1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
31 #include <linux/cma.h>
34 #include <asm/pgalloc.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/node.h>
41 #include <linux/userfaultfd_k.h>
42 #include <linux/page_owner.h>
45 int hugetlb_max_hstate __read_mostly;
46 unsigned int default_hstate_idx;
47 struct hstate hstates[HUGE_MAX_HSTATE];
50 static struct cma *hugetlb_cma[MAX_NUMNODES];
52 static unsigned long hugetlb_cma_size __initdata;
55 * Minimum page order among possible hugepage sizes, set to a proper value
58 static unsigned int minimum_order __read_mostly = UINT_MAX;
60 __initdata LIST_HEAD(huge_boot_pages);
62 /* for command line parsing */
63 static struct hstate * __initdata parsed_hstate;
64 static unsigned long __initdata default_hstate_max_huge_pages;
65 static bool __initdata parsed_valid_hugepagesz = true;
66 static bool __initdata parsed_default_hugepagesz;
69 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
70 * free_huge_pages, and surplus_huge_pages.
72 DEFINE_SPINLOCK(hugetlb_lock);
75 * Serializes faults on the same logical page. This is used to
76 * prevent spurious OOMs when the hugepage pool is fully utilized.
78 static int num_fault_mutexes;
79 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
81 /* Forward declaration */
82 static int hugetlb_acct_memory(struct hstate *h, long delta);
84 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
86 bool free = (spool->count == 0) && (spool->used_hpages == 0);
88 spin_unlock(&spool->lock);
90 /* If no pages are used, and no other handles to the subpool
91 * remain, give up any reservations based on minimum size and
94 if (spool->min_hpages != -1)
95 hugetlb_acct_memory(spool->hstate,
101 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
104 struct hugepage_subpool *spool;
106 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
110 spin_lock_init(&spool->lock);
112 spool->max_hpages = max_hpages;
114 spool->min_hpages = min_hpages;
116 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
120 spool->rsv_hpages = min_hpages;
125 void hugepage_put_subpool(struct hugepage_subpool *spool)
127 spin_lock(&spool->lock);
128 BUG_ON(!spool->count);
130 unlock_or_release_subpool(spool);
134 * Subpool accounting for allocating and reserving pages.
135 * Return -ENOMEM if there are not enough resources to satisfy the
136 * the request. Otherwise, return the number of pages by which the
137 * global pools must be adjusted (upward). The returned value may
138 * only be different than the passed value (delta) in the case where
139 * a subpool minimum size must be maintained.
141 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
149 spin_lock(&spool->lock);
151 if (spool->max_hpages != -1) { /* maximum size accounting */
152 if ((spool->used_hpages + delta) <= spool->max_hpages)
153 spool->used_hpages += delta;
160 /* minimum size accounting */
161 if (spool->min_hpages != -1 && spool->rsv_hpages) {
162 if (delta > spool->rsv_hpages) {
164 * Asking for more reserves than those already taken on
165 * behalf of subpool. Return difference.
167 ret = delta - spool->rsv_hpages;
168 spool->rsv_hpages = 0;
170 ret = 0; /* reserves already accounted for */
171 spool->rsv_hpages -= delta;
176 spin_unlock(&spool->lock);
181 * Subpool accounting for freeing and unreserving pages.
182 * Return the number of global page reservations that must be dropped.
183 * The return value may only be different than the passed value (delta)
184 * in the case where a subpool minimum size must be maintained.
186 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
194 spin_lock(&spool->lock);
196 if (spool->max_hpages != -1) /* maximum size accounting */
197 spool->used_hpages -= delta;
199 /* minimum size accounting */
200 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
201 if (spool->rsv_hpages + delta <= spool->min_hpages)
204 ret = spool->rsv_hpages + delta - spool->min_hpages;
206 spool->rsv_hpages += delta;
207 if (spool->rsv_hpages > spool->min_hpages)
208 spool->rsv_hpages = spool->min_hpages;
212 * If hugetlbfs_put_super couldn't free spool due to an outstanding
213 * quota reference, free it now.
215 unlock_or_release_subpool(spool);
220 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
222 return HUGETLBFS_SB(inode->i_sb)->spool;
225 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
227 return subpool_inode(file_inode(vma->vm_file));
230 /* Helper that removes a struct file_region from the resv_map cache and returns
233 static struct file_region *
234 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
236 struct file_region *nrg = NULL;
238 VM_BUG_ON(resv->region_cache_count <= 0);
240 resv->region_cache_count--;
241 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
243 list_del(&nrg->link);
251 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
252 struct file_region *rg)
254 #ifdef CONFIG_CGROUP_HUGETLB
255 nrg->reservation_counter = rg->reservation_counter;
262 /* Helper that records hugetlb_cgroup uncharge info. */
263 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
265 struct resv_map *resv,
266 struct file_region *nrg)
268 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter =
271 &h_cg->rsvd_hugepage[hstate_index(h)];
272 nrg->css = &h_cg->css;
273 if (!resv->pages_per_hpage)
274 resv->pages_per_hpage = pages_per_huge_page(h);
275 /* pages_per_hpage should be the same for all entries in
278 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
280 nrg->reservation_counter = NULL;
286 static bool has_same_uncharge_info(struct file_region *rg,
287 struct file_region *org)
289 #ifdef CONFIG_CGROUP_HUGETLB
291 rg->reservation_counter == org->reservation_counter &&
299 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
301 struct file_region *nrg = NULL, *prg = NULL;
303 prg = list_prev_entry(rg, link);
304 if (&prg->link != &resv->regions && prg->to == rg->from &&
305 has_same_uncharge_info(prg, rg)) {
311 coalesce_file_region(resv, prg);
315 nrg = list_next_entry(rg, link);
316 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
317 has_same_uncharge_info(nrg, rg)) {
318 nrg->from = rg->from;
323 coalesce_file_region(resv, nrg);
328 /* Must be called with resv->lock held. Calling this with count_only == true
329 * will count the number of pages to be added but will not modify the linked
330 * list. If regions_needed != NULL and count_only == true, then regions_needed
331 * will indicate the number of file_regions needed in the cache to carry out to
332 * add the regions for this range.
334 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
335 struct hugetlb_cgroup *h_cg,
336 struct hstate *h, long *regions_needed,
340 struct list_head *head = &resv->regions;
341 long last_accounted_offset = f;
342 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
347 /* In this loop, we essentially handle an entry for the range
348 * [last_accounted_offset, rg->from), at every iteration, with some
351 list_for_each_entry_safe(rg, trg, head, link) {
352 /* Skip irrelevant regions that start before our range. */
354 /* If this region ends after the last accounted offset,
355 * then we need to update last_accounted_offset.
357 if (rg->to > last_accounted_offset)
358 last_accounted_offset = rg->to;
362 /* When we find a region that starts beyond our range, we've
368 /* Add an entry for last_accounted_offset -> rg->from, and
369 * update last_accounted_offset.
371 if (rg->from > last_accounted_offset) {
372 add += rg->from - last_accounted_offset;
374 nrg = get_file_region_entry_from_cache(
375 resv, last_accounted_offset, rg->from);
376 record_hugetlb_cgroup_uncharge_info(h_cg, h,
378 list_add(&nrg->link, rg->link.prev);
379 coalesce_file_region(resv, nrg);
380 } else if (regions_needed)
381 *regions_needed += 1;
384 last_accounted_offset = rg->to;
387 /* Handle the case where our range extends beyond
388 * last_accounted_offset.
390 if (last_accounted_offset < t) {
391 add += t - last_accounted_offset;
393 nrg = get_file_region_entry_from_cache(
394 resv, last_accounted_offset, t);
395 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
396 list_add(&nrg->link, rg->link.prev);
397 coalesce_file_region(resv, nrg);
398 } else if (regions_needed)
399 *regions_needed += 1;
406 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
408 static int allocate_file_region_entries(struct resv_map *resv,
410 __must_hold(&resv->lock)
412 struct list_head allocated_regions;
413 int to_allocate = 0, i = 0;
414 struct file_region *trg = NULL, *rg = NULL;
416 VM_BUG_ON(regions_needed < 0);
418 INIT_LIST_HEAD(&allocated_regions);
421 * Check for sufficient descriptors in the cache to accommodate
422 * the number of in progress add operations plus regions_needed.
424 * This is a while loop because when we drop the lock, some other call
425 * to region_add or region_del may have consumed some region_entries,
426 * so we keep looping here until we finally have enough entries for
427 * (adds_in_progress + regions_needed).
429 while (resv->region_cache_count <
430 (resv->adds_in_progress + regions_needed)) {
431 to_allocate = resv->adds_in_progress + regions_needed -
432 resv->region_cache_count;
434 /* At this point, we should have enough entries in the cache
435 * for all the existings adds_in_progress. We should only be
436 * needing to allocate for regions_needed.
438 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
440 spin_unlock(&resv->lock);
441 for (i = 0; i < to_allocate; i++) {
442 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
445 list_add(&trg->link, &allocated_regions);
448 spin_lock(&resv->lock);
450 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
452 list_add(&rg->link, &resv->region_cache);
453 resv->region_cache_count++;
460 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
468 * Add the huge page range represented by [f, t) to the reserve
469 * map. Regions will be taken from the cache to fill in this range.
470 * Sufficient regions should exist in the cache due to the previous
471 * call to region_chg with the same range, but in some cases the cache will not
472 * have sufficient entries due to races with other code doing region_add or
473 * region_del. The extra needed entries will be allocated.
475 * regions_needed is the out value provided by a previous call to region_chg.
477 * Return the number of new huge pages added to the map. This number is greater
478 * than or equal to zero. If file_region entries needed to be allocated for
479 * this operation and we were not able to allocate, it returns -ENOMEM.
480 * region_add of regions of length 1 never allocate file_regions and cannot
481 * fail; region_chg will always allocate at least 1 entry and a region_add for
482 * 1 page will only require at most 1 entry.
484 static long region_add(struct resv_map *resv, long f, long t,
485 long in_regions_needed, struct hstate *h,
486 struct hugetlb_cgroup *h_cg)
488 long add = 0, actual_regions_needed = 0;
490 spin_lock(&resv->lock);
493 /* Count how many regions are actually needed to execute this add. */
494 add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed,
498 * Check for sufficient descriptors in the cache to accommodate
499 * this add operation. Note that actual_regions_needed may be greater
500 * than in_regions_needed, as the resv_map may have been modified since
501 * the region_chg call. In this case, we need to make sure that we
502 * allocate extra entries, such that we have enough for all the
503 * existing adds_in_progress, plus the excess needed for this
506 if (actual_regions_needed > in_regions_needed &&
507 resv->region_cache_count <
508 resv->adds_in_progress +
509 (actual_regions_needed - in_regions_needed)) {
510 /* region_add operation of range 1 should never need to
511 * allocate file_region entries.
513 VM_BUG_ON(t - f <= 1);
515 if (allocate_file_region_entries(
516 resv, actual_regions_needed - in_regions_needed)) {
523 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false);
525 resv->adds_in_progress -= in_regions_needed;
527 spin_unlock(&resv->lock);
533 * Examine the existing reserve map and determine how many
534 * huge pages in the specified range [f, t) are NOT currently
535 * represented. This routine is called before a subsequent
536 * call to region_add that will actually modify the reserve
537 * map to add the specified range [f, t). region_chg does
538 * not change the number of huge pages represented by the
539 * map. A number of new file_region structures is added to the cache as a
540 * placeholder, for the subsequent region_add call to use. At least 1
541 * file_region structure is added.
543 * out_regions_needed is the number of regions added to the
544 * resv->adds_in_progress. This value needs to be provided to a follow up call
545 * to region_add or region_abort for proper accounting.
547 * Returns the number of huge pages that need to be added to the existing
548 * reservation map for the range [f, t). This number is greater or equal to
549 * zero. -ENOMEM is returned if a new file_region structure or cache entry
550 * is needed and can not be allocated.
552 static long region_chg(struct resv_map *resv, long f, long t,
553 long *out_regions_needed)
557 spin_lock(&resv->lock);
559 /* Count how many hugepages in this range are NOT respresented. */
560 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
561 out_regions_needed, true);
563 if (*out_regions_needed == 0)
564 *out_regions_needed = 1;
566 if (allocate_file_region_entries(resv, *out_regions_needed))
569 resv->adds_in_progress += *out_regions_needed;
571 spin_unlock(&resv->lock);
576 * Abort the in progress add operation. The adds_in_progress field
577 * of the resv_map keeps track of the operations in progress between
578 * calls to region_chg and region_add. Operations are sometimes
579 * aborted after the call to region_chg. In such cases, region_abort
580 * is called to decrement the adds_in_progress counter. regions_needed
581 * is the value returned by the region_chg call, it is used to decrement
582 * the adds_in_progress counter.
584 * NOTE: The range arguments [f, t) are not needed or used in this
585 * routine. They are kept to make reading the calling code easier as
586 * arguments will match the associated region_chg call.
588 static void region_abort(struct resv_map *resv, long f, long t,
591 spin_lock(&resv->lock);
592 VM_BUG_ON(!resv->region_cache_count);
593 resv->adds_in_progress -= regions_needed;
594 spin_unlock(&resv->lock);
598 * Delete the specified range [f, t) from the reserve map. If the
599 * t parameter is LONG_MAX, this indicates that ALL regions after f
600 * should be deleted. Locate the regions which intersect [f, t)
601 * and either trim, delete or split the existing regions.
603 * Returns the number of huge pages deleted from the reserve map.
604 * In the normal case, the return value is zero or more. In the
605 * case where a region must be split, a new region descriptor must
606 * be allocated. If the allocation fails, -ENOMEM will be returned.
607 * NOTE: If the parameter t == LONG_MAX, then we will never split
608 * a region and possibly return -ENOMEM. Callers specifying
609 * t == LONG_MAX do not need to check for -ENOMEM error.
611 static long region_del(struct resv_map *resv, long f, long t)
613 struct list_head *head = &resv->regions;
614 struct file_region *rg, *trg;
615 struct file_region *nrg = NULL;
619 spin_lock(&resv->lock);
620 list_for_each_entry_safe(rg, trg, head, link) {
622 * Skip regions before the range to be deleted. file_region
623 * ranges are normally of the form [from, to). However, there
624 * may be a "placeholder" entry in the map which is of the form
625 * (from, to) with from == to. Check for placeholder entries
626 * at the beginning of the range to be deleted.
628 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
634 if (f > rg->from && t < rg->to) { /* Must split region */
636 * Check for an entry in the cache before dropping
637 * lock and attempting allocation.
640 resv->region_cache_count > resv->adds_in_progress) {
641 nrg = list_first_entry(&resv->region_cache,
644 list_del(&nrg->link);
645 resv->region_cache_count--;
649 spin_unlock(&resv->lock);
650 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
658 /* New entry for end of split region */
662 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
664 INIT_LIST_HEAD(&nrg->link);
666 /* Original entry is trimmed */
669 hugetlb_cgroup_uncharge_file_region(
670 resv, rg, nrg->to - nrg->from);
672 list_add(&nrg->link, &rg->link);
677 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
678 del += rg->to - rg->from;
679 hugetlb_cgroup_uncharge_file_region(resv, rg,
686 if (f <= rg->from) { /* Trim beginning of region */
690 hugetlb_cgroup_uncharge_file_region(resv, rg,
692 } else { /* Trim end of region */
696 hugetlb_cgroup_uncharge_file_region(resv, rg,
701 spin_unlock(&resv->lock);
707 * A rare out of memory error was encountered which prevented removal of
708 * the reserve map region for a page. The huge page itself was free'ed
709 * and removed from the page cache. This routine will adjust the subpool
710 * usage count, and the global reserve count if needed. By incrementing
711 * these counts, the reserve map entry which could not be deleted will
712 * appear as a "reserved" entry instead of simply dangling with incorrect
715 void hugetlb_fix_reserve_counts(struct inode *inode)
717 struct hugepage_subpool *spool = subpool_inode(inode);
720 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
722 struct hstate *h = hstate_inode(inode);
724 hugetlb_acct_memory(h, 1);
729 * Count and return the number of huge pages in the reserve map
730 * that intersect with the range [f, t).
732 static long region_count(struct resv_map *resv, long f, long t)
734 struct list_head *head = &resv->regions;
735 struct file_region *rg;
738 spin_lock(&resv->lock);
739 /* Locate each segment we overlap with, and count that overlap. */
740 list_for_each_entry(rg, head, link) {
749 seg_from = max(rg->from, f);
750 seg_to = min(rg->to, t);
752 chg += seg_to - seg_from;
754 spin_unlock(&resv->lock);
760 * Convert the address within this vma to the page offset within
761 * the mapping, in pagecache page units; huge pages here.
763 static pgoff_t vma_hugecache_offset(struct hstate *h,
764 struct vm_area_struct *vma, unsigned long address)
766 return ((address - vma->vm_start) >> huge_page_shift(h)) +
767 (vma->vm_pgoff >> huge_page_order(h));
770 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
771 unsigned long address)
773 return vma_hugecache_offset(hstate_vma(vma), vma, address);
775 EXPORT_SYMBOL_GPL(linear_hugepage_index);
778 * Return the size of the pages allocated when backing a VMA. In the majority
779 * cases this will be same size as used by the page table entries.
781 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
783 if (vma->vm_ops && vma->vm_ops->pagesize)
784 return vma->vm_ops->pagesize(vma);
787 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
790 * Return the page size being used by the MMU to back a VMA. In the majority
791 * of cases, the page size used by the kernel matches the MMU size. On
792 * architectures where it differs, an architecture-specific 'strong'
793 * version of this symbol is required.
795 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
797 return vma_kernel_pagesize(vma);
801 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
802 * bits of the reservation map pointer, which are always clear due to
805 #define HPAGE_RESV_OWNER (1UL << 0)
806 #define HPAGE_RESV_UNMAPPED (1UL << 1)
807 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
810 * These helpers are used to track how many pages are reserved for
811 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
812 * is guaranteed to have their future faults succeed.
814 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
815 * the reserve counters are updated with the hugetlb_lock held. It is safe
816 * to reset the VMA at fork() time as it is not in use yet and there is no
817 * chance of the global counters getting corrupted as a result of the values.
819 * The private mapping reservation is represented in a subtly different
820 * manner to a shared mapping. A shared mapping has a region map associated
821 * with the underlying file, this region map represents the backing file
822 * pages which have ever had a reservation assigned which this persists even
823 * after the page is instantiated. A private mapping has a region map
824 * associated with the original mmap which is attached to all VMAs which
825 * reference it, this region map represents those offsets which have consumed
826 * reservation ie. where pages have been instantiated.
828 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
830 return (unsigned long)vma->vm_private_data;
833 static void set_vma_private_data(struct vm_area_struct *vma,
836 vma->vm_private_data = (void *)value;
840 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
841 struct hugetlb_cgroup *h_cg,
844 #ifdef CONFIG_CGROUP_HUGETLB
846 resv_map->reservation_counter = NULL;
847 resv_map->pages_per_hpage = 0;
848 resv_map->css = NULL;
850 resv_map->reservation_counter =
851 &h_cg->rsvd_hugepage[hstate_index(h)];
852 resv_map->pages_per_hpage = pages_per_huge_page(h);
853 resv_map->css = &h_cg->css;
858 struct resv_map *resv_map_alloc(void)
860 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
861 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
863 if (!resv_map || !rg) {
869 kref_init(&resv_map->refs);
870 spin_lock_init(&resv_map->lock);
871 INIT_LIST_HEAD(&resv_map->regions);
873 resv_map->adds_in_progress = 0;
875 * Initialize these to 0. On shared mappings, 0's here indicate these
876 * fields don't do cgroup accounting. On private mappings, these will be
877 * re-initialized to the proper values, to indicate that hugetlb cgroup
878 * reservations are to be un-charged from here.
880 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
882 INIT_LIST_HEAD(&resv_map->region_cache);
883 list_add(&rg->link, &resv_map->region_cache);
884 resv_map->region_cache_count = 1;
889 void resv_map_release(struct kref *ref)
891 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
892 struct list_head *head = &resv_map->region_cache;
893 struct file_region *rg, *trg;
895 /* Clear out any active regions before we release the map. */
896 region_del(resv_map, 0, LONG_MAX);
898 /* ... and any entries left in the cache */
899 list_for_each_entry_safe(rg, trg, head, link) {
904 VM_BUG_ON(resv_map->adds_in_progress);
909 static inline struct resv_map *inode_resv_map(struct inode *inode)
912 * At inode evict time, i_mapping may not point to the original
913 * address space within the inode. This original address space
914 * contains the pointer to the resv_map. So, always use the
915 * address space embedded within the inode.
916 * The VERY common case is inode->mapping == &inode->i_data but,
917 * this may not be true for device special inodes.
919 return (struct resv_map *)(&inode->i_data)->private_data;
922 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
924 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
925 if (vma->vm_flags & VM_MAYSHARE) {
926 struct address_space *mapping = vma->vm_file->f_mapping;
927 struct inode *inode = mapping->host;
929 return inode_resv_map(inode);
932 return (struct resv_map *)(get_vma_private_data(vma) &
937 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
939 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
940 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
942 set_vma_private_data(vma, (get_vma_private_data(vma) &
943 HPAGE_RESV_MASK) | (unsigned long)map);
946 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
948 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
949 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
951 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
954 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
956 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958 return (get_vma_private_data(vma) & flag) != 0;
961 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
962 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
964 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
965 if (!(vma->vm_flags & VM_MAYSHARE))
966 vma->vm_private_data = (void *)0;
969 /* Returns true if the VMA has associated reserve pages */
970 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
972 if (vma->vm_flags & VM_NORESERVE) {
974 * This address is already reserved by other process(chg == 0),
975 * so, we should decrement reserved count. Without decrementing,
976 * reserve count remains after releasing inode, because this
977 * allocated page will go into page cache and is regarded as
978 * coming from reserved pool in releasing step. Currently, we
979 * don't have any other solution to deal with this situation
980 * properly, so add work-around here.
982 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
988 /* Shared mappings always use reserves */
989 if (vma->vm_flags & VM_MAYSHARE) {
991 * We know VM_NORESERVE is not set. Therefore, there SHOULD
992 * be a region map for all pages. The only situation where
993 * there is no region map is if a hole was punched via
994 * fallocate. In this case, there really are no reserves to
995 * use. This situation is indicated if chg != 0.
1004 * Only the process that called mmap() has reserves for
1007 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1009 * Like the shared case above, a hole punch or truncate
1010 * could have been performed on the private mapping.
1011 * Examine the value of chg to determine if reserves
1012 * actually exist or were previously consumed.
1013 * Very Subtle - The value of chg comes from a previous
1014 * call to vma_needs_reserves(). The reserve map for
1015 * private mappings has different (opposite) semantics
1016 * than that of shared mappings. vma_needs_reserves()
1017 * has already taken this difference in semantics into
1018 * account. Therefore, the meaning of chg is the same
1019 * as in the shared case above. Code could easily be
1020 * combined, but keeping it separate draws attention to
1021 * subtle differences.
1032 static void enqueue_huge_page(struct hstate *h, struct page *page)
1034 int nid = page_to_nid(page);
1035 list_move(&page->lru, &h->hugepage_freelists[nid]);
1036 h->free_huge_pages++;
1037 h->free_huge_pages_node[nid]++;
1040 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1044 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
1045 if (!PageHWPoison(page))
1048 * if 'non-isolated free hugepage' not found on the list,
1049 * the allocation fails.
1051 if (&h->hugepage_freelists[nid] == &page->lru)
1053 list_move(&page->lru, &h->hugepage_activelist);
1054 set_page_refcounted(page);
1055 h->free_huge_pages--;
1056 h->free_huge_pages_node[nid]--;
1060 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1063 unsigned int cpuset_mems_cookie;
1064 struct zonelist *zonelist;
1067 int node = NUMA_NO_NODE;
1069 zonelist = node_zonelist(nid, gfp_mask);
1072 cpuset_mems_cookie = read_mems_allowed_begin();
1073 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1076 if (!cpuset_zone_allowed(zone, gfp_mask))
1079 * no need to ask again on the same node. Pool is node rather than
1082 if (zone_to_nid(zone) == node)
1084 node = zone_to_nid(zone);
1086 page = dequeue_huge_page_node_exact(h, node);
1090 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1096 /* Movability of hugepages depends on migration support. */
1097 static inline gfp_t htlb_alloc_mask(struct hstate *h)
1099 if (hugepage_movable_supported(h))
1100 return GFP_HIGHUSER_MOVABLE;
1102 return GFP_HIGHUSER;
1105 static struct page *dequeue_huge_page_vma(struct hstate *h,
1106 struct vm_area_struct *vma,
1107 unsigned long address, int avoid_reserve,
1111 struct mempolicy *mpol;
1113 nodemask_t *nodemask;
1117 * A child process with MAP_PRIVATE mappings created by their parent
1118 * have no page reserves. This check ensures that reservations are
1119 * not "stolen". The child may still get SIGKILLed
1121 if (!vma_has_reserves(vma, chg) &&
1122 h->free_huge_pages - h->resv_huge_pages == 0)
1125 /* If reserves cannot be used, ensure enough pages are in the pool */
1126 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1129 gfp_mask = htlb_alloc_mask(h);
1130 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1131 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1132 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1133 SetPagePrivate(page);
1134 h->resv_huge_pages--;
1137 mpol_cond_put(mpol);
1145 * common helper functions for hstate_next_node_to_{alloc|free}.
1146 * We may have allocated or freed a huge page based on a different
1147 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1148 * be outside of *nodes_allowed. Ensure that we use an allowed
1149 * node for alloc or free.
1151 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1153 nid = next_node_in(nid, *nodes_allowed);
1154 VM_BUG_ON(nid >= MAX_NUMNODES);
1159 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1161 if (!node_isset(nid, *nodes_allowed))
1162 nid = next_node_allowed(nid, nodes_allowed);
1167 * returns the previously saved node ["this node"] from which to
1168 * allocate a persistent huge page for the pool and advance the
1169 * next node from which to allocate, handling wrap at end of node
1172 static int hstate_next_node_to_alloc(struct hstate *h,
1173 nodemask_t *nodes_allowed)
1177 VM_BUG_ON(!nodes_allowed);
1179 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1180 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1186 * helper for free_pool_huge_page() - return the previously saved
1187 * node ["this node"] from which to free a huge page. Advance the
1188 * next node id whether or not we find a free huge page to free so
1189 * that the next attempt to free addresses the next node.
1191 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1195 VM_BUG_ON(!nodes_allowed);
1197 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1198 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1203 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1204 for (nr_nodes = nodes_weight(*mask); \
1206 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1209 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1210 for (nr_nodes = nodes_weight(*mask); \
1212 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1215 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1216 static void destroy_compound_gigantic_page(struct page *page,
1220 int nr_pages = 1 << order;
1221 struct page *p = page + 1;
1223 atomic_set(compound_mapcount_ptr(page), 0);
1224 if (hpage_pincount_available(page))
1225 atomic_set(compound_pincount_ptr(page), 0);
1227 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1228 clear_compound_head(p);
1229 set_page_refcounted(p);
1232 set_compound_order(page, 0);
1233 __ClearPageHead(page);
1236 static void free_gigantic_page(struct page *page, unsigned int order)
1239 * If the page isn't allocated using the cma allocator,
1240 * cma_release() returns false.
1243 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1247 free_contig_range(page_to_pfn(page), 1 << order);
1250 #ifdef CONFIG_CONTIG_ALLOC
1251 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1252 int nid, nodemask_t *nodemask)
1254 unsigned long nr_pages = 1UL << huge_page_order(h);
1261 for_each_node_mask(node, *nodemask) {
1262 if (!hugetlb_cma[node])
1265 page = cma_alloc(hugetlb_cma[node], nr_pages,
1266 huge_page_order(h), true);
1273 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1276 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1277 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1278 #else /* !CONFIG_CONTIG_ALLOC */
1279 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1280 int nid, nodemask_t *nodemask)
1284 #endif /* CONFIG_CONTIG_ALLOC */
1286 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1287 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1288 int nid, nodemask_t *nodemask)
1292 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1293 static inline void destroy_compound_gigantic_page(struct page *page,
1294 unsigned int order) { }
1297 static void update_and_free_page(struct hstate *h, struct page *page)
1301 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1305 h->nr_huge_pages_node[page_to_nid(page)]--;
1306 for (i = 0; i < pages_per_huge_page(h); i++) {
1307 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1308 1 << PG_referenced | 1 << PG_dirty |
1309 1 << PG_active | 1 << PG_private |
1312 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1313 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1314 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1315 set_page_refcounted(page);
1316 if (hstate_is_gigantic(h)) {
1318 * Temporarily drop the hugetlb_lock, because
1319 * we might block in free_gigantic_page().
1321 spin_unlock(&hugetlb_lock);
1322 destroy_compound_gigantic_page(page, huge_page_order(h));
1323 free_gigantic_page(page, huge_page_order(h));
1324 spin_lock(&hugetlb_lock);
1326 __free_pages(page, huge_page_order(h));
1330 struct hstate *size_to_hstate(unsigned long size)
1334 for_each_hstate(h) {
1335 if (huge_page_size(h) == size)
1342 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1343 * to hstate->hugepage_activelist.)
1345 * This function can be called for tail pages, but never returns true for them.
1347 bool page_huge_active(struct page *page)
1349 VM_BUG_ON_PAGE(!PageHuge(page), page);
1350 return PageHead(page) && PagePrivate(&page[1]);
1353 /* never called for tail page */
1354 static void set_page_huge_active(struct page *page)
1356 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1357 SetPagePrivate(&page[1]);
1360 static void clear_page_huge_active(struct page *page)
1362 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1363 ClearPagePrivate(&page[1]);
1367 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1370 static inline bool PageHugeTemporary(struct page *page)
1372 if (!PageHuge(page))
1375 return (unsigned long)page[2].mapping == -1U;
1378 static inline void SetPageHugeTemporary(struct page *page)
1380 page[2].mapping = (void *)-1U;
1383 static inline void ClearPageHugeTemporary(struct page *page)
1385 page[2].mapping = NULL;
1388 static void __free_huge_page(struct page *page)
1391 * Can't pass hstate in here because it is called from the
1392 * compound page destructor.
1394 struct hstate *h = page_hstate(page);
1395 int nid = page_to_nid(page);
1396 struct hugepage_subpool *spool =
1397 (struct hugepage_subpool *)page_private(page);
1398 bool restore_reserve;
1400 VM_BUG_ON_PAGE(page_count(page), page);
1401 VM_BUG_ON_PAGE(page_mapcount(page), page);
1403 set_page_private(page, 0);
1404 page->mapping = NULL;
1405 restore_reserve = PagePrivate(page);
1406 ClearPagePrivate(page);
1409 * If PagePrivate() was set on page, page allocation consumed a
1410 * reservation. If the page was associated with a subpool, there
1411 * would have been a page reserved in the subpool before allocation
1412 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1413 * reservtion, do not call hugepage_subpool_put_pages() as this will
1414 * remove the reserved page from the subpool.
1416 if (!restore_reserve) {
1418 * A return code of zero implies that the subpool will be
1419 * under its minimum size if the reservation is not restored
1420 * after page is free. Therefore, force restore_reserve
1423 if (hugepage_subpool_put_pages(spool, 1) == 0)
1424 restore_reserve = true;
1427 spin_lock(&hugetlb_lock);
1428 clear_page_huge_active(page);
1429 hugetlb_cgroup_uncharge_page(hstate_index(h),
1430 pages_per_huge_page(h), page);
1431 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1432 pages_per_huge_page(h), page);
1433 if (restore_reserve)
1434 h->resv_huge_pages++;
1436 if (PageHugeTemporary(page)) {
1437 list_del(&page->lru);
1438 ClearPageHugeTemporary(page);
1439 update_and_free_page(h, page);
1440 } else if (h->surplus_huge_pages_node[nid]) {
1441 /* remove the page from active list */
1442 list_del(&page->lru);
1443 update_and_free_page(h, page);
1444 h->surplus_huge_pages--;
1445 h->surplus_huge_pages_node[nid]--;
1447 arch_clear_hugepage_flags(page);
1448 enqueue_huge_page(h, page);
1450 spin_unlock(&hugetlb_lock);
1454 * As free_huge_page() can be called from a non-task context, we have
1455 * to defer the actual freeing in a workqueue to prevent potential
1456 * hugetlb_lock deadlock.
1458 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1459 * be freed and frees them one-by-one. As the page->mapping pointer is
1460 * going to be cleared in __free_huge_page() anyway, it is reused as the
1461 * llist_node structure of a lockless linked list of huge pages to be freed.
1463 static LLIST_HEAD(hpage_freelist);
1465 static void free_hpage_workfn(struct work_struct *work)
1467 struct llist_node *node;
1470 node = llist_del_all(&hpage_freelist);
1473 page = container_of((struct address_space **)node,
1474 struct page, mapping);
1476 __free_huge_page(page);
1479 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1481 void free_huge_page(struct page *page)
1484 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1488 * Only call schedule_work() if hpage_freelist is previously
1489 * empty. Otherwise, schedule_work() had been called but the
1490 * workfn hasn't retrieved the list yet.
1492 if (llist_add((struct llist_node *)&page->mapping,
1494 schedule_work(&free_hpage_work);
1498 __free_huge_page(page);
1501 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1503 INIT_LIST_HEAD(&page->lru);
1504 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1505 spin_lock(&hugetlb_lock);
1506 set_hugetlb_cgroup(page, NULL);
1507 set_hugetlb_cgroup_rsvd(page, NULL);
1509 h->nr_huge_pages_node[nid]++;
1510 spin_unlock(&hugetlb_lock);
1513 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1516 int nr_pages = 1 << order;
1517 struct page *p = page + 1;
1519 /* we rely on prep_new_huge_page to set the destructor */
1520 set_compound_order(page, order);
1521 __ClearPageReserved(page);
1522 __SetPageHead(page);
1523 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1525 * For gigantic hugepages allocated through bootmem at
1526 * boot, it's safer to be consistent with the not-gigantic
1527 * hugepages and clear the PG_reserved bit from all tail pages
1528 * too. Otherwise drivers using get_user_pages() to access tail
1529 * pages may get the reference counting wrong if they see
1530 * PG_reserved set on a tail page (despite the head page not
1531 * having PG_reserved set). Enforcing this consistency between
1532 * head and tail pages allows drivers to optimize away a check
1533 * on the head page when they need know if put_page() is needed
1534 * after get_user_pages().
1536 __ClearPageReserved(p);
1537 set_page_count(p, 0);
1538 set_compound_head(p, page);
1540 atomic_set(compound_mapcount_ptr(page), -1);
1542 if (hpage_pincount_available(page))
1543 atomic_set(compound_pincount_ptr(page), 0);
1547 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1548 * transparent huge pages. See the PageTransHuge() documentation for more
1551 int PageHuge(struct page *page)
1553 if (!PageCompound(page))
1556 page = compound_head(page);
1557 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1559 EXPORT_SYMBOL_GPL(PageHuge);
1562 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1563 * normal or transparent huge pages.
1565 int PageHeadHuge(struct page *page_head)
1567 if (!PageHead(page_head))
1570 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1574 * Find address_space associated with hugetlbfs page.
1575 * Upon entry page is locked and page 'was' mapped although mapped state
1576 * could change. If necessary, use anon_vma to find vma and associated
1577 * address space. The returned mapping may be stale, but it can not be
1578 * invalid as page lock (which is held) is required to destroy mapping.
1580 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1582 struct anon_vma *anon_vma;
1583 pgoff_t pgoff_start, pgoff_end;
1584 struct anon_vma_chain *avc;
1585 struct address_space *mapping = page_mapping(hpage);
1587 /* Simple file based mapping */
1592 * Even anonymous hugetlbfs mappings are associated with an
1593 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1594 * code). Find a vma associated with the anonymous vma, and
1595 * use the file pointer to get address_space.
1597 anon_vma = page_lock_anon_vma_read(hpage);
1599 return mapping; /* NULL */
1601 /* Use first found vma */
1602 pgoff_start = page_to_pgoff(hpage);
1603 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
1604 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1605 pgoff_start, pgoff_end) {
1606 struct vm_area_struct *vma = avc->vma;
1608 mapping = vma->vm_file->f_mapping;
1612 anon_vma_unlock_read(anon_vma);
1617 * Find and lock address space (mapping) in write mode.
1619 * Upon entry, the page is locked which allows us to find the mapping
1620 * even in the case of an anon page. However, locking order dictates
1621 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1622 * specific. So, we first try to lock the sema while still holding the
1623 * page lock. If this works, great! If not, then we need to drop the
1624 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1625 * course, need to revalidate state along the way.
1627 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1629 struct address_space *mapping, *mapping2;
1631 mapping = _get_hugetlb_page_mapping(hpage);
1637 * If no contention, take lock and return
1639 if (i_mmap_trylock_write(mapping))
1643 * Must drop page lock and wait on mapping sema.
1644 * Note: Once page lock is dropped, mapping could become invalid.
1645 * As a hack, increase map count until we lock page again.
1647 atomic_inc(&hpage->_mapcount);
1649 i_mmap_lock_write(mapping);
1651 atomic_add_negative(-1, &hpage->_mapcount);
1653 /* verify page is still mapped */
1654 if (!page_mapped(hpage)) {
1655 i_mmap_unlock_write(mapping);
1660 * Get address space again and verify it is the same one
1661 * we locked. If not, drop lock and retry.
1663 mapping2 = _get_hugetlb_page_mapping(hpage);
1664 if (mapping2 != mapping) {
1665 i_mmap_unlock_write(mapping);
1673 pgoff_t __basepage_index(struct page *page)
1675 struct page *page_head = compound_head(page);
1676 pgoff_t index = page_index(page_head);
1677 unsigned long compound_idx;
1679 if (!PageHuge(page_head))
1680 return page_index(page);
1682 if (compound_order(page_head) >= MAX_ORDER)
1683 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1685 compound_idx = page - page_head;
1687 return (index << compound_order(page_head)) + compound_idx;
1690 static struct page *alloc_buddy_huge_page(struct hstate *h,
1691 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1692 nodemask_t *node_alloc_noretry)
1694 int order = huge_page_order(h);
1696 bool alloc_try_hard = true;
1699 * By default we always try hard to allocate the page with
1700 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1701 * a loop (to adjust global huge page counts) and previous allocation
1702 * failed, do not continue to try hard on the same node. Use the
1703 * node_alloc_noretry bitmap to manage this state information.
1705 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1706 alloc_try_hard = false;
1707 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1709 gfp_mask |= __GFP_RETRY_MAYFAIL;
1710 if (nid == NUMA_NO_NODE)
1711 nid = numa_mem_id();
1712 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1714 __count_vm_event(HTLB_BUDDY_PGALLOC);
1716 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1719 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1720 * indicates an overall state change. Clear bit so that we resume
1721 * normal 'try hard' allocations.
1723 if (node_alloc_noretry && page && !alloc_try_hard)
1724 node_clear(nid, *node_alloc_noretry);
1727 * If we tried hard to get a page but failed, set bit so that
1728 * subsequent attempts will not try as hard until there is an
1729 * overall state change.
1731 if (node_alloc_noretry && !page && alloc_try_hard)
1732 node_set(nid, *node_alloc_noretry);
1738 * Common helper to allocate a fresh hugetlb page. All specific allocators
1739 * should use this function to get new hugetlb pages
1741 static struct page *alloc_fresh_huge_page(struct hstate *h,
1742 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1743 nodemask_t *node_alloc_noretry)
1747 if (hstate_is_gigantic(h))
1748 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1750 page = alloc_buddy_huge_page(h, gfp_mask,
1751 nid, nmask, node_alloc_noretry);
1755 if (hstate_is_gigantic(h))
1756 prep_compound_gigantic_page(page, huge_page_order(h));
1757 prep_new_huge_page(h, page, page_to_nid(page));
1763 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1766 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1767 nodemask_t *node_alloc_noretry)
1771 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1773 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1774 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1775 node_alloc_noretry);
1783 put_page(page); /* free it into the hugepage allocator */
1789 * Free huge page from pool from next node to free.
1790 * Attempt to keep persistent huge pages more or less
1791 * balanced over allowed nodes.
1792 * Called with hugetlb_lock locked.
1794 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1800 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1802 * If we're returning unused surplus pages, only examine
1803 * nodes with surplus pages.
1805 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1806 !list_empty(&h->hugepage_freelists[node])) {
1808 list_entry(h->hugepage_freelists[node].next,
1810 list_del(&page->lru);
1811 h->free_huge_pages--;
1812 h->free_huge_pages_node[node]--;
1814 h->surplus_huge_pages--;
1815 h->surplus_huge_pages_node[node]--;
1817 update_and_free_page(h, page);
1827 * Dissolve a given free hugepage into free buddy pages. This function does
1828 * nothing for in-use hugepages and non-hugepages.
1829 * This function returns values like below:
1831 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1832 * (allocated or reserved.)
1833 * 0: successfully dissolved free hugepages or the page is not a
1834 * hugepage (considered as already dissolved)
1836 int dissolve_free_huge_page(struct page *page)
1840 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1841 if (!PageHuge(page))
1844 spin_lock(&hugetlb_lock);
1845 if (!PageHuge(page)) {
1850 if (!page_count(page)) {
1851 struct page *head = compound_head(page);
1852 struct hstate *h = page_hstate(head);
1853 int nid = page_to_nid(head);
1854 if (h->free_huge_pages - h->resv_huge_pages == 0)
1857 * Move PageHWPoison flag from head page to the raw error page,
1858 * which makes any subpages rather than the error page reusable.
1860 if (PageHWPoison(head) && page != head) {
1861 SetPageHWPoison(page);
1862 ClearPageHWPoison(head);
1864 list_del(&head->lru);
1865 h->free_huge_pages--;
1866 h->free_huge_pages_node[nid]--;
1867 h->max_huge_pages--;
1868 update_and_free_page(h, head);
1872 spin_unlock(&hugetlb_lock);
1877 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1878 * make specified memory blocks removable from the system.
1879 * Note that this will dissolve a free gigantic hugepage completely, if any
1880 * part of it lies within the given range.
1881 * Also note that if dissolve_free_huge_page() returns with an error, all
1882 * free hugepages that were dissolved before that error are lost.
1884 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1890 if (!hugepages_supported())
1893 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1894 page = pfn_to_page(pfn);
1895 rc = dissolve_free_huge_page(page);
1904 * Allocates a fresh surplus page from the page allocator.
1906 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1907 int nid, nodemask_t *nmask)
1909 struct page *page = NULL;
1911 if (hstate_is_gigantic(h))
1914 spin_lock(&hugetlb_lock);
1915 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1917 spin_unlock(&hugetlb_lock);
1919 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1923 spin_lock(&hugetlb_lock);
1925 * We could have raced with the pool size change.
1926 * Double check that and simply deallocate the new page
1927 * if we would end up overcommiting the surpluses. Abuse
1928 * temporary page to workaround the nasty free_huge_page
1931 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1932 SetPageHugeTemporary(page);
1933 spin_unlock(&hugetlb_lock);
1937 h->surplus_huge_pages++;
1938 h->surplus_huge_pages_node[page_to_nid(page)]++;
1942 spin_unlock(&hugetlb_lock);
1947 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1948 int nid, nodemask_t *nmask)
1952 if (hstate_is_gigantic(h))
1955 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1960 * We do not account these pages as surplus because they are only
1961 * temporary and will be released properly on the last reference
1963 SetPageHugeTemporary(page);
1969 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1972 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1973 struct vm_area_struct *vma, unsigned long addr)
1976 struct mempolicy *mpol;
1977 gfp_t gfp_mask = htlb_alloc_mask(h);
1979 nodemask_t *nodemask;
1981 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1982 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1983 mpol_cond_put(mpol);
1988 /* page migration callback function */
1989 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1991 gfp_t gfp_mask = htlb_alloc_mask(h);
1992 struct page *page = NULL;
1994 if (nid != NUMA_NO_NODE)
1995 gfp_mask |= __GFP_THISNODE;
1997 spin_lock(&hugetlb_lock);
1998 if (h->free_huge_pages - h->resv_huge_pages > 0)
1999 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
2000 spin_unlock(&hugetlb_lock);
2003 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
2008 /* page migration callback function */
2009 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2012 gfp_t gfp_mask = htlb_alloc_mask(h);
2014 spin_lock(&hugetlb_lock);
2015 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2018 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2020 spin_unlock(&hugetlb_lock);
2024 spin_unlock(&hugetlb_lock);
2026 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2029 /* mempolicy aware migration callback */
2030 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2031 unsigned long address)
2033 struct mempolicy *mpol;
2034 nodemask_t *nodemask;
2039 gfp_mask = htlb_alloc_mask(h);
2040 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2041 page = alloc_huge_page_nodemask(h, node, nodemask);
2042 mpol_cond_put(mpol);
2048 * Increase the hugetlb pool such that it can accommodate a reservation
2051 static int gather_surplus_pages(struct hstate *h, int delta)
2052 __must_hold(&hugetlb_lock)
2054 struct list_head surplus_list;
2055 struct page *page, *tmp;
2057 int needed, allocated;
2058 bool alloc_ok = true;
2060 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2062 h->resv_huge_pages += delta;
2067 INIT_LIST_HEAD(&surplus_list);
2071 spin_unlock(&hugetlb_lock);
2072 for (i = 0; i < needed; i++) {
2073 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2074 NUMA_NO_NODE, NULL);
2079 list_add(&page->lru, &surplus_list);
2085 * After retaking hugetlb_lock, we need to recalculate 'needed'
2086 * because either resv_huge_pages or free_huge_pages may have changed.
2088 spin_lock(&hugetlb_lock);
2089 needed = (h->resv_huge_pages + delta) -
2090 (h->free_huge_pages + allocated);
2095 * We were not able to allocate enough pages to
2096 * satisfy the entire reservation so we free what
2097 * we've allocated so far.
2102 * The surplus_list now contains _at_least_ the number of extra pages
2103 * needed to accommodate the reservation. Add the appropriate number
2104 * of pages to the hugetlb pool and free the extras back to the buddy
2105 * allocator. Commit the entire reservation here to prevent another
2106 * process from stealing the pages as they are added to the pool but
2107 * before they are reserved.
2109 needed += allocated;
2110 h->resv_huge_pages += delta;
2113 /* Free the needed pages to the hugetlb pool */
2114 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2118 * This page is now managed by the hugetlb allocator and has
2119 * no users -- drop the buddy allocator's reference.
2121 put_page_testzero(page);
2122 VM_BUG_ON_PAGE(page_count(page), page);
2123 enqueue_huge_page(h, page);
2126 spin_unlock(&hugetlb_lock);
2128 /* Free unnecessary surplus pages to the buddy allocator */
2129 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2131 spin_lock(&hugetlb_lock);
2137 * This routine has two main purposes:
2138 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2139 * in unused_resv_pages. This corresponds to the prior adjustments made
2140 * to the associated reservation map.
2141 * 2) Free any unused surplus pages that may have been allocated to satisfy
2142 * the reservation. As many as unused_resv_pages may be freed.
2144 * Called with hugetlb_lock held. However, the lock could be dropped (and
2145 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2146 * we must make sure nobody else can claim pages we are in the process of
2147 * freeing. Do this by ensuring resv_huge_page always is greater than the
2148 * number of huge pages we plan to free when dropping the lock.
2150 static void return_unused_surplus_pages(struct hstate *h,
2151 unsigned long unused_resv_pages)
2153 unsigned long nr_pages;
2155 /* Cannot return gigantic pages currently */
2156 if (hstate_is_gigantic(h))
2160 * Part (or even all) of the reservation could have been backed
2161 * by pre-allocated pages. Only free surplus pages.
2163 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2166 * We want to release as many surplus pages as possible, spread
2167 * evenly across all nodes with memory. Iterate across these nodes
2168 * until we can no longer free unreserved surplus pages. This occurs
2169 * when the nodes with surplus pages have no free pages.
2170 * free_pool_huge_page() will balance the the freed pages across the
2171 * on-line nodes with memory and will handle the hstate accounting.
2173 * Note that we decrement resv_huge_pages as we free the pages. If
2174 * we drop the lock, resv_huge_pages will still be sufficiently large
2175 * to cover subsequent pages we may free.
2177 while (nr_pages--) {
2178 h->resv_huge_pages--;
2179 unused_resv_pages--;
2180 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2182 cond_resched_lock(&hugetlb_lock);
2186 /* Fully uncommit the reservation */
2187 h->resv_huge_pages -= unused_resv_pages;
2192 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2193 * are used by the huge page allocation routines to manage reservations.
2195 * vma_needs_reservation is called to determine if the huge page at addr
2196 * within the vma has an associated reservation. If a reservation is
2197 * needed, the value 1 is returned. The caller is then responsible for
2198 * managing the global reservation and subpool usage counts. After
2199 * the huge page has been allocated, vma_commit_reservation is called
2200 * to add the page to the reservation map. If the page allocation fails,
2201 * the reservation must be ended instead of committed. vma_end_reservation
2202 * is called in such cases.
2204 * In the normal case, vma_commit_reservation returns the same value
2205 * as the preceding vma_needs_reservation call. The only time this
2206 * is not the case is if a reserve map was changed between calls. It
2207 * is the responsibility of the caller to notice the difference and
2208 * take appropriate action.
2210 * vma_add_reservation is used in error paths where a reservation must
2211 * be restored when a newly allocated huge page must be freed. It is
2212 * to be called after calling vma_needs_reservation to determine if a
2213 * reservation exists.
2215 enum vma_resv_mode {
2221 static long __vma_reservation_common(struct hstate *h,
2222 struct vm_area_struct *vma, unsigned long addr,
2223 enum vma_resv_mode mode)
2225 struct resv_map *resv;
2228 long dummy_out_regions_needed;
2230 resv = vma_resv_map(vma);
2234 idx = vma_hugecache_offset(h, vma, addr);
2236 case VMA_NEEDS_RESV:
2237 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2238 /* We assume that vma_reservation_* routines always operate on
2239 * 1 page, and that adding to resv map a 1 page entry can only
2240 * ever require 1 region.
2242 VM_BUG_ON(dummy_out_regions_needed != 1);
2244 case VMA_COMMIT_RESV:
2245 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2246 /* region_add calls of range 1 should never fail. */
2250 region_abort(resv, idx, idx + 1, 1);
2254 if (vma->vm_flags & VM_MAYSHARE) {
2255 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2256 /* region_add calls of range 1 should never fail. */
2259 region_abort(resv, idx, idx + 1, 1);
2260 ret = region_del(resv, idx, idx + 1);
2267 if (vma->vm_flags & VM_MAYSHARE)
2269 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2271 * In most cases, reserves always exist for private mappings.
2272 * However, a file associated with mapping could have been
2273 * hole punched or truncated after reserves were consumed.
2274 * As subsequent fault on such a range will not use reserves.
2275 * Subtle - The reserve map for private mappings has the
2276 * opposite meaning than that of shared mappings. If NO
2277 * entry is in the reserve map, it means a reservation exists.
2278 * If an entry exists in the reserve map, it means the
2279 * reservation has already been consumed. As a result, the
2280 * return value of this routine is the opposite of the
2281 * value returned from reserve map manipulation routines above.
2289 return ret < 0 ? ret : 0;
2292 static long vma_needs_reservation(struct hstate *h,
2293 struct vm_area_struct *vma, unsigned long addr)
2295 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2298 static long vma_commit_reservation(struct hstate *h,
2299 struct vm_area_struct *vma, unsigned long addr)
2301 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2304 static void vma_end_reservation(struct hstate *h,
2305 struct vm_area_struct *vma, unsigned long addr)
2307 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2310 static long vma_add_reservation(struct hstate *h,
2311 struct vm_area_struct *vma, unsigned long addr)
2313 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2317 * This routine is called to restore a reservation on error paths. In the
2318 * specific error paths, a huge page was allocated (via alloc_huge_page)
2319 * and is about to be freed. If a reservation for the page existed,
2320 * alloc_huge_page would have consumed the reservation and set PagePrivate
2321 * in the newly allocated page. When the page is freed via free_huge_page,
2322 * the global reservation count will be incremented if PagePrivate is set.
2323 * However, free_huge_page can not adjust the reserve map. Adjust the
2324 * reserve map here to be consistent with global reserve count adjustments
2325 * to be made by free_huge_page.
2327 static void restore_reserve_on_error(struct hstate *h,
2328 struct vm_area_struct *vma, unsigned long address,
2331 if (unlikely(PagePrivate(page))) {
2332 long rc = vma_needs_reservation(h, vma, address);
2334 if (unlikely(rc < 0)) {
2336 * Rare out of memory condition in reserve map
2337 * manipulation. Clear PagePrivate so that
2338 * global reserve count will not be incremented
2339 * by free_huge_page. This will make it appear
2340 * as though the reservation for this page was
2341 * consumed. This may prevent the task from
2342 * faulting in the page at a later time. This
2343 * is better than inconsistent global huge page
2344 * accounting of reserve counts.
2346 ClearPagePrivate(page);
2348 rc = vma_add_reservation(h, vma, address);
2349 if (unlikely(rc < 0))
2351 * See above comment about rare out of
2354 ClearPagePrivate(page);
2356 vma_end_reservation(h, vma, address);
2360 struct page *alloc_huge_page(struct vm_area_struct *vma,
2361 unsigned long addr, int avoid_reserve)
2363 struct hugepage_subpool *spool = subpool_vma(vma);
2364 struct hstate *h = hstate_vma(vma);
2366 long map_chg, map_commit;
2369 struct hugetlb_cgroup *h_cg;
2370 bool deferred_reserve;
2372 idx = hstate_index(h);
2374 * Examine the region/reserve map to determine if the process
2375 * has a reservation for the page to be allocated. A return
2376 * code of zero indicates a reservation exists (no change).
2378 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2380 return ERR_PTR(-ENOMEM);
2383 * Processes that did not create the mapping will have no
2384 * reserves as indicated by the region/reserve map. Check
2385 * that the allocation will not exceed the subpool limit.
2386 * Allocations for MAP_NORESERVE mappings also need to be
2387 * checked against any subpool limit.
2389 if (map_chg || avoid_reserve) {
2390 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2392 vma_end_reservation(h, vma, addr);
2393 return ERR_PTR(-ENOSPC);
2397 * Even though there was no reservation in the region/reserve
2398 * map, there could be reservations associated with the
2399 * subpool that can be used. This would be indicated if the
2400 * return value of hugepage_subpool_get_pages() is zero.
2401 * However, if avoid_reserve is specified we still avoid even
2402 * the subpool reservations.
2408 /* If this allocation is not consuming a reservation, charge it now.
2410 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2411 if (deferred_reserve) {
2412 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2413 idx, pages_per_huge_page(h), &h_cg);
2415 goto out_subpool_put;
2418 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2420 goto out_uncharge_cgroup_reservation;
2422 spin_lock(&hugetlb_lock);
2424 * glb_chg is passed to indicate whether or not a page must be taken
2425 * from the global free pool (global change). gbl_chg == 0 indicates
2426 * a reservation exists for the allocation.
2428 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2430 spin_unlock(&hugetlb_lock);
2431 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2433 goto out_uncharge_cgroup;
2434 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2435 SetPagePrivate(page);
2436 h->resv_huge_pages--;
2438 spin_lock(&hugetlb_lock);
2439 list_move(&page->lru, &h->hugepage_activelist);
2442 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2443 /* If allocation is not consuming a reservation, also store the
2444 * hugetlb_cgroup pointer on the page.
2446 if (deferred_reserve) {
2447 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2451 spin_unlock(&hugetlb_lock);
2453 set_page_private(page, (unsigned long)spool);
2455 map_commit = vma_commit_reservation(h, vma, addr);
2456 if (unlikely(map_chg > map_commit)) {
2458 * The page was added to the reservation map between
2459 * vma_needs_reservation and vma_commit_reservation.
2460 * This indicates a race with hugetlb_reserve_pages.
2461 * Adjust for the subpool count incremented above AND
2462 * in hugetlb_reserve_pages for the same page. Also,
2463 * the reservation count added in hugetlb_reserve_pages
2464 * no longer applies.
2468 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2469 hugetlb_acct_memory(h, -rsv_adjust);
2473 out_uncharge_cgroup:
2474 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2475 out_uncharge_cgroup_reservation:
2476 if (deferred_reserve)
2477 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2480 if (map_chg || avoid_reserve)
2481 hugepage_subpool_put_pages(spool, 1);
2482 vma_end_reservation(h, vma, addr);
2483 return ERR_PTR(-ENOSPC);
2486 int alloc_bootmem_huge_page(struct hstate *h)
2487 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2488 int __alloc_bootmem_huge_page(struct hstate *h)
2490 struct huge_bootmem_page *m;
2493 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2496 addr = memblock_alloc_try_nid_raw(
2497 huge_page_size(h), huge_page_size(h),
2498 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2501 * Use the beginning of the huge page to store the
2502 * huge_bootmem_page struct (until gather_bootmem
2503 * puts them into the mem_map).
2512 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2513 /* Put them into a private list first because mem_map is not up yet */
2514 INIT_LIST_HEAD(&m->list);
2515 list_add(&m->list, &huge_boot_pages);
2520 static void __init prep_compound_huge_page(struct page *page,
2523 if (unlikely(order > (MAX_ORDER - 1)))
2524 prep_compound_gigantic_page(page, order);
2526 prep_compound_page(page, order);
2529 /* Put bootmem huge pages into the standard lists after mem_map is up */
2530 static void __init gather_bootmem_prealloc(void)
2532 struct huge_bootmem_page *m;
2534 list_for_each_entry(m, &huge_boot_pages, list) {
2535 struct page *page = virt_to_page(m);
2536 struct hstate *h = m->hstate;
2538 WARN_ON(page_count(page) != 1);
2539 prep_compound_huge_page(page, h->order);
2540 WARN_ON(PageReserved(page));
2541 prep_new_huge_page(h, page, page_to_nid(page));
2542 put_page(page); /* free it into the hugepage allocator */
2545 * If we had gigantic hugepages allocated at boot time, we need
2546 * to restore the 'stolen' pages to totalram_pages in order to
2547 * fix confusing memory reports from free(1) and another
2548 * side-effects, like CommitLimit going negative.
2550 if (hstate_is_gigantic(h))
2551 adjust_managed_page_count(page, 1 << h->order);
2556 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2559 nodemask_t *node_alloc_noretry;
2561 if (!hstate_is_gigantic(h)) {
2563 * Bit mask controlling how hard we retry per-node allocations.
2564 * Ignore errors as lower level routines can deal with
2565 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2566 * time, we are likely in bigger trouble.
2568 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2571 /* allocations done at boot time */
2572 node_alloc_noretry = NULL;
2575 /* bit mask controlling how hard we retry per-node allocations */
2576 if (node_alloc_noretry)
2577 nodes_clear(*node_alloc_noretry);
2579 for (i = 0; i < h->max_huge_pages; ++i) {
2580 if (hstate_is_gigantic(h)) {
2581 if (hugetlb_cma_size) {
2582 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2585 if (!alloc_bootmem_huge_page(h))
2587 } else if (!alloc_pool_huge_page(h,
2588 &node_states[N_MEMORY],
2589 node_alloc_noretry))
2593 if (i < h->max_huge_pages) {
2596 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2597 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2598 h->max_huge_pages, buf, i);
2599 h->max_huge_pages = i;
2602 kfree(node_alloc_noretry);
2605 static void __init hugetlb_init_hstates(void)
2609 for_each_hstate(h) {
2610 if (minimum_order > huge_page_order(h))
2611 minimum_order = huge_page_order(h);
2613 /* oversize hugepages were init'ed in early boot */
2614 if (!hstate_is_gigantic(h))
2615 hugetlb_hstate_alloc_pages(h);
2617 VM_BUG_ON(minimum_order == UINT_MAX);
2620 static void __init report_hugepages(void)
2624 for_each_hstate(h) {
2627 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2628 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2629 buf, h->free_huge_pages);
2633 #ifdef CONFIG_HIGHMEM
2634 static void try_to_free_low(struct hstate *h, unsigned long count,
2635 nodemask_t *nodes_allowed)
2639 if (hstate_is_gigantic(h))
2642 for_each_node_mask(i, *nodes_allowed) {
2643 struct page *page, *next;
2644 struct list_head *freel = &h->hugepage_freelists[i];
2645 list_for_each_entry_safe(page, next, freel, lru) {
2646 if (count >= h->nr_huge_pages)
2648 if (PageHighMem(page))
2650 list_del(&page->lru);
2651 update_and_free_page(h, page);
2652 h->free_huge_pages--;
2653 h->free_huge_pages_node[page_to_nid(page)]--;
2658 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2659 nodemask_t *nodes_allowed)
2665 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2666 * balanced by operating on them in a round-robin fashion.
2667 * Returns 1 if an adjustment was made.
2669 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2674 VM_BUG_ON(delta != -1 && delta != 1);
2677 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2678 if (h->surplus_huge_pages_node[node])
2682 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2683 if (h->surplus_huge_pages_node[node] <
2684 h->nr_huge_pages_node[node])
2691 h->surplus_huge_pages += delta;
2692 h->surplus_huge_pages_node[node] += delta;
2696 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2697 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2698 nodemask_t *nodes_allowed)
2700 unsigned long min_count, ret;
2701 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2704 * Bit mask controlling how hard we retry per-node allocations.
2705 * If we can not allocate the bit mask, do not attempt to allocate
2706 * the requested huge pages.
2708 if (node_alloc_noretry)
2709 nodes_clear(*node_alloc_noretry);
2713 spin_lock(&hugetlb_lock);
2716 * Check for a node specific request.
2717 * Changing node specific huge page count may require a corresponding
2718 * change to the global count. In any case, the passed node mask
2719 * (nodes_allowed) will restrict alloc/free to the specified node.
2721 if (nid != NUMA_NO_NODE) {
2722 unsigned long old_count = count;
2724 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2726 * User may have specified a large count value which caused the
2727 * above calculation to overflow. In this case, they wanted
2728 * to allocate as many huge pages as possible. Set count to
2729 * largest possible value to align with their intention.
2731 if (count < old_count)
2736 * Gigantic pages runtime allocation depend on the capability for large
2737 * page range allocation.
2738 * If the system does not provide this feature, return an error when
2739 * the user tries to allocate gigantic pages but let the user free the
2740 * boottime allocated gigantic pages.
2742 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2743 if (count > persistent_huge_pages(h)) {
2744 spin_unlock(&hugetlb_lock);
2745 NODEMASK_FREE(node_alloc_noretry);
2748 /* Fall through to decrease pool */
2752 * Increase the pool size
2753 * First take pages out of surplus state. Then make up the
2754 * remaining difference by allocating fresh huge pages.
2756 * We might race with alloc_surplus_huge_page() here and be unable
2757 * to convert a surplus huge page to a normal huge page. That is
2758 * not critical, though, it just means the overall size of the
2759 * pool might be one hugepage larger than it needs to be, but
2760 * within all the constraints specified by the sysctls.
2762 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2763 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2767 while (count > persistent_huge_pages(h)) {
2769 * If this allocation races such that we no longer need the
2770 * page, free_huge_page will handle it by freeing the page
2771 * and reducing the surplus.
2773 spin_unlock(&hugetlb_lock);
2775 /* yield cpu to avoid soft lockup */
2778 ret = alloc_pool_huge_page(h, nodes_allowed,
2779 node_alloc_noretry);
2780 spin_lock(&hugetlb_lock);
2784 /* Bail for signals. Probably ctrl-c from user */
2785 if (signal_pending(current))
2790 * Decrease the pool size
2791 * First return free pages to the buddy allocator (being careful
2792 * to keep enough around to satisfy reservations). Then place
2793 * pages into surplus state as needed so the pool will shrink
2794 * to the desired size as pages become free.
2796 * By placing pages into the surplus state independent of the
2797 * overcommit value, we are allowing the surplus pool size to
2798 * exceed overcommit. There are few sane options here. Since
2799 * alloc_surplus_huge_page() is checking the global counter,
2800 * though, we'll note that we're not allowed to exceed surplus
2801 * and won't grow the pool anywhere else. Not until one of the
2802 * sysctls are changed, or the surplus pages go out of use.
2804 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2805 min_count = max(count, min_count);
2806 try_to_free_low(h, min_count, nodes_allowed);
2807 while (min_count < persistent_huge_pages(h)) {
2808 if (!free_pool_huge_page(h, nodes_allowed, 0))
2810 cond_resched_lock(&hugetlb_lock);
2812 while (count < persistent_huge_pages(h)) {
2813 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2817 h->max_huge_pages = persistent_huge_pages(h);
2818 spin_unlock(&hugetlb_lock);
2820 NODEMASK_FREE(node_alloc_noretry);
2825 #define HSTATE_ATTR_RO(_name) \
2826 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2828 #define HSTATE_ATTR(_name) \
2829 static struct kobj_attribute _name##_attr = \
2830 __ATTR(_name, 0644, _name##_show, _name##_store)
2832 static struct kobject *hugepages_kobj;
2833 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2835 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2837 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2841 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2842 if (hstate_kobjs[i] == kobj) {
2844 *nidp = NUMA_NO_NODE;
2848 return kobj_to_node_hstate(kobj, nidp);
2851 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2852 struct kobj_attribute *attr, char *buf)
2855 unsigned long nr_huge_pages;
2858 h = kobj_to_hstate(kobj, &nid);
2859 if (nid == NUMA_NO_NODE)
2860 nr_huge_pages = h->nr_huge_pages;
2862 nr_huge_pages = h->nr_huge_pages_node[nid];
2864 return sprintf(buf, "%lu\n", nr_huge_pages);
2867 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2868 struct hstate *h, int nid,
2869 unsigned long count, size_t len)
2872 nodemask_t nodes_allowed, *n_mask;
2874 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2877 if (nid == NUMA_NO_NODE) {
2879 * global hstate attribute
2881 if (!(obey_mempolicy &&
2882 init_nodemask_of_mempolicy(&nodes_allowed)))
2883 n_mask = &node_states[N_MEMORY];
2885 n_mask = &nodes_allowed;
2888 * Node specific request. count adjustment happens in
2889 * set_max_huge_pages() after acquiring hugetlb_lock.
2891 init_nodemask_of_node(&nodes_allowed, nid);
2892 n_mask = &nodes_allowed;
2895 err = set_max_huge_pages(h, count, nid, n_mask);
2897 return err ? err : len;
2900 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2901 struct kobject *kobj, const char *buf,
2905 unsigned long count;
2909 err = kstrtoul(buf, 10, &count);
2913 h = kobj_to_hstate(kobj, &nid);
2914 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2917 static ssize_t nr_hugepages_show(struct kobject *kobj,
2918 struct kobj_attribute *attr, char *buf)
2920 return nr_hugepages_show_common(kobj, attr, buf);
2923 static ssize_t nr_hugepages_store(struct kobject *kobj,
2924 struct kobj_attribute *attr, const char *buf, size_t len)
2926 return nr_hugepages_store_common(false, kobj, buf, len);
2928 HSTATE_ATTR(nr_hugepages);
2933 * hstate attribute for optionally mempolicy-based constraint on persistent
2934 * huge page alloc/free.
2936 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2937 struct kobj_attribute *attr, char *buf)
2939 return nr_hugepages_show_common(kobj, attr, buf);
2942 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2943 struct kobj_attribute *attr, const char *buf, size_t len)
2945 return nr_hugepages_store_common(true, kobj, buf, len);
2947 HSTATE_ATTR(nr_hugepages_mempolicy);
2951 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2952 struct kobj_attribute *attr, char *buf)
2954 struct hstate *h = kobj_to_hstate(kobj, NULL);
2955 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2958 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2959 struct kobj_attribute *attr, const char *buf, size_t count)
2962 unsigned long input;
2963 struct hstate *h = kobj_to_hstate(kobj, NULL);
2965 if (hstate_is_gigantic(h))
2968 err = kstrtoul(buf, 10, &input);
2972 spin_lock(&hugetlb_lock);
2973 h->nr_overcommit_huge_pages = input;
2974 spin_unlock(&hugetlb_lock);
2978 HSTATE_ATTR(nr_overcommit_hugepages);
2980 static ssize_t free_hugepages_show(struct kobject *kobj,
2981 struct kobj_attribute *attr, char *buf)
2984 unsigned long free_huge_pages;
2987 h = kobj_to_hstate(kobj, &nid);
2988 if (nid == NUMA_NO_NODE)
2989 free_huge_pages = h->free_huge_pages;
2991 free_huge_pages = h->free_huge_pages_node[nid];
2993 return sprintf(buf, "%lu\n", free_huge_pages);
2995 HSTATE_ATTR_RO(free_hugepages);
2997 static ssize_t resv_hugepages_show(struct kobject *kobj,
2998 struct kobj_attribute *attr, char *buf)
3000 struct hstate *h = kobj_to_hstate(kobj, NULL);
3001 return sprintf(buf, "%lu\n", h->resv_huge_pages);
3003 HSTATE_ATTR_RO(resv_hugepages);
3005 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3006 struct kobj_attribute *attr, char *buf)
3009 unsigned long surplus_huge_pages;
3012 h = kobj_to_hstate(kobj, &nid);
3013 if (nid == NUMA_NO_NODE)
3014 surplus_huge_pages = h->surplus_huge_pages;
3016 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3018 return sprintf(buf, "%lu\n", surplus_huge_pages);
3020 HSTATE_ATTR_RO(surplus_hugepages);
3022 static struct attribute *hstate_attrs[] = {
3023 &nr_hugepages_attr.attr,
3024 &nr_overcommit_hugepages_attr.attr,
3025 &free_hugepages_attr.attr,
3026 &resv_hugepages_attr.attr,
3027 &surplus_hugepages_attr.attr,
3029 &nr_hugepages_mempolicy_attr.attr,
3034 static const struct attribute_group hstate_attr_group = {
3035 .attrs = hstate_attrs,
3038 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3039 struct kobject **hstate_kobjs,
3040 const struct attribute_group *hstate_attr_group)
3043 int hi = hstate_index(h);
3045 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3046 if (!hstate_kobjs[hi])
3049 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3051 kobject_put(hstate_kobjs[hi]);
3056 static void __init hugetlb_sysfs_init(void)
3061 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3062 if (!hugepages_kobj)
3065 for_each_hstate(h) {
3066 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3067 hstate_kobjs, &hstate_attr_group);
3069 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3076 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3077 * with node devices in node_devices[] using a parallel array. The array
3078 * index of a node device or _hstate == node id.
3079 * This is here to avoid any static dependency of the node device driver, in
3080 * the base kernel, on the hugetlb module.
3082 struct node_hstate {
3083 struct kobject *hugepages_kobj;
3084 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3086 static struct node_hstate node_hstates[MAX_NUMNODES];
3089 * A subset of global hstate attributes for node devices
3091 static struct attribute *per_node_hstate_attrs[] = {
3092 &nr_hugepages_attr.attr,
3093 &free_hugepages_attr.attr,
3094 &surplus_hugepages_attr.attr,
3098 static const struct attribute_group per_node_hstate_attr_group = {
3099 .attrs = per_node_hstate_attrs,
3103 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3104 * Returns node id via non-NULL nidp.
3106 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3110 for (nid = 0; nid < nr_node_ids; nid++) {
3111 struct node_hstate *nhs = &node_hstates[nid];
3113 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3114 if (nhs->hstate_kobjs[i] == kobj) {
3126 * Unregister hstate attributes from a single node device.
3127 * No-op if no hstate attributes attached.
3129 static void hugetlb_unregister_node(struct node *node)
3132 struct node_hstate *nhs = &node_hstates[node->dev.id];
3134 if (!nhs->hugepages_kobj)
3135 return; /* no hstate attributes */
3137 for_each_hstate(h) {
3138 int idx = hstate_index(h);
3139 if (nhs->hstate_kobjs[idx]) {
3140 kobject_put(nhs->hstate_kobjs[idx]);
3141 nhs->hstate_kobjs[idx] = NULL;
3145 kobject_put(nhs->hugepages_kobj);
3146 nhs->hugepages_kobj = NULL;
3151 * Register hstate attributes for a single node device.
3152 * No-op if attributes already registered.
3154 static void hugetlb_register_node(struct node *node)
3157 struct node_hstate *nhs = &node_hstates[node->dev.id];
3160 if (nhs->hugepages_kobj)
3161 return; /* already allocated */
3163 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3165 if (!nhs->hugepages_kobj)
3168 for_each_hstate(h) {
3169 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3171 &per_node_hstate_attr_group);
3173 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3174 h->name, node->dev.id);
3175 hugetlb_unregister_node(node);
3182 * hugetlb init time: register hstate attributes for all registered node
3183 * devices of nodes that have memory. All on-line nodes should have
3184 * registered their associated device by this time.
3186 static void __init hugetlb_register_all_nodes(void)
3190 for_each_node_state(nid, N_MEMORY) {
3191 struct node *node = node_devices[nid];
3192 if (node->dev.id == nid)
3193 hugetlb_register_node(node);
3197 * Let the node device driver know we're here so it can
3198 * [un]register hstate attributes on node hotplug.
3200 register_hugetlbfs_with_node(hugetlb_register_node,
3201 hugetlb_unregister_node);
3203 #else /* !CONFIG_NUMA */
3205 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3213 static void hugetlb_register_all_nodes(void) { }
3217 static int __init hugetlb_init(void)
3221 if (!hugepages_supported()) {
3222 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3223 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3228 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3229 * architectures depend on setup being done here.
3231 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3232 if (!parsed_default_hugepagesz) {
3234 * If we did not parse a default huge page size, set
3235 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3236 * number of huge pages for this default size was implicitly
3237 * specified, set that here as well.
3238 * Note that the implicit setting will overwrite an explicit
3239 * setting. A warning will be printed in this case.
3241 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3242 if (default_hstate_max_huge_pages) {
3243 if (default_hstate.max_huge_pages) {
3246 string_get_size(huge_page_size(&default_hstate),
3247 1, STRING_UNITS_2, buf, 32);
3248 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3249 default_hstate.max_huge_pages, buf);
3250 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3251 default_hstate_max_huge_pages);
3253 default_hstate.max_huge_pages =
3254 default_hstate_max_huge_pages;
3258 hugetlb_cma_check();
3259 hugetlb_init_hstates();
3260 gather_bootmem_prealloc();
3263 hugetlb_sysfs_init();
3264 hugetlb_register_all_nodes();
3265 hugetlb_cgroup_file_init();
3268 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3270 num_fault_mutexes = 1;
3272 hugetlb_fault_mutex_table =
3273 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3275 BUG_ON(!hugetlb_fault_mutex_table);
3277 for (i = 0; i < num_fault_mutexes; i++)
3278 mutex_init(&hugetlb_fault_mutex_table[i]);
3281 subsys_initcall(hugetlb_init);
3283 /* Overwritten by architectures with more huge page sizes */
3284 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3286 return size == HPAGE_SIZE;
3289 void __init hugetlb_add_hstate(unsigned int order)
3294 if (size_to_hstate(PAGE_SIZE << order)) {
3297 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3299 h = &hstates[hugetlb_max_hstate++];
3301 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3302 h->nr_huge_pages = 0;
3303 h->free_huge_pages = 0;
3304 for (i = 0; i < MAX_NUMNODES; ++i)
3305 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3306 INIT_LIST_HEAD(&h->hugepage_activelist);
3307 h->next_nid_to_alloc = first_memory_node;
3308 h->next_nid_to_free = first_memory_node;
3309 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3310 huge_page_size(h)/1024);
3316 * hugepages command line processing
3317 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3318 * specification. If not, ignore the hugepages value. hugepages can also
3319 * be the first huge page command line option in which case it implicitly
3320 * specifies the number of huge pages for the default size.
3322 static int __init hugepages_setup(char *s)
3325 static unsigned long *last_mhp;
3327 if (!parsed_valid_hugepagesz) {
3328 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3329 parsed_valid_hugepagesz = true;
3334 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3335 * yet, so this hugepages= parameter goes to the "default hstate".
3336 * Otherwise, it goes with the previously parsed hugepagesz or
3337 * default_hugepagesz.
3339 else if (!hugetlb_max_hstate)
3340 mhp = &default_hstate_max_huge_pages;
3342 mhp = &parsed_hstate->max_huge_pages;
3344 if (mhp == last_mhp) {
3345 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3349 if (sscanf(s, "%lu", mhp) <= 0)
3353 * Global state is always initialized later in hugetlb_init.
3354 * But we need to allocate >= MAX_ORDER hstates here early to still
3355 * use the bootmem allocator.
3357 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3358 hugetlb_hstate_alloc_pages(parsed_hstate);
3364 __setup("hugepages=", hugepages_setup);
3367 * hugepagesz command line processing
3368 * A specific huge page size can only be specified once with hugepagesz.
3369 * hugepagesz is followed by hugepages on the command line. The global
3370 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3371 * hugepagesz argument was valid.
3373 static int __init hugepagesz_setup(char *s)
3378 parsed_valid_hugepagesz = false;
3379 size = (unsigned long)memparse(s, NULL);
3381 if (!arch_hugetlb_valid_size(size)) {
3382 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3386 h = size_to_hstate(size);
3389 * hstate for this size already exists. This is normally
3390 * an error, but is allowed if the existing hstate is the
3391 * default hstate. More specifically, it is only allowed if
3392 * the number of huge pages for the default hstate was not
3393 * previously specified.
3395 if (!parsed_default_hugepagesz || h != &default_hstate ||
3396 default_hstate.max_huge_pages) {
3397 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3402 * No need to call hugetlb_add_hstate() as hstate already
3403 * exists. But, do set parsed_hstate so that a following
3404 * hugepages= parameter will be applied to this hstate.
3407 parsed_valid_hugepagesz = true;
3411 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3412 parsed_valid_hugepagesz = true;
3415 __setup("hugepagesz=", hugepagesz_setup);
3418 * default_hugepagesz command line input
3419 * Only one instance of default_hugepagesz allowed on command line.
3421 static int __init default_hugepagesz_setup(char *s)
3425 parsed_valid_hugepagesz = false;
3426 if (parsed_default_hugepagesz) {
3427 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3431 size = (unsigned long)memparse(s, NULL);
3433 if (!arch_hugetlb_valid_size(size)) {
3434 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3438 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3439 parsed_valid_hugepagesz = true;
3440 parsed_default_hugepagesz = true;
3441 default_hstate_idx = hstate_index(size_to_hstate(size));
3444 * The number of default huge pages (for this size) could have been
3445 * specified as the first hugetlb parameter: hugepages=X. If so,
3446 * then default_hstate_max_huge_pages is set. If the default huge
3447 * page size is gigantic (>= MAX_ORDER), then the pages must be
3448 * allocated here from bootmem allocator.
3450 if (default_hstate_max_huge_pages) {
3451 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3452 if (hstate_is_gigantic(&default_hstate))
3453 hugetlb_hstate_alloc_pages(&default_hstate);
3454 default_hstate_max_huge_pages = 0;
3459 __setup("default_hugepagesz=", default_hugepagesz_setup);
3461 static unsigned int allowed_mems_nr(struct hstate *h)
3464 unsigned int nr = 0;
3465 nodemask_t *mpol_allowed;
3466 unsigned int *array = h->free_huge_pages_node;
3467 gfp_t gfp_mask = htlb_alloc_mask(h);
3469 mpol_allowed = policy_nodemask_current(gfp_mask);
3471 for_each_node_mask(node, cpuset_current_mems_allowed) {
3472 if (!mpol_allowed ||
3473 (mpol_allowed && node_isset(node, *mpol_allowed)))
3480 #ifdef CONFIG_SYSCTL
3481 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3482 struct ctl_table *table, int write,
3483 void *buffer, size_t *length, loff_t *ppos)
3485 struct hstate *h = &default_hstate;
3486 unsigned long tmp = h->max_huge_pages;
3489 if (!hugepages_supported())
3493 table->maxlen = sizeof(unsigned long);
3494 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3499 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3500 NUMA_NO_NODE, tmp, *length);
3505 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3506 void *buffer, size_t *length, loff_t *ppos)
3509 return hugetlb_sysctl_handler_common(false, table, write,
3510 buffer, length, ppos);
3514 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3515 void *buffer, size_t *length, loff_t *ppos)
3517 return hugetlb_sysctl_handler_common(true, table, write,
3518 buffer, length, ppos);
3520 #endif /* CONFIG_NUMA */
3522 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3523 void *buffer, size_t *length, loff_t *ppos)
3525 struct hstate *h = &default_hstate;
3529 if (!hugepages_supported())
3532 tmp = h->nr_overcommit_huge_pages;
3534 if (write && hstate_is_gigantic(h))
3538 table->maxlen = sizeof(unsigned long);
3539 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3544 spin_lock(&hugetlb_lock);
3545 h->nr_overcommit_huge_pages = tmp;
3546 spin_unlock(&hugetlb_lock);
3552 #endif /* CONFIG_SYSCTL */
3554 void hugetlb_report_meminfo(struct seq_file *m)
3557 unsigned long total = 0;
3559 if (!hugepages_supported())
3562 for_each_hstate(h) {
3563 unsigned long count = h->nr_huge_pages;
3565 total += (PAGE_SIZE << huge_page_order(h)) * count;
3567 if (h == &default_hstate)
3569 "HugePages_Total: %5lu\n"
3570 "HugePages_Free: %5lu\n"
3571 "HugePages_Rsvd: %5lu\n"
3572 "HugePages_Surp: %5lu\n"
3573 "Hugepagesize: %8lu kB\n",
3577 h->surplus_huge_pages,
3578 (PAGE_SIZE << huge_page_order(h)) / 1024);
3581 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3584 int hugetlb_report_node_meminfo(int nid, char *buf)
3586 struct hstate *h = &default_hstate;
3587 if (!hugepages_supported())
3590 "Node %d HugePages_Total: %5u\n"
3591 "Node %d HugePages_Free: %5u\n"
3592 "Node %d HugePages_Surp: %5u\n",
3593 nid, h->nr_huge_pages_node[nid],
3594 nid, h->free_huge_pages_node[nid],
3595 nid, h->surplus_huge_pages_node[nid]);
3598 void hugetlb_show_meminfo(void)
3603 if (!hugepages_supported())
3606 for_each_node_state(nid, N_MEMORY)
3608 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3610 h->nr_huge_pages_node[nid],
3611 h->free_huge_pages_node[nid],
3612 h->surplus_huge_pages_node[nid],
3613 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3616 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3618 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3619 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3622 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3623 unsigned long hugetlb_total_pages(void)
3626 unsigned long nr_total_pages = 0;
3629 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3630 return nr_total_pages;
3633 static int hugetlb_acct_memory(struct hstate *h, long delta)
3637 spin_lock(&hugetlb_lock);
3639 * When cpuset is configured, it breaks the strict hugetlb page
3640 * reservation as the accounting is done on a global variable. Such
3641 * reservation is completely rubbish in the presence of cpuset because
3642 * the reservation is not checked against page availability for the
3643 * current cpuset. Application can still potentially OOM'ed by kernel
3644 * with lack of free htlb page in cpuset that the task is in.
3645 * Attempt to enforce strict accounting with cpuset is almost
3646 * impossible (or too ugly) because cpuset is too fluid that
3647 * task or memory node can be dynamically moved between cpusets.
3649 * The change of semantics for shared hugetlb mapping with cpuset is
3650 * undesirable. However, in order to preserve some of the semantics,
3651 * we fall back to check against current free page availability as
3652 * a best attempt and hopefully to minimize the impact of changing
3653 * semantics that cpuset has.
3655 * Apart from cpuset, we also have memory policy mechanism that
3656 * also determines from which node the kernel will allocate memory
3657 * in a NUMA system. So similar to cpuset, we also should consider
3658 * the memory policy of the current task. Similar to the description
3662 if (gather_surplus_pages(h, delta) < 0)
3665 if (delta > allowed_mems_nr(h)) {
3666 return_unused_surplus_pages(h, delta);
3673 return_unused_surplus_pages(h, (unsigned long) -delta);
3676 spin_unlock(&hugetlb_lock);
3680 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3682 struct resv_map *resv = vma_resv_map(vma);
3685 * This new VMA should share its siblings reservation map if present.
3686 * The VMA will only ever have a valid reservation map pointer where
3687 * it is being copied for another still existing VMA. As that VMA
3688 * has a reference to the reservation map it cannot disappear until
3689 * after this open call completes. It is therefore safe to take a
3690 * new reference here without additional locking.
3692 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3693 kref_get(&resv->refs);
3696 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3698 struct hstate *h = hstate_vma(vma);
3699 struct resv_map *resv = vma_resv_map(vma);
3700 struct hugepage_subpool *spool = subpool_vma(vma);
3701 unsigned long reserve, start, end;
3704 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3707 start = vma_hugecache_offset(h, vma, vma->vm_start);
3708 end = vma_hugecache_offset(h, vma, vma->vm_end);
3710 reserve = (end - start) - region_count(resv, start, end);
3711 hugetlb_cgroup_uncharge_counter(resv, start, end);
3714 * Decrement reserve counts. The global reserve count may be
3715 * adjusted if the subpool has a minimum size.
3717 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3718 hugetlb_acct_memory(h, -gbl_reserve);
3721 kref_put(&resv->refs, resv_map_release);
3724 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3726 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3731 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3733 struct hstate *hstate = hstate_vma(vma);
3735 return 1UL << huge_page_shift(hstate);
3739 * We cannot handle pagefaults against hugetlb pages at all. They cause
3740 * handle_mm_fault() to try to instantiate regular-sized pages in the
3741 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3744 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3751 * When a new function is introduced to vm_operations_struct and added
3752 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3753 * This is because under System V memory model, mappings created via
3754 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3755 * their original vm_ops are overwritten with shm_vm_ops.
3757 const struct vm_operations_struct hugetlb_vm_ops = {
3758 .fault = hugetlb_vm_op_fault,
3759 .open = hugetlb_vm_op_open,
3760 .close = hugetlb_vm_op_close,
3761 .split = hugetlb_vm_op_split,
3762 .pagesize = hugetlb_vm_op_pagesize,
3765 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3771 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3772 vma->vm_page_prot)));
3774 entry = huge_pte_wrprotect(mk_huge_pte(page,
3775 vma->vm_page_prot));
3777 entry = pte_mkyoung(entry);
3778 entry = pte_mkhuge(entry);
3779 entry = arch_make_huge_pte(entry, vma, page, writable);
3784 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3785 unsigned long address, pte_t *ptep)
3789 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3790 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3791 update_mmu_cache(vma, address, ptep);
3794 bool is_hugetlb_entry_migration(pte_t pte)
3798 if (huge_pte_none(pte) || pte_present(pte))
3800 swp = pte_to_swp_entry(pte);
3801 if (non_swap_entry(swp) && is_migration_entry(swp))
3807 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3811 if (huge_pte_none(pte) || pte_present(pte))
3813 swp = pte_to_swp_entry(pte);
3814 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3820 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3821 struct vm_area_struct *vma)
3823 pte_t *src_pte, *dst_pte, entry, dst_entry;
3824 struct page *ptepage;
3827 struct hstate *h = hstate_vma(vma);
3828 unsigned long sz = huge_page_size(h);
3829 struct address_space *mapping = vma->vm_file->f_mapping;
3830 struct mmu_notifier_range range;
3833 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3836 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3839 mmu_notifier_invalidate_range_start(&range);
3842 * For shared mappings i_mmap_rwsem must be held to call
3843 * huge_pte_alloc, otherwise the returned ptep could go
3844 * away if part of a shared pmd and another thread calls
3847 i_mmap_lock_read(mapping);
3850 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3851 spinlock_t *src_ptl, *dst_ptl;
3852 src_pte = huge_pte_offset(src, addr, sz);
3855 dst_pte = huge_pte_alloc(dst, addr, sz);
3862 * If the pagetables are shared don't copy or take references.
3863 * dst_pte == src_pte is the common case of src/dest sharing.
3865 * However, src could have 'unshared' and dst shares with
3866 * another vma. If dst_pte !none, this implies sharing.
3867 * Check here before taking page table lock, and once again
3868 * after taking the lock below.
3870 dst_entry = huge_ptep_get(dst_pte);
3871 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3874 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3875 src_ptl = huge_pte_lockptr(h, src, src_pte);
3876 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3877 entry = huge_ptep_get(src_pte);
3878 dst_entry = huge_ptep_get(dst_pte);
3879 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3881 * Skip if src entry none. Also, skip in the
3882 * unlikely case dst entry !none as this implies
3883 * sharing with another vma.
3886 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3887 is_hugetlb_entry_hwpoisoned(entry))) {
3888 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3890 if (is_write_migration_entry(swp_entry) && cow) {
3892 * COW mappings require pages in both
3893 * parent and child to be set to read.
3895 make_migration_entry_read(&swp_entry);
3896 entry = swp_entry_to_pte(swp_entry);
3897 set_huge_swap_pte_at(src, addr, src_pte,
3900 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3904 * No need to notify as we are downgrading page
3905 * table protection not changing it to point
3908 * See Documentation/vm/mmu_notifier.rst
3910 huge_ptep_set_wrprotect(src, addr, src_pte);
3912 entry = huge_ptep_get(src_pte);
3913 ptepage = pte_page(entry);
3915 page_dup_rmap(ptepage, true);
3916 set_huge_pte_at(dst, addr, dst_pte, entry);
3917 hugetlb_count_add(pages_per_huge_page(h), dst);
3919 spin_unlock(src_ptl);
3920 spin_unlock(dst_ptl);
3924 mmu_notifier_invalidate_range_end(&range);
3926 i_mmap_unlock_read(mapping);
3931 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3932 unsigned long start, unsigned long end,
3933 struct page *ref_page)
3935 struct mm_struct *mm = vma->vm_mm;
3936 unsigned long address;
3941 struct hstate *h = hstate_vma(vma);
3942 unsigned long sz = huge_page_size(h);
3943 struct mmu_notifier_range range;
3945 WARN_ON(!is_vm_hugetlb_page(vma));
3946 BUG_ON(start & ~huge_page_mask(h));
3947 BUG_ON(end & ~huge_page_mask(h));
3950 * This is a hugetlb vma, all the pte entries should point
3953 tlb_change_page_size(tlb, sz);
3954 tlb_start_vma(tlb, vma);
3957 * If sharing possible, alert mmu notifiers of worst case.
3959 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3961 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3962 mmu_notifier_invalidate_range_start(&range);
3964 for (; address < end; address += sz) {
3965 ptep = huge_pte_offset(mm, address, sz);
3969 ptl = huge_pte_lock(h, mm, ptep);
3970 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3973 * We just unmapped a page of PMDs by clearing a PUD.
3974 * The caller's TLB flush range should cover this area.
3979 pte = huge_ptep_get(ptep);
3980 if (huge_pte_none(pte)) {
3986 * Migrating hugepage or HWPoisoned hugepage is already
3987 * unmapped and its refcount is dropped, so just clear pte here.
3989 if (unlikely(!pte_present(pte))) {
3990 huge_pte_clear(mm, address, ptep, sz);
3995 page = pte_page(pte);
3997 * If a reference page is supplied, it is because a specific
3998 * page is being unmapped, not a range. Ensure the page we
3999 * are about to unmap is the actual page of interest.
4002 if (page != ref_page) {
4007 * Mark the VMA as having unmapped its page so that
4008 * future faults in this VMA will fail rather than
4009 * looking like data was lost
4011 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4014 pte = huge_ptep_get_and_clear(mm, address, ptep);
4015 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4016 if (huge_pte_dirty(pte))
4017 set_page_dirty(page);
4019 hugetlb_count_sub(pages_per_huge_page(h), mm);
4020 page_remove_rmap(page, true);
4023 tlb_remove_page_size(tlb, page, huge_page_size(h));
4025 * Bail out after unmapping reference page if supplied
4030 mmu_notifier_invalidate_range_end(&range);
4031 tlb_end_vma(tlb, vma);
4034 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4035 struct vm_area_struct *vma, unsigned long start,
4036 unsigned long end, struct page *ref_page)
4038 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4041 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4042 * test will fail on a vma being torn down, and not grab a page table
4043 * on its way out. We're lucky that the flag has such an appropriate
4044 * name, and can in fact be safely cleared here. We could clear it
4045 * before the __unmap_hugepage_range above, but all that's necessary
4046 * is to clear it before releasing the i_mmap_rwsem. This works
4047 * because in the context this is called, the VMA is about to be
4048 * destroyed and the i_mmap_rwsem is held.
4050 vma->vm_flags &= ~VM_MAYSHARE;
4053 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4054 unsigned long end, struct page *ref_page)
4056 struct mm_struct *mm;
4057 struct mmu_gather tlb;
4058 unsigned long tlb_start = start;
4059 unsigned long tlb_end = end;
4062 * If shared PMDs were possibly used within this vma range, adjust
4063 * start/end for worst case tlb flushing.
4064 * Note that we can not be sure if PMDs are shared until we try to
4065 * unmap pages. However, we want to make sure TLB flushing covers
4066 * the largest possible range.
4068 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4072 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4073 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4074 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4078 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4079 * mappping it owns the reserve page for. The intention is to unmap the page
4080 * from other VMAs and let the children be SIGKILLed if they are faulting the
4083 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4084 struct page *page, unsigned long address)
4086 struct hstate *h = hstate_vma(vma);
4087 struct vm_area_struct *iter_vma;
4088 struct address_space *mapping;
4092 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4093 * from page cache lookup which is in HPAGE_SIZE units.
4095 address = address & huge_page_mask(h);
4096 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4098 mapping = vma->vm_file->f_mapping;
4101 * Take the mapping lock for the duration of the table walk. As
4102 * this mapping should be shared between all the VMAs,
4103 * __unmap_hugepage_range() is called as the lock is already held
4105 i_mmap_lock_write(mapping);
4106 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4107 /* Do not unmap the current VMA */
4108 if (iter_vma == vma)
4112 * Shared VMAs have their own reserves and do not affect
4113 * MAP_PRIVATE accounting but it is possible that a shared
4114 * VMA is using the same page so check and skip such VMAs.
4116 if (iter_vma->vm_flags & VM_MAYSHARE)
4120 * Unmap the page from other VMAs without their own reserves.
4121 * They get marked to be SIGKILLed if they fault in these
4122 * areas. This is because a future no-page fault on this VMA
4123 * could insert a zeroed page instead of the data existing
4124 * from the time of fork. This would look like data corruption
4126 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4127 unmap_hugepage_range(iter_vma, address,
4128 address + huge_page_size(h), page);
4130 i_mmap_unlock_write(mapping);
4134 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4135 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4136 * cannot race with other handlers or page migration.
4137 * Keep the pte_same checks anyway to make transition from the mutex easier.
4139 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4140 unsigned long address, pte_t *ptep,
4141 struct page *pagecache_page, spinlock_t *ptl)
4144 struct hstate *h = hstate_vma(vma);
4145 struct page *old_page, *new_page;
4146 int outside_reserve = 0;
4148 unsigned long haddr = address & huge_page_mask(h);
4149 struct mmu_notifier_range range;
4151 pte = huge_ptep_get(ptep);
4152 old_page = pte_page(pte);
4155 /* If no-one else is actually using this page, avoid the copy
4156 * and just make the page writable */
4157 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4158 page_move_anon_rmap(old_page, vma);
4159 set_huge_ptep_writable(vma, haddr, ptep);
4164 * If the process that created a MAP_PRIVATE mapping is about to
4165 * perform a COW due to a shared page count, attempt to satisfy
4166 * the allocation without using the existing reserves. The pagecache
4167 * page is used to determine if the reserve at this address was
4168 * consumed or not. If reserves were used, a partial faulted mapping
4169 * at the time of fork() could consume its reserves on COW instead
4170 * of the full address range.
4172 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4173 old_page != pagecache_page)
4174 outside_reserve = 1;
4179 * Drop page table lock as buddy allocator may be called. It will
4180 * be acquired again before returning to the caller, as expected.
4183 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4185 if (IS_ERR(new_page)) {
4187 * If a process owning a MAP_PRIVATE mapping fails to COW,
4188 * it is due to references held by a child and an insufficient
4189 * huge page pool. To guarantee the original mappers
4190 * reliability, unmap the page from child processes. The child
4191 * may get SIGKILLed if it later faults.
4193 if (outside_reserve) {
4195 BUG_ON(huge_pte_none(pte));
4196 unmap_ref_private(mm, vma, old_page, haddr);
4197 BUG_ON(huge_pte_none(pte));
4199 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4201 pte_same(huge_ptep_get(ptep), pte)))
4202 goto retry_avoidcopy;
4204 * race occurs while re-acquiring page table
4205 * lock, and our job is done.
4210 ret = vmf_error(PTR_ERR(new_page));
4211 goto out_release_old;
4215 * When the original hugepage is shared one, it does not have
4216 * anon_vma prepared.
4218 if (unlikely(anon_vma_prepare(vma))) {
4220 goto out_release_all;
4223 copy_user_huge_page(new_page, old_page, address, vma,
4224 pages_per_huge_page(h));
4225 __SetPageUptodate(new_page);
4227 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4228 haddr + huge_page_size(h));
4229 mmu_notifier_invalidate_range_start(&range);
4232 * Retake the page table lock to check for racing updates
4233 * before the page tables are altered
4236 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4237 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4238 ClearPagePrivate(new_page);
4241 huge_ptep_clear_flush(vma, haddr, ptep);
4242 mmu_notifier_invalidate_range(mm, range.start, range.end);
4243 set_huge_pte_at(mm, haddr, ptep,
4244 make_huge_pte(vma, new_page, 1));
4245 page_remove_rmap(old_page, true);
4246 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4247 set_page_huge_active(new_page);
4248 /* Make the old page be freed below */
4249 new_page = old_page;
4252 mmu_notifier_invalidate_range_end(&range);
4254 restore_reserve_on_error(h, vma, haddr, new_page);
4259 spin_lock(ptl); /* Caller expects lock to be held */
4263 /* Return the pagecache page at a given address within a VMA */
4264 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4265 struct vm_area_struct *vma, unsigned long address)
4267 struct address_space *mapping;
4270 mapping = vma->vm_file->f_mapping;
4271 idx = vma_hugecache_offset(h, vma, address);
4273 return find_lock_page(mapping, idx);
4277 * Return whether there is a pagecache page to back given address within VMA.
4278 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4280 static bool hugetlbfs_pagecache_present(struct hstate *h,
4281 struct vm_area_struct *vma, unsigned long address)
4283 struct address_space *mapping;
4287 mapping = vma->vm_file->f_mapping;
4288 idx = vma_hugecache_offset(h, vma, address);
4290 page = find_get_page(mapping, idx);
4293 return page != NULL;
4296 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4299 struct inode *inode = mapping->host;
4300 struct hstate *h = hstate_inode(inode);
4301 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4305 ClearPagePrivate(page);
4308 * set page dirty so that it will not be removed from cache/file
4309 * by non-hugetlbfs specific code paths.
4311 set_page_dirty(page);
4313 spin_lock(&inode->i_lock);
4314 inode->i_blocks += blocks_per_huge_page(h);
4315 spin_unlock(&inode->i_lock);
4319 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4320 struct vm_area_struct *vma,
4321 struct address_space *mapping, pgoff_t idx,
4322 unsigned long address, pte_t *ptep, unsigned int flags)
4324 struct hstate *h = hstate_vma(vma);
4325 vm_fault_t ret = VM_FAULT_SIGBUS;
4331 unsigned long haddr = address & huge_page_mask(h);
4332 bool new_page = false;
4335 * Currently, we are forced to kill the process in the event the
4336 * original mapper has unmapped pages from the child due to a failed
4337 * COW. Warn that such a situation has occurred as it may not be obvious
4339 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4340 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4346 * We can not race with truncation due to holding i_mmap_rwsem.
4347 * i_size is modified when holding i_mmap_rwsem, so check here
4348 * once for faults beyond end of file.
4350 size = i_size_read(mapping->host) >> huge_page_shift(h);
4355 page = find_lock_page(mapping, idx);
4358 * Check for page in userfault range
4360 if (userfaultfd_missing(vma)) {
4362 struct vm_fault vmf = {
4367 * Hard to debug if it ends up being
4368 * used by a callee that assumes
4369 * something about the other
4370 * uninitialized fields... same as in
4376 * hugetlb_fault_mutex and i_mmap_rwsem must be
4377 * dropped before handling userfault. Reacquire
4378 * after handling fault to make calling code simpler.
4380 hash = hugetlb_fault_mutex_hash(mapping, idx);
4381 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4382 i_mmap_unlock_read(mapping);
4383 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4384 i_mmap_lock_read(mapping);
4385 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4389 page = alloc_huge_page(vma, haddr, 0);
4392 * Returning error will result in faulting task being
4393 * sent SIGBUS. The hugetlb fault mutex prevents two
4394 * tasks from racing to fault in the same page which
4395 * could result in false unable to allocate errors.
4396 * Page migration does not take the fault mutex, but
4397 * does a clear then write of pte's under page table
4398 * lock. Page fault code could race with migration,
4399 * notice the clear pte and try to allocate a page
4400 * here. Before returning error, get ptl and make
4401 * sure there really is no pte entry.
4403 ptl = huge_pte_lock(h, mm, ptep);
4404 if (!huge_pte_none(huge_ptep_get(ptep))) {
4410 ret = vmf_error(PTR_ERR(page));
4413 clear_huge_page(page, address, pages_per_huge_page(h));
4414 __SetPageUptodate(page);
4417 if (vma->vm_flags & VM_MAYSHARE) {
4418 int err = huge_add_to_page_cache(page, mapping, idx);
4427 if (unlikely(anon_vma_prepare(vma))) {
4429 goto backout_unlocked;
4435 * If memory error occurs between mmap() and fault, some process
4436 * don't have hwpoisoned swap entry for errored virtual address.
4437 * So we need to block hugepage fault by PG_hwpoison bit check.
4439 if (unlikely(PageHWPoison(page))) {
4440 ret = VM_FAULT_HWPOISON |
4441 VM_FAULT_SET_HINDEX(hstate_index(h));
4442 goto backout_unlocked;
4447 * If we are going to COW a private mapping later, we examine the
4448 * pending reservations for this page now. This will ensure that
4449 * any allocations necessary to record that reservation occur outside
4452 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4453 if (vma_needs_reservation(h, vma, haddr) < 0) {
4455 goto backout_unlocked;
4457 /* Just decrements count, does not deallocate */
4458 vma_end_reservation(h, vma, haddr);
4461 ptl = huge_pte_lock(h, mm, ptep);
4463 if (!huge_pte_none(huge_ptep_get(ptep)))
4467 ClearPagePrivate(page);
4468 hugepage_add_new_anon_rmap(page, vma, haddr);
4470 page_dup_rmap(page, true);
4471 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4472 && (vma->vm_flags & VM_SHARED)));
4473 set_huge_pte_at(mm, haddr, ptep, new_pte);
4475 hugetlb_count_add(pages_per_huge_page(h), mm);
4476 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4477 /* Optimization, do the COW without a second fault */
4478 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4484 * Only make newly allocated pages active. Existing pages found
4485 * in the pagecache could be !page_huge_active() if they have been
4486 * isolated for migration.
4489 set_page_huge_active(page);
4499 restore_reserve_on_error(h, vma, haddr, page);
4505 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4507 unsigned long key[2];
4510 key[0] = (unsigned long) mapping;
4513 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4515 return hash & (num_fault_mutexes - 1);
4519 * For uniprocesor systems we always use a single mutex, so just
4520 * return 0 and avoid the hashing overhead.
4522 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4528 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4529 unsigned long address, unsigned int flags)
4536 struct page *page = NULL;
4537 struct page *pagecache_page = NULL;
4538 struct hstate *h = hstate_vma(vma);
4539 struct address_space *mapping;
4540 int need_wait_lock = 0;
4541 unsigned long haddr = address & huge_page_mask(h);
4543 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4546 * Since we hold no locks, ptep could be stale. That is
4547 * OK as we are only making decisions based on content and
4548 * not actually modifying content here.
4550 entry = huge_ptep_get(ptep);
4551 if (unlikely(is_hugetlb_entry_migration(entry))) {
4552 migration_entry_wait_huge(vma, mm, ptep);
4554 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4555 return VM_FAULT_HWPOISON_LARGE |
4556 VM_FAULT_SET_HINDEX(hstate_index(h));
4560 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4561 * until finished with ptep. This serves two purposes:
4562 * 1) It prevents huge_pmd_unshare from being called elsewhere
4563 * and making the ptep no longer valid.
4564 * 2) It synchronizes us with i_size modifications during truncation.
4566 * ptep could have already be assigned via huge_pte_offset. That
4567 * is OK, as huge_pte_alloc will return the same value unless
4568 * something has changed.
4570 mapping = vma->vm_file->f_mapping;
4571 i_mmap_lock_read(mapping);
4572 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4574 i_mmap_unlock_read(mapping);
4575 return VM_FAULT_OOM;
4579 * Serialize hugepage allocation and instantiation, so that we don't
4580 * get spurious allocation failures if two CPUs race to instantiate
4581 * the same page in the page cache.
4583 idx = vma_hugecache_offset(h, vma, haddr);
4584 hash = hugetlb_fault_mutex_hash(mapping, idx);
4585 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4587 entry = huge_ptep_get(ptep);
4588 if (huge_pte_none(entry)) {
4589 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4596 * entry could be a migration/hwpoison entry at this point, so this
4597 * check prevents the kernel from going below assuming that we have
4598 * an active hugepage in pagecache. This goto expects the 2nd page
4599 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4600 * properly handle it.
4602 if (!pte_present(entry))
4606 * If we are going to COW the mapping later, we examine the pending
4607 * reservations for this page now. This will ensure that any
4608 * allocations necessary to record that reservation occur outside the
4609 * spinlock. For private mappings, we also lookup the pagecache
4610 * page now as it is used to determine if a reservation has been
4613 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4614 if (vma_needs_reservation(h, vma, haddr) < 0) {
4618 /* Just decrements count, does not deallocate */
4619 vma_end_reservation(h, vma, haddr);
4621 if (!(vma->vm_flags & VM_MAYSHARE))
4622 pagecache_page = hugetlbfs_pagecache_page(h,
4626 ptl = huge_pte_lock(h, mm, ptep);
4628 /* Check for a racing update before calling hugetlb_cow */
4629 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4633 * hugetlb_cow() requires page locks of pte_page(entry) and
4634 * pagecache_page, so here we need take the former one
4635 * when page != pagecache_page or !pagecache_page.
4637 page = pte_page(entry);
4638 if (page != pagecache_page)
4639 if (!trylock_page(page)) {
4646 if (flags & FAULT_FLAG_WRITE) {
4647 if (!huge_pte_write(entry)) {
4648 ret = hugetlb_cow(mm, vma, address, ptep,
4649 pagecache_page, ptl);
4652 entry = huge_pte_mkdirty(entry);
4654 entry = pte_mkyoung(entry);
4655 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4656 flags & FAULT_FLAG_WRITE))
4657 update_mmu_cache(vma, haddr, ptep);
4659 if (page != pagecache_page)
4665 if (pagecache_page) {
4666 unlock_page(pagecache_page);
4667 put_page(pagecache_page);
4670 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4671 i_mmap_unlock_read(mapping);
4673 * Generally it's safe to hold refcount during waiting page lock. But
4674 * here we just wait to defer the next page fault to avoid busy loop and
4675 * the page is not used after unlocked before returning from the current
4676 * page fault. So we are safe from accessing freed page, even if we wait
4677 * here without taking refcount.
4680 wait_on_page_locked(page);
4685 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4686 * modifications for huge pages.
4688 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4690 struct vm_area_struct *dst_vma,
4691 unsigned long dst_addr,
4692 unsigned long src_addr,
4693 struct page **pagep)
4695 struct address_space *mapping;
4698 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4699 struct hstate *h = hstate_vma(dst_vma);
4707 page = alloc_huge_page(dst_vma, dst_addr, 0);
4711 ret = copy_huge_page_from_user(page,
4712 (const void __user *) src_addr,
4713 pages_per_huge_page(h), false);
4715 /* fallback to copy_from_user outside mmap_lock */
4716 if (unlikely(ret)) {
4719 /* don't free the page */
4728 * The memory barrier inside __SetPageUptodate makes sure that
4729 * preceding stores to the page contents become visible before
4730 * the set_pte_at() write.
4732 __SetPageUptodate(page);
4734 mapping = dst_vma->vm_file->f_mapping;
4735 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4738 * If shared, add to page cache
4741 size = i_size_read(mapping->host) >> huge_page_shift(h);
4744 goto out_release_nounlock;
4747 * Serialization between remove_inode_hugepages() and
4748 * huge_add_to_page_cache() below happens through the
4749 * hugetlb_fault_mutex_table that here must be hold by
4752 ret = huge_add_to_page_cache(page, mapping, idx);
4754 goto out_release_nounlock;
4757 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4761 * Recheck the i_size after holding PT lock to make sure not
4762 * to leave any page mapped (as page_mapped()) beyond the end
4763 * of the i_size (remove_inode_hugepages() is strict about
4764 * enforcing that). If we bail out here, we'll also leave a
4765 * page in the radix tree in the vm_shared case beyond the end
4766 * of the i_size, but remove_inode_hugepages() will take care
4767 * of it as soon as we drop the hugetlb_fault_mutex_table.
4769 size = i_size_read(mapping->host) >> huge_page_shift(h);
4772 goto out_release_unlock;
4775 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4776 goto out_release_unlock;
4779 page_dup_rmap(page, true);
4781 ClearPagePrivate(page);
4782 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4785 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4786 if (dst_vma->vm_flags & VM_WRITE)
4787 _dst_pte = huge_pte_mkdirty(_dst_pte);
4788 _dst_pte = pte_mkyoung(_dst_pte);
4790 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4792 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4793 dst_vma->vm_flags & VM_WRITE);
4794 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4796 /* No need to invalidate - it was non-present before */
4797 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4800 set_page_huge_active(page);
4810 out_release_nounlock:
4815 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4816 struct page **pages, struct vm_area_struct **vmas,
4817 unsigned long *position, unsigned long *nr_pages,
4818 long i, unsigned int flags, int *locked)
4820 unsigned long pfn_offset;
4821 unsigned long vaddr = *position;
4822 unsigned long remainder = *nr_pages;
4823 struct hstate *h = hstate_vma(vma);
4826 while (vaddr < vma->vm_end && remainder) {
4828 spinlock_t *ptl = NULL;
4833 * If we have a pending SIGKILL, don't keep faulting pages and
4834 * potentially allocating memory.
4836 if (fatal_signal_pending(current)) {
4842 * Some archs (sparc64, sh*) have multiple pte_ts to
4843 * each hugepage. We have to make sure we get the
4844 * first, for the page indexing below to work.
4846 * Note that page table lock is not held when pte is null.
4848 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4851 ptl = huge_pte_lock(h, mm, pte);
4852 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4855 * When coredumping, it suits get_dump_page if we just return
4856 * an error where there's an empty slot with no huge pagecache
4857 * to back it. This way, we avoid allocating a hugepage, and
4858 * the sparse dumpfile avoids allocating disk blocks, but its
4859 * huge holes still show up with zeroes where they need to be.
4861 if (absent && (flags & FOLL_DUMP) &&
4862 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4870 * We need call hugetlb_fault for both hugepages under migration
4871 * (in which case hugetlb_fault waits for the migration,) and
4872 * hwpoisoned hugepages (in which case we need to prevent the
4873 * caller from accessing to them.) In order to do this, we use
4874 * here is_swap_pte instead of is_hugetlb_entry_migration and
4875 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4876 * both cases, and because we can't follow correct pages
4877 * directly from any kind of swap entries.
4879 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4880 ((flags & FOLL_WRITE) &&
4881 !huge_pte_write(huge_ptep_get(pte)))) {
4883 unsigned int fault_flags = 0;
4887 if (flags & FOLL_WRITE)
4888 fault_flags |= FAULT_FLAG_WRITE;
4890 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4891 FAULT_FLAG_KILLABLE;
4892 if (flags & FOLL_NOWAIT)
4893 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4894 FAULT_FLAG_RETRY_NOWAIT;
4895 if (flags & FOLL_TRIED) {
4897 * Note: FAULT_FLAG_ALLOW_RETRY and
4898 * FAULT_FLAG_TRIED can co-exist
4900 fault_flags |= FAULT_FLAG_TRIED;
4902 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4903 if (ret & VM_FAULT_ERROR) {
4904 err = vm_fault_to_errno(ret, flags);
4908 if (ret & VM_FAULT_RETRY) {
4910 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4914 * VM_FAULT_RETRY must not return an
4915 * error, it will return zero
4918 * No need to update "position" as the
4919 * caller will not check it after
4920 * *nr_pages is set to 0.
4927 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4928 page = pte_page(huge_ptep_get(pte));
4931 * If subpage information not requested, update counters
4932 * and skip the same_page loop below.
4934 if (!pages && !vmas && !pfn_offset &&
4935 (vaddr + huge_page_size(h) < vma->vm_end) &&
4936 (remainder >= pages_per_huge_page(h))) {
4937 vaddr += huge_page_size(h);
4938 remainder -= pages_per_huge_page(h);
4939 i += pages_per_huge_page(h);
4946 pages[i] = mem_map_offset(page, pfn_offset);
4948 * try_grab_page() should always succeed here, because:
4949 * a) we hold the ptl lock, and b) we've just checked
4950 * that the huge page is present in the page tables. If
4951 * the huge page is present, then the tail pages must
4952 * also be present. The ptl prevents the head page and
4953 * tail pages from being rearranged in any way. So this
4954 * page must be available at this point, unless the page
4955 * refcount overflowed:
4957 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4972 if (vaddr < vma->vm_end && remainder &&
4973 pfn_offset < pages_per_huge_page(h)) {
4975 * We use pfn_offset to avoid touching the pageframes
4976 * of this compound page.
4982 *nr_pages = remainder;
4984 * setting position is actually required only if remainder is
4985 * not zero but it's faster not to add a "if (remainder)"
4993 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4995 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4998 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5001 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5002 unsigned long address, unsigned long end, pgprot_t newprot)
5004 struct mm_struct *mm = vma->vm_mm;
5005 unsigned long start = address;
5008 struct hstate *h = hstate_vma(vma);
5009 unsigned long pages = 0;
5010 bool shared_pmd = false;
5011 struct mmu_notifier_range range;
5014 * In the case of shared PMDs, the area to flush could be beyond
5015 * start/end. Set range.start/range.end to cover the maximum possible
5016 * range if PMD sharing is possible.
5018 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5019 0, vma, mm, start, end);
5020 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5022 BUG_ON(address >= end);
5023 flush_cache_range(vma, range.start, range.end);
5025 mmu_notifier_invalidate_range_start(&range);
5026 i_mmap_lock_write(vma->vm_file->f_mapping);
5027 for (; address < end; address += huge_page_size(h)) {
5029 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5032 ptl = huge_pte_lock(h, mm, ptep);
5033 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5039 pte = huge_ptep_get(ptep);
5040 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5044 if (unlikely(is_hugetlb_entry_migration(pte))) {
5045 swp_entry_t entry = pte_to_swp_entry(pte);
5047 if (is_write_migration_entry(entry)) {
5050 make_migration_entry_read(&entry);
5051 newpte = swp_entry_to_pte(entry);
5052 set_huge_swap_pte_at(mm, address, ptep,
5053 newpte, huge_page_size(h));
5059 if (!huge_pte_none(pte)) {
5062 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5063 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5064 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5065 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5071 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5072 * may have cleared our pud entry and done put_page on the page table:
5073 * once we release i_mmap_rwsem, another task can do the final put_page
5074 * and that page table be reused and filled with junk. If we actually
5075 * did unshare a page of pmds, flush the range corresponding to the pud.
5078 flush_hugetlb_tlb_range(vma, range.start, range.end);
5080 flush_hugetlb_tlb_range(vma, start, end);
5082 * No need to call mmu_notifier_invalidate_range() we are downgrading
5083 * page table protection not changing it to point to a new page.
5085 * See Documentation/vm/mmu_notifier.rst
5087 i_mmap_unlock_write(vma->vm_file->f_mapping);
5088 mmu_notifier_invalidate_range_end(&range);
5090 return pages << h->order;
5093 int hugetlb_reserve_pages(struct inode *inode,
5095 struct vm_area_struct *vma,
5096 vm_flags_t vm_flags)
5098 long ret, chg, add = -1;
5099 struct hstate *h = hstate_inode(inode);
5100 struct hugepage_subpool *spool = subpool_inode(inode);
5101 struct resv_map *resv_map;
5102 struct hugetlb_cgroup *h_cg = NULL;
5103 long gbl_reserve, regions_needed = 0;
5105 /* This should never happen */
5107 VM_WARN(1, "%s called with a negative range\n", __func__);
5112 * Only apply hugepage reservation if asked. At fault time, an
5113 * attempt will be made for VM_NORESERVE to allocate a page
5114 * without using reserves
5116 if (vm_flags & VM_NORESERVE)
5120 * Shared mappings base their reservation on the number of pages that
5121 * are already allocated on behalf of the file. Private mappings need
5122 * to reserve the full area even if read-only as mprotect() may be
5123 * called to make the mapping read-write. Assume !vma is a shm mapping
5125 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5127 * resv_map can not be NULL as hugetlb_reserve_pages is only
5128 * called for inodes for which resv_maps were created (see
5129 * hugetlbfs_get_inode).
5131 resv_map = inode_resv_map(inode);
5133 chg = region_chg(resv_map, from, to, ®ions_needed);
5136 /* Private mapping. */
5137 resv_map = resv_map_alloc();
5143 set_vma_resv_map(vma, resv_map);
5144 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5152 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5153 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5160 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5161 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5164 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5168 * There must be enough pages in the subpool for the mapping. If
5169 * the subpool has a minimum size, there may be some global
5170 * reservations already in place (gbl_reserve).
5172 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5173 if (gbl_reserve < 0) {
5175 goto out_uncharge_cgroup;
5179 * Check enough hugepages are available for the reservation.
5180 * Hand the pages back to the subpool if there are not
5182 ret = hugetlb_acct_memory(h, gbl_reserve);
5188 * Account for the reservations made. Shared mappings record regions
5189 * that have reservations as they are shared by multiple VMAs.
5190 * When the last VMA disappears, the region map says how much
5191 * the reservation was and the page cache tells how much of
5192 * the reservation was consumed. Private mappings are per-VMA and
5193 * only the consumed reservations are tracked. When the VMA
5194 * disappears, the original reservation is the VMA size and the
5195 * consumed reservations are stored in the map. Hence, nothing
5196 * else has to be done for private mappings here
5198 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5199 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5201 if (unlikely(add < 0)) {
5202 hugetlb_acct_memory(h, -gbl_reserve);
5204 } else if (unlikely(chg > add)) {
5206 * pages in this range were added to the reserve
5207 * map between region_chg and region_add. This
5208 * indicates a race with alloc_huge_page. Adjust
5209 * the subpool and reserve counts modified above
5210 * based on the difference.
5214 hugetlb_cgroup_uncharge_cgroup_rsvd(
5216 (chg - add) * pages_per_huge_page(h), h_cg);
5218 rsv_adjust = hugepage_subpool_put_pages(spool,
5220 hugetlb_acct_memory(h, -rsv_adjust);
5225 /* put back original number of pages, chg */
5226 (void)hugepage_subpool_put_pages(spool, chg);
5227 out_uncharge_cgroup:
5228 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5229 chg * pages_per_huge_page(h), h_cg);
5231 if (!vma || vma->vm_flags & VM_MAYSHARE)
5232 /* Only call region_abort if the region_chg succeeded but the
5233 * region_add failed or didn't run.
5235 if (chg >= 0 && add < 0)
5236 region_abort(resv_map, from, to, regions_needed);
5237 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5238 kref_put(&resv_map->refs, resv_map_release);
5242 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5245 struct hstate *h = hstate_inode(inode);
5246 struct resv_map *resv_map = inode_resv_map(inode);
5248 struct hugepage_subpool *spool = subpool_inode(inode);
5252 * Since this routine can be called in the evict inode path for all
5253 * hugetlbfs inodes, resv_map could be NULL.
5256 chg = region_del(resv_map, start, end);
5258 * region_del() can fail in the rare case where a region
5259 * must be split and another region descriptor can not be
5260 * allocated. If end == LONG_MAX, it will not fail.
5266 spin_lock(&inode->i_lock);
5267 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5268 spin_unlock(&inode->i_lock);
5271 * If the subpool has a minimum size, the number of global
5272 * reservations to be released may be adjusted.
5274 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5275 hugetlb_acct_memory(h, -gbl_reserve);
5280 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5281 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5282 struct vm_area_struct *vma,
5283 unsigned long addr, pgoff_t idx)
5285 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5287 unsigned long sbase = saddr & PUD_MASK;
5288 unsigned long s_end = sbase + PUD_SIZE;
5290 /* Allow segments to share if only one is marked locked */
5291 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5292 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5295 * match the virtual addresses, permission and the alignment of the
5298 if (pmd_index(addr) != pmd_index(saddr) ||
5299 vm_flags != svm_flags ||
5300 sbase < svma->vm_start || svma->vm_end < s_end)
5306 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5308 unsigned long base = addr & PUD_MASK;
5309 unsigned long end = base + PUD_SIZE;
5312 * check on proper vm_flags and page table alignment
5314 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5320 * Determine if start,end range within vma could be mapped by shared pmd.
5321 * If yes, adjust start and end to cover range associated with possible
5322 * shared pmd mappings.
5324 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5325 unsigned long *start, unsigned long *end)
5327 unsigned long a_start, a_end;
5329 if (!(vma->vm_flags & VM_MAYSHARE))
5332 /* Extend the range to be PUD aligned for a worst case scenario */
5333 a_start = ALIGN_DOWN(*start, PUD_SIZE);
5334 a_end = ALIGN(*end, PUD_SIZE);
5337 * Intersect the range with the vma range, since pmd sharing won't be
5338 * across vma after all
5340 *start = max(vma->vm_start, a_start);
5341 *end = min(vma->vm_end, a_end);
5345 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5346 * and returns the corresponding pte. While this is not necessary for the
5347 * !shared pmd case because we can allocate the pmd later as well, it makes the
5348 * code much cleaner.
5350 * This routine must be called with i_mmap_rwsem held in at least read mode.
5351 * For hugetlbfs, this prevents removal of any page table entries associated
5352 * with the address space. This is important as we are setting up sharing
5353 * based on existing page table entries (mappings).
5355 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5357 struct vm_area_struct *vma = find_vma(mm, addr);
5358 struct address_space *mapping = vma->vm_file->f_mapping;
5359 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5361 struct vm_area_struct *svma;
5362 unsigned long saddr;
5367 if (!vma_shareable(vma, addr))
5368 return (pte_t *)pmd_alloc(mm, pud, addr);
5370 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5374 saddr = page_table_shareable(svma, vma, addr, idx);
5376 spte = huge_pte_offset(svma->vm_mm, saddr,
5377 vma_mmu_pagesize(svma));
5379 get_page(virt_to_page(spte));
5388 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5389 if (pud_none(*pud)) {
5390 pud_populate(mm, pud,
5391 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5394 put_page(virt_to_page(spte));
5398 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5403 * unmap huge page backed by shared pte.
5405 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5406 * indicated by page_count > 1, unmap is achieved by clearing pud and
5407 * decrementing the ref count. If count == 1, the pte page is not shared.
5409 * Called with page table lock held and i_mmap_rwsem held in write mode.
5411 * returns: 1 successfully unmapped a shared pte page
5412 * 0 the underlying pte page is not shared, or it is the last user
5414 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5415 unsigned long *addr, pte_t *ptep)
5417 pgd_t *pgd = pgd_offset(mm, *addr);
5418 p4d_t *p4d = p4d_offset(pgd, *addr);
5419 pud_t *pud = pud_offset(p4d, *addr);
5421 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5422 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5423 if (page_count(virt_to_page(ptep)) == 1)
5427 put_page(virt_to_page(ptep));
5429 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5432 #define want_pmd_share() (1)
5433 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5434 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5439 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5440 unsigned long *addr, pte_t *ptep)
5445 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5446 unsigned long *start, unsigned long *end)
5449 #define want_pmd_share() (0)
5450 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5452 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5453 pte_t *huge_pte_alloc(struct mm_struct *mm,
5454 unsigned long addr, unsigned long sz)
5461 pgd = pgd_offset(mm, addr);
5462 p4d = p4d_alloc(mm, pgd, addr);
5465 pud = pud_alloc(mm, p4d, addr);
5467 if (sz == PUD_SIZE) {
5470 BUG_ON(sz != PMD_SIZE);
5471 if (want_pmd_share() && pud_none(*pud))
5472 pte = huge_pmd_share(mm, addr, pud);
5474 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5477 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5483 * huge_pte_offset() - Walk the page table to resolve the hugepage
5484 * entry at address @addr
5486 * Return: Pointer to page table entry (PUD or PMD) for
5487 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5488 * size @sz doesn't match the hugepage size at this level of the page
5491 pte_t *huge_pte_offset(struct mm_struct *mm,
5492 unsigned long addr, unsigned long sz)
5499 pgd = pgd_offset(mm, addr);
5500 if (!pgd_present(*pgd))
5502 p4d = p4d_offset(pgd, addr);
5503 if (!p4d_present(*p4d))
5506 pud = pud_offset(p4d, addr);
5508 /* must be pud huge, non-present or none */
5509 return (pte_t *)pud;
5510 if (!pud_present(*pud))
5512 /* must have a valid entry and size to go further */
5514 pmd = pmd_offset(pud, addr);
5515 /* must be pmd huge, non-present or none */
5516 return (pte_t *)pmd;
5519 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5522 * These functions are overwritable if your architecture needs its own
5525 struct page * __weak
5526 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5529 return ERR_PTR(-EINVAL);
5532 struct page * __weak
5533 follow_huge_pd(struct vm_area_struct *vma,
5534 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5536 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5540 struct page * __weak
5541 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5542 pmd_t *pmd, int flags)
5544 struct page *page = NULL;
5548 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5549 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5550 (FOLL_PIN | FOLL_GET)))
5554 ptl = pmd_lockptr(mm, pmd);
5557 * make sure that the address range covered by this pmd is not
5558 * unmapped from other threads.
5560 if (!pmd_huge(*pmd))
5562 pte = huge_ptep_get((pte_t *)pmd);
5563 if (pte_present(pte)) {
5564 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5566 * try_grab_page() should always succeed here, because: a) we
5567 * hold the pmd (ptl) lock, and b) we've just checked that the
5568 * huge pmd (head) page is present in the page tables. The ptl
5569 * prevents the head page and tail pages from being rearranged
5570 * in any way. So this page must be available at this point,
5571 * unless the page refcount overflowed:
5573 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5578 if (is_hugetlb_entry_migration(pte)) {
5580 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5584 * hwpoisoned entry is treated as no_page_table in
5585 * follow_page_mask().
5593 struct page * __weak
5594 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5595 pud_t *pud, int flags)
5597 if (flags & (FOLL_GET | FOLL_PIN))
5600 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5603 struct page * __weak
5604 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5606 if (flags & (FOLL_GET | FOLL_PIN))
5609 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5612 bool isolate_huge_page(struct page *page, struct list_head *list)
5616 VM_BUG_ON_PAGE(!PageHead(page), page);
5617 spin_lock(&hugetlb_lock);
5618 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5622 clear_page_huge_active(page);
5623 list_move_tail(&page->lru, list);
5625 spin_unlock(&hugetlb_lock);
5629 void putback_active_hugepage(struct page *page)
5631 VM_BUG_ON_PAGE(!PageHead(page), page);
5632 spin_lock(&hugetlb_lock);
5633 set_page_huge_active(page);
5634 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5635 spin_unlock(&hugetlb_lock);
5639 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5641 struct hstate *h = page_hstate(oldpage);
5643 hugetlb_cgroup_migrate(oldpage, newpage);
5644 set_page_owner_migrate_reason(newpage, reason);
5647 * transfer temporary state of the new huge page. This is
5648 * reverse to other transitions because the newpage is going to
5649 * be final while the old one will be freed so it takes over
5650 * the temporary status.
5652 * Also note that we have to transfer the per-node surplus state
5653 * here as well otherwise the global surplus count will not match
5656 if (PageHugeTemporary(newpage)) {
5657 int old_nid = page_to_nid(oldpage);
5658 int new_nid = page_to_nid(newpage);
5660 SetPageHugeTemporary(oldpage);
5661 ClearPageHugeTemporary(newpage);
5663 spin_lock(&hugetlb_lock);
5664 if (h->surplus_huge_pages_node[old_nid]) {
5665 h->surplus_huge_pages_node[old_nid]--;
5666 h->surplus_huge_pages_node[new_nid]++;
5668 spin_unlock(&hugetlb_lock);
5673 static bool cma_reserve_called __initdata;
5675 static int __init cmdline_parse_hugetlb_cma(char *p)
5677 hugetlb_cma_size = memparse(p, &p);
5681 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5683 void __init hugetlb_cma_reserve(int order)
5685 unsigned long size, reserved, per_node;
5688 cma_reserve_called = true;
5690 if (!hugetlb_cma_size)
5693 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5694 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5695 (PAGE_SIZE << order) / SZ_1M);
5700 * If 3 GB area is requested on a machine with 4 numa nodes,
5701 * let's allocate 1 GB on first three nodes and ignore the last one.
5703 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5704 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5705 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5708 for_each_node_state(nid, N_ONLINE) {
5711 size = min(per_node, hugetlb_cma_size - reserved);
5712 size = round_up(size, PAGE_SIZE << order);
5714 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5715 0, false, "hugetlb",
5716 &hugetlb_cma[nid], nid);
5718 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5724 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5727 if (reserved >= hugetlb_cma_size)
5732 void __init hugetlb_cma_check(void)
5734 if (!hugetlb_cma_size || cma_reserve_called)
5737 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5740 #endif /* CONFIG_CMA */