist for private mappings.
2456 * However, a file associated with mapping could have been
2457 * hole punched or truncated after reserves were consumed.
2458 * As subsequent fault on such a range will not use reserves.
2459 * Subtle - The reserve map for private mappings has the
2460 * opposite meaning than that of shared mappings. If NO
2461 * entry is in the reserve map, it means a reservation exists.
2462 * If an entry exists in the reserve map, it means the
2463 * reservation has already been consumed. As a result, the
2464 * return value of this routine is the opposite of the
2465 * value returned from reserve map manipulation routines above.
2474 static long vma_needs_reservation(struct hstate *h,
2475 struct vm_area_struct *vma, unsigned long addr)
2477 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2480 static long vma_commit_reservation(struct hstate *h,
2481 struct vm_area_struct *vma, unsigned long addr)
2483 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2486 static void vma_end_reservation(struct hstate *h,
2487 struct vm_area_struct *vma, unsigned long addr)
2489 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2492 static long vma_add_reservation(struct hstate *h,
2493 struct vm_area_struct *vma, unsigned long addr)
2495 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2498 static long vma_del_reservation(struct hstate *h,
2499 struct vm_area_struct *vma, unsigned long addr)
2501 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2505 * This routine is called to restore reservation information on error paths.
2506 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2507 * the hugetlb mutex should remain held when calling this routine.
2509 * It handles two specific cases:
2510 * 1) A reservation was in place and the page consumed the reservation.
2511 * HPageRestoreReserve is set in the page.
2512 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2513 * not set. However, alloc_huge_page always updates the reserve map.
2515 * In case 1, free_huge_page later in the error path will increment the
2516 * global reserve count. But, free_huge_page does not have enough context
2517 * to adjust the reservation map. This case deals primarily with private
2518 * mappings. Adjust the reserve map here to be consistent with global
2519 * reserve count adjustments to be made by free_huge_page. Make sure the
2520 * reserve map indicates there is a reservation present.
2522 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2524 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2525 unsigned long address, struct page *page)
2527 long rc = vma_needs_reservation(h, vma, address);
2529 if (HPageRestoreReserve(page)) {
2530 if (unlikely(rc < 0))
2532 * Rare out of memory condition in reserve map
2533 * manipulation. Clear HPageRestoreReserve so that
2534 * global reserve count will not be incremented
2535 * by free_huge_page. This will make it appear
2536 * as though the reservation for this page was
2537 * consumed. This may prevent the task from
2538 * faulting in the page at a later time. This
2539 * is better than inconsistent global huge page
2540 * accounting of reserve counts.
2542 ClearHPageRestoreReserve(page);
2544 (void)vma_add_reservation(h, vma, address);
2546 vma_end_reservation(h, vma, address);
2550 * This indicates there is an entry in the reserve map
2551 * not added by alloc_huge_page. We know it was added
2552 * before the alloc_huge_page call, otherwise
2553 * HPageRestoreReserve would be set on the page.
2554 * Remove the entry so that a subsequent allocation
2555 * does not consume a reservation.
2557 rc = vma_del_reservation(h, vma, address);
2560 * VERY rare out of memory condition. Since
2561 * we can not delete the entry, set
2562 * HPageRestoreReserve so that the reserve
2563 * count will be incremented when the page
2564 * is freed. This reserve will be consumed
2565 * on a subsequent allocation.
2567 SetHPageRestoreReserve(page);
2568 } else if (rc < 0) {
2570 * Rare out of memory condition from
2571 * vma_needs_reservation call. Memory allocation is
2572 * only attempted if a new entry is needed. Therefore,
2573 * this implies there is not an entry in the
2576 * For shared mappings, no entry in the map indicates
2577 * no reservation. We are done.
2579 if (!(vma->vm_flags & VM_MAYSHARE))
2581 * For private mappings, no entry indicates
2582 * a reservation is present. Since we can
2583 * not add an entry, set SetHPageRestoreReserve
2584 * on the page so reserve count will be
2585 * incremented when freed. This reserve will
2586 * be consumed on a subsequent allocation.
2588 SetHPageRestoreReserve(page);
2591 * No reservation present, do nothing
2593 vma_end_reservation(h, vma, address);
2598 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2599 * @h: struct hstate old page belongs to
2600 * @old_page: Old page to dissolve
2601 * @list: List to isolate the page in case we need to
2602 * Returns 0 on success, otherwise negated error.
2604 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2605 struct list_head *list)
2607 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2608 int nid = page_to_nid(old_page);
2609 bool alloc_retry = false;
2610 struct page *new_page;
2614 * Before dissolving the page, we need to allocate a new one for the
2615 * pool to remain stable. Here, we allocate the page and 'prep' it
2616 * by doing everything but actually updating counters and adding to
2617 * the pool. This simplifies and let us do most of the processing
2621 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2625 * If all goes well, this page will be directly added to the free
2626 * list in the pool. For this the ref count needs to be zero.
2627 * Attempt to drop now, and retry once if needed. It is VERY
2628 * unlikely there is another ref on the page.
2630 * If someone else has a reference to the page, it will be freed
2631 * when they drop their ref. Abuse temporary page flag to accomplish
2632 * this. Retry once if there is an inflated ref count.
2634 SetHPageTemporary(new_page);
2635 if (!put_page_testzero(new_page)) {
2642 ClearHPageTemporary(new_page);
2644 __prep_new_huge_page(h, new_page);
2647 spin_lock_irq(&hugetlb_lock);
2648 if (!PageHuge(old_page)) {
2650 * Freed from under us. Drop new_page too.
2653 } else if (page_count(old_page)) {
2655 * Someone has grabbed the page, try to isolate it here.
2656 * Fail with -EBUSY if not possible.
2658 spin_unlock_irq(&hugetlb_lock);
2659 ret = isolate_hugetlb(old_page, list);
2660 spin_lock_irq(&hugetlb_lock);
2662 } else if (!HPageFreed(old_page)) {
2664 * Page's refcount is 0 but it has not been enqueued in the
2665 * freelist yet. Race window is small, so we can succeed here if
2668 spin_unlock_irq(&hugetlb_lock);
2673 * Ok, old_page is still a genuine free hugepage. Remove it from
2674 * the freelist and decrease the counters. These will be
2675 * incremented again when calling __prep_account_new_huge_page()
2676 * and enqueue_huge_page() for new_page. The counters will remain
2677 * stable since this happens under the lock.
2679 remove_hugetlb_page(h, old_page, false);
2682 * Ref count on new page is already zero as it was dropped
2683 * earlier. It can be directly added to the pool free list.
2685 __prep_account_new_huge_page(h, nid);
2686 enqueue_huge_page(h, new_page);
2689 * Pages have been replaced, we can safely free the old one.
2691 spin_unlock_irq(&hugetlb_lock);
2692 update_and_free_page(h, old_page, false);
2698 spin_unlock_irq(&hugetlb_lock);
2699 /* Page has a zero ref count, but needs a ref to be freed */
2700 set_page_refcounted(new_page);
2701 update_and_free_page(h, new_page, false);
2706 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2713 * The page might have been dissolved from under our feet, so make sure
2714 * to carefully check the state under the lock.
2715 * Return success when racing as if we dissolved the page ourselves.
2717 spin_lock_irq(&hugetlb_lock);
2718 if (PageHuge(page)) {
2719 head = compound_head(page);
2720 h = page_hstate(head);
2722 spin_unlock_irq(&hugetlb_lock);
2725 spin_unlock_irq(&hugetlb_lock);
2728 * Fence off gigantic pages as there is a cyclic dependency between
2729 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2730 * of bailing out right away without further retrying.
2732 if (hstate_is_gigantic(h))
2735 if (page_count(head) && !isolate_hugetlb(head, list))
2737 else if (!page_count(head))
2738 ret = alloc_and_dissolve_huge_page(h, head, list);
2743 struct page *alloc_huge_page(struct vm_area_struct *vma,
2744 unsigned long addr, int avoid_reserve)
2746 struct hugepage_subpool *spool = subpool_vma(vma);
2747 struct hstate *h = hstate_vma(vma);
2749 long map_chg, map_commit;
2752 struct hugetlb_cgroup *h_cg;
2753 bool deferred_reserve;
2755 idx = hstate_index(h);
2757 * Examine the region/reserve map to determine if the process
2758 * has a reservation for the page to be allocated. A return
2759 * code of zero indicates a reservation exists (no change).
2761 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2763 return ERR_PTR(-ENOMEM);
2766 * Processes that did not create the mapping will have no
2767 * reserves as indicated by the region/reserve map. Check
2768 * that the allocation will not exceed the subpool limit.
2769 * Allocations for MAP_NORESERVE mappings also need to be
2770 * checked against any subpool limit.
2772 if (map_chg || avoid_reserve) {
2773 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2775 vma_end_reservation(h, vma, addr);
2776 return ERR_PTR(-ENOSPC);
2780 * Even though there was no reservation in the region/reserve
2781 * map, there could be reservations associated with the
2782 * subpool that can be used. This would be indicated if the
2783 * return value of hugepage_subpool_get_pages() is zero.
2784 * However, if avoid_reserve is specified we still avoid even
2785 * the subpool reservations.
2791 /* If this allocation is not consuming a reservation, charge it now.
2793 deferred_reserve = map_chg || avoid_reserve;
2794 if (deferred_reserve) {
2795 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2796 idx, pages_per_huge_page(h), &h_cg);
2798 goto out_subpool_put;
2801 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2803 goto out_uncharge_cgroup_reservation;
2805 spin_lock_irq(&hugetlb_lock);
2807 * glb_chg is passed to indicate whether or not a page must be taken
2808 * from the global free pool (global change). gbl_chg == 0 indicates
2809 * a reservation exists for the allocation.
2811 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2813 spin_unlock_irq(&hugetlb_lock);
2814 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2816 goto out_uncharge_cgroup;
2817 spin_lock_irq(&hugetlb_lock);
2818 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2819 SetHPageRestoreReserve(page);
2820 h->resv_huge_pages--;
2822 list_add(&page->lru, &h->hugepage_activelist);
2825 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2826 /* If allocation is not consuming a reservation, also store the
2827 * hugetlb_cgroup pointer on the page.
2829 if (deferred_reserve) {
2830 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2834 spin_unlock_irq(&hugetlb_lock);
2836 hugetlb_set_page_subpool(page, spool);
2838 map_commit = vma_commit_reservation(h, vma, addr);
2839 if (unlikely(map_chg > map_commit)) {
2841 * The page was added to the reservation map between
2842 * vma_needs_reservation and vma_commit_reservation.
2843 * This indicates a race with hugetlb_reserve_pages.
2844 * Adjust for the subpool count incremented above AND
2845 * in hugetlb_reserve_pages for the same page. Also,
2846 * the reservation count added in hugetlb_reserve_pages
2847 * no longer applies.
2851 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2852 hugetlb_acct_memory(h, -rsv_adjust);
2853 if (deferred_reserve)
2854 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2855 pages_per_huge_page(h), page);
2859 out_uncharge_cgroup:
2860 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2861 out_uncharge_cgroup_reservation:
2862 if (deferred_reserve)
2863 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2866 if (map_chg || avoid_reserve)
2867 hugepage_subpool_put_pages(spool, 1);
2868 vma_end_reservation(h, vma, addr);
2869 return ERR_PTR(-ENOSPC);
2872 int alloc_bootmem_huge_page(struct hstate *h)
2873 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2874 int __alloc_bootmem_huge_page(struct hstate *h)
2876 struct huge_bootmem_page *m;
2879 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2882 addr = memblock_alloc_try_nid_raw(
2883 huge_page_size(h), huge_page_size(h),
2884 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2887 * Use the beginning of the huge page to store the
2888 * huge_bootmem_page struct (until gather_bootmem
2889 * puts them into the mem_map).
2898 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2899 /* Put them into a private list first because mem_map is not up yet */
2900 INIT_LIST_HEAD(&m->list);
2901 list_add(&m->list, &huge_boot_pages);
2907 * Put bootmem huge pages into the standard lists after mem_map is up.
2908 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2910 static void __init gather_bootmem_prealloc(void)
2912 struct huge_bootmem_page *m;
2914 list_for_each_entry(m, &huge_boot_pages, list) {
2915 struct page *page = virt_to_page(m);
2916 struct hstate *h = m->hstate;
2918 VM_BUG_ON(!hstate_is_gigantic(h));
2919 WARN_ON(page_count(page) != 1);
2920 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2921 WARN_ON(PageReserved(page));
2922 prep_new_huge_page(h, page, page_to_nid(page));
2923 put_page(page); /* add to the hugepage allocator */
2925 /* VERY unlikely inflated ref count on a tail page */
2926 free_gigantic_page(page, huge_page_order(h));
2930 * We need to restore the 'stolen' pages to totalram_pages
2931 * in order to fix confusing memory reports from free(1) and
2932 * other side-effects, like CommitLimit going negative.
2934 adjust_managed_page_count(page, pages_per_huge_page(h));
2939 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2942 nodemask_t *node_alloc_noretry;
2944 if (!hstate_is_gigantic(h)) {
2946 * Bit mask controlling how hard we retry per-node allocations.
2947 * Ignore errors as lower level routines can deal with
2948 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2949 * time, we are likely in bigger trouble.
2951 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2954 /* allocations done at boot time */
2955 node_alloc_noretry = NULL;
2958 /* bit mask controlling how hard we retry per-node allocations */
2959 if (node_alloc_noretry)
2960 nodes_clear(*node_alloc_noretry);
2962 for (i = 0; i < h->max_huge_pages; ++i) {
2963 if (hstate_is_gigantic(h)) {
2964 if (hugetlb_cma_size) {
2965 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2968 if (!alloc_bootmem_huge_page(h))
2970 } else if (!alloc_pool_huge_page(h,
2971 &node_states[N_MEMORY],
2972 node_alloc_noretry))
2976 if (i < h->max_huge_pages) {
2979 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2980 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2981 h->max_huge_pages, buf, i);
2982 h->max_huge_pages = i;
2985 kfree(node_alloc_noretry);
2988 static void __init hugetlb_init_hstates(void)
2992 for_each_hstate(h) {
2993 if (minimum_order > huge_page_order(h))
2994 minimum_order = huge_page_order(h);
2996 /* oversize hugepages were init'ed in early boot */
2997 if (!hstate_is_gigantic(h))
2998 hugetlb_hstate_alloc_pages(h);
3000 VM_BUG_ON(minimum_order == UINT_MAX);
3003 static void __init report_hugepages(void)
3007 for_each_hstate(h) {
3010 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3011 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3012 buf, h->free_huge_pages);
3016 #ifdef CONFIG_HIGHMEM
3017 static void try_to_free_low(struct hstate *h, unsigned long count,
3018 nodemask_t *nodes_allowed)
3021 LIST_HEAD(page_list);
3023 lockdep_assert_held(&hugetlb_lock);
3024 if (hstate_is_gigantic(h))
3028 * Collect pages to be freed on a list, and free after dropping lock
3030 for_each_node_mask(i, *nodes_allowed) {
3031 struct page *page, *next;
3032 struct list_head *freel = &h->hugepage_freelists[i];
3033 list_for_each_entry_safe(page, next, freel, lru) {
3034 if (count >= h->nr_huge_pages)
3036 if (PageHighMem(page))
3038 remove_hugetlb_page(h, page, false);
3039 list_add(&page->lru, &page_list);
3044 spin_unlock_irq(&hugetlb_lock);
3045 update_and_free_pages_bulk(h, &page_list);
3046 spin_lock_irq(&hugetlb_lock);
3049 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3050 nodemask_t *nodes_allowed)
3056 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3057 * balanced by operating on them in a round-robin fashion.
3058 * Returns 1 if an adjustment was made.
3060 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3065 lockdep_assert_held(&hugetlb_lock);
3066 VM_BUG_ON(delta != -1 && delta != 1);
3069 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3070 if (h->surplus_huge_pages_node[node])
3074 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3075 if (h->surplus_huge_pages_node[node] <
3076 h->nr_huge_pages_node[node])
3083 h->surplus_huge_pages += delta;
3084 h->surplus_huge_pages_node[node] += delta;
3088 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3089 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3090 nodemask_t *nodes_allowed)
3092 unsigned long min_count, ret;
3094 LIST_HEAD(page_list);
3095 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3098 * Bit mask controlling how hard we retry per-node allocations.
3099 * If we can not allocate the bit mask, do not attempt to allocate
3100 * the requested huge pages.
3102 if (node_alloc_noretry)
3103 nodes_clear(*node_alloc_noretry);
3108 * resize_lock mutex prevents concurrent adjustments to number of
3109 * pages in hstate via the proc/sysfs interfaces.
3111 mutex_lock(&h->resize_lock);
3112 flush_free_hpage_work(h);
3113 spin_lock_irq(&hugetlb_lock);
3116 * Check for a node specific request.
3117 * Changing node specific huge page count may require a corresponding
3118 * change to the global count. In any case, the passed node mask
3119 * (nodes_allowed) will restrict alloc/free to the specified node.
3121 if (nid != NUMA_NO_NODE) {
3122 unsigned long old_count = count;
3124 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3126 * User may have specified a large count value which caused the
3127 * above calculation to overflow. In this case, they wanted
3128 * to allocate as many huge pages as possible. Set count to
3129 * largest possible value to align with their intention.
3131 if (count < old_count)
3136 * Gigantic pages runtime allocation depend on the capability for large
3137 * page range allocation.
3138 * If the system does not provide this feature, return an error when
3139 * the user tries to allocate gigantic pages but let the user free the
3140 * boottime allocated gigantic pages.
3142 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3143 if (count > persistent_huge_pages(h)) {
3144 spin_unlock_irq(&hugetlb_lock);
3145 mutex_unlock(&h->resize_lock);
3146 NODEMASK_FREE(node_alloc_noretry);
3149 /* Fall through to decrease pool */
3153 * Increase the pool size
3154 * First take pages out of surplus state. Then make up the
3155 * remaining difference by allocating fresh huge pages.
3157 * We might race with alloc_surplus_huge_page() here and be unable
3158 * to convert a surplus huge page to a normal huge page. That is
3159 * not critical, though, it just means the overall size of the
3160 * pool might be one hugepage larger than it needs to be, but
3161 * within all the constraints specified by the sysctls.
3163 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3164 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3168 while (count > persistent_huge_pages(h)) {
3170 * If this allocation races such that we no longer need the
3171 * page, free_huge_page will handle it by freeing the page
3172 * and reducing the surplus.
3174 spin_unlock_irq(&hugetlb_lock);
3176 /* yield cpu to avoid soft lockup */
3179 ret = alloc_pool_huge_page(h, nodes_allowed,
3180 node_alloc_noretry);
3181 spin_lock_irq(&hugetlb_lock);
3185 /* Bail for signals. Probably ctrl-c from user */
3186 if (signal_pending(current))
3191 * Decrease the pool size
3192 * First return free pages to the buddy allocator (being careful
3193 * to keep enough around to satisfy reservations). Then place
3194 * pages into surplus state as needed so the pool will shrink
3195 * to the desired size as pages become free.
3197 * By placing pages into the surplus state independent of the
3198 * overcommit value, we are allowing the surplus pool size to
3199 * exceed overcommit. There are few sane options here. Since
3200 * alloc_surplus_huge_page() is checking the global counter,
3201 * though, we'll note that we're not allowed to exceed surplus
3202 * and won't grow the pool anywhere else. Not until one of the
3203 * sysctls are changed, or the surplus pages go out of use.
3205 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3206 min_count = max(count, min_count);
3207 try_to_free_low(h, min_count, nodes_allowed);
3210 * Collect pages to be removed on list without dropping lock
3212 while (min_count < persistent_huge_pages(h)) {
3213 page = remove_pool_huge_page(h, nodes_allowed, 0);
3217 list_add(&page->lru, &page_list);
3219 /* free the pages after dropping lock */
3220 spin_unlock_irq(&hugetlb_lock);
3221 update_and_free_pages_bulk(h, &page_list);
3222 flush_free_hpage_work(h);
3223 spin_lock_irq(&hugetlb_lock);
3225 while (count < persistent_huge_pages(h)) {
3226 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3230 h->max_huge_pages = persistent_huge_pages(h);
3231 spin_unlock_irq(&hugetlb_lock);
3232 mutex_unlock(&h->resize_lock);
3234 NODEMASK_FREE(node_alloc_noretry);
3239 #define HSTATE_ATTR_RO(_name) \
3240 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3242 #define HSTATE_ATTR(_name) \
3243 static struct kobj_attribute _name##_attr = \
3244 __ATTR(_name, 0644, _name##_show, _name##_store)
3246 static struct kobject *hugepages_kobj;
3247 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3249 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3251 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3255 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3256 if (hstate_kobjs[i] == kobj) {
3258 *nidp = NUMA_NO_NODE;
3262 return kobj_to_node_hstate(kobj, nidp);
3265 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3266 struct kobj_attribute *attr, char *buf)
3269 unsigned long nr_huge_pages;
3272 h = kobj_to_hstate(kobj, &nid);
3273 if (nid == NUMA_NO_NODE)
3274 nr_huge_pages = h->nr_huge_pages;
3276 nr_huge_pages = h->nr_huge_pages_node[nid];
3278 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3281 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3282 struct hstate *h, int nid,
3283 unsigned long count, size_t len)
3286 nodemask_t nodes_allowed, *n_mask;
3288 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3291 if (nid == NUMA_NO_NODE) {
3293 * global hstate attribute
3295 if (!(obey_mempolicy &&
3296 init_nodemask_of_mempolicy(&nodes_allowed)))
3297 n_mask = &node_states[N_MEMORY];
3299 n_mask = &nodes_allowed;
3302 * Node specific request. count adjustment happens in
3303 * set_max_huge_pages() after acquiring hugetlb_lock.
3305 init_nodemask_of_node(&nodes_allowed, nid);
3306 n_mask = &nodes_allowed;
3309 err = set_max_huge_pages(h, count, nid, n_mask);
3311 return err ? err : len;
3314 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3315 struct kobject *kobj, const char *buf,
3319 unsigned long count;
3323 err = kstrtoul(buf, 10, &count);
3327 h = kobj_to_hstate(kobj, &nid);
3328 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3331 static ssize_t nr_hugepages_show(struct kobject *kobj,
3332 struct kobj_attribute *attr, char *buf)
3334 return nr_hugepages_show_common(kobj, attr, buf);
3337 static ssize_t nr_hugepages_store(struct kobject *kobj,
3338 struct kobj_attribute *attr, const char *buf, size_t len)
3340 return nr_hugepages_store_common(false, kobj, buf, len);
3342 HSTATE_ATTR(nr_hugepages);
3347 * hstate attribute for optionally mempolicy-based constraint on persistent
3348 * huge page alloc/free.
3350 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3351 struct kobj_attribute *attr,
3354 return nr_hugepages_show_common(kobj, attr, buf);
3357 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3358 struct kobj_attribute *attr, const char *buf, size_t len)
3360 return nr_hugepages_store_common(true, kobj, buf, len);
3362 HSTATE_ATTR(nr_hugepages_mempolicy);
3366 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3367 struct kobj_attribute *attr, char *buf)
3369 struct hstate *h = kobj_to_hstate(kobj, NULL);
3370 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3373 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3374 struct kobj_attribute *attr, const char *buf, size_t count)
3377 unsigned long input;
3378 struct hstate *h = kobj_to_hstate(kobj, NULL);
3380 if (hstate_is_gigantic(h))
3383 err = kstrtoul(buf, 10, &input);
3387 spin_lock_irq(&hugetlb_lock);
3388 h->nr_overcommit_huge_pages = input;
3389 spin_unlock_irq(&hugetlb_lock);
3393 HSTATE_ATTR(nr_overcommit_hugepages);
3395 static ssize_t free_hugepages_show(struct kobject *kobj,
3396 struct kobj_attribute *attr, char *buf)
3399 unsigned long free_huge_pages;
3402 h = kobj_to_hstate(kobj, &nid);
3403 if (nid == NUMA_NO_NODE)
3404 free_huge_pages = h->free_huge_pages;
3406 free_huge_pages = h->free_huge_pages_node[nid];
3408 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3410 HSTATE_ATTR_RO(free_hugepages);
3412 static ssize_t resv_hugepages_show(struct kobject *kobj,
3413 struct kobj_attribute *attr, char *buf)
3415 struct hstate *h = kobj_to_hstate(kobj, NULL);
3416 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3418 HSTATE_ATTR_RO(resv_hugepages);
3420 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3421 struct kobj_attribute *attr, char *buf)
3424 unsigned long surplus_huge_pages;
3427 h = kobj_to_hstate(kobj, &nid);
3428 if (nid == NUMA_NO_NODE)
3429 surplus_huge_pages = h->surplus_huge_pages;
3431 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3433 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3435 HSTATE_ATTR_RO(surplus_hugepages);
3437 static struct attribute *hstate_attrs[] = {
3438 &nr_hugepages_attr.attr,
3439 &nr_overcommit_hugepages_attr.attr,
3440 &free_hugepages_attr.attr,
3441 &resv_hugepages_attr.attr,
3442 &surplus_hugepages_attr.attr,
3444 &nr_hugepages_mempolicy_attr.attr,
3449 static const struct attribute_group hstate_attr_group = {
3450 .attrs = hstate_attrs,
3453 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3454 struct kobject **hstate_kobjs,
3455 const struct attribute_group *hstate_attr_group)
3458 int hi = hstate_index(h);
3460 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3461 if (!hstate_kobjs[hi])
3464 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3466 kobject_put(hstate_kobjs[hi]);
3467 hstate_kobjs[hi] = NULL;
3473 static void __init hugetlb_sysfs_init(void)
3478 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3479 if (!hugepages_kobj)
3482 for_each_hstate(h) {
3483 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3484 hstate_kobjs, &hstate_attr_group);
3486 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3493 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3494 * with node devices in node_devices[] using a parallel array. The array
3495 * index of a node device or _hstate == node id.
3496 * This is here to avoid any static dependency of the node device driver, in
3497 * the base kernel, on the hugetlb module.
3499 struct node_hstate {
3500 struct kobject *hugepages_kobj;
3501 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3503 static struct node_hstate node_hstates[MAX_NUMNODES];
3506 * A subset of global hstate attributes for node devices
3508 static struct attribute *per_node_hstate_attrs[] = {
3509 &nr_hugepages_attr.attr,
3510 &free_hugepages_attr.attr,
3511 &surplus_hugepages_attr.attr,
3515 static const struct attribute_group per_node_hstate_attr_group = {
3516 .attrs = per_node_hstate_attrs,
3520 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3521 * Returns node id via non-NULL nidp.
3523 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3527 for (nid = 0; nid < nr_node_ids; nid++) {
3528 struct node_hstate *nhs = &node_hstates[nid];
3530 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3531 if (nhs->hstate_kobjs[i] == kobj) {
3543 * Unregister hstate attributes from a single node device.
3544 * No-op if no hstate attributes attached.
3546 static void hugetlb_unregister_node(struct node *node)
3549 struct node_hstate *nhs = &node_hstates[node->dev.id];
3551 if (!nhs->hugepages_kobj)
3552 return; /* no hstate attributes */
3554 for_each_hstate(h) {
3555 int idx = hstate_index(h);
3556 if (nhs->hstate_kobjs[idx]) {
3557 kobject_put(nhs->hstate_kobjs[idx]);
3558 nhs->hstate_kobjs[idx] = NULL;
3562 kobject_put(nhs->hugepages_kobj);
3563 nhs->hugepages_kobj = NULL;
3568 * Register hstate attributes for a single node device.
3569 * No-op if attributes already registered.
3571 static void hugetlb_register_node(struct node *node)
3574 struct node_hstate *nhs = &node_hstates[node->dev.id];
3577 if (nhs->hugepages_kobj)
3578 return; /* already allocated */
3580 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3582 if (!nhs->hugepages_kobj)
3585 for_each_hstate(h) {
3586 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3588 &per_node_hstate_attr_group);
3590 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3591 h->name, node->dev.id);
3592 hugetlb_unregister_node(node);
3599 * hugetlb init time: register hstate attributes for all registered node
3600 * devices of nodes that have memory. All on-line nodes should have
3601 * registered their associated device by this time.
3603 static void __init hugetlb_register_all_nodes(void)
3607 for_each_node_state(nid, N_MEMORY) {
3608 struct node *node = node_devices[nid];
3609 if (node->dev.id == nid)
3610 hugetlb_register_node(node);
3614 * Let the node device driver know we're here so it can
3615 * [un]register hstate attributes on node hotplug.
3617 register_hugetlbfs_with_node(hugetlb_register_node,
3618 hugetlb_unregister_node);
3620 #else /* !CONFIG_NUMA */
3622 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3630 static void hugetlb_register_all_nodes(void) { }
3634 static int __init hugetlb_init(void)
3638 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3641 if (!hugepages_supported()) {
3642 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3643 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3648 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3649 * architectures depend on setup being done here.
3651 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3652 if (!parsed_default_hugepagesz) {
3654 * If we did not parse a default huge page size, set
3655 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3656 * number of huge pages for this default size was implicitly
3657 * specified, set that here as well.
3658 * Note that the implicit setting will overwrite an explicit
3659 * setting. A warning will be printed in this case.
3661 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3662 if (default_hstate_max_huge_pages) {
3663 if (default_hstate.max_huge_pages) {
3666 string_get_size(huge_page_size(&default_hstate),
3667 1, STRING_UNITS_2, buf, 32);
3668 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3669 default_hstate.max_huge_pages, buf);
3670 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3671 default_hstate_max_huge_pages);
3673 default_hstate.max_huge_pages =
3674 default_hstate_max_huge_pages;
3678 hugetlb_cma_check();
3679 hugetlb_init_hstates();
3680 gather_bootmem_prealloc();
3683 hugetlb_sysfs_init();
3684 hugetlb_register_all_nodes();
3685 hugetlb_cgroup_file_init();
3688 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3690 num_fault_mutexes = 1;
3692 hugetlb_fault_mutex_table =
3693 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3695 BUG_ON(!hugetlb_fault_mutex_table);
3697 for (i = 0; i < num_fault_mutexes; i++)
3698 mutex_init(&hugetlb_fault_mutex_table[i]);
3701 subsys_initcall(hugetlb_init);
3703 /* Overwritten by architectures with more huge page sizes */
3704 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3706 return size == HPAGE_SIZE;
3709 void __init hugetlb_add_hstate(unsigned int order)
3714 if (size_to_hstate(PAGE_SIZE << order)) {
3717 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3719 h = &hstates[hugetlb_max_hstate++];
3720 mutex_init(&h->resize_lock);
3722 h->mask = ~(huge_page_size(h) - 1);
3723 for (i = 0; i < MAX_NUMNODES; ++i)
3724 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3725 INIT_LIST_HEAD(&h->hugepage_activelist);
3726 h->next_nid_to_alloc = first_memory_node;
3727 h->next_nid_to_free = first_memory_node;
3728 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3729 huge_page_size(h)/1024);
3730 hugetlb_vmemmap_init(h);
3736 * hugepages command line processing
3737 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3738 * specification. If not, ignore the hugepages value. hugepages can also
3739 * be the first huge page command line option in which case it implicitly
3740 * specifies the number of huge pages for the default size.
3742 static int __init hugepages_setup(char *s)
3745 static unsigned long *last_mhp;
3747 if (!parsed_valid_hugepagesz) {
3748 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3749 parsed_valid_hugepagesz = true;
3754 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3755 * yet, so this hugepages= parameter goes to the "default hstate".
3756 * Otherwise, it goes with the previously parsed hugepagesz or
3757 * default_hugepagesz.
3759 else if (!hugetlb_max_hstate)
3760 mhp = &default_hstate_max_huge_pages;
3762 mhp = &parsed_hstate->max_huge_pages;
3764 if (mhp == last_mhp) {
3765 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3769 if (sscanf(s, "%lu", mhp) <= 0)
3773 * Global state is always initialized later in hugetlb_init.
3774 * But we need to allocate gigantic hstates here early to still
3775 * use the bootmem allocator.
3777 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3778 hugetlb_hstate_alloc_pages(parsed_hstate);
3784 __setup("hugepages=", hugepages_setup);
3787 * hugepagesz command line processing
3788 * A specific huge page size can only be specified once with hugepagesz.
3789 * hugepagesz is followed by hugepages on the command line. The global
3790 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3791 * hugepagesz argument was valid.
3793 static int __init hugepagesz_setup(char *s)
3798 parsed_valid_hugepagesz = false;
3799 size = (unsigned long)memparse(s, NULL);
3801 if (!arch_hugetlb_valid_size(size)) {
3802 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3806 h = size_to_hstate(size);
3809 * hstate for this size already exists. This is normally
3810 * an error, but is allowed if the existing hstate is the
3811 * default hstate. More specifically, it is only allowed if
3812 * the number of huge pages for the default hstate was not
3813 * previously specified.
3815 if (!parsed_default_hugepagesz || h != &default_hstate ||
3816 default_hstate.max_huge_pages) {
3817 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3822 * No need to call hugetlb_add_hstate() as hstate already
3823 * exists. But, do set parsed_hstate so that a following
3824 * hugepages= parameter will be applied to this hstate.
3827 parsed_valid_hugepagesz = true;
3831 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3832 parsed_valid_hugepagesz = true;
3835 __setup("hugepagesz=", hugepagesz_setup);
3838 * default_hugepagesz command line input
3839 * Only one instance of default_hugepagesz allowed on command line.
3841 static int __init default_hugepagesz_setup(char *s)
3845 parsed_valid_hugepagesz = false;
3846 if (parsed_default_hugepagesz) {
3847 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3851 size = (unsigned long)memparse(s, NULL);
3853 if (!arch_hugetlb_valid_size(size)) {
3854 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3858 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3859 parsed_valid_hugepagesz = true;
3860 parsed_default_hugepagesz = true;
3861 default_hstate_idx = hstate_index(size_to_hstate(size));
3864 * The number of default huge pages (for this size) could have been
3865 * specified as the first hugetlb parameter: hugepages=X. If so,
3866 * then default_hstate_max_huge_pages is set. If the default huge
3867 * page size is gigantic (>= MAX_ORDER), then the pages must be
3868 * allocated here from bootmem allocator.
3870 if (default_hstate_max_huge_pages) {
3871 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3872 if (hstate_is_gigantic(&default_hstate))
3873 hugetlb_hstate_alloc_pages(&default_hstate);
3874 default_hstate_max_huge_pages = 0;
3879 __setup("default_hugepagesz=", default_hugepagesz_setup);
3881 static unsigned int allowed_mems_nr(struct hstate *h)
3884 unsigned int nr = 0;
3885 nodemask_t *mpol_allowed;
3886 unsigned int *array = h->free_huge_pages_node;
3887 gfp_t gfp_mask = htlb_alloc_mask(h);
3889 mpol_allowed = policy_nodemask_current(gfp_mask);
3891 for_each_node_mask(node, cpuset_current_mems_allowed) {
3892 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3899 #ifdef CONFIG_SYSCTL
3900 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3901 void *buffer, size_t *length,
3902 loff_t *ppos, unsigned long *out)
3904 struct ctl_table dup_table;
3907 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3908 * can duplicate the @table and alter the duplicate of it.
3911 dup_table.data = out;
3913 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3916 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3917 struct ctl_table *table, int write,
3918 void *buffer, size_t *length, loff_t *ppos)
3920 struct hstate *h = &default_hstate;
3921 unsigned long tmp = h->max_huge_pages;
3924 if (!hugepages_supported())
3927 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3933 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3934 NUMA_NO_NODE, tmp, *length);
3939 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3940 void *buffer, size_t *length, loff_t *ppos)
3943 return hugetlb_sysctl_handler_common(false, table, write,
3944 buffer, length, ppos);
3948 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3949 void *buffer, size_t *length, loff_t *ppos)
3951 return hugetlb_sysctl_handler_common(true, table, write,
3952 buffer, length, ppos);
3954 #endif /* CONFIG_NUMA */
3956 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3957 void *buffer, size_t *length, loff_t *ppos)
3959 struct hstate *h = &default_hstate;
3963 if (!hugepages_supported())
3966 tmp = h->nr_overcommit_huge_pages;
3968 if (write && hstate_is_gigantic(h))
3971 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3977 spin_lock_irq(&hugetlb_lock);
3978 h->nr_overcommit_huge_pages = tmp;
3979 spin_unlock_irq(&hugetlb_lock);
3985 #endif /* CONFIG_SYSCTL */
3987 void hugetlb_report_meminfo(struct seq_file *m)
3990 unsigned long total = 0;
3992 if (!hugepages_supported())
3995 for_each_hstate(h) {
3996 unsigned long count = h->nr_huge_pages;
3998 total += huge_page_size(h) * count;
4000 if (h == &default_hstate)
4002 "HugePages_Total: %5lu\n"
4003 "HugePages_Free: %5lu\n"
4004 "HugePages_Rsvd: %5lu\n"
4005 "HugePages_Surp: %5lu\n"
4006 "Hugepagesize: %8lu kB\n",
4010 h->surplus_huge_pages,
4011 huge_page_size(h) / SZ_1K);
4014 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4017 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4019 struct hstate *h = &default_hstate;
4021 if (!hugepages_supported())
4024 return sysfs_emit_at(buf, len,
4025 "Node %d HugePages_Total: %5u\n"
4026 "Node %d HugePages_Free: %5u\n"
4027 "Node %d HugePages_Surp: %5u\n",
4028 nid, h->nr_huge_pages_node[nid],
4029 nid, h->free_huge_pages_node[nid],
4030 nid, h->surplus_huge_pages_node[nid]);
4033 void hugetlb_show_meminfo(void)
4038 if (!hugepages_supported())
4041 for_each_node_state(nid, N_MEMORY)
4043 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4045 h->nr_huge_pages_node[nid],
4046 h->free_huge_pages_node[nid],
4047 h->surplus_huge_pages_node[nid],
4048 huge_page_size(h) / SZ_1K);
4051 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4053 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4054 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4057 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4058 unsigned long hugetlb_total_pages(void)
4061 unsigned long nr_total_pages = 0;
4064 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4065 return nr_total_pages;
4068 static int hugetlb_acct_memory(struct hstate *h, long delta)
4075 spin_lock_irq(&hugetlb_lock);
4077 * When cpuset is configured, it breaks the strict hugetlb page
4078 * reservation as the accounting is done on a global variable. Such
4079 * reservation is completely rubbish in the presence of cpuset because
4080 * the reservation is not checked against page availability for the
4081 * current cpuset. Application can still potentially OOM'ed by kernel
4082 * with lack of free htlb page in cpuset that the task is in.
4083 * Attempt to enforce strict accounting with cpuset is almost
4084 * impossible (or too ugly) because cpuset is too fluid that
4085 * task or memory node can be dynamically moved between cpusets.
4087 * The change of semantics for shared hugetlb mapping with cpuset is
4088 * undesirable. However, in order to preserve some of the semantics,
4089 * we fall back to check against current free page availability as
4090 * a best attempt and hopefully to minimize the impact of changing
4091 * semantics that cpuset has.
4093 * Apart from cpuset, we also have memory policy mechanism that
4094 * also determines from which node the kernel will allocate memory
4095 * in a NUMA system. So similar to cpuset, we also should consider
4096 * the memory policy of the current task. Similar to the description
4100 if (gather_surplus_pages(h, delta) < 0)
4103 if (delta > allowed_mems_nr(h)) {
4104 return_unused_surplus_pages(h, delta);
4111 return_unused_surplus_pages(h, (unsigned long) -delta);
4114 spin_unlock_irq(&hugetlb_lock);
4118 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4120 struct resv_map *resv = vma_resv_map(vma);
4123 * This new VMA should share its siblings reservation map if present.
4124 * The VMA will only ever have a valid reservation map pointer where
4125 * it is being copied for another still existing VMA. As that VMA
4126 * has a reference to the reservation map it cannot disappear until
4127 * after this open call completes. It is therefore safe to take a
4128 * new reference here without additional locking.
4130 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4131 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4132 kref_get(&resv->refs);
4136 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4138 struct hstate *h = hstate_vma(vma);
4139 struct resv_map *resv = vma_resv_map(vma);
4140 struct hugepage_subpool *spool = subpool_vma(vma);
4141 unsigned long reserve, start, end;
4144 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4147 start = vma_hugecache_offset(h, vma, vma->vm_start);
4148 end = vma_hugecache_offset(h, vma, vma->vm_end);
4150 reserve = (end - start) - region_count(resv, start, end);
4151 hugetlb_cgroup_uncharge_counter(resv, start, end);
4154 * Decrement reserve counts. The global reserve count may be
4155 * adjusted if the subpool has a minimum size.
4157 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4158 hugetlb_acct_memory(h, -gbl_reserve);
4161 kref_put(&resv->refs, resv_map_release);
4164 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4166 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4170 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4171 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4172 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4174 if (addr & ~PUD_MASK) {
4176 * hugetlb_vm_op_split is called right before we attempt to
4177 * split the VMA. We will need to unshare PMDs in the old and
4178 * new VMAs, so let's unshare before we split.
4180 unsigned long floor = addr & PUD_MASK;
4181 unsigned long ceil = floor + PUD_SIZE;
4183 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4184 hugetlb_unshare_pmds(vma, floor, ceil);
4190 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4192 return huge_page_size(hstate_vma(vma));
4196 * We cannot handle pagefaults against hugetlb pages at all. They cause
4197 * handle_mm_fault() to try to instantiate regular-sized pages in the
4198 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4201 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4208 * When a new function is introduced to vm_operations_struct and added
4209 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4210 * This is because under System V memory model, mappings created via
4211 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4212 * their original vm_ops are overwritten with shm_vm_ops.
4214 const struct vm_operations_struct hugetlb_vm_ops = {
4215 .fault = hugetlb_vm_op_fault,
4216 .open = hugetlb_vm_op_open,
4217 .close = hugetlb_vm_op_close,
4218 .may_split = hugetlb_vm_op_split,
4219 .pagesize = hugetlb_vm_op_pagesize,
4222 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4226 unsigned int shift = huge_page_shift(hstate_vma(vma));
4229 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4230 vma->vm_page_prot)));
4232 entry = huge_pte_wrprotect(mk_huge_pte(page,
4233 vma->vm_page_prot));
4235 entry = pte_mkyoung(entry);
4236 entry = pte_mkhuge(entry);
4237 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4242 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4243 unsigned long address, pte_t *ptep)
4247 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4248 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4249 update_mmu_cache(vma, address, ptep);
4252 bool is_hugetlb_entry_migration(pte_t pte)
4256 if (huge_pte_none(pte) || pte_present(pte))
4258 swp = pte_to_swp_entry(pte);
4259 if (is_migration_entry(swp))
4265 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4269 if (huge_pte_none(pte) || pte_present(pte))
4271 swp = pte_to_swp_entry(pte);
4272 if (is_hwpoison_entry(swp))
4279 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4280 struct page *new_page)
4282 __SetPageUptodate(new_page);
4283 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4284 hugepage_add_new_anon_rmap(new_page, vma, addr);
4285 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4286 ClearHPageRestoreReserve(new_page);
4287 SetHPageMigratable(new_page);
4290 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4291 struct vm_area_struct *vma)
4293 pte_t *src_pte, *dst_pte, entry, dst_entry;
4294 struct page *ptepage;
4296 bool cow = is_cow_mapping(vma->vm_flags);
4297 struct hstate *h = hstate_vma(vma);
4298 unsigned long sz = huge_page_size(h);
4299 unsigned long npages = pages_per_huge_page(h);
4300 struct address_space *mapping = vma->vm_file->f_mapping;
4301 struct mmu_notifier_range range;
4305 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4308 mmu_notifier_invalidate_range_start(&range);
4311 * For shared mappings i_mmap_rwsem must be held to call
4312 * huge_pte_alloc, otherwise the returned ptep could go
4313 * away if part of a shared pmd and another thread calls
4316 i_mmap_lock_read(mapping);
4319 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4320 spinlock_t *src_ptl, *dst_ptl;
4321 src_pte = huge_pte_offset(src, addr, sz);
4324 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4331 * If the pagetables are shared don't copy or take references.
4332 * dst_pte == src_pte is the common case of src/dest sharing.
4334 * However, src could have 'unshared' and dst shares with
4335 * another vma. If dst_pte !none, this implies sharing.
4336 * Check here before taking page table lock, and once again
4337 * after taking the lock below.
4339 dst_entry = huge_ptep_get(dst_pte);
4340 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4343 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4344 src_ptl = huge_pte_lockptr(h, src, src_pte);
4345 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4346 entry = huge_ptep_get(src_pte);
4347 dst_entry = huge_ptep_get(dst_pte);
4349 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4351 * Skip if src entry none. Also, skip in the
4352 * unlikely case dst entry !none as this implies
4353 * sharing with another vma.
4356 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4357 is_hugetlb_entry_hwpoisoned(entry))) {
4358 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4360 if (is_writable_migration_entry(swp_entry) && cow) {
4362 * COW mappings require pages in both
4363 * parent and child to be set to read.
4365 swp_entry = make_readable_migration_entry(
4366 swp_offset(swp_entry));
4367 entry = swp_entry_to_pte(swp_entry);
4368 set_huge_swap_pte_at(src, addr, src_pte,
4371 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4373 entry = huge_ptep_get(src_pte);
4374 ptepage = pte_page(entry);
4378 * This is a rare case where we see pinned hugetlb
4379 * pages while they're prone to COW. We need to do the
4380 * COW earlier during fork.
4382 * When pre-allocating the page or copying data, we
4383 * need to be without the pgtable locks since we could
4384 * sleep during the process.
4386 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4387 pte_t src_pte_old = entry;
4390 spin_unlock(src_ptl);
4391 spin_unlock(dst_ptl);
4392 /* Do not use reserve as it's private owned */
4393 new = alloc_huge_page(vma, addr, 1);
4399 copy_user_huge_page(new, ptepage, addr, vma,
4403 /* Install the new huge page if src pte stable */
4404 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4405 src_ptl = huge_pte_lockptr(h, src, src_pte);
4406 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4407 entry = huge_ptep_get(src_pte);
4408 if (!pte_same(src_pte_old, entry)) {
4409 restore_reserve_on_error(h, vma, addr,
4412 /* dst_entry won't change as in child */
4415 hugetlb_install_page(vma, dst_pte, addr, new);
4416 spin_unlock(src_ptl);
4417 spin_unlock(dst_ptl);
4423 * No need to notify as we are downgrading page
4424 * table protection not changing it to point
4427 * See Documentation/vm/mmu_notifier.rst
4429 huge_ptep_set_wrprotect(src, addr, src_pte);
4430 entry = huge_pte_wrprotect(entry);
4433 page_dup_rmap(ptepage, true);
4434 set_huge_pte_at(dst, addr, dst_pte, entry);
4435 hugetlb_count_add(npages, dst);
4437 spin_unlock(src_ptl);
4438 spin_unlock(dst_ptl);
4442 mmu_notifier_invalidate_range_end(&range);
4444 i_mmap_unlock_read(mapping);
4449 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4450 unsigned long start, unsigned long end,
4451 struct page *ref_page)
4453 struct mm_struct *mm = vma->vm_mm;
4454 unsigned long address;
4459 struct hstate *h = hstate_vma(vma);
4460 unsigned long sz = huge_page_size(h);
4461 struct mmu_notifier_range range;
4462 bool force_flush = false;
4464 WARN_ON(!is_vm_hugetlb_page(vma));
4465 BUG_ON(start & ~huge_page_mask(h));
4466 BUG_ON(end & ~huge_page_mask(h));
4469 * This is a hugetlb vma, all the pte entries should point
4472 tlb_change_page_size(tlb, sz);
4473 tlb_start_vma(tlb, vma);
4476 * If sharing possible, alert mmu notifiers of worst case.
4478 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4480 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4481 mmu_notifier_invalidate_range_start(&range);
4483 for (; address < end; address += sz) {
4484 ptep = huge_pte_offset(mm, address, sz);
4488 ptl = huge_pte_lock(h, mm, ptep);
4489 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4491 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
4496 pte = huge_ptep_get(ptep);
4497 if (huge_pte_none(pte)) {
4503 * Migrating hugepage or HWPoisoned hugepage is already
4504 * unmapped and its refcount is dropped, so just clear pte here.
4506 if (unlikely(!pte_present(pte))) {
4507 huge_pte_clear(mm, address, ptep, sz);
4512 page = pte_page(pte);
4514 * If a reference page is supplied, it is because a specific
4515 * page is being unmapped, not a range. Ensure the page we
4516 * are about to unmap is the actual page of interest.
4519 if (page != ref_page) {
4524 * Mark the VMA as having unmapped its page so that
4525 * future faults in this VMA will fail rather than
4526 * looking like data was lost
4528 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4531 pte = huge_ptep_get_and_clear(mm, address, ptep);
4532 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4533 if (huge_pte_dirty(pte))
4534 set_page_dirty(page);
4536 hugetlb_count_sub(pages_per_huge_page(h), mm);
4537 page_remove_rmap(page, true);
4540 tlb_remove_page_size(tlb, page, huge_page_size(h));
4542 * Bail out after unmapping reference page if supplied
4547 mmu_notifier_invalidate_range_end(&range);
4548 tlb_end_vma(tlb, vma);
4551 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
4552 * could defer the flush until now, since by holding i_mmap_rwsem we
4553 * guaranteed that the last refernece would not be dropped. But we must
4554 * do the flushing before we return, as otherwise i_mmap_rwsem will be
4555 * dropped and the last reference to the shared PMDs page might be
4558 * In theory we could defer the freeing of the PMD pages as well, but
4559 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
4560 * detect sharing, so we cannot defer the release of the page either.
4561 * Instead, do flush now.
4564 tlb_flush_mmu_tlbonly(tlb);
4567 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4568 struct vm_area_struct *vma, unsigned long start,
4569 unsigned long end, struct page *ref_page)
4571 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4574 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4575 * test will fail on a vma being torn down, and not grab a page table
4576 * on its way out. We're lucky that the flag has such an appropriate
4577 * name, and can in fact be safely cleared here. We could clear it
4578 * before the __unmap_hugepage_range above, but all that's necessary
4579 * is to clear it before releasing the i_mmap_rwsem. This works
4580 * because in the context this is called, the VMA is about to be
4581 * destroyed and the i_mmap_rwsem is held.
4583 vma->vm_flags &= ~VM_MAYSHARE;
4586 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4587 unsigned long end, struct page *ref_page)
4589 struct mmu_gather tlb;
4591 tlb_gather_mmu(&tlb, vma->vm_mm);
4592 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4593 tlb_finish_mmu(&tlb);
4597 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4598 * mapping it owns the reserve page for. The intention is to unmap the page
4599 * from other VMAs and let the children be SIGKILLed if they are faulting the
4602 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4603 struct page *page, unsigned long address)
4605 struct hstate *h = hstate_vma(vma);
4606 struct vm_area_struct *iter_vma;
4607 struct address_space *mapping;
4611 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4612 * from page cache lookup which is in HPAGE_SIZE units.
4614 address = address & huge_page_mask(h);
4615 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4617 mapping = vma->vm_file->f_mapping;
4620 * Take the mapping lock for the duration of the table walk. As
4621 * this mapping should be shared between all the VMAs,
4622 * __unmap_hugepage_range() is called as the lock is already held
4624 i_mmap_lock_write(mapping);
4625 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4626 /* Do not unmap the current VMA */
4627 if (iter_vma == vma)
4631 * Shared VMAs have their own reserves and do not affect
4632 * MAP_PRIVATE accounting but it is possible that a shared
4633 * VMA is using the same page so check and skip such VMAs.
4635 if (iter_vma->vm_flags & VM_MAYSHARE)
4639 * Unmap the page from other VMAs without their own reserves.
4640 * They get marked to be SIGKILLed if they fault in these
4641 * areas. This is because a future no-page fault on this VMA
4642 * could insert a zeroed page instead of the data existing
4643 * from the time of fork. This would look like data corruption
4645 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4646 unmap_hugepage_range(iter_vma, address,
4647 address + huge_page_size(h), page);
4649 i_mmap_unlock_write(mapping);
4653 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4654 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4655 * cannot race with other handlers or page migration.
4656 * Keep the pte_same checks anyway to make transition from the mutex easier.
4658 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4659 unsigned long address, pte_t *ptep,
4660 struct page *pagecache_page, spinlock_t *ptl)
4663 struct hstate *h = hstate_vma(vma);
4664 struct page *old_page, *new_page;
4665 int outside_reserve = 0;
4667 unsigned long haddr = address & huge_page_mask(h);
4668 struct mmu_notifier_range range;
4670 pte = huge_ptep_get(ptep);
4671 old_page = pte_page(pte);
4674 /* If no-one else is actually using this page, avoid the copy
4675 * and just make the page writable */
4676 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4677 page_move_anon_rmap(old_page, vma);
4678 set_huge_ptep_writable(vma, haddr, ptep);
4683 * If the process that created a MAP_PRIVATE mapping is about to
4684 * perform a COW due to a shared page count, attempt to satisfy
4685 * the allocation without using the existing reserves. The pagecache
4686 * page is used to determine if the reserve at this address was
4687 * consumed or not. If reserves were used, a partial faulted mapping
4688 * at the time of fork() could consume its reserves on COW instead
4689 * of the full address range.
4691 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4692 old_page != pagecache_page)
4693 outside_reserve = 1;
4698 * Drop page table lock as buddy allocator may be called. It will
4699 * be acquired again before returning to the caller, as expected.
4702 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4704 if (IS_ERR(new_page)) {
4706 * If a process owning a MAP_PRIVATE mapping fails to COW,
4707 * it is due to references held by a child and an insufficient
4708 * huge page pool. To guarantee the original mappers
4709 * reliability, unmap the page from child processes. The child
4710 * may get SIGKILLed if it later faults.
4712 if (outside_reserve) {
4713 struct address_space *mapping = vma->vm_file->f_mapping;
4718 BUG_ON(huge_pte_none(pte));
4720 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4721 * unmapping. unmapping needs to hold i_mmap_rwsem
4722 * in write mode. Dropping i_mmap_rwsem in read mode
4723 * here is OK as COW mappings do not interact with
4726 * Reacquire both after unmap operation.
4728 idx = vma_hugecache_offset(h, vma, haddr);
4729 hash = hugetlb_fault_mutex_hash(mapping, idx);
4730 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4731 i_mmap_unlock_read(mapping);
4733 unmap_ref_private(mm, vma, old_page, haddr);
4735 i_mmap_lock_read(mapping);
4736 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4738 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4740 pte_same(huge_ptep_get(ptep), pte)))
4741 goto retry_avoidcopy;
4743 * race occurs while re-acquiring page table
4744 * lock, and our job is done.
4749 ret = vmf_error(PTR_ERR(new_page));
4750 goto out_release_old;
4754 * When the original hugepage is shared one, it does not have
4755 * anon_vma prepared.
4757 if (unlikely(anon_vma_prepare(vma))) {
4759 goto out_release_all;
4762 copy_user_huge_page(new_page, old_page, address, vma,
4763 pages_per_huge_page(h));
4764 __SetPageUptodate(new_page);
4766 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4767 haddr + huge_page_size(h));
4768 mmu_notifier_invalidate_range_start(&range);
4771 * Retake the page table lock to check for racing updates
4772 * before the page tables are altered
4775 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4776 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4777 ClearHPageRestoreReserve(new_page);
4780 huge_ptep_clear_flush(vma, haddr, ptep);
4781 mmu_notifier_invalidate_range(mm, range.start, range.end);
4782 set_huge_pte_at(mm, haddr, ptep,
4783 make_huge_pte(vma, new_page, 1));
4784 page_remove_rmap(old_page, true);
4785 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4786 SetHPageMigratable(new_page);
4787 /* Make the old page be freed below */
4788 new_page = old_page;
4791 mmu_notifier_invalidate_range_end(&range);
4793 /* No restore in case of successful pagetable update (Break COW) */
4794 if (new_page != old_page)
4795 restore_reserve_on_error(h, vma, haddr, new_page);
4800 spin_lock(ptl); /* Caller expects lock to be held */
4804 /* Return the pagecache page at a given address within a VMA */
4805 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4806 struct vm_area_struct *vma, unsigned long address)
4808 struct address_space *mapping;
4811 mapping = vma->vm_file->f_mapping;
4812 idx = vma_hugecache_offset(h, vma, address);
4814 return find_lock_page(mapping, idx);
4818 * Return whether there is a pagecache page to back given address within VMA.
4819 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4821 static bool hugetlbfs_pagecache_present(struct hstate *h,
4822 struct vm_area_struct *vma, unsigned long address)
4824 struct address_space *mapping;
4828 mapping = vma->vm_file->f_mapping;
4829 idx = vma_hugecache_offset(h, vma, address);
4831 page = find_get_page(mapping, idx);
4834 return page != NULL;
4837 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4840 struct inode *inode = mapping->host;
4841 struct hstate *h = hstate_inode(inode);
4842 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4846 ClearHPageRestoreReserve(page);
4849 * set page dirty so that it will not be removed from cache/file
4850 * by non-hugetlbfs specific code paths.
4852 set_page_dirty(page);
4854 spin_lock(&inode->i_lock);
4855 inode->i_blocks += blocks_per_huge_page(h);
4856 spin_unlock(&inode->i_lock);
4860 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4861 struct address_space *mapping,
4864 unsigned long haddr,
4865 unsigned long reason)
4868 struct vm_fault vmf = {
4874 * Hard to debug if it ends up being
4875 * used by a callee that assumes
4876 * something about the other
4877 * uninitialized fields... same as in
4883 * vma_lock and hugetlb_fault_mutex must be dropped before handling
4884 * userfault. Also mmap_lock will be dropped during handling
4885 * userfault, any vma operation should be careful from here.
4887 hash = hugetlb_fault_mutex_hash(mapping, idx);
4888 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4889 i_mmap_unlock_read(mapping);
4890 return handle_userfault(&vmf, reason);
4893 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4894 struct vm_area_struct *vma,
4895 struct address_space *mapping, pgoff_t idx,
4896 unsigned long address, pte_t *ptep, unsigned int flags)
4898 struct hstate *h = hstate_vma(vma);
4899 vm_fault_t ret = VM_FAULT_SIGBUS;
4905 unsigned long haddr = address & huge_page_mask(h);
4906 bool new_page, new_pagecache_page = false;
4907 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
4910 * Currently, we are forced to kill the process in the event the
4911 * original mapper has unmapped pages from the child due to a failed
4912 * COW. Warn that such a situation has occurred as it may not be obvious
4914 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4915 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4921 * We can not race with truncation due to holding i_mmap_rwsem.
4922 * i_size is modified when holding i_mmap_rwsem, so check here
4923 * once for faults beyond end of file.
4925 size = i_size_read(mapping->host) >> huge_page_shift(h);
4931 page = find_lock_page(mapping, idx);
4933 /* Check for page in userfault range */
4934 if (userfaultfd_missing(vma))
4935 return hugetlb_handle_userfault(vma, mapping, idx,
4939 page = alloc_huge_page(vma, haddr, 0);
4942 * Returning error will result in faulting task being
4943 * sent SIGBUS. The hugetlb fault mutex prevents two
4944 * tasks from racing to fault in the same page which
4945 * could result in false unable to allocate errors.
4946 * Page migration does not take the fault mutex, but
4947 * does a clear then write of pte's under page table
4948 * lock. Page fault code could race with migration,
4949 * notice the clear pte and try to allocate a page
4950 * here. Before returning error, get ptl and make
4951 * sure there really is no pte entry.
4953 ptl = huge_pte_lock(h, mm, ptep);
4955 if (huge_pte_none(huge_ptep_get(ptep)))
4956 ret = vmf_error(PTR_ERR(page));
4960 clear_huge_page(page, address, pages_per_huge_page(h));
4961 __SetPageUptodate(page);
4964 if (vma->vm_flags & VM_MAYSHARE) {
4965 int err = huge_add_to_page_cache(page, mapping, idx);
4972 new_pagecache_page = true;
4975 if (unlikely(anon_vma_prepare(vma))) {
4977 goto backout_unlocked;
4983 * If memory error occurs between mmap() and fault, some process
4984 * don't have hwpoisoned swap entry for errored virtual address.
4985 * So we need to block hugepage fault by PG_hwpoison bit check.
4987 if (unlikely(PageHWPoison(page))) {
4988 ret = VM_FAULT_HWPOISON_LARGE |
4989 VM_FAULT_SET_HINDEX(hstate_index(h));
4990 goto backout_unlocked;
4993 /* Check for page in userfault range. */
4994 if (userfaultfd_minor(vma)) {
4997 return hugetlb_handle_userfault(vma, mapping, idx,
5004 * If we are going to COW a private mapping later, we examine the
5005 * pending reservations for this page now. This will ensure that
5006 * any allocations necessary to record that reservation occur outside
5009 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5010 if (vma_needs_reservation(h, vma, haddr) < 0) {
5012 goto backout_unlocked;
5014 /* Just decrements count, does not deallocate */
5015 vma_end_reservation(h, vma, haddr);
5018 ptl = huge_pte_lock(h, mm, ptep);
5020 if (!huge_pte_none(huge_ptep_get(ptep)))
5024 ClearHPageRestoreReserve(page);
5025 hugepage_add_new_anon_rmap(page, vma, haddr);
5027 page_dup_rmap(page, true);
5028 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5029 && (vma->vm_flags & VM_SHARED)));
5030 set_huge_pte_at(mm, haddr, ptep, new_pte);
5032 hugetlb_count_add(pages_per_huge_page(h), mm);
5033 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5034 /* Optimization, do the COW without a second fault */
5035 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
5041 * Only set HPageMigratable in newly allocated pages. Existing pages
5042 * found in the pagecache may not have HPageMigratableset if they have
5043 * been isolated for migration.
5046 SetHPageMigratable(page);
5050 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5051 i_mmap_unlock_read(mapping);
5058 /* restore reserve for newly allocated pages not in page cache */
5059 if (new_page && !new_pagecache_page)
5060 restore_reserve_on_error(h, vma, haddr, page);
5066 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5068 unsigned long key[2];
5071 key[0] = (unsigned long) mapping;
5074 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5076 return hash & (num_fault_mutexes - 1);
5080 * For uniprocessor systems we always use a single mutex, so just
5081 * return 0 and avoid the hashing overhead.
5083 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5089 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5090 unsigned long address, unsigned int flags)
5097 struct page *page = NULL;
5098 struct page *pagecache_page = NULL;
5099 struct hstate *h = hstate_vma(vma);
5100 struct address_space *mapping;
5101 int need_wait_lock = 0;
5102 unsigned long haddr = address & huge_page_mask(h);
5104 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5107 * Since we hold no locks, ptep could be stale. That is
5108 * OK as we are only making decisions based on content and
5109 * not actually modifying content here.
5111 entry = huge_ptep_get(ptep);
5112 if (unlikely(is_hugetlb_entry_migration(entry))) {
5113 migration_entry_wait_huge(vma, mm, ptep);
5115 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5116 return VM_FAULT_HWPOISON_LARGE |
5117 VM_FAULT_SET_HINDEX(hstate_index(h));
5121 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5122 * until finished with ptep. This serves two purposes:
5123 * 1) It prevents huge_pmd_unshare from being called elsewhere
5124 * and making the ptep no longer valid.
5125 * 2) It synchronizes us with i_size modifications during truncation.
5127 * ptep could have already be assigned via huge_pte_offset. That
5128 * is OK, as huge_pte_alloc will return the same value unless
5129 * something has changed.
5131 mapping = vma->vm_file->f_mapping;
5132 i_mmap_lock_read(mapping);
5133 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5135 i_mmap_unlock_read(mapping);
5136 return VM_FAULT_OOM;
5140 * Serialize hugepage allocation and instantiation, so that we don't
5141 * get spurious allocation failures if two CPUs race to instantiate
5142 * the same page in the page cache.
5144 idx = vma_hugecache_offset(h, vma, haddr);
5145 hash = hugetlb_fault_mutex_hash(mapping, idx);
5146 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5148 entry = huge_ptep_get(ptep);
5149 if (huge_pte_none(entry))
5151 * hugetlb_no_page will drop vma lock and hugetlb fault
5152 * mutex internally, which make us return immediately.
5154 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5159 * entry could be a migration/hwpoison entry at this point, so this
5160 * check prevents the kernel from going below assuming that we have
5161 * an active hugepage in pagecache. This goto expects the 2nd page
5162 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5163 * properly handle it.
5165 if (!pte_present(entry))
5169 * If we are going to COW the mapping later, we examine the pending
5170 * reservations for this page now. This will ensure that any
5171 * allocations necessary to record that reservation occur outside the
5172 * spinlock. For private mappings, we also lookup the pagecache
5173 * page now as it is used to determine if a reservation has been
5176 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5177 if (vma_needs_reservation(h, vma, haddr) < 0) {
5181 /* Just decrements count, does not deallocate */
5182 vma_end_reservation(h, vma, haddr);
5184 if (!(vma->vm_flags & VM_MAYSHARE))
5185 pagecache_page = hugetlbfs_pagecache_page(h,
5189 ptl = huge_pte_lock(h, mm, ptep);
5191 /* Check for a racing update before calling hugetlb_cow */
5192 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5196 * hugetlb_cow() requires page locks of pte_page(entry) and
5197 * pagecache_page, so here we need take the former one
5198 * when page != pagecache_page or !pagecache_page.
5200 page = pte_page(entry);
5201 if (page != pagecache_page)
5202 if (!trylock_page(page)) {
5209 if (flags & FAULT_FLAG_WRITE) {
5210 if (!huge_pte_write(entry)) {
5211 ret = hugetlb_cow(mm, vma, address, ptep,
5212 pagecache_page, ptl);
5215 entry = huge_pte_mkdirty(entry);
5217 entry = pte_mkyoung(entry);
5218 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5219 flags & FAULT_FLAG_WRITE))
5220 update_mmu_cache(vma, haddr, ptep);
5222 if (page != pagecache_page)
5228 if (pagecache_page) {
5229 unlock_page(pagecache_page);
5230 put_page(pagecache_page);
5233 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5234 i_mmap_unlock_read(mapping);
5236 * Generally it's safe to hold refcount during waiting page lock. But
5237 * here we just wait to defer the next page fault to avoid busy loop and
5238 * the page is not used after unlocked before returning from the current
5239 * page fault. So we are safe from accessing freed page, even if we wait
5240 * here without taking refcount.
5243 wait_on_page_locked(page);
5247 #ifdef CONFIG_USERFAULTFD
5249 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5250 * modifications for huge pages.
5252 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5254 struct vm_area_struct *dst_vma,
5255 unsigned long dst_addr,
5256 unsigned long src_addr,
5257 enum mcopy_atomic_mode mode,
5258 struct page **pagep)
5260 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5261 struct hstate *h = hstate_vma(dst_vma);
5262 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5263 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5265 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5271 bool page_in_pagecache = false;
5275 page = find_lock_page(mapping, idx);
5278 page_in_pagecache = true;
5279 } else if (!*pagep) {
5280 /* If a page already exists, then it's UFFDIO_COPY for
5281 * a non-missing case. Return -EEXIST.
5284 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5289 page = alloc_huge_page(dst_vma, dst_addr, 0);
5295 ret = copy_huge_page_from_user(page,
5296 (const void __user *) src_addr,
5297 pages_per_huge_page(h), false);
5299 /* fallback to copy_from_user outside mmap_lock */
5300 if (unlikely(ret)) {
5302 /* Free the allocated page which may have
5303 * consumed a reservation.
5305 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5308 /* Allocate a temporary page to hold the copied
5311 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5317 /* Set the outparam pagep and return to the caller to
5318 * copy the contents outside the lock. Don't free the
5325 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5332 page = alloc_huge_page(dst_vma, dst_addr, 0);
5339 copy_huge_page(page, *pagep);
5345 * The memory barrier inside __SetPageUptodate makes sure that
5346 * preceding stores to the page contents become visible before
5347 * the set_pte_at() write.
5349 __SetPageUptodate(page);
5351 /* Add shared, newly allocated pages to the page cache. */
5352 if (vm_shared && !is_continue) {
5353 size = i_size_read(mapping->host) >> huge_page_shift(h);
5356 goto out_release_nounlock;
5359 * Serialization between remove_inode_hugepages() and
5360 * huge_add_to_page_cache() below happens through the
5361 * hugetlb_fault_mutex_table that here must be hold by
5364 ret = huge_add_to_page_cache(page, mapping, idx);
5366 goto out_release_nounlock;
5367 page_in_pagecache = true;
5370 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5374 if (PageHWPoison(page))
5375 goto out_release_unlock;
5378 * Recheck the i_size after holding PT lock to make sure not
5379 * to leave any page mapped (as page_mapped()) beyond the end
5380 * of the i_size (remove_inode_hugepages() is strict about
5381 * enforcing that). If we bail out here, we'll also leave a
5382 * page in the radix tree in the vm_shared case beyond the end
5383 * of the i_size, but remove_inode_hugepages() will take care
5384 * of it as soon as we drop the hugetlb_fault_mutex_table.
5386 size = i_size_read(mapping->host) >> huge_page_shift(h);
5389 goto out_release_unlock;
5392 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5393 goto out_release_unlock;
5395 if (page_in_pagecache) {
5396 page_dup_rmap(page, true);
5398 ClearHPageRestoreReserve(page);
5399 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5402 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5403 if (is_continue && !vm_shared)
5406 writable = dst_vma->vm_flags & VM_WRITE;
5408 _dst_pte = make_huge_pte(dst_vma, page, writable);
5410 _dst_pte = huge_pte_mkdirty(_dst_pte);
5411 _dst_pte = pte_mkyoung(_dst_pte);
5413 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5415 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5416 dst_vma->vm_flags & VM_WRITE);
5417 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5419 /* No need to invalidate - it was non-present before */
5420 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5424 SetHPageMigratable(page);
5425 if (vm_shared || is_continue)
5432 if (vm_shared || is_continue)
5434 out_release_nounlock:
5435 if (!page_in_pagecache)
5436 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5440 #endif /* CONFIG_USERFAULTFD */
5442 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5443 int refs, struct page **pages,
5444 struct vm_area_struct **vmas)
5448 for (nr = 0; nr < refs; nr++) {
5450 pages[nr] = mem_map_offset(page, nr);
5456 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5457 struct page **pages, struct vm_area_struct **vmas,
5458 unsigned long *position, unsigned long *nr_pages,
5459 long i, unsigned int flags, int *locked)
5461 unsigned long pfn_offset;
5462 unsigned long vaddr = *position;
5463 unsigned long remainder = *nr_pages;
5464 struct hstate *h = hstate_vma(vma);
5465 int err = -EFAULT, refs;
5467 while (vaddr < vma->vm_end && remainder) {
5469 spinlock_t *ptl = NULL;
5474 * If we have a pending SIGKILL, don't keep faulting pages and
5475 * potentially allocating memory.
5477 if (fatal_signal_pending(current)) {
5483 * Some archs (sparc64, sh*) have multiple pte_ts to
5484 * each hugepage. We have to make sure we get the
5485 * first, for the page indexing below to work.
5487 * Note that page table lock is not held when pte is null.
5489 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5492 ptl = huge_pte_lock(h, mm, pte);
5493 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5496 * When coredumping, it suits get_dump_page if we just return
5497 * an error where there's an empty slot with no huge pagecache
5498 * to back it. This way, we avoid allocating a hugepage, and
5499 * the sparse dumpfile avoids allocating disk blocks, but its
5500 * huge holes still show up with zeroes where they need to be.
5502 if (absent && (flags & FOLL_DUMP) &&
5503 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5511 * We need call hugetlb_fault for both hugepages under migration
5512 * (in which case hugetlb_fault waits for the migration,) and
5513 * hwpoisoned hugepages (in which case we need to prevent the
5514 * caller from accessing to them.) In order to do this, we use
5515 * here is_swap_pte instead of is_hugetlb_entry_migration and
5516 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5517 * both cases, and because we can't follow correct pages
5518 * directly from any kind of swap entries.
5520 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5521 ((flags & FOLL_WRITE) &&
5522 !huge_pte_write(huge_ptep_get(pte)))) {
5524 unsigned int fault_flags = 0;
5528 if (flags & FOLL_WRITE)
5529 fault_flags |= FAULT_FLAG_WRITE;
5531 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5532 FAULT_FLAG_KILLABLE;
5533 if (flags & FOLL_NOWAIT)
5534 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5535 FAULT_FLAG_RETRY_NOWAIT;
5536 if (flags & FOLL_TRIED) {
5538 * Note: FAULT_FLAG_ALLOW_RETRY and
5539 * FAULT_FLAG_TRIED can co-exist
5541 fault_flags |= FAULT_FLAG_TRIED;
5543 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5544 if (ret & VM_FAULT_ERROR) {
5545 err = vm_fault_to_errno(ret, flags);
5549 if (ret & VM_FAULT_RETRY) {
5551 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5555 * VM_FAULT_RETRY must not return an
5556 * error, it will return zero
5559 * No need to update "position" as the
5560 * caller will not check it after
5561 * *nr_pages is set to 0.
5568 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5569 page = pte_page(huge_ptep_get(pte));
5572 * If subpage information not requested, update counters
5573 * and skip the same_page loop below.
5575 if (!pages && !vmas && !pfn_offset &&
5576 (vaddr + huge_page_size(h) < vma->vm_end) &&
5577 (remainder >= pages_per_huge_page(h))) {
5578 vaddr += huge_page_size(h);
5579 remainder -= pages_per_huge_page(h);
5580 i += pages_per_huge_page(h);
5585 /* vaddr may not be aligned to PAGE_SIZE */
5586 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
5587 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
5590 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5592 likely(pages) ? pages + i : NULL,
5593 vmas ? vmas + i : NULL);
5597 * try_grab_compound_head() should always succeed here,
5598 * because: a) we hold the ptl lock, and b) we've just
5599 * checked that the huge page is present in the page
5600 * tables. If the huge page is present, then the tail
5601 * pages must also be present. The ptl prevents the
5602 * head page and tail pages from being rearranged in
5603 * any way. So this page must be available at this
5604 * point, unless the page refcount overflowed:
5606 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5616 vaddr += (refs << PAGE_SHIFT);
5622 *nr_pages = remainder;
5624 * setting position is actually required only if remainder is
5625 * not zero but it's faster not to add a "if (remainder)"
5633 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5634 unsigned long address, unsigned long end, pgprot_t newprot)
5636 struct mm_struct *mm = vma->vm_mm;
5637 unsigned long start = address;
5640 struct hstate *h = hstate_vma(vma);
5641 unsigned long pages = 0;
5642 bool shared_pmd = false;
5643 struct mmu_notifier_range range;
5646 * In the case of shared PMDs, the area to flush could be beyond
5647 * start/end. Set range.start/range.end to cover the maximum possible
5648 * range if PMD sharing is possible.
5650 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5651 0, vma, mm, start, end);
5652 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5654 BUG_ON(address >= end);
5655 flush_cache_range(vma, range.start, range.end);
5657 mmu_notifier_invalidate_range_start(&range);
5658 i_mmap_lock_write(vma->vm_file->f_mapping);
5659 for (; address < end; address += huge_page_size(h)) {
5661 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5664 ptl = huge_pte_lock(h, mm, ptep);
5665 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5671 pte = huge_ptep_get(ptep);
5672 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5676 if (unlikely(is_hugetlb_entry_migration(pte))) {
5677 swp_entry_t entry = pte_to_swp_entry(pte);
5679 if (is_writable_migration_entry(entry)) {
5682 entry = make_readable_migration_entry(
5684 newpte = swp_entry_to_pte(entry);
5685 set_huge_swap_pte_at(mm, address, ptep,
5686 newpte, huge_page_size(h));
5692 if (!huge_pte_none(pte)) {
5694 unsigned int shift = huge_page_shift(hstate_vma(vma));
5696 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5697 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5698 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5699 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5705 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5706 * may have cleared our pud entry and done put_page on the page table:
5707 * once we release i_mmap_rwsem, another task can do the final put_page
5708 * and that page table be reused and filled with junk. If we actually
5709 * did unshare a page of pmds, flush the range corresponding to the pud.
5712 flush_hugetlb_tlb_range(vma, range.start, range.end);
5714 flush_hugetlb_tlb_range(vma, start, end);
5716 * No need to call mmu_notifier_invalidate_range() we are downgrading
5717 * page table protection not changing it to point to a new page.
5719 * See Documentation/vm/mmu_notifier.rst
5721 i_mmap_unlock_write(vma->vm_file->f_mapping);
5722 mmu_notifier_invalidate_range_end(&range);
5724 return pages << h->order;
5727 /* Return true if reservation was successful, false otherwise. */
5728 bool hugetlb_reserve_pages(struct inode *inode,
5730 struct vm_area_struct *vma,
5731 vm_flags_t vm_flags)
5734 struct hstate *h = hstate_inode(inode);
5735 struct hugepage_subpool *spool = subpool_inode(inode);
5736 struct resv_map *resv_map;
5737 struct hugetlb_cgroup *h_cg = NULL;
5738 long gbl_reserve, regions_needed = 0;
5740 /* This should never happen */
5742 VM_WARN(1, "%s called with a negative range\n", __func__);
5747 * Only apply hugepage reservation if asked. At fault time, an
5748 * attempt will be made for VM_NORESERVE to allocate a page
5749 * without using reserves
5751 if (vm_flags & VM_NORESERVE)
5755 * Shared mappings base their reservation on the number of pages that
5756 * are already allocated on behalf of the file. Private mappings need
5757 * to reserve the full area even if read-only as mprotect() may be
5758 * called to make the mapping read-write. Assume !vma is a shm mapping
5760 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5762 * resv_map can not be NULL as hugetlb_reserve_pages is only
5763 * called for inodes for which resv_maps were created (see
5764 * hugetlbfs_get_inode).
5766 resv_map = inode_resv_map(inode);
5768 chg = region_chg(resv_map, from, to, ®ions_needed);
5771 /* Private mapping. */
5772 resv_map = resv_map_alloc();
5778 set_vma_resv_map(vma, resv_map);
5779 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5785 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5786 chg * pages_per_huge_page(h), &h_cg) < 0)
5789 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5790 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5793 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5797 * There must be enough pages in the subpool for the mapping. If
5798 * the subpool has a minimum size, there may be some global
5799 * reservations already in place (gbl_reserve).
5801 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5802 if (gbl_reserve < 0)
5803 goto out_uncharge_cgroup;
5806 * Check enough hugepages are available for the reservation.
5807 * Hand the pages back to the subpool if there are not
5809 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5813 * Account for the reservations made. Shared mappings record regions
5814 * that have reservations as they are shared by multiple VMAs.
5815 * When the last VMA disappears, the region map says how much
5816 * the reservation was and the page cache tells how much of
5817 * the reservation was consumed. Private mappings are per-VMA and
5818 * only the consumed reservations are tracked. When the VMA
5819 * disappears, the original reservation is the VMA size and the
5820 * consumed reservations are stored in the map. Hence, nothing
5821 * else has to be done for private mappings here
5823 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5824 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5826 if (unlikely(add < 0)) {
5827 hugetlb_acct_memory(h, -gbl_reserve);
5829 } else if (unlikely(chg > add)) {
5831 * pages in this range were added to the reserve
5832 * map between region_chg and region_add. This
5833 * indicates a race with alloc_huge_page. Adjust
5834 * the subpool and reserve counts modified above
5835 * based on the difference.
5840 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5841 * reference to h_cg->css. See comment below for detail.
5843 hugetlb_cgroup_uncharge_cgroup_rsvd(
5845 (chg - add) * pages_per_huge_page(h), h_cg);
5847 rsv_adjust = hugepage_subpool_put_pages(spool,
5849 hugetlb_acct_memory(h, -rsv_adjust);
5852 * The file_regions will hold their own reference to
5853 * h_cg->css. So we should release the reference held
5854 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5857 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5863 /* put back original number of pages, chg */
5864 (void)hugepage_subpool_put_pages(spool, chg);
5865 out_uncharge_cgroup:
5866 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5867 chg * pages_per_huge_page(h), h_cg);
5869 if (!vma || vma->vm_flags & VM_MAYSHARE)
5870 /* Only call region_abort if the region_chg succeeded but the
5871 * region_add failed or didn't run.
5873 if (chg >= 0 && add < 0)
5874 region_abort(resv_map, from, to, regions_needed);
5875 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5876 kref_put(&resv_map->refs, resv_map_release);
5880 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5883 struct hstate *h = hstate_inode(inode);
5884 struct resv_map *resv_map = inode_resv_map(inode);
5886 struct hugepage_subpool *spool = subpool_inode(inode);
5890 * Since this routine can be called in the evict inode path for all
5891 * hugetlbfs inodes, resv_map could be NULL.
5894 chg = region_del(resv_map, start, end);
5896 * region_del() can fail in the rare case where a region
5897 * must be split and another region descriptor can not be
5898 * allocated. If end == LONG_MAX, it will not fail.
5904 spin_lock(&inode->i_lock);
5905 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5906 spin_unlock(&inode->i_lock);
5909 * If the subpool has a minimum size, the number of global
5910 * reservations to be released may be adjusted.
5912 * Note that !resv_map implies freed == 0. So (chg - freed)
5913 * won't go negative.
5915 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5916 hugetlb_acct_memory(h, -gbl_reserve);
5921 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5922 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5923 struct vm_area_struct *vma,
5924 unsigned long addr, pgoff_t idx)
5926 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5928 unsigned long sbase = saddr & PUD_MASK;
5929 unsigned long s_end = sbase + PUD_SIZE;
5931 /* Allow segments to share if only one is marked locked */
5932 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5933 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5936 * match the virtual addresses, permission and the alignment of the
5939 if (pmd_index(addr) != pmd_index(saddr) ||
5940 vm_flags != svm_flags ||
5941 !range_in_vma(svma, sbase, s_end))
5947 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5949 unsigned long base = addr & PUD_MASK;
5950 unsigned long end = base + PUD_SIZE;
5953 * check on proper vm_flags and page table alignment
5955 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5960 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5962 #ifdef CONFIG_USERFAULTFD
5963 if (uffd_disable_huge_pmd_share(vma))
5966 return vma_shareable(vma, addr);
5970 * Determine if start,end range within vma could be mapped by shared pmd.
5971 * If yes, adjust start and end to cover range associated with possible
5972 * shared pmd mappings.
5974 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5975 unsigned long *start, unsigned long *end)
5977 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5978 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5981 * vma needs to span at least one aligned PUD size, and the range
5982 * must be at least partially within in.
5984 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5985 (*end <= v_start) || (*start >= v_end))
5988 /* Extend the range to be PUD aligned for a worst case scenario */
5989 if (*start > v_start)
5990 *start = ALIGN_DOWN(*start, PUD_SIZE);
5993 *end = ALIGN(*end, PUD_SIZE);
5997 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5998 * and returns the corresponding pte. While this is not necessary for the
5999 * !shared pmd case because we can allocate the pmd later as well, it makes the
6000 * code much cleaner.
6002 * This routine must be called with i_mmap_rwsem held in at least read mode if
6003 * sharing is possible. For hugetlbfs, this prevents removal of any page
6004 * table entries associated with the address space. This is important as we
6005 * are setting up sharing based on existing page table entries (mappings).
6007 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
6008 * huge_pte_alloc know that sharing is not possible and do not take
6009 * i_mmap_rwsem as a performance optimization. This is handled by the
6010 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
6011 * only required for subsequent processing.
6013 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6014 unsigned long addr, pud_t *pud)
6016 struct address_space *mapping = vma->vm_file->f_mapping;
6017 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6019 struct vm_area_struct *svma;
6020 unsigned long saddr;
6025 i_mmap_assert_locked(mapping);
6026 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6030 saddr = page_table_shareable(svma, vma, addr, idx);
6032 spte = huge_pte_offset(svma->vm_mm, saddr,
6033 vma_mmu_pagesize(svma));
6035 get_page(virt_to_page(spte));
6044 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6045 if (pud_none(*pud)) {
6046 pud_populate(mm, pud,
6047 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6050 put_page(virt_to_page(spte));
6054 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6059 * unmap huge page backed by shared pte.
6061 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6062 * indicated by page_count > 1, unmap is achieved by clearing pud and
6063 * decrementing the ref count. If count == 1, the pte page is not shared.
6065 * Called with page table lock held and i_mmap_rwsem held in write mode.
6067 * returns: 1 successfully unmapped a shared pte page
6068 * 0 the underlying pte page is not shared, or it is the last user
6070 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6071 unsigned long *addr, pte_t *ptep)
6073 pgd_t *pgd = pgd_offset(mm, *addr);
6074 p4d_t *p4d = p4d_offset(pgd, *addr);
6075 pud_t *pud = pud_offset(p4d, *addr);
6077 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6078 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6079 if (page_count(virt_to_page(ptep)) == 1)
6083 put_page(virt_to_page(ptep));
6086 * This update of passed address optimizes loops sequentially
6087 * processing addresses in increments of huge page size (PMD_SIZE
6088 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
6089 * Update address to the 'last page' in the cleared area so that
6090 * calling loop can move to first page past this area.
6092 *addr |= PUD_SIZE - PMD_SIZE;
6096 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6097 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6098 unsigned long addr, pud_t *pud)
6103 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6104 unsigned long *addr, pte_t *ptep)
6109 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6110 unsigned long *start, unsigned long *end)
6114 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6118 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6120 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6121 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6122 unsigned long addr, unsigned long sz)
6129 pgd = pgd_offset(mm, addr);
6130 p4d = p4d_alloc(mm, pgd, addr);
6133 pud = pud_alloc(mm, p4d, addr);
6135 if (sz == PUD_SIZE) {
6138 BUG_ON(sz != PMD_SIZE);
6139 if (want_pmd_share(vma, addr) && pud_none(*pud))
6140 pte = huge_pmd_share(mm, vma, addr, pud);
6142 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6145 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6151 * huge_pte_offset() - Walk the page table to resolve the hugepage
6152 * entry at address @addr
6154 * Return: Pointer to page table entry (PUD or PMD) for
6155 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6156 * size @sz doesn't match the hugepage size at this level of the page
6159 pte_t *huge_pte_offset(struct mm_struct *mm,
6160 unsigned long addr, unsigned long sz)
6167 pgd = pgd_offset(mm, addr);
6168 if (!pgd_present(*pgd))
6170 p4d = p4d_offset(pgd, addr);
6171 if (!p4d_present(*p4d))
6174 pud = pud_offset(p4d, addr);
6176 /* must be pud huge, non-present or none */
6177 return (pte_t *)pud;
6178 if (!pud_present(*pud))
6180 /* must have a valid entry and size to go further */
6182 pmd = pmd_offset(pud, addr);
6183 /* must be pmd huge, non-present or none */
6184 return (pte_t *)pmd;
6187 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6190 * These functions are overwritable if your architecture needs its own
6193 struct page * __weak
6194 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6197 return ERR_PTR(-EINVAL);
6200 struct page * __weak
6201 follow_huge_pd(struct vm_area_struct *vma,
6202 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6204 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6208 struct page * __weak
6209 follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags)
6211 struct hstate *h = hstate_vma(vma);
6212 struct mm_struct *mm = vma->vm_mm;
6213 struct page *page = NULL;
6217 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6218 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6219 (FOLL_PIN | FOLL_GET)))
6223 ptep = huge_pte_offset(mm, address, huge_page_size(h));
6227 ptl = huge_pte_lock(h, mm, ptep);
6228 pte = huge_ptep_get(ptep);
6229 if (pte_present(pte)) {
6230 page = pte_page(pte) +
6231 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
6233 * try_grab_page() should always succeed here, because: a) we
6234 * hold the pmd (ptl) lock, and b) we've just checked that the
6235 * huge pmd (head) page is present in the page tables. The ptl
6236 * prevents the head page and tail pages from being rearranged
6237 * in any way. So this page must be available at this point,
6238 * unless the page refcount overflowed:
6240 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6245 if (is_hugetlb_entry_migration(pte)) {
6247 __migration_entry_wait(mm, ptep, ptl);
6251 * hwpoisoned entry is treated as no_page_table in
6252 * follow_page_mask().
6260 struct page * __weak
6261 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6262 pud_t *pud, int flags)
6264 if (flags & (FOLL_GET | FOLL_PIN))
6267 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6270 struct page * __weak
6271 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6273 if (flags & (FOLL_GET | FOLL_PIN))
6276 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6279 int isolate_hugetlb(struct page *page, struct list_head *list)
6283 spin_lock_irq(&hugetlb_lock);
6284 if (!PageHeadHuge(page) ||
6285 !HPageMigratable(page) ||
6286 !get_page_unless_zero(page)) {
6290 ClearHPageMigratable(page);
6291 list_move_tail(&page->lru, list);
6293 spin_unlock_irq(&hugetlb_lock);
6297 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6302 spin_lock_irq(&hugetlb_lock);
6303 if (PageHeadHuge(page)) {
6305 if (HPageFreed(page) || HPageMigratable(page))
6306 ret = get_page_unless_zero(page);
6310 spin_unlock_irq(&hugetlb_lock);
6314 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
6318 spin_lock_irq(&hugetlb_lock);
6319 ret = __get_huge_page_for_hwpoison(pfn, flags);
6320 spin_unlock_irq(&hugetlb_lock);
6324 void putback_active_hugepage(struct page *page)
6326 spin_lock_irq(&hugetlb_lock);
6327 SetHPageMigratable(page);
6328 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6329 spin_unlock_irq(&hugetlb_lock);
6333 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6335 struct hstate *h = page_hstate(oldpage);
6337 hugetlb_cgroup_migrate(oldpage, newpage);
6338 set_page_owner_migrate_reason(newpage, reason);
6341 * transfer temporary state of the new huge page. This is
6342 * reverse to other transitions because the newpage is going to
6343 * be final while the old one will be freed so it takes over
6344 * the temporary status.
6346 * Also note that we have to transfer the per-node surplus state
6347 * here as well otherwise the global surplus count will not match
6350 if (HPageTemporary(newpage)) {
6351 int old_nid = page_to_nid(oldpage);
6352 int new_nid = page_to_nid(newpage);
6354 SetHPageTemporary(oldpage);
6355 ClearHPageTemporary(newpage);
6358 * There is no need to transfer the per-node surplus state
6359 * when we do not cross the node.
6361 if (new_nid == old_nid)
6363 spin_lock_irq(&hugetlb_lock);
6364 if (h->surplus_huge_pages_node[old_nid]) {
6365 h->surplus_huge_pages_node[old_nid]--;
6366 h->surplus_huge_pages_node[new_nid]++;
6368 spin_unlock_irq(&hugetlb_lock);
6372 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
6373 unsigned long start,
6376 struct hstate *h = hstate_vma(vma);
6377 unsigned long sz = huge_page_size(h);
6378 struct mm_struct *mm = vma->vm_mm;
6379 struct mmu_notifier_range range;
6380 unsigned long address;
6384 if (!(vma->vm_flags & VM_MAYSHARE))
6391 * No need to call adjust_range_if_pmd_sharing_possible(), because
6392 * we have already done the PUD_SIZE alignment.
6394 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6396 mmu_notifier_invalidate_range_start(&range);
6397 i_mmap_lock_write(vma->vm_file->f_mapping);
6398 for (address = start; address < end; address += PUD_SIZE) {
6399 unsigned long tmp = address;
6401 ptep = huge_pte_offset(mm, address, sz);
6404 ptl = huge_pte_lock(h, mm, ptep);
6405 /* We don't want 'address' to be changed */
6406 huge_pmd_unshare(mm, vma, &tmp, ptep);
6409 flush_hugetlb_tlb_range(vma, start, end);
6410 i_mmap_unlock_write(vma->vm_file->f_mapping);
6412 * No need to call mmu_notifier_invalidate_range(), see
6413 * Documentation/vm/mmu_notifier.rst.
6415 mmu_notifier_invalidate_range_end(&range);
6419 * This function will unconditionally remove all the shared pmd pgtable entries
6420 * within the specific vma for a hugetlbfs memory range.
6422 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6424 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
6425 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
6429 static bool cma_reserve_called __initdata;
6431 static int __init cmdline_parse_hugetlb_cma(char *p)
6433 hugetlb_cma_size = memparse(p, &p);
6437 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6439 void __init hugetlb_cma_reserve(int order)
6441 unsigned long size, reserved, per_node;
6444 cma_reserve_called = true;
6446 if (!hugetlb_cma_size)
6449 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6450 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6451 (PAGE_SIZE << order) / SZ_1M);
6456 * If 3 GB area is requested on a machine with 4 numa nodes,
6457 * let's allocate 1 GB on first three nodes and ignore the last one.
6459 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6460 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6461 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6464 for_each_node_state(nid, N_ONLINE) {
6466 char name[CMA_MAX_NAME];
6468 size = min(per_node, hugetlb_cma_size - reserved);
6469 size = round_up(size, PAGE_SIZE << order);
6471 snprintf(name, sizeof(name), "hugetlb%d", nid);
6472 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6474 &hugetlb_cma[nid], nid);
6476 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6482 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6485 if (reserved >= hugetlb_cma_size)
6490 void __init hugetlb_cma_check(void)
6492 if (!hugetlb_cma_size || cma_reserve_called)
6495 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6498 #endif /* CONFIG_CMA */