1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
85 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
87 bool free = (spool->count == 0) && (spool->used_hpages == 0);
89 spin_unlock(&spool->lock);
91 /* If no pages are used, and no other handles to the subpool
92 * remain, give up any reservations based on minimum size and
95 if (spool->min_hpages != -1)
96 hugetlb_acct_memory(spool->hstate,
102 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
105 struct hugepage_subpool *spool;
107 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
111 spin_lock_init(&spool->lock);
113 spool->max_hpages = max_hpages;
115 spool->min_hpages = min_hpages;
117 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
121 spool->rsv_hpages = min_hpages;
126 void hugepage_put_subpool(struct hugepage_subpool *spool)
128 spin_lock(&spool->lock);
129 BUG_ON(!spool->count);
131 unlock_or_release_subpool(spool);
135 * Subpool accounting for allocating and reserving pages.
136 * Return -ENOMEM if there are not enough resources to satisfy the
137 * request. Otherwise, return the number of pages by which the
138 * global pools must be adjusted (upward). The returned value may
139 * only be different than the passed value (delta) in the case where
140 * a subpool minimum size must be maintained.
142 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
150 spin_lock(&spool->lock);
152 if (spool->max_hpages != -1) { /* maximum size accounting */
153 if ((spool->used_hpages + delta) <= spool->max_hpages)
154 spool->used_hpages += delta;
161 /* minimum size accounting */
162 if (spool->min_hpages != -1 && spool->rsv_hpages) {
163 if (delta > spool->rsv_hpages) {
165 * Asking for more reserves than those already taken on
166 * behalf of subpool. Return difference.
168 ret = delta - spool->rsv_hpages;
169 spool->rsv_hpages = 0;
171 ret = 0; /* reserves already accounted for */
172 spool->rsv_hpages -= delta;
177 spin_unlock(&spool->lock);
182 * Subpool accounting for freeing and unreserving pages.
183 * Return the number of global page reservations that must be dropped.
184 * The return value may only be different than the passed value (delta)
185 * in the case where a subpool minimum size must be maintained.
187 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
195 spin_lock(&spool->lock);
197 if (spool->max_hpages != -1) /* maximum size accounting */
198 spool->used_hpages -= delta;
200 /* minimum size accounting */
201 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
202 if (spool->rsv_hpages + delta <= spool->min_hpages)
205 ret = spool->rsv_hpages + delta - spool->min_hpages;
207 spool->rsv_hpages += delta;
208 if (spool->rsv_hpages > spool->min_hpages)
209 spool->rsv_hpages = spool->min_hpages;
213 * If hugetlbfs_put_super couldn't free spool due to an outstanding
214 * quota reference, free it now.
216 unlock_or_release_subpool(spool);
221 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
223 return HUGETLBFS_SB(inode->i_sb)->spool;
226 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
228 return subpool_inode(file_inode(vma->vm_file));
231 /* Helper that removes a struct file_region from the resv_map cache and returns
234 static struct file_region *
235 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
237 struct file_region *nrg = NULL;
239 VM_BUG_ON(resv->region_cache_count <= 0);
241 resv->region_cache_count--;
242 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)) {
314 nrg = list_next_entry(rg, link);
315 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
316 has_same_uncharge_info(nrg, rg)) {
317 nrg->from = rg->from;
325 * Must be called with resv->lock held.
327 * Calling this with regions_needed != NULL will count the number of pages
328 * to be added but will not modify the linked list. And regions_needed will
329 * indicate the number of file_regions needed in the cache to carry out to add
330 * the regions for this range.
332 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
333 struct hugetlb_cgroup *h_cg,
334 struct hstate *h, long *regions_needed)
337 struct list_head *head = &resv->regions;
338 long last_accounted_offset = f;
339 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
344 /* In this loop, we essentially handle an entry for the range
345 * [last_accounted_offset, rg->from), at every iteration, with some
348 list_for_each_entry_safe(rg, trg, head, link) {
349 /* Skip irrelevant regions that start before our range. */
351 /* If this region ends after the last accounted offset,
352 * then we need to update last_accounted_offset.
354 if (rg->to > last_accounted_offset)
355 last_accounted_offset = rg->to;
359 /* When we find a region that starts beyond our range, we've
365 /* Add an entry for last_accounted_offset -> rg->from, and
366 * update last_accounted_offset.
368 if (rg->from > last_accounted_offset) {
369 add += rg->from - last_accounted_offset;
370 if (!regions_needed) {
371 nrg = get_file_region_entry_from_cache(
372 resv, last_accounted_offset, rg->from);
373 record_hugetlb_cgroup_uncharge_info(h_cg, h,
375 list_add(&nrg->link, rg->link.prev);
376 coalesce_file_region(resv, nrg);
378 *regions_needed += 1;
381 last_accounted_offset = rg->to;
384 /* Handle the case where our range extends beyond
385 * last_accounted_offset.
387 if (last_accounted_offset < t) {
388 add += t - last_accounted_offset;
389 if (!regions_needed) {
390 nrg = get_file_region_entry_from_cache(
391 resv, last_accounted_offset, t);
392 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
393 list_add(&nrg->link, rg->link.prev);
394 coalesce_file_region(resv, nrg);
396 *regions_needed += 1;
403 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
405 static int allocate_file_region_entries(struct resv_map *resv,
407 __must_hold(&resv->lock)
409 struct list_head allocated_regions;
410 int to_allocate = 0, i = 0;
411 struct file_region *trg = NULL, *rg = NULL;
413 VM_BUG_ON(regions_needed < 0);
415 INIT_LIST_HEAD(&allocated_regions);
418 * Check for sufficient descriptors in the cache to accommodate
419 * the number of in progress add operations plus regions_needed.
421 * This is a while loop because when we drop the lock, some other call
422 * to region_add or region_del may have consumed some region_entries,
423 * so we keep looping here until we finally have enough entries for
424 * (adds_in_progress + regions_needed).
426 while (resv->region_cache_count <
427 (resv->adds_in_progress + regions_needed)) {
428 to_allocate = resv->adds_in_progress + regions_needed -
429 resv->region_cache_count;
431 /* At this point, we should have enough entries in the cache
432 * for all the existings adds_in_progress. We should only be
433 * needing to allocate for regions_needed.
435 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
437 spin_unlock(&resv->lock);
438 for (i = 0; i < to_allocate; i++) {
439 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
442 list_add(&trg->link, &allocated_regions);
445 spin_lock(&resv->lock);
447 list_splice(&allocated_regions, &resv->region_cache);
448 resv->region_cache_count += to_allocate;
454 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
462 * Add the huge page range represented by [f, t) to the reserve
463 * map. Regions will be taken from the cache to fill in this range.
464 * Sufficient regions should exist in the cache due to the previous
465 * call to region_chg with the same range, but in some cases the cache will not
466 * have sufficient entries due to races with other code doing region_add or
467 * region_del. The extra needed entries will be allocated.
469 * regions_needed is the out value provided by a previous call to region_chg.
471 * Return the number of new huge pages added to the map. This number is greater
472 * than or equal to zero. If file_region entries needed to be allocated for
473 * this operation and we were not able to allocate, it returns -ENOMEM.
474 * region_add of regions of length 1 never allocate file_regions and cannot
475 * fail; region_chg will always allocate at least 1 entry and a region_add for
476 * 1 page will only require at most 1 entry.
478 static long region_add(struct resv_map *resv, long f, long t,
479 long in_regions_needed, struct hstate *h,
480 struct hugetlb_cgroup *h_cg)
482 long add = 0, actual_regions_needed = 0;
484 spin_lock(&resv->lock);
487 /* Count how many regions are actually needed to execute this add. */
488 add_reservation_in_range(resv, f, t, NULL, NULL,
489 &actual_regions_needed);
492 * Check for sufficient descriptors in the cache to accommodate
493 * this add operation. Note that actual_regions_needed may be greater
494 * than in_regions_needed, as the resv_map may have been modified since
495 * the region_chg call. In this case, we need to make sure that we
496 * allocate extra entries, such that we have enough for all the
497 * existing adds_in_progress, plus the excess needed for this
500 if (actual_regions_needed > in_regions_needed &&
501 resv->region_cache_count <
502 resv->adds_in_progress +
503 (actual_regions_needed - in_regions_needed)) {
504 /* region_add operation of range 1 should never need to
505 * allocate file_region entries.
507 VM_BUG_ON(t - f <= 1);
509 if (allocate_file_region_entries(
510 resv, actual_regions_needed - in_regions_needed)) {
517 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
519 resv->adds_in_progress -= in_regions_needed;
521 spin_unlock(&resv->lock);
527 * Examine the existing reserve map and determine how many
528 * huge pages in the specified range [f, t) are NOT currently
529 * represented. This routine is called before a subsequent
530 * call to region_add that will actually modify the reserve
531 * map to add the specified range [f, t). region_chg does
532 * not change the number of huge pages represented by the
533 * map. A number of new file_region structures is added to the cache as a
534 * placeholder, for the subsequent region_add call to use. At least 1
535 * file_region structure is added.
537 * out_regions_needed is the number of regions added to the
538 * resv->adds_in_progress. This value needs to be provided to a follow up call
539 * to region_add or region_abort for proper accounting.
541 * Returns the number of huge pages that need to be added to the existing
542 * reservation map for the range [f, t). This number is greater or equal to
543 * zero. -ENOMEM is returned if a new file_region structure or cache entry
544 * is needed and can not be allocated.
546 static long region_chg(struct resv_map *resv, long f, long t,
547 long *out_regions_needed)
551 spin_lock(&resv->lock);
553 /* Count how many hugepages in this range are NOT represented. */
554 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
557 if (*out_regions_needed == 0)
558 *out_regions_needed = 1;
560 if (allocate_file_region_entries(resv, *out_regions_needed))
563 resv->adds_in_progress += *out_regions_needed;
565 spin_unlock(&resv->lock);
570 * Abort the in progress add operation. The adds_in_progress field
571 * of the resv_map keeps track of the operations in progress between
572 * calls to region_chg and region_add. Operations are sometimes
573 * aborted after the call to region_chg. In such cases, region_abort
574 * is called to decrement the adds_in_progress counter. regions_needed
575 * is the value returned by the region_chg call, it is used to decrement
576 * the adds_in_progress counter.
578 * NOTE: The range arguments [f, t) are not needed or used in this
579 * routine. They are kept to make reading the calling code easier as
580 * arguments will match the associated region_chg call.
582 static void region_abort(struct resv_map *resv, long f, long t,
585 spin_lock(&resv->lock);
586 VM_BUG_ON(!resv->region_cache_count);
587 resv->adds_in_progress -= regions_needed;
588 spin_unlock(&resv->lock);
592 * Delete the specified range [f, t) from the reserve map. If the
593 * t parameter is LONG_MAX, this indicates that ALL regions after f
594 * should be deleted. Locate the regions which intersect [f, t)
595 * and either trim, delete or split the existing regions.
597 * Returns the number of huge pages deleted from the reserve map.
598 * In the normal case, the return value is zero or more. In the
599 * case where a region must be split, a new region descriptor must
600 * be allocated. If the allocation fails, -ENOMEM will be returned.
601 * NOTE: If the parameter t == LONG_MAX, then we will never split
602 * a region and possibly return -ENOMEM. Callers specifying
603 * t == LONG_MAX do not need to check for -ENOMEM error.
605 static long region_del(struct resv_map *resv, long f, long t)
607 struct list_head *head = &resv->regions;
608 struct file_region *rg, *trg;
609 struct file_region *nrg = NULL;
613 spin_lock(&resv->lock);
614 list_for_each_entry_safe(rg, trg, head, link) {
616 * Skip regions before the range to be deleted. file_region
617 * ranges are normally of the form [from, to). However, there
618 * may be a "placeholder" entry in the map which is of the form
619 * (from, to) with from == to. Check for placeholder entries
620 * at the beginning of the range to be deleted.
622 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
628 if (f > rg->from && t < rg->to) { /* Must split region */
630 * Check for an entry in the cache before dropping
631 * lock and attempting allocation.
634 resv->region_cache_count > resv->adds_in_progress) {
635 nrg = list_first_entry(&resv->region_cache,
638 list_del(&nrg->link);
639 resv->region_cache_count--;
643 spin_unlock(&resv->lock);
644 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
652 /* New entry for end of split region */
656 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
658 INIT_LIST_HEAD(&nrg->link);
660 /* Original entry is trimmed */
663 hugetlb_cgroup_uncharge_file_region(
664 resv, rg, nrg->to - nrg->from);
666 list_add(&nrg->link, &rg->link);
671 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
672 del += rg->to - rg->from;
673 hugetlb_cgroup_uncharge_file_region(resv, rg,
680 if (f <= rg->from) { /* Trim beginning of region */
684 hugetlb_cgroup_uncharge_file_region(resv, rg,
686 } else { /* Trim end of region */
690 hugetlb_cgroup_uncharge_file_region(resv, rg,
695 spin_unlock(&resv->lock);
701 * A rare out of memory error was encountered which prevented removal of
702 * the reserve map region for a page. The huge page itself was free'ed
703 * and removed from the page cache. This routine will adjust the subpool
704 * usage count, and the global reserve count if needed. By incrementing
705 * these counts, the reserve map entry which could not be deleted will
706 * appear as a "reserved" entry instead of simply dangling with incorrect
709 void hugetlb_fix_reserve_counts(struct inode *inode)
711 struct hugepage_subpool *spool = subpool_inode(inode);
714 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
716 struct hstate *h = hstate_inode(inode);
718 hugetlb_acct_memory(h, 1);
723 * Count and return the number of huge pages in the reserve map
724 * that intersect with the range [f, t).
726 static long region_count(struct resv_map *resv, long f, long t)
728 struct list_head *head = &resv->regions;
729 struct file_region *rg;
732 spin_lock(&resv->lock);
733 /* Locate each segment we overlap with, and count that overlap. */
734 list_for_each_entry(rg, head, link) {
743 seg_from = max(rg->from, f);
744 seg_to = min(rg->to, t);
746 chg += seg_to - seg_from;
748 spin_unlock(&resv->lock);
754 * Convert the address within this vma to the page offset within
755 * the mapping, in pagecache page units; huge pages here.
757 static pgoff_t vma_hugecache_offset(struct hstate *h,
758 struct vm_area_struct *vma, unsigned long address)
760 return ((address - vma->vm_start) >> huge_page_shift(h)) +
761 (vma->vm_pgoff >> huge_page_order(h));
764 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
765 unsigned long address)
767 return vma_hugecache_offset(hstate_vma(vma), vma, address);
769 EXPORT_SYMBOL_GPL(linear_hugepage_index);
772 * Return the size of the pages allocated when backing a VMA. In the majority
773 * cases this will be same size as used by the page table entries.
775 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
777 if (vma->vm_ops && vma->vm_ops->pagesize)
778 return vma->vm_ops->pagesize(vma);
781 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
784 * Return the page size being used by the MMU to back a VMA. In the majority
785 * of cases, the page size used by the kernel matches the MMU size. On
786 * architectures where it differs, an architecture-specific 'strong'
787 * version of this symbol is required.
789 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
791 return vma_kernel_pagesize(vma);
795 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
796 * bits of the reservation map pointer, which are always clear due to
799 #define HPAGE_RESV_OWNER (1UL << 0)
800 #define HPAGE_RESV_UNMAPPED (1UL << 1)
801 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
804 * These helpers are used to track how many pages are reserved for
805 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
806 * is guaranteed to have their future faults succeed.
808 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
809 * the reserve counters are updated with the hugetlb_lock held. It is safe
810 * to reset the VMA at fork() time as it is not in use yet and there is no
811 * chance of the global counters getting corrupted as a result of the values.
813 * The private mapping reservation is represented in a subtly different
814 * manner to a shared mapping. A shared mapping has a region map associated
815 * with the underlying file, this region map represents the backing file
816 * pages which have ever had a reservation assigned which this persists even
817 * after the page is instantiated. A private mapping has a region map
818 * associated with the original mmap which is attached to all VMAs which
819 * reference it, this region map represents those offsets which have consumed
820 * reservation ie. where pages have been instantiated.
822 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
824 return (unsigned long)vma->vm_private_data;
827 static void set_vma_private_data(struct vm_area_struct *vma,
830 vma->vm_private_data = (void *)value;
834 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
835 struct hugetlb_cgroup *h_cg,
838 #ifdef CONFIG_CGROUP_HUGETLB
840 resv_map->reservation_counter = NULL;
841 resv_map->pages_per_hpage = 0;
842 resv_map->css = NULL;
844 resv_map->reservation_counter =
845 &h_cg->rsvd_hugepage[hstate_index(h)];
846 resv_map->pages_per_hpage = pages_per_huge_page(h);
847 resv_map->css = &h_cg->css;
852 struct resv_map *resv_map_alloc(void)
854 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
855 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
857 if (!resv_map || !rg) {
863 kref_init(&resv_map->refs);
864 spin_lock_init(&resv_map->lock);
865 INIT_LIST_HEAD(&resv_map->regions);
867 resv_map->adds_in_progress = 0;
869 * Initialize these to 0. On shared mappings, 0's here indicate these
870 * fields don't do cgroup accounting. On private mappings, these will be
871 * re-initialized to the proper values, to indicate that hugetlb cgroup
872 * reservations are to be un-charged from here.
874 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
876 INIT_LIST_HEAD(&resv_map->region_cache);
877 list_add(&rg->link, &resv_map->region_cache);
878 resv_map->region_cache_count = 1;
883 void resv_map_release(struct kref *ref)
885 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
886 struct list_head *head = &resv_map->region_cache;
887 struct file_region *rg, *trg;
889 /* Clear out any active regions before we release the map. */
890 region_del(resv_map, 0, LONG_MAX);
892 /* ... and any entries left in the cache */
893 list_for_each_entry_safe(rg, trg, head, link) {
898 VM_BUG_ON(resv_map->adds_in_progress);
903 static inline struct resv_map *inode_resv_map(struct inode *inode)
906 * At inode evict time, i_mapping may not point to the original
907 * address space within the inode. This original address space
908 * contains the pointer to the resv_map. So, always use the
909 * address space embedded within the inode.
910 * The VERY common case is inode->mapping == &inode->i_data but,
911 * this may not be true for device special inodes.
913 return (struct resv_map *)(&inode->i_data)->private_data;
916 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
918 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
919 if (vma->vm_flags & VM_MAYSHARE) {
920 struct address_space *mapping = vma->vm_file->f_mapping;
921 struct inode *inode = mapping->host;
923 return inode_resv_map(inode);
926 return (struct resv_map *)(get_vma_private_data(vma) &
931 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
933 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
934 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
936 set_vma_private_data(vma, (get_vma_private_data(vma) &
937 HPAGE_RESV_MASK) | (unsigned long)map);
940 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
942 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
943 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
945 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
948 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
950 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
952 return (get_vma_private_data(vma) & flag) != 0;
955 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
956 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
958 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
959 if (!(vma->vm_flags & VM_MAYSHARE))
960 vma->vm_private_data = (void *)0;
963 /* Returns true if the VMA has associated reserve pages */
964 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
966 if (vma->vm_flags & VM_NORESERVE) {
968 * This address is already reserved by other process(chg == 0),
969 * so, we should decrement reserved count. Without decrementing,
970 * reserve count remains after releasing inode, because this
971 * allocated page will go into page cache and is regarded as
972 * coming from reserved pool in releasing step. Currently, we
973 * don't have any other solution to deal with this situation
974 * properly, so add work-around here.
976 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
982 /* Shared mappings always use reserves */
983 if (vma->vm_flags & VM_MAYSHARE) {
985 * We know VM_NORESERVE is not set. Therefore, there SHOULD
986 * be a region map for all pages. The only situation where
987 * there is no region map is if a hole was punched via
988 * fallocate. In this case, there really are no reserves to
989 * use. This situation is indicated if chg != 0.
998 * Only the process that called mmap() has reserves for
1001 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1003 * Like the shared case above, a hole punch or truncate
1004 * could have been performed on the private mapping.
1005 * Examine the value of chg to determine if reserves
1006 * actually exist or were previously consumed.
1007 * Very Subtle - The value of chg comes from a previous
1008 * call to vma_needs_reserves(). The reserve map for
1009 * private mappings has different (opposite) semantics
1010 * than that of shared mappings. vma_needs_reserves()
1011 * has already taken this difference in semantics into
1012 * account. Therefore, the meaning of chg is the same
1013 * as in the shared case above. Code could easily be
1014 * combined, but keeping it separate draws attention to
1015 * subtle differences.
1026 static void enqueue_huge_page(struct hstate *h, struct page *page)
1028 int nid = page_to_nid(page);
1029 list_move(&page->lru, &h->hugepage_freelists[nid]);
1030 h->free_huge_pages++;
1031 h->free_huge_pages_node[nid]++;
1034 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1037 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1039 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1040 if (nocma && is_migrate_cma_page(page))
1043 if (PageHWPoison(page))
1046 list_move(&page->lru, &h->hugepage_activelist);
1047 set_page_refcounted(page);
1048 h->free_huge_pages--;
1049 h->free_huge_pages_node[nid]--;
1056 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1059 unsigned int cpuset_mems_cookie;
1060 struct zonelist *zonelist;
1063 int node = NUMA_NO_NODE;
1065 zonelist = node_zonelist(nid, gfp_mask);
1068 cpuset_mems_cookie = read_mems_allowed_begin();
1069 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1072 if (!cpuset_zone_allowed(zone, gfp_mask))
1075 * no need to ask again on the same node. Pool is node rather than
1078 if (zone_to_nid(zone) == node)
1080 node = zone_to_nid(zone);
1082 page = dequeue_huge_page_node_exact(h, node);
1086 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1092 static struct page *dequeue_huge_page_vma(struct hstate *h,
1093 struct vm_area_struct *vma,
1094 unsigned long address, int avoid_reserve,
1098 struct mempolicy *mpol;
1100 nodemask_t *nodemask;
1104 * A child process with MAP_PRIVATE mappings created by their parent
1105 * have no page reserves. This check ensures that reservations are
1106 * not "stolen". The child may still get SIGKILLed
1108 if (!vma_has_reserves(vma, chg) &&
1109 h->free_huge_pages - h->resv_huge_pages == 0)
1112 /* If reserves cannot be used, ensure enough pages are in the pool */
1113 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1116 gfp_mask = htlb_alloc_mask(h);
1117 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1118 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1119 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1120 SetPagePrivate(page);
1121 h->resv_huge_pages--;
1124 mpol_cond_put(mpol);
1132 * common helper functions for hstate_next_node_to_{alloc|free}.
1133 * We may have allocated or freed a huge page based on a different
1134 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1135 * be outside of *nodes_allowed. Ensure that we use an allowed
1136 * node for alloc or free.
1138 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1140 nid = next_node_in(nid, *nodes_allowed);
1141 VM_BUG_ON(nid >= MAX_NUMNODES);
1146 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1148 if (!node_isset(nid, *nodes_allowed))
1149 nid = next_node_allowed(nid, nodes_allowed);
1154 * returns the previously saved node ["this node"] from which to
1155 * allocate a persistent huge page for the pool and advance the
1156 * next node from which to allocate, handling wrap at end of node
1159 static int hstate_next_node_to_alloc(struct hstate *h,
1160 nodemask_t *nodes_allowed)
1164 VM_BUG_ON(!nodes_allowed);
1166 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1167 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1173 * helper for free_pool_huge_page() - return the previously saved
1174 * node ["this node"] from which to free a huge page. Advance the
1175 * next node id whether or not we find a free huge page to free so
1176 * that the next attempt to free addresses the next node.
1178 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1182 VM_BUG_ON(!nodes_allowed);
1184 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1185 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1190 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1191 for (nr_nodes = nodes_weight(*mask); \
1193 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1196 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1197 for (nr_nodes = nodes_weight(*mask); \
1199 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1202 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1203 static void destroy_compound_gigantic_page(struct page *page,
1207 int nr_pages = 1 << order;
1208 struct page *p = page + 1;
1210 atomic_set(compound_mapcount_ptr(page), 0);
1211 if (hpage_pincount_available(page))
1212 atomic_set(compound_pincount_ptr(page), 0);
1214 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1215 clear_compound_head(p);
1216 set_page_refcounted(p);
1219 set_compound_order(page, 0);
1220 __ClearPageHead(page);
1223 static void free_gigantic_page(struct page *page, unsigned int order)
1226 * If the page isn't allocated using the cma allocator,
1227 * cma_release() returns false.
1230 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1234 free_contig_range(page_to_pfn(page), 1 << order);
1237 #ifdef CONFIG_CONTIG_ALLOC
1238 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1239 int nid, nodemask_t *nodemask)
1241 unsigned long nr_pages = 1UL << huge_page_order(h);
1242 if (nid == NUMA_NO_NODE)
1243 nid = numa_mem_id();
1250 if (hugetlb_cma[nid]) {
1251 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1252 huge_page_order(h), true);
1257 if (!(gfp_mask & __GFP_THISNODE)) {
1258 for_each_node_mask(node, *nodemask) {
1259 if (node == nid || !hugetlb_cma[node])
1262 page = cma_alloc(hugetlb_cma[node], nr_pages,
1263 huge_page_order(h), true);
1271 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1274 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1275 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1276 #else /* !CONFIG_CONTIG_ALLOC */
1277 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1278 int nid, nodemask_t *nodemask)
1282 #endif /* CONFIG_CONTIG_ALLOC */
1284 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1285 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1286 int nid, nodemask_t *nodemask)
1290 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1291 static inline void destroy_compound_gigantic_page(struct page *page,
1292 unsigned int order) { }
1295 static void update_and_free_page(struct hstate *h, struct page *page)
1299 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1303 h->nr_huge_pages_node[page_to_nid(page)]--;
1304 for (i = 0; i < pages_per_huge_page(h); i++) {
1305 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1306 1 << PG_referenced | 1 << PG_dirty |
1307 1 << PG_active | 1 << PG_private |
1310 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1311 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1312 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1313 set_page_refcounted(page);
1314 if (hstate_is_gigantic(h)) {
1316 * Temporarily drop the hugetlb_lock, because
1317 * we might block in free_gigantic_page().
1319 spin_unlock(&hugetlb_lock);
1320 destroy_compound_gigantic_page(page, huge_page_order(h));
1321 free_gigantic_page(page, huge_page_order(h));
1322 spin_lock(&hugetlb_lock);
1324 __free_pages(page, huge_page_order(h));
1328 struct hstate *size_to_hstate(unsigned long size)
1332 for_each_hstate(h) {
1333 if (huge_page_size(h) == size)
1340 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1341 * to hstate->hugepage_activelist.)
1343 * This function can be called for tail pages, but never returns true for them.
1345 bool page_huge_active(struct page *page)
1347 VM_BUG_ON_PAGE(!PageHuge(page), page);
1348 return PageHead(page) && PagePrivate(&page[1]);
1351 /* never called for tail page */
1352 static void set_page_huge_active(struct page *page)
1354 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1355 SetPagePrivate(&page[1]);
1358 static void clear_page_huge_active(struct page *page)
1360 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1361 ClearPagePrivate(&page[1]);
1365 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1368 static inline bool PageHugeTemporary(struct page *page)
1370 if (!PageHuge(page))
1373 return (unsigned long)page[2].mapping == -1U;
1376 static inline void SetPageHugeTemporary(struct page *page)
1378 page[2].mapping = (void *)-1U;
1381 static inline void ClearPageHugeTemporary(struct page *page)
1383 page[2].mapping = NULL;
1386 static void __free_huge_page(struct page *page)
1389 * Can't pass hstate in here because it is called from the
1390 * compound page destructor.
1392 struct hstate *h = page_hstate(page);
1393 int nid = page_to_nid(page);
1394 struct hugepage_subpool *spool =
1395 (struct hugepage_subpool *)page_private(page);
1396 bool restore_reserve;
1398 VM_BUG_ON_PAGE(page_count(page), page);
1399 VM_BUG_ON_PAGE(page_mapcount(page), page);
1401 set_page_private(page, 0);
1402 page->mapping = NULL;
1403 restore_reserve = PagePrivate(page);
1404 ClearPagePrivate(page);
1407 * If PagePrivate() was set on page, page allocation consumed a
1408 * reservation. If the page was associated with a subpool, there
1409 * would have been a page reserved in the subpool before allocation
1410 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1411 * reservtion, do not call hugepage_subpool_put_pages() as this will
1412 * remove the reserved page from the subpool.
1414 if (!restore_reserve) {
1416 * A return code of zero implies that the subpool will be
1417 * under its minimum size if the reservation is not restored
1418 * after page is free. Therefore, force restore_reserve
1421 if (hugepage_subpool_put_pages(spool, 1) == 0)
1422 restore_reserve = true;
1425 spin_lock(&hugetlb_lock);
1426 clear_page_huge_active(page);
1427 hugetlb_cgroup_uncharge_page(hstate_index(h),
1428 pages_per_huge_page(h), page);
1429 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1430 pages_per_huge_page(h), page);
1431 if (restore_reserve)
1432 h->resv_huge_pages++;
1434 if (PageHugeTemporary(page)) {
1435 list_del(&page->lru);
1436 ClearPageHugeTemporary(page);
1437 update_and_free_page(h, page);
1438 } else if (h->surplus_huge_pages_node[nid]) {
1439 /* remove the page from active list */
1440 list_del(&page->lru);
1441 update_and_free_page(h, page);
1442 h->surplus_huge_pages--;
1443 h->surplus_huge_pages_node[nid]--;
1445 arch_clear_hugepage_flags(page);
1446 enqueue_huge_page(h, page);
1448 spin_unlock(&hugetlb_lock);
1452 * As free_huge_page() can be called from a non-task context, we have
1453 * to defer the actual freeing in a workqueue to prevent potential
1454 * hugetlb_lock deadlock.
1456 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1457 * be freed and frees them one-by-one. As the page->mapping pointer is
1458 * going to be cleared in __free_huge_page() anyway, it is reused as the
1459 * llist_node structure of a lockless linked list of huge pages to be freed.
1461 static LLIST_HEAD(hpage_freelist);
1463 static void free_hpage_workfn(struct work_struct *work)
1465 struct llist_node *node;
1468 node = llist_del_all(&hpage_freelist);
1471 page = container_of((struct address_space **)node,
1472 struct page, mapping);
1474 __free_huge_page(page);
1477 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1479 void free_huge_page(struct page *page)
1482 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1486 * Only call schedule_work() if hpage_freelist is previously
1487 * empty. Otherwise, schedule_work() had been called but the
1488 * workfn hasn't retrieved the list yet.
1490 if (llist_add((struct llist_node *)&page->mapping,
1492 schedule_work(&free_hpage_work);
1496 __free_huge_page(page);
1499 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1501 INIT_LIST_HEAD(&page->lru);
1502 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1503 set_hugetlb_cgroup(page, NULL);
1504 set_hugetlb_cgroup_rsvd(page, NULL);
1505 spin_lock(&hugetlb_lock);
1507 h->nr_huge_pages_node[nid]++;
1508 spin_unlock(&hugetlb_lock);
1511 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1514 int nr_pages = 1 << order;
1515 struct page *p = page + 1;
1517 /* we rely on prep_new_huge_page to set the destructor */
1518 set_compound_order(page, order);
1519 __ClearPageReserved(page);
1520 __SetPageHead(page);
1521 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1523 * For gigantic hugepages allocated through bootmem at
1524 * boot, it's safer to be consistent with the not-gigantic
1525 * hugepages and clear the PG_reserved bit from all tail pages
1526 * too. Otherwise drivers using get_user_pages() to access tail
1527 * pages may get the reference counting wrong if they see
1528 * PG_reserved set on a tail page (despite the head page not
1529 * having PG_reserved set). Enforcing this consistency between
1530 * head and tail pages allows drivers to optimize away a check
1531 * on the head page when they need know if put_page() is needed
1532 * after get_user_pages().
1534 __ClearPageReserved(p);
1535 set_page_count(p, 0);
1536 set_compound_head(p, page);
1538 atomic_set(compound_mapcount_ptr(page), -1);
1540 if (hpage_pincount_available(page))
1541 atomic_set(compound_pincount_ptr(page), 0);
1545 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1546 * transparent huge pages. See the PageTransHuge() documentation for more
1549 int PageHuge(struct page *page)
1551 if (!PageCompound(page))
1554 page = compound_head(page);
1555 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1557 EXPORT_SYMBOL_GPL(PageHuge);
1560 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1561 * normal or transparent huge pages.
1563 int PageHeadHuge(struct page *page_head)
1565 if (!PageHead(page_head))
1568 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1572 * Find address_space associated with hugetlbfs page.
1573 * Upon entry page is locked and page 'was' mapped although mapped state
1574 * could change. If necessary, use anon_vma to find vma and associated
1575 * address space. The returned mapping may be stale, but it can not be
1576 * invalid as page lock (which is held) is required to destroy mapping.
1578 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1580 struct anon_vma *anon_vma;
1581 pgoff_t pgoff_start, pgoff_end;
1582 struct anon_vma_chain *avc;
1583 struct address_space *mapping = page_mapping(hpage);
1585 /* Simple file based mapping */
1590 * Even anonymous hugetlbfs mappings are associated with an
1591 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1592 * code). Find a vma associated with the anonymous vma, and
1593 * use the file pointer to get address_space.
1595 anon_vma = page_lock_anon_vma_read(hpage);
1597 return mapping; /* NULL */
1599 /* Use first found vma */
1600 pgoff_start = page_to_pgoff(hpage);
1601 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
1602 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1603 pgoff_start, pgoff_end) {
1604 struct vm_area_struct *vma = avc->vma;
1606 mapping = vma->vm_file->f_mapping;
1610 anon_vma_unlock_read(anon_vma);
1615 * Find and lock address space (mapping) in write mode.
1617 * Upon entry, the page is locked which allows us to find the mapping
1618 * even in the case of an anon page. However, locking order dictates
1619 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1620 * specific. So, we first try to lock the sema while still holding the
1621 * page lock. If this works, great! If not, then we need to drop the
1622 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1623 * course, need to revalidate state along the way.
1625 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1627 struct address_space *mapping, *mapping2;
1629 mapping = _get_hugetlb_page_mapping(hpage);
1635 * If no contention, take lock and return
1637 if (i_mmap_trylock_write(mapping))
1641 * Must drop page lock and wait on mapping sema.
1642 * Note: Once page lock is dropped, mapping could become invalid.
1643 * As a hack, increase map count until we lock page again.
1645 atomic_inc(&hpage->_mapcount);
1647 i_mmap_lock_write(mapping);
1649 atomic_add_negative(-1, &hpage->_mapcount);
1651 /* verify page is still mapped */
1652 if (!page_mapped(hpage)) {
1653 i_mmap_unlock_write(mapping);
1658 * Get address space again and verify it is the same one
1659 * we locked. If not, drop lock and retry.
1661 mapping2 = _get_hugetlb_page_mapping(hpage);
1662 if (mapping2 != mapping) {
1663 i_mmap_unlock_write(mapping);
1671 pgoff_t __basepage_index(struct page *page)
1673 struct page *page_head = compound_head(page);
1674 pgoff_t index = page_index(page_head);
1675 unsigned long compound_idx;
1677 if (!PageHuge(page_head))
1678 return page_index(page);
1680 if (compound_order(page_head) >= MAX_ORDER)
1681 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1683 compound_idx = page - page_head;
1685 return (index << compound_order(page_head)) + compound_idx;
1688 static struct page *alloc_buddy_huge_page(struct hstate *h,
1689 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1690 nodemask_t *node_alloc_noretry)
1692 int order = huge_page_order(h);
1694 bool alloc_try_hard = true;
1697 * By default we always try hard to allocate the page with
1698 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1699 * a loop (to adjust global huge page counts) and previous allocation
1700 * failed, do not continue to try hard on the same node. Use the
1701 * node_alloc_noretry bitmap to manage this state information.
1703 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1704 alloc_try_hard = false;
1705 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1707 gfp_mask |= __GFP_RETRY_MAYFAIL;
1708 if (nid == NUMA_NO_NODE)
1709 nid = numa_mem_id();
1710 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1712 __count_vm_event(HTLB_BUDDY_PGALLOC);
1714 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1717 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1718 * indicates an overall state change. Clear bit so that we resume
1719 * normal 'try hard' allocations.
1721 if (node_alloc_noretry && page && !alloc_try_hard)
1722 node_clear(nid, *node_alloc_noretry);
1725 * If we tried hard to get a page but failed, set bit so that
1726 * subsequent attempts will not try as hard until there is an
1727 * overall state change.
1729 if (node_alloc_noretry && !page && alloc_try_hard)
1730 node_set(nid, *node_alloc_noretry);
1736 * Common helper to allocate a fresh hugetlb page. All specific allocators
1737 * should use this function to get new hugetlb pages
1739 static struct page *alloc_fresh_huge_page(struct hstate *h,
1740 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1741 nodemask_t *node_alloc_noretry)
1745 if (hstate_is_gigantic(h))
1746 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1748 page = alloc_buddy_huge_page(h, gfp_mask,
1749 nid, nmask, node_alloc_noretry);
1753 if (hstate_is_gigantic(h))
1754 prep_compound_gigantic_page(page, huge_page_order(h));
1755 prep_new_huge_page(h, page, page_to_nid(page));
1761 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1764 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1765 nodemask_t *node_alloc_noretry)
1769 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1771 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1772 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1773 node_alloc_noretry);
1781 put_page(page); /* free it into the hugepage allocator */
1787 * Free huge page from pool from next node to free.
1788 * Attempt to keep persistent huge pages more or less
1789 * balanced over allowed nodes.
1790 * Called with hugetlb_lock locked.
1792 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1798 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1800 * If we're returning unused surplus pages, only examine
1801 * nodes with surplus pages.
1803 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1804 !list_empty(&h->hugepage_freelists[node])) {
1806 list_entry(h->hugepage_freelists[node].next,
1808 list_del(&page->lru);
1809 h->free_huge_pages--;
1810 h->free_huge_pages_node[node]--;
1812 h->surplus_huge_pages--;
1813 h->surplus_huge_pages_node[node]--;
1815 update_and_free_page(h, page);
1825 * Dissolve a given free hugepage into free buddy pages. This function does
1826 * nothing for in-use hugepages and non-hugepages.
1827 * This function returns values like below:
1829 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1830 * (allocated or reserved.)
1831 * 0: successfully dissolved free hugepages or the page is not a
1832 * hugepage (considered as already dissolved)
1834 int dissolve_free_huge_page(struct page *page)
1838 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1839 if (!PageHuge(page))
1842 spin_lock(&hugetlb_lock);
1843 if (!PageHuge(page)) {
1848 if (!page_count(page)) {
1849 struct page *head = compound_head(page);
1850 struct hstate *h = page_hstate(head);
1851 int nid = page_to_nid(head);
1852 if (h->free_huge_pages - h->resv_huge_pages == 0)
1855 * Move PageHWPoison flag from head page to the raw error page,
1856 * which makes any subpages rather than the error page reusable.
1858 if (PageHWPoison(head) && page != head) {
1859 SetPageHWPoison(page);
1860 ClearPageHWPoison(head);
1862 list_del(&head->lru);
1863 h->free_huge_pages--;
1864 h->free_huge_pages_node[nid]--;
1865 h->max_huge_pages--;
1866 update_and_free_page(h, head);
1870 spin_unlock(&hugetlb_lock);
1875 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1876 * make specified memory blocks removable from the system.
1877 * Note that this will dissolve a free gigantic hugepage completely, if any
1878 * part of it lies within the given range.
1879 * Also note that if dissolve_free_huge_page() returns with an error, all
1880 * free hugepages that were dissolved before that error are lost.
1882 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1888 if (!hugepages_supported())
1891 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1892 page = pfn_to_page(pfn);
1893 rc = dissolve_free_huge_page(page);
1902 * Allocates a fresh surplus page from the page allocator.
1904 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1905 int nid, nodemask_t *nmask)
1907 struct page *page = NULL;
1909 if (hstate_is_gigantic(h))
1912 spin_lock(&hugetlb_lock);
1913 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1915 spin_unlock(&hugetlb_lock);
1917 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1921 spin_lock(&hugetlb_lock);
1923 * We could have raced with the pool size change.
1924 * Double check that and simply deallocate the new page
1925 * if we would end up overcommiting the surpluses. Abuse
1926 * temporary page to workaround the nasty free_huge_page
1929 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1930 SetPageHugeTemporary(page);
1931 spin_unlock(&hugetlb_lock);
1935 h->surplus_huge_pages++;
1936 h->surplus_huge_pages_node[page_to_nid(page)]++;
1940 spin_unlock(&hugetlb_lock);
1945 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1946 int nid, nodemask_t *nmask)
1950 if (hstate_is_gigantic(h))
1953 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1958 * We do not account these pages as surplus because they are only
1959 * temporary and will be released properly on the last reference
1961 SetPageHugeTemporary(page);
1967 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1970 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1971 struct vm_area_struct *vma, unsigned long addr)
1974 struct mempolicy *mpol;
1975 gfp_t gfp_mask = htlb_alloc_mask(h);
1977 nodemask_t *nodemask;
1979 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1980 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1981 mpol_cond_put(mpol);
1986 /* page migration callback function */
1987 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1988 nodemask_t *nmask, gfp_t gfp_mask)
1990 spin_lock(&hugetlb_lock);
1991 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1994 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1996 spin_unlock(&hugetlb_lock);
2000 spin_unlock(&hugetlb_lock);
2002 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2005 /* mempolicy aware migration callback */
2006 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2007 unsigned long address)
2009 struct mempolicy *mpol;
2010 nodemask_t *nodemask;
2015 gfp_mask = htlb_alloc_mask(h);
2016 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2017 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2018 mpol_cond_put(mpol);
2024 * Increase the hugetlb pool such that it can accommodate a reservation
2027 static int gather_surplus_pages(struct hstate *h, int delta)
2028 __must_hold(&hugetlb_lock)
2030 struct list_head surplus_list;
2031 struct page *page, *tmp;
2033 int needed, allocated;
2034 bool alloc_ok = true;
2036 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2038 h->resv_huge_pages += delta;
2043 INIT_LIST_HEAD(&surplus_list);
2047 spin_unlock(&hugetlb_lock);
2048 for (i = 0; i < needed; i++) {
2049 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2050 NUMA_NO_NODE, NULL);
2055 list_add(&page->lru, &surplus_list);
2061 * After retaking hugetlb_lock, we need to recalculate 'needed'
2062 * because either resv_huge_pages or free_huge_pages may have changed.
2064 spin_lock(&hugetlb_lock);
2065 needed = (h->resv_huge_pages + delta) -
2066 (h->free_huge_pages + allocated);
2071 * We were not able to allocate enough pages to
2072 * satisfy the entire reservation so we free what
2073 * we've allocated so far.
2078 * The surplus_list now contains _at_least_ the number of extra pages
2079 * needed to accommodate the reservation. Add the appropriate number
2080 * of pages to the hugetlb pool and free the extras back to the buddy
2081 * allocator. Commit the entire reservation here to prevent another
2082 * process from stealing the pages as they are added to the pool but
2083 * before they are reserved.
2085 needed += allocated;
2086 h->resv_huge_pages += delta;
2089 /* Free the needed pages to the hugetlb pool */
2090 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2094 * This page is now managed by the hugetlb allocator and has
2095 * no users -- drop the buddy allocator's reference.
2097 put_page_testzero(page);
2098 VM_BUG_ON_PAGE(page_count(page), page);
2099 enqueue_huge_page(h, page);
2102 spin_unlock(&hugetlb_lock);
2104 /* Free unnecessary surplus pages to the buddy allocator */
2105 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2107 spin_lock(&hugetlb_lock);
2113 * This routine has two main purposes:
2114 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2115 * in unused_resv_pages. This corresponds to the prior adjustments made
2116 * to the associated reservation map.
2117 * 2) Free any unused surplus pages that may have been allocated to satisfy
2118 * the reservation. As many as unused_resv_pages may be freed.
2120 * Called with hugetlb_lock held. However, the lock could be dropped (and
2121 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2122 * we must make sure nobody else can claim pages we are in the process of
2123 * freeing. Do this by ensuring resv_huge_page always is greater than the
2124 * number of huge pages we plan to free when dropping the lock.
2126 static void return_unused_surplus_pages(struct hstate *h,
2127 unsigned long unused_resv_pages)
2129 unsigned long nr_pages;
2131 /* Cannot return gigantic pages currently */
2132 if (hstate_is_gigantic(h))
2136 * Part (or even all) of the reservation could have been backed
2137 * by pre-allocated pages. Only free surplus pages.
2139 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2142 * We want to release as many surplus pages as possible, spread
2143 * evenly across all nodes with memory. Iterate across these nodes
2144 * until we can no longer free unreserved surplus pages. This occurs
2145 * when the nodes with surplus pages have no free pages.
2146 * free_pool_huge_page() will balance the freed pages across the
2147 * on-line nodes with memory and will handle the hstate accounting.
2149 * Note that we decrement resv_huge_pages as we free the pages. If
2150 * we drop the lock, resv_huge_pages will still be sufficiently large
2151 * to cover subsequent pages we may free.
2153 while (nr_pages--) {
2154 h->resv_huge_pages--;
2155 unused_resv_pages--;
2156 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2158 cond_resched_lock(&hugetlb_lock);
2162 /* Fully uncommit the reservation */
2163 h->resv_huge_pages -= unused_resv_pages;
2168 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2169 * are used by the huge page allocation routines to manage reservations.
2171 * vma_needs_reservation is called to determine if the huge page at addr
2172 * within the vma has an associated reservation. If a reservation is
2173 * needed, the value 1 is returned. The caller is then responsible for
2174 * managing the global reservation and subpool usage counts. After
2175 * the huge page has been allocated, vma_commit_reservation is called
2176 * to add the page to the reservation map. If the page allocation fails,
2177 * the reservation must be ended instead of committed. vma_end_reservation
2178 * is called in such cases.
2180 * In the normal case, vma_commit_reservation returns the same value
2181 * as the preceding vma_needs_reservation call. The only time this
2182 * is not the case is if a reserve map was changed between calls. It
2183 * is the responsibility of the caller to notice the difference and
2184 * take appropriate action.
2186 * vma_add_reservation is used in error paths where a reservation must
2187 * be restored when a newly allocated huge page must be freed. It is
2188 * to be called after calling vma_needs_reservation to determine if a
2189 * reservation exists.
2191 enum vma_resv_mode {
2197 static long __vma_reservation_common(struct hstate *h,
2198 struct vm_area_struct *vma, unsigned long addr,
2199 enum vma_resv_mode mode)
2201 struct resv_map *resv;
2204 long dummy_out_regions_needed;
2206 resv = vma_resv_map(vma);
2210 idx = vma_hugecache_offset(h, vma, addr);
2212 case VMA_NEEDS_RESV:
2213 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2214 /* We assume that vma_reservation_* routines always operate on
2215 * 1 page, and that adding to resv map a 1 page entry can only
2216 * ever require 1 region.
2218 VM_BUG_ON(dummy_out_regions_needed != 1);
2220 case VMA_COMMIT_RESV:
2221 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2222 /* region_add calls of range 1 should never fail. */
2226 region_abort(resv, idx, idx + 1, 1);
2230 if (vma->vm_flags & VM_MAYSHARE) {
2231 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2232 /* region_add calls of range 1 should never fail. */
2235 region_abort(resv, idx, idx + 1, 1);
2236 ret = region_del(resv, idx, idx + 1);
2243 if (vma->vm_flags & VM_MAYSHARE)
2245 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2247 * In most cases, reserves always exist for private mappings.
2248 * However, a file associated with mapping could have been
2249 * hole punched or truncated after reserves were consumed.
2250 * As subsequent fault on such a range will not use reserves.
2251 * Subtle - The reserve map for private mappings has the
2252 * opposite meaning than that of shared mappings. If NO
2253 * entry is in the reserve map, it means a reservation exists.
2254 * If an entry exists in the reserve map, it means the
2255 * reservation has already been consumed. As a result, the
2256 * return value of this routine is the opposite of the
2257 * value returned from reserve map manipulation routines above.
2265 return ret < 0 ? ret : 0;
2268 static long vma_needs_reservation(struct hstate *h,
2269 struct vm_area_struct *vma, unsigned long addr)
2271 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2274 static long vma_commit_reservation(struct hstate *h,
2275 struct vm_area_struct *vma, unsigned long addr)
2277 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2280 static void vma_end_reservation(struct hstate *h,
2281 struct vm_area_struct *vma, unsigned long addr)
2283 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2286 static long vma_add_reservation(struct hstate *h,
2287 struct vm_area_struct *vma, unsigned long addr)
2289 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2293 * This routine is called to restore a reservation on error paths. In the
2294 * specific error paths, a huge page was allocated (via alloc_huge_page)
2295 * and is about to be freed. If a reservation for the page existed,
2296 * alloc_huge_page would have consumed the reservation and set PagePrivate
2297 * in the newly allocated page. When the page is freed via free_huge_page,
2298 * the global reservation count will be incremented if PagePrivate is set.
2299 * However, free_huge_page can not adjust the reserve map. Adjust the
2300 * reserve map here to be consistent with global reserve count adjustments
2301 * to be made by free_huge_page.
2303 static void restore_reserve_on_error(struct hstate *h,
2304 struct vm_area_struct *vma, unsigned long address,
2307 if (unlikely(PagePrivate(page))) {
2308 long rc = vma_needs_reservation(h, vma, address);
2310 if (unlikely(rc < 0)) {
2312 * Rare out of memory condition in reserve map
2313 * manipulation. Clear PagePrivate so that
2314 * global reserve count will not be incremented
2315 * by free_huge_page. This will make it appear
2316 * as though the reservation for this page was
2317 * consumed. This may prevent the task from
2318 * faulting in the page at a later time. This
2319 * is better than inconsistent global huge page
2320 * accounting of reserve counts.
2322 ClearPagePrivate(page);
2324 rc = vma_add_reservation(h, vma, address);
2325 if (unlikely(rc < 0))
2327 * See above comment about rare out of
2330 ClearPagePrivate(page);
2332 vma_end_reservation(h, vma, address);
2336 struct page *alloc_huge_page(struct vm_area_struct *vma,
2337 unsigned long addr, int avoid_reserve)
2339 struct hugepage_subpool *spool = subpool_vma(vma);
2340 struct hstate *h = hstate_vma(vma);
2342 long map_chg, map_commit;
2345 struct hugetlb_cgroup *h_cg;
2346 bool deferred_reserve;
2348 idx = hstate_index(h);
2350 * Examine the region/reserve map to determine if the process
2351 * has a reservation for the page to be allocated. A return
2352 * code of zero indicates a reservation exists (no change).
2354 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2356 return ERR_PTR(-ENOMEM);
2359 * Processes that did not create the mapping will have no
2360 * reserves as indicated by the region/reserve map. Check
2361 * that the allocation will not exceed the subpool limit.
2362 * Allocations for MAP_NORESERVE mappings also need to be
2363 * checked against any subpool limit.
2365 if (map_chg || avoid_reserve) {
2366 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2368 vma_end_reservation(h, vma, addr);
2369 return ERR_PTR(-ENOSPC);
2373 * Even though there was no reservation in the region/reserve
2374 * map, there could be reservations associated with the
2375 * subpool that can be used. This would be indicated if the
2376 * return value of hugepage_subpool_get_pages() is zero.
2377 * However, if avoid_reserve is specified we still avoid even
2378 * the subpool reservations.
2384 /* If this allocation is not consuming a reservation, charge it now.
2386 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2387 if (deferred_reserve) {
2388 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2389 idx, pages_per_huge_page(h), &h_cg);
2391 goto out_subpool_put;
2394 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2396 goto out_uncharge_cgroup_reservation;
2398 spin_lock(&hugetlb_lock);
2400 * glb_chg is passed to indicate whether or not a page must be taken
2401 * from the global free pool (global change). gbl_chg == 0 indicates
2402 * a reservation exists for the allocation.
2404 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2406 spin_unlock(&hugetlb_lock);
2407 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2409 goto out_uncharge_cgroup;
2410 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2411 SetPagePrivate(page);
2412 h->resv_huge_pages--;
2414 spin_lock(&hugetlb_lock);
2415 list_add(&page->lru, &h->hugepage_activelist);
2418 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2419 /* If allocation is not consuming a reservation, also store the
2420 * hugetlb_cgroup pointer on the page.
2422 if (deferred_reserve) {
2423 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2427 spin_unlock(&hugetlb_lock);
2429 set_page_private(page, (unsigned long)spool);
2431 map_commit = vma_commit_reservation(h, vma, addr);
2432 if (unlikely(map_chg > map_commit)) {
2434 * The page was added to the reservation map between
2435 * vma_needs_reservation and vma_commit_reservation.
2436 * This indicates a race with hugetlb_reserve_pages.
2437 * Adjust for the subpool count incremented above AND
2438 * in hugetlb_reserve_pages for the same page. Also,
2439 * the reservation count added in hugetlb_reserve_pages
2440 * no longer applies.
2444 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2445 hugetlb_acct_memory(h, -rsv_adjust);
2449 out_uncharge_cgroup:
2450 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2451 out_uncharge_cgroup_reservation:
2452 if (deferred_reserve)
2453 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2456 if (map_chg || avoid_reserve)
2457 hugepage_subpool_put_pages(spool, 1);
2458 vma_end_reservation(h, vma, addr);
2459 return ERR_PTR(-ENOSPC);
2462 int alloc_bootmem_huge_page(struct hstate *h)
2463 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2464 int __alloc_bootmem_huge_page(struct hstate *h)
2466 struct huge_bootmem_page *m;
2469 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2472 addr = memblock_alloc_try_nid_raw(
2473 huge_page_size(h), huge_page_size(h),
2474 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2477 * Use the beginning of the huge page to store the
2478 * huge_bootmem_page struct (until gather_bootmem
2479 * puts them into the mem_map).
2488 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2489 /* Put them into a private list first because mem_map is not up yet */
2490 INIT_LIST_HEAD(&m->list);
2491 list_add(&m->list, &huge_boot_pages);
2496 static void __init prep_compound_huge_page(struct page *page,
2499 if (unlikely(order > (MAX_ORDER - 1)))
2500 prep_compound_gigantic_page(page, order);
2502 prep_compound_page(page, order);
2505 /* Put bootmem huge pages into the standard lists after mem_map is up */
2506 static void __init gather_bootmem_prealloc(void)
2508 struct huge_bootmem_page *m;
2510 list_for_each_entry(m, &huge_boot_pages, list) {
2511 struct page *page = virt_to_page(m);
2512 struct hstate *h = m->hstate;
2514 WARN_ON(page_count(page) != 1);
2515 prep_compound_huge_page(page, h->order);
2516 WARN_ON(PageReserved(page));
2517 prep_new_huge_page(h, page, page_to_nid(page));
2518 put_page(page); /* free it into the hugepage allocator */
2521 * If we had gigantic hugepages allocated at boot time, we need
2522 * to restore the 'stolen' pages to totalram_pages in order to
2523 * fix confusing memory reports from free(1) and another
2524 * side-effects, like CommitLimit going negative.
2526 if (hstate_is_gigantic(h))
2527 adjust_managed_page_count(page, 1 << h->order);
2532 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2535 nodemask_t *node_alloc_noretry;
2537 if (!hstate_is_gigantic(h)) {
2539 * Bit mask controlling how hard we retry per-node allocations.
2540 * Ignore errors as lower level routines can deal with
2541 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2542 * time, we are likely in bigger trouble.
2544 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2547 /* allocations done at boot time */
2548 node_alloc_noretry = NULL;
2551 /* bit mask controlling how hard we retry per-node allocations */
2552 if (node_alloc_noretry)
2553 nodes_clear(*node_alloc_noretry);
2555 for (i = 0; i < h->max_huge_pages; ++i) {
2556 if (hstate_is_gigantic(h)) {
2557 if (hugetlb_cma_size) {
2558 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2561 if (!alloc_bootmem_huge_page(h))
2563 } else if (!alloc_pool_huge_page(h,
2564 &node_states[N_MEMORY],
2565 node_alloc_noretry))
2569 if (i < h->max_huge_pages) {
2572 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2573 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2574 h->max_huge_pages, buf, i);
2575 h->max_huge_pages = i;
2578 kfree(node_alloc_noretry);
2581 static void __init hugetlb_init_hstates(void)
2585 for_each_hstate(h) {
2586 if (minimum_order > huge_page_order(h))
2587 minimum_order = huge_page_order(h);
2589 /* oversize hugepages were init'ed in early boot */
2590 if (!hstate_is_gigantic(h))
2591 hugetlb_hstate_alloc_pages(h);
2593 VM_BUG_ON(minimum_order == UINT_MAX);
2596 static void __init report_hugepages(void)
2600 for_each_hstate(h) {
2603 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2604 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2605 buf, h->free_huge_pages);
2609 #ifdef CONFIG_HIGHMEM
2610 static void try_to_free_low(struct hstate *h, unsigned long count,
2611 nodemask_t *nodes_allowed)
2615 if (hstate_is_gigantic(h))
2618 for_each_node_mask(i, *nodes_allowed) {
2619 struct page *page, *next;
2620 struct list_head *freel = &h->hugepage_freelists[i];
2621 list_for_each_entry_safe(page, next, freel, lru) {
2622 if (count >= h->nr_huge_pages)
2624 if (PageHighMem(page))
2626 list_del(&page->lru);
2627 update_and_free_page(h, page);
2628 h->free_huge_pages--;
2629 h->free_huge_pages_node[page_to_nid(page)]--;
2634 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2635 nodemask_t *nodes_allowed)
2641 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2642 * balanced by operating on them in a round-robin fashion.
2643 * Returns 1 if an adjustment was made.
2645 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2650 VM_BUG_ON(delta != -1 && delta != 1);
2653 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2654 if (h->surplus_huge_pages_node[node])
2658 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2659 if (h->surplus_huge_pages_node[node] <
2660 h->nr_huge_pages_node[node])
2667 h->surplus_huge_pages += delta;
2668 h->surplus_huge_pages_node[node] += delta;
2672 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2673 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2674 nodemask_t *nodes_allowed)
2676 unsigned long min_count, ret;
2677 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2680 * Bit mask controlling how hard we retry per-node allocations.
2681 * If we can not allocate the bit mask, do not attempt to allocate
2682 * the requested huge pages.
2684 if (node_alloc_noretry)
2685 nodes_clear(*node_alloc_noretry);
2689 spin_lock(&hugetlb_lock);
2692 * Check for a node specific request.
2693 * Changing node specific huge page count may require a corresponding
2694 * change to the global count. In any case, the passed node mask
2695 * (nodes_allowed) will restrict alloc/free to the specified node.
2697 if (nid != NUMA_NO_NODE) {
2698 unsigned long old_count = count;
2700 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2702 * User may have specified a large count value which caused the
2703 * above calculation to overflow. In this case, they wanted
2704 * to allocate as many huge pages as possible. Set count to
2705 * largest possible value to align with their intention.
2707 if (count < old_count)
2712 * Gigantic pages runtime allocation depend on the capability for large
2713 * page range allocation.
2714 * If the system does not provide this feature, return an error when
2715 * the user tries to allocate gigantic pages but let the user free the
2716 * boottime allocated gigantic pages.
2718 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2719 if (count > persistent_huge_pages(h)) {
2720 spin_unlock(&hugetlb_lock);
2721 NODEMASK_FREE(node_alloc_noretry);
2724 /* Fall through to decrease pool */
2728 * Increase the pool size
2729 * First take pages out of surplus state. Then make up the
2730 * remaining difference by allocating fresh huge pages.
2732 * We might race with alloc_surplus_huge_page() here and be unable
2733 * to convert a surplus huge page to a normal huge page. That is
2734 * not critical, though, it just means the overall size of the
2735 * pool might be one hugepage larger than it needs to be, but
2736 * within all the constraints specified by the sysctls.
2738 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2739 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2743 while (count > persistent_huge_pages(h)) {
2745 * If this allocation races such that we no longer need the
2746 * page, free_huge_page will handle it by freeing the page
2747 * and reducing the surplus.
2749 spin_unlock(&hugetlb_lock);
2751 /* yield cpu to avoid soft lockup */
2754 ret = alloc_pool_huge_page(h, nodes_allowed,
2755 node_alloc_noretry);
2756 spin_lock(&hugetlb_lock);
2760 /* Bail for signals. Probably ctrl-c from user */
2761 if (signal_pending(current))
2766 * Decrease the pool size
2767 * First return free pages to the buddy allocator (being careful
2768 * to keep enough around to satisfy reservations). Then place
2769 * pages into surplus state as needed so the pool will shrink
2770 * to the desired size as pages become free.
2772 * By placing pages into the surplus state independent of the
2773 * overcommit value, we are allowing the surplus pool size to
2774 * exceed overcommit. There are few sane options here. Since
2775 * alloc_surplus_huge_page() is checking the global counter,
2776 * though, we'll note that we're not allowed to exceed surplus
2777 * and won't grow the pool anywhere else. Not until one of the
2778 * sysctls are changed, or the surplus pages go out of use.
2780 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2781 min_count = max(count, min_count);
2782 try_to_free_low(h, min_count, nodes_allowed);
2783 while (min_count < persistent_huge_pages(h)) {
2784 if (!free_pool_huge_page(h, nodes_allowed, 0))
2786 cond_resched_lock(&hugetlb_lock);
2788 while (count < persistent_huge_pages(h)) {
2789 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2793 h->max_huge_pages = persistent_huge_pages(h);
2794 spin_unlock(&hugetlb_lock);
2796 NODEMASK_FREE(node_alloc_noretry);
2801 #define HSTATE_ATTR_RO(_name) \
2802 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2804 #define HSTATE_ATTR(_name) \
2805 static struct kobj_attribute _name##_attr = \
2806 __ATTR(_name, 0644, _name##_show, _name##_store)
2808 static struct kobject *hugepages_kobj;
2809 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2811 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2813 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2817 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2818 if (hstate_kobjs[i] == kobj) {
2820 *nidp = NUMA_NO_NODE;
2824 return kobj_to_node_hstate(kobj, nidp);
2827 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2828 struct kobj_attribute *attr, char *buf)
2831 unsigned long nr_huge_pages;
2834 h = kobj_to_hstate(kobj, &nid);
2835 if (nid == NUMA_NO_NODE)
2836 nr_huge_pages = h->nr_huge_pages;
2838 nr_huge_pages = h->nr_huge_pages_node[nid];
2840 return sprintf(buf, "%lu\n", nr_huge_pages);
2843 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2844 struct hstate *h, int nid,
2845 unsigned long count, size_t len)
2848 nodemask_t nodes_allowed, *n_mask;
2850 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2853 if (nid == NUMA_NO_NODE) {
2855 * global hstate attribute
2857 if (!(obey_mempolicy &&
2858 init_nodemask_of_mempolicy(&nodes_allowed)))
2859 n_mask = &node_states[N_MEMORY];
2861 n_mask = &nodes_allowed;
2864 * Node specific request. count adjustment happens in
2865 * set_max_huge_pages() after acquiring hugetlb_lock.
2867 init_nodemask_of_node(&nodes_allowed, nid);
2868 n_mask = &nodes_allowed;
2871 err = set_max_huge_pages(h, count, nid, n_mask);
2873 return err ? err : len;
2876 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2877 struct kobject *kobj, const char *buf,
2881 unsigned long count;
2885 err = kstrtoul(buf, 10, &count);
2889 h = kobj_to_hstate(kobj, &nid);
2890 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2893 static ssize_t nr_hugepages_show(struct kobject *kobj,
2894 struct kobj_attribute *attr, char *buf)
2896 return nr_hugepages_show_common(kobj, attr, buf);
2899 static ssize_t nr_hugepages_store(struct kobject *kobj,
2900 struct kobj_attribute *attr, const char *buf, size_t len)
2902 return nr_hugepages_store_common(false, kobj, buf, len);
2904 HSTATE_ATTR(nr_hugepages);
2909 * hstate attribute for optionally mempolicy-based constraint on persistent
2910 * huge page alloc/free.
2912 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2913 struct kobj_attribute *attr, char *buf)
2915 return nr_hugepages_show_common(kobj, attr, buf);
2918 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2919 struct kobj_attribute *attr, const char *buf, size_t len)
2921 return nr_hugepages_store_common(true, kobj, buf, len);
2923 HSTATE_ATTR(nr_hugepages_mempolicy);
2927 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2928 struct kobj_attribute *attr, char *buf)
2930 struct hstate *h = kobj_to_hstate(kobj, NULL);
2931 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2934 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2935 struct kobj_attribute *attr, const char *buf, size_t count)
2938 unsigned long input;
2939 struct hstate *h = kobj_to_hstate(kobj, NULL);
2941 if (hstate_is_gigantic(h))
2944 err = kstrtoul(buf, 10, &input);
2948 spin_lock(&hugetlb_lock);
2949 h->nr_overcommit_huge_pages = input;
2950 spin_unlock(&hugetlb_lock);
2954 HSTATE_ATTR(nr_overcommit_hugepages);
2956 static ssize_t free_hugepages_show(struct kobject *kobj,
2957 struct kobj_attribute *attr, char *buf)
2960 unsigned long free_huge_pages;
2963 h = kobj_to_hstate(kobj, &nid);
2964 if (nid == NUMA_NO_NODE)
2965 free_huge_pages = h->free_huge_pages;
2967 free_huge_pages = h->free_huge_pages_node[nid];
2969 return sprintf(buf, "%lu\n", free_huge_pages);
2971 HSTATE_ATTR_RO(free_hugepages);
2973 static ssize_t resv_hugepages_show(struct kobject *kobj,
2974 struct kobj_attribute *attr, char *buf)
2976 struct hstate *h = kobj_to_hstate(kobj, NULL);
2977 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2979 HSTATE_ATTR_RO(resv_hugepages);
2981 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2982 struct kobj_attribute *attr, char *buf)
2985 unsigned long surplus_huge_pages;
2988 h = kobj_to_hstate(kobj, &nid);
2989 if (nid == NUMA_NO_NODE)
2990 surplus_huge_pages = h->surplus_huge_pages;
2992 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2994 return sprintf(buf, "%lu\n", surplus_huge_pages);
2996 HSTATE_ATTR_RO(surplus_hugepages);
2998 static struct attribute *hstate_attrs[] = {
2999 &nr_hugepages_attr.attr,
3000 &nr_overcommit_hugepages_attr.attr,
3001 &free_hugepages_attr.attr,
3002 &resv_hugepages_attr.attr,
3003 &surplus_hugepages_attr.attr,
3005 &nr_hugepages_mempolicy_attr.attr,
3010 static const struct attribute_group hstate_attr_group = {
3011 .attrs = hstate_attrs,
3014 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3015 struct kobject **hstate_kobjs,
3016 const struct attribute_group *hstate_attr_group)
3019 int hi = hstate_index(h);
3021 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3022 if (!hstate_kobjs[hi])
3025 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3027 kobject_put(hstate_kobjs[hi]);
3032 static void __init hugetlb_sysfs_init(void)
3037 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3038 if (!hugepages_kobj)
3041 for_each_hstate(h) {
3042 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3043 hstate_kobjs, &hstate_attr_group);
3045 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3052 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3053 * with node devices in node_devices[] using a parallel array. The array
3054 * index of a node device or _hstate == node id.
3055 * This is here to avoid any static dependency of the node device driver, in
3056 * the base kernel, on the hugetlb module.
3058 struct node_hstate {
3059 struct kobject *hugepages_kobj;
3060 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3062 static struct node_hstate node_hstates[MAX_NUMNODES];
3065 * A subset of global hstate attributes for node devices
3067 static struct attribute *per_node_hstate_attrs[] = {
3068 &nr_hugepages_attr.attr,
3069 &free_hugepages_attr.attr,
3070 &surplus_hugepages_attr.attr,
3074 static const struct attribute_group per_node_hstate_attr_group = {
3075 .attrs = per_node_hstate_attrs,
3079 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3080 * Returns node id via non-NULL nidp.
3082 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3086 for (nid = 0; nid < nr_node_ids; nid++) {
3087 struct node_hstate *nhs = &node_hstates[nid];
3089 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3090 if (nhs->hstate_kobjs[i] == kobj) {
3102 * Unregister hstate attributes from a single node device.
3103 * No-op if no hstate attributes attached.
3105 static void hugetlb_unregister_node(struct node *node)
3108 struct node_hstate *nhs = &node_hstates[node->dev.id];
3110 if (!nhs->hugepages_kobj)
3111 return; /* no hstate attributes */
3113 for_each_hstate(h) {
3114 int idx = hstate_index(h);
3115 if (nhs->hstate_kobjs[idx]) {
3116 kobject_put(nhs->hstate_kobjs[idx]);
3117 nhs->hstate_kobjs[idx] = NULL;
3121 kobject_put(nhs->hugepages_kobj);
3122 nhs->hugepages_kobj = NULL;
3127 * Register hstate attributes for a single node device.
3128 * No-op if attributes already registered.
3130 static void hugetlb_register_node(struct node *node)
3133 struct node_hstate *nhs = &node_hstates[node->dev.id];
3136 if (nhs->hugepages_kobj)
3137 return; /* already allocated */
3139 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3141 if (!nhs->hugepages_kobj)
3144 for_each_hstate(h) {
3145 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3147 &per_node_hstate_attr_group);
3149 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3150 h->name, node->dev.id);
3151 hugetlb_unregister_node(node);
3158 * hugetlb init time: register hstate attributes for all registered node
3159 * devices of nodes that have memory. All on-line nodes should have
3160 * registered their associated device by this time.
3162 static void __init hugetlb_register_all_nodes(void)
3166 for_each_node_state(nid, N_MEMORY) {
3167 struct node *node = node_devices[nid];
3168 if (node->dev.id == nid)
3169 hugetlb_register_node(node);
3173 * Let the node device driver know we're here so it can
3174 * [un]register hstate attributes on node hotplug.
3176 register_hugetlbfs_with_node(hugetlb_register_node,
3177 hugetlb_unregister_node);
3179 #else /* !CONFIG_NUMA */
3181 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3189 static void hugetlb_register_all_nodes(void) { }
3193 static int __init hugetlb_init(void)
3197 if (!hugepages_supported()) {
3198 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3199 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3204 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3205 * architectures depend on setup being done here.
3207 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3208 if (!parsed_default_hugepagesz) {
3210 * If we did not parse a default huge page size, set
3211 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3212 * number of huge pages for this default size was implicitly
3213 * specified, set that here as well.
3214 * Note that the implicit setting will overwrite an explicit
3215 * setting. A warning will be printed in this case.
3217 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3218 if (default_hstate_max_huge_pages) {
3219 if (default_hstate.max_huge_pages) {
3222 string_get_size(huge_page_size(&default_hstate),
3223 1, STRING_UNITS_2, buf, 32);
3224 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3225 default_hstate.max_huge_pages, buf);
3226 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3227 default_hstate_max_huge_pages);
3229 default_hstate.max_huge_pages =
3230 default_hstate_max_huge_pages;
3234 hugetlb_cma_check();
3235 hugetlb_init_hstates();
3236 gather_bootmem_prealloc();
3239 hugetlb_sysfs_init();
3240 hugetlb_register_all_nodes();
3241 hugetlb_cgroup_file_init();
3244 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3246 num_fault_mutexes = 1;
3248 hugetlb_fault_mutex_table =
3249 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3251 BUG_ON(!hugetlb_fault_mutex_table);
3253 for (i = 0; i < num_fault_mutexes; i++)
3254 mutex_init(&hugetlb_fault_mutex_table[i]);
3257 subsys_initcall(hugetlb_init);
3259 /* Overwritten by architectures with more huge page sizes */
3260 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3262 return size == HPAGE_SIZE;
3265 void __init hugetlb_add_hstate(unsigned int order)
3270 if (size_to_hstate(PAGE_SIZE << order)) {
3273 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3275 h = &hstates[hugetlb_max_hstate++];
3277 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3278 h->nr_huge_pages = 0;
3279 h->free_huge_pages = 0;
3280 for (i = 0; i < MAX_NUMNODES; ++i)
3281 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3282 INIT_LIST_HEAD(&h->hugepage_activelist);
3283 h->next_nid_to_alloc = first_memory_node;
3284 h->next_nid_to_free = first_memory_node;
3285 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3286 huge_page_size(h)/1024);
3292 * hugepages command line processing
3293 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3294 * specification. If not, ignore the hugepages value. hugepages can also
3295 * be the first huge page command line option in which case it implicitly
3296 * specifies the number of huge pages for the default size.
3298 static int __init hugepages_setup(char *s)
3301 static unsigned long *last_mhp;
3303 if (!parsed_valid_hugepagesz) {
3304 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3305 parsed_valid_hugepagesz = true;
3310 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3311 * yet, so this hugepages= parameter goes to the "default hstate".
3312 * Otherwise, it goes with the previously parsed hugepagesz or
3313 * default_hugepagesz.
3315 else if (!hugetlb_max_hstate)
3316 mhp = &default_hstate_max_huge_pages;
3318 mhp = &parsed_hstate->max_huge_pages;
3320 if (mhp == last_mhp) {
3321 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3325 if (sscanf(s, "%lu", mhp) <= 0)
3329 * Global state is always initialized later in hugetlb_init.
3330 * But we need to allocate >= MAX_ORDER hstates here early to still
3331 * use the bootmem allocator.
3333 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3334 hugetlb_hstate_alloc_pages(parsed_hstate);
3340 __setup("hugepages=", hugepages_setup);
3343 * hugepagesz command line processing
3344 * A specific huge page size can only be specified once with hugepagesz.
3345 * hugepagesz is followed by hugepages on the command line. The global
3346 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3347 * hugepagesz argument was valid.
3349 static int __init hugepagesz_setup(char *s)
3354 parsed_valid_hugepagesz = false;
3355 size = (unsigned long)memparse(s, NULL);
3357 if (!arch_hugetlb_valid_size(size)) {
3358 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3362 h = size_to_hstate(size);
3365 * hstate for this size already exists. This is normally
3366 * an error, but is allowed if the existing hstate is the
3367 * default hstate. More specifically, it is only allowed if
3368 * the number of huge pages for the default hstate was not
3369 * previously specified.
3371 if (!parsed_default_hugepagesz || h != &default_hstate ||
3372 default_hstate.max_huge_pages) {
3373 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3378 * No need to call hugetlb_add_hstate() as hstate already
3379 * exists. But, do set parsed_hstate so that a following
3380 * hugepages= parameter will be applied to this hstate.
3383 parsed_valid_hugepagesz = true;
3387 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3388 parsed_valid_hugepagesz = true;
3391 __setup("hugepagesz=", hugepagesz_setup);
3394 * default_hugepagesz command line input
3395 * Only one instance of default_hugepagesz allowed on command line.
3397 static int __init default_hugepagesz_setup(char *s)
3401 parsed_valid_hugepagesz = false;
3402 if (parsed_default_hugepagesz) {
3403 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3407 size = (unsigned long)memparse(s, NULL);
3409 if (!arch_hugetlb_valid_size(size)) {
3410 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3414 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3415 parsed_valid_hugepagesz = true;
3416 parsed_default_hugepagesz = true;
3417 default_hstate_idx = hstate_index(size_to_hstate(size));
3420 * The number of default huge pages (for this size) could have been
3421 * specified as the first hugetlb parameter: hugepages=X. If so,
3422 * then default_hstate_max_huge_pages is set. If the default huge
3423 * page size is gigantic (>= MAX_ORDER), then the pages must be
3424 * allocated here from bootmem allocator.
3426 if (default_hstate_max_huge_pages) {
3427 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3428 if (hstate_is_gigantic(&default_hstate))
3429 hugetlb_hstate_alloc_pages(&default_hstate);
3430 default_hstate_max_huge_pages = 0;
3435 __setup("default_hugepagesz=", default_hugepagesz_setup);
3437 static unsigned int allowed_mems_nr(struct hstate *h)
3440 unsigned int nr = 0;
3441 nodemask_t *mpol_allowed;
3442 unsigned int *array = h->free_huge_pages_node;
3443 gfp_t gfp_mask = htlb_alloc_mask(h);
3445 mpol_allowed = policy_nodemask_current(gfp_mask);
3447 for_each_node_mask(node, cpuset_current_mems_allowed) {
3448 if (!mpol_allowed ||
3449 (mpol_allowed && node_isset(node, *mpol_allowed)))
3456 #ifdef CONFIG_SYSCTL
3457 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3458 void *buffer, size_t *length,
3459 loff_t *ppos, unsigned long *out)
3461 struct ctl_table dup_table;
3464 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3465 * can duplicate the @table and alter the duplicate of it.
3468 dup_table.data = out;
3470 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3473 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3474 struct ctl_table *table, int write,
3475 void *buffer, size_t *length, loff_t *ppos)
3477 struct hstate *h = &default_hstate;
3478 unsigned long tmp = h->max_huge_pages;
3481 if (!hugepages_supported())
3484 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3490 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3491 NUMA_NO_NODE, tmp, *length);
3496 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3497 void *buffer, size_t *length, loff_t *ppos)
3500 return hugetlb_sysctl_handler_common(false, table, write,
3501 buffer, length, ppos);
3505 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3506 void *buffer, size_t *length, loff_t *ppos)
3508 return hugetlb_sysctl_handler_common(true, table, write,
3509 buffer, length, ppos);
3511 #endif /* CONFIG_NUMA */
3513 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3514 void *buffer, size_t *length, loff_t *ppos)
3516 struct hstate *h = &default_hstate;
3520 if (!hugepages_supported())
3523 tmp = h->nr_overcommit_huge_pages;
3525 if (write && hstate_is_gigantic(h))
3528 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3534 spin_lock(&hugetlb_lock);
3535 h->nr_overcommit_huge_pages = tmp;
3536 spin_unlock(&hugetlb_lock);
3542 #endif /* CONFIG_SYSCTL */
3544 void hugetlb_report_meminfo(struct seq_file *m)
3547 unsigned long total = 0;
3549 if (!hugepages_supported())
3552 for_each_hstate(h) {
3553 unsigned long count = h->nr_huge_pages;
3555 total += (PAGE_SIZE << huge_page_order(h)) * count;
3557 if (h == &default_hstate)
3559 "HugePages_Total: %5lu\n"
3560 "HugePages_Free: %5lu\n"
3561 "HugePages_Rsvd: %5lu\n"
3562 "HugePages_Surp: %5lu\n"
3563 "Hugepagesize: %8lu kB\n",
3567 h->surplus_huge_pages,
3568 (PAGE_SIZE << huge_page_order(h)) / 1024);
3571 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3574 int hugetlb_report_node_meminfo(int nid, char *buf)
3576 struct hstate *h = &default_hstate;
3577 if (!hugepages_supported())
3580 "Node %d HugePages_Total: %5u\n"
3581 "Node %d HugePages_Free: %5u\n"
3582 "Node %d HugePages_Surp: %5u\n",
3583 nid, h->nr_huge_pages_node[nid],
3584 nid, h->free_huge_pages_node[nid],
3585 nid, h->surplus_huge_pages_node[nid]);
3588 void hugetlb_show_meminfo(void)
3593 if (!hugepages_supported())
3596 for_each_node_state(nid, N_MEMORY)
3598 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3600 h->nr_huge_pages_node[nid],
3601 h->free_huge_pages_node[nid],
3602 h->surplus_huge_pages_node[nid],
3603 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3606 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3608 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3609 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3612 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3613 unsigned long hugetlb_total_pages(void)
3616 unsigned long nr_total_pages = 0;
3619 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3620 return nr_total_pages;
3623 static int hugetlb_acct_memory(struct hstate *h, long delta)
3627 spin_lock(&hugetlb_lock);
3629 * When cpuset is configured, it breaks the strict hugetlb page
3630 * reservation as the accounting is done on a global variable. Such
3631 * reservation is completely rubbish in the presence of cpuset because
3632 * the reservation is not checked against page availability for the
3633 * current cpuset. Application can still potentially OOM'ed by kernel
3634 * with lack of free htlb page in cpuset that the task is in.
3635 * Attempt to enforce strict accounting with cpuset is almost
3636 * impossible (or too ugly) because cpuset is too fluid that
3637 * task or memory node can be dynamically moved between cpusets.
3639 * The change of semantics for shared hugetlb mapping with cpuset is
3640 * undesirable. However, in order to preserve some of the semantics,
3641 * we fall back to check against current free page availability as
3642 * a best attempt and hopefully to minimize the impact of changing
3643 * semantics that cpuset has.
3645 * Apart from cpuset, we also have memory policy mechanism that
3646 * also determines from which node the kernel will allocate memory
3647 * in a NUMA system. So similar to cpuset, we also should consider
3648 * the memory policy of the current task. Similar to the description
3652 if (gather_surplus_pages(h, delta) < 0)
3655 if (delta > allowed_mems_nr(h)) {
3656 return_unused_surplus_pages(h, delta);
3663 return_unused_surplus_pages(h, (unsigned long) -delta);
3666 spin_unlock(&hugetlb_lock);
3670 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3672 struct resv_map *resv = vma_resv_map(vma);
3675 * This new VMA should share its siblings reservation map if present.
3676 * The VMA will only ever have a valid reservation map pointer where
3677 * it is being copied for another still existing VMA. As that VMA
3678 * has a reference to the reservation map it cannot disappear until
3679 * after this open call completes. It is therefore safe to take a
3680 * new reference here without additional locking.
3682 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3683 kref_get(&resv->refs);
3686 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3688 struct hstate *h = hstate_vma(vma);
3689 struct resv_map *resv = vma_resv_map(vma);
3690 struct hugepage_subpool *spool = subpool_vma(vma);
3691 unsigned long reserve, start, end;
3694 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3697 start = vma_hugecache_offset(h, vma, vma->vm_start);
3698 end = vma_hugecache_offset(h, vma, vma->vm_end);
3700 reserve = (end - start) - region_count(resv, start, end);
3701 hugetlb_cgroup_uncharge_counter(resv, start, end);
3704 * Decrement reserve counts. The global reserve count may be
3705 * adjusted if the subpool has a minimum size.
3707 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3708 hugetlb_acct_memory(h, -gbl_reserve);
3711 kref_put(&resv->refs, resv_map_release);
3714 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3716 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3721 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3723 struct hstate *hstate = hstate_vma(vma);
3725 return 1UL << huge_page_shift(hstate);
3729 * We cannot handle pagefaults against hugetlb pages at all. They cause
3730 * handle_mm_fault() to try to instantiate regular-sized pages in the
3731 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3734 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3741 * When a new function is introduced to vm_operations_struct and added
3742 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3743 * This is because under System V memory model, mappings created via
3744 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3745 * their original vm_ops are overwritten with shm_vm_ops.
3747 const struct vm_operations_struct hugetlb_vm_ops = {
3748 .fault = hugetlb_vm_op_fault,
3749 .open = hugetlb_vm_op_open,
3750 .close = hugetlb_vm_op_close,
3751 .split = hugetlb_vm_op_split,
3752 .pagesize = hugetlb_vm_op_pagesize,
3755 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3761 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3762 vma->vm_page_prot)));
3764 entry = huge_pte_wrprotect(mk_huge_pte(page,
3765 vma->vm_page_prot));
3767 entry = pte_mkyoung(entry);
3768 entry = pte_mkhuge(entry);
3769 entry = arch_make_huge_pte(entry, vma, page, writable);
3774 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3775 unsigned long address, pte_t *ptep)
3779 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3780 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3781 update_mmu_cache(vma, address, ptep);
3784 bool is_hugetlb_entry_migration(pte_t pte)
3788 if (huge_pte_none(pte) || pte_present(pte))
3790 swp = pte_to_swp_entry(pte);
3791 if (is_migration_entry(swp))
3797 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3801 if (huge_pte_none(pte) || pte_present(pte))
3803 swp = pte_to_swp_entry(pte);
3804 if (is_hwpoison_entry(swp))
3810 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3811 struct vm_area_struct *vma)
3813 pte_t *src_pte, *dst_pte, entry, dst_entry;
3814 struct page *ptepage;
3817 struct hstate *h = hstate_vma(vma);
3818 unsigned long sz = huge_page_size(h);
3819 struct address_space *mapping = vma->vm_file->f_mapping;
3820 struct mmu_notifier_range range;
3823 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3826 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3829 mmu_notifier_invalidate_range_start(&range);
3832 * For shared mappings i_mmap_rwsem must be held to call
3833 * huge_pte_alloc, otherwise the returned ptep could go
3834 * away if part of a shared pmd and another thread calls
3837 i_mmap_lock_read(mapping);
3840 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3841 spinlock_t *src_ptl, *dst_ptl;
3842 src_pte = huge_pte_offset(src, addr, sz);
3845 dst_pte = huge_pte_alloc(dst, addr, sz);
3852 * If the pagetables are shared don't copy or take references.
3853 * dst_pte == src_pte is the common case of src/dest sharing.
3855 * However, src could have 'unshared' and dst shares with
3856 * another vma. If dst_pte !none, this implies sharing.
3857 * Check here before taking page table lock, and once again
3858 * after taking the lock below.
3860 dst_entry = huge_ptep_get(dst_pte);
3861 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3864 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3865 src_ptl = huge_pte_lockptr(h, src, src_pte);
3866 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3867 entry = huge_ptep_get(src_pte);
3868 dst_entry = huge_ptep_get(dst_pte);
3869 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3871 * Skip if src entry none. Also, skip in the
3872 * unlikely case dst entry !none as this implies
3873 * sharing with another vma.
3876 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3877 is_hugetlb_entry_hwpoisoned(entry))) {
3878 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3880 if (is_write_migration_entry(swp_entry) && cow) {
3882 * COW mappings require pages in both
3883 * parent and child to be set to read.
3885 make_migration_entry_read(&swp_entry);
3886 entry = swp_entry_to_pte(swp_entry);
3887 set_huge_swap_pte_at(src, addr, src_pte,
3890 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3894 * No need to notify as we are downgrading page
3895 * table protection not changing it to point
3898 * See Documentation/vm/mmu_notifier.rst
3900 huge_ptep_set_wrprotect(src, addr, src_pte);
3902 entry = huge_ptep_get(src_pte);
3903 ptepage = pte_page(entry);
3905 page_dup_rmap(ptepage, true);
3906 set_huge_pte_at(dst, addr, dst_pte, entry);
3907 hugetlb_count_add(pages_per_huge_page(h), dst);
3909 spin_unlock(src_ptl);
3910 spin_unlock(dst_ptl);
3914 mmu_notifier_invalidate_range_end(&range);
3916 i_mmap_unlock_read(mapping);
3921 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3922 unsigned long start, unsigned long end,
3923 struct page *ref_page)
3925 struct mm_struct *mm = vma->vm_mm;
3926 unsigned long address;
3931 struct hstate *h = hstate_vma(vma);
3932 unsigned long sz = huge_page_size(h);
3933 struct mmu_notifier_range range;
3935 WARN_ON(!is_vm_hugetlb_page(vma));
3936 BUG_ON(start & ~huge_page_mask(h));
3937 BUG_ON(end & ~huge_page_mask(h));
3940 * This is a hugetlb vma, all the pte entries should point
3943 tlb_change_page_size(tlb, sz);
3944 tlb_start_vma(tlb, vma);
3947 * If sharing possible, alert mmu notifiers of worst case.
3949 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3951 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3952 mmu_notifier_invalidate_range_start(&range);
3954 for (; address < end; address += sz) {
3955 ptep = huge_pte_offset(mm, address, sz);
3959 ptl = huge_pte_lock(h, mm, ptep);
3960 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3963 * We just unmapped a page of PMDs by clearing a PUD.
3964 * The caller's TLB flush range should cover this area.
3969 pte = huge_ptep_get(ptep);
3970 if (huge_pte_none(pte)) {
3976 * Migrating hugepage or HWPoisoned hugepage is already
3977 * unmapped and its refcount is dropped, so just clear pte here.
3979 if (unlikely(!pte_present(pte))) {
3980 huge_pte_clear(mm, address, ptep, sz);
3985 page = pte_page(pte);
3987 * If a reference page is supplied, it is because a specific
3988 * page is being unmapped, not a range. Ensure the page we
3989 * are about to unmap is the actual page of interest.
3992 if (page != ref_page) {
3997 * Mark the VMA as having unmapped its page so that
3998 * future faults in this VMA will fail rather than
3999 * looking like data was lost
4001 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4004 pte = huge_ptep_get_and_clear(mm, address, ptep);
4005 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4006 if (huge_pte_dirty(pte))
4007 set_page_dirty(page);
4009 hugetlb_count_sub(pages_per_huge_page(h), mm);
4010 page_remove_rmap(page, true);
4013 tlb_remove_page_size(tlb, page, huge_page_size(h));
4015 * Bail out after unmapping reference page if supplied
4020 mmu_notifier_invalidate_range_end(&range);
4021 tlb_end_vma(tlb, vma);
4024 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4025 struct vm_area_struct *vma, unsigned long start,
4026 unsigned long end, struct page *ref_page)
4028 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4031 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4032 * test will fail on a vma being torn down, and not grab a page table
4033 * on its way out. We're lucky that the flag has such an appropriate
4034 * name, and can in fact be safely cleared here. We could clear it
4035 * before the __unmap_hugepage_range above, but all that's necessary
4036 * is to clear it before releasing the i_mmap_rwsem. This works
4037 * because in the context this is called, the VMA is about to be
4038 * destroyed and the i_mmap_rwsem is held.
4040 vma->vm_flags &= ~VM_MAYSHARE;
4043 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4044 unsigned long end, struct page *ref_page)
4046 struct mm_struct *mm;
4047 struct mmu_gather tlb;
4048 unsigned long tlb_start = start;
4049 unsigned long tlb_end = end;
4052 * If shared PMDs were possibly used within this vma range, adjust
4053 * start/end for worst case tlb flushing.
4054 * Note that we can not be sure if PMDs are shared until we try to
4055 * unmap pages. However, we want to make sure TLB flushing covers
4056 * the largest possible range.
4058 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4062 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4063 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4064 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4068 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4069 * mappping it owns the reserve page for. The intention is to unmap the page
4070 * from other VMAs and let the children be SIGKILLed if they are faulting the
4073 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4074 struct page *page, unsigned long address)
4076 struct hstate *h = hstate_vma(vma);
4077 struct vm_area_struct *iter_vma;
4078 struct address_space *mapping;
4082 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4083 * from page cache lookup which is in HPAGE_SIZE units.
4085 address = address & huge_page_mask(h);
4086 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4088 mapping = vma->vm_file->f_mapping;
4091 * Take the mapping lock for the duration of the table walk. As
4092 * this mapping should be shared between all the VMAs,
4093 * __unmap_hugepage_range() is called as the lock is already held
4095 i_mmap_lock_write(mapping);
4096 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4097 /* Do not unmap the current VMA */
4098 if (iter_vma == vma)
4102 * Shared VMAs have their own reserves and do not affect
4103 * MAP_PRIVATE accounting but it is possible that a shared
4104 * VMA is using the same page so check and skip such VMAs.
4106 if (iter_vma->vm_flags & VM_MAYSHARE)
4110 * Unmap the page from other VMAs without their own reserves.
4111 * They get marked to be SIGKILLed if they fault in these
4112 * areas. This is because a future no-page fault on this VMA
4113 * could insert a zeroed page instead of the data existing
4114 * from the time of fork. This would look like data corruption
4116 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4117 unmap_hugepage_range(iter_vma, address,
4118 address + huge_page_size(h), page);
4120 i_mmap_unlock_write(mapping);
4124 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4125 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4126 * cannot race with other handlers or page migration.
4127 * Keep the pte_same checks anyway to make transition from the mutex easier.
4129 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4130 unsigned long address, pte_t *ptep,
4131 struct page *pagecache_page, spinlock_t *ptl)
4134 struct hstate *h = hstate_vma(vma);
4135 struct page *old_page, *new_page;
4136 int outside_reserve = 0;
4138 unsigned long haddr = address & huge_page_mask(h);
4139 struct mmu_notifier_range range;
4141 pte = huge_ptep_get(ptep);
4142 old_page = pte_page(pte);
4145 /* If no-one else is actually using this page, avoid the copy
4146 * and just make the page writable */
4147 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4148 page_move_anon_rmap(old_page, vma);
4149 set_huge_ptep_writable(vma, haddr, ptep);
4154 * If the process that created a MAP_PRIVATE mapping is about to
4155 * perform a COW due to a shared page count, attempt to satisfy
4156 * the allocation without using the existing reserves. The pagecache
4157 * page is used to determine if the reserve at this address was
4158 * consumed or not. If reserves were used, a partial faulted mapping
4159 * at the time of fork() could consume its reserves on COW instead
4160 * of the full address range.
4162 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4163 old_page != pagecache_page)
4164 outside_reserve = 1;
4169 * Drop page table lock as buddy allocator may be called. It will
4170 * be acquired again before returning to the caller, as expected.
4173 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4175 if (IS_ERR(new_page)) {
4177 * If a process owning a MAP_PRIVATE mapping fails to COW,
4178 * it is due to references held by a child and an insufficient
4179 * huge page pool. To guarantee the original mappers
4180 * reliability, unmap the page from child processes. The child
4181 * may get SIGKILLed if it later faults.
4183 if (outside_reserve) {
4185 BUG_ON(huge_pte_none(pte));
4186 unmap_ref_private(mm, vma, old_page, haddr);
4187 BUG_ON(huge_pte_none(pte));
4189 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4191 pte_same(huge_ptep_get(ptep), pte)))
4192 goto retry_avoidcopy;
4194 * race occurs while re-acquiring page table
4195 * lock, and our job is done.
4200 ret = vmf_error(PTR_ERR(new_page));
4201 goto out_release_old;
4205 * When the original hugepage is shared one, it does not have
4206 * anon_vma prepared.
4208 if (unlikely(anon_vma_prepare(vma))) {
4210 goto out_release_all;
4213 copy_user_huge_page(new_page, old_page, address, vma,
4214 pages_per_huge_page(h));
4215 __SetPageUptodate(new_page);
4217 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4218 haddr + huge_page_size(h));
4219 mmu_notifier_invalidate_range_start(&range);
4222 * Retake the page table lock to check for racing updates
4223 * before the page tables are altered
4226 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4227 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4228 ClearPagePrivate(new_page);
4231 huge_ptep_clear_flush(vma, haddr, ptep);
4232 mmu_notifier_invalidate_range(mm, range.start, range.end);
4233 set_huge_pte_at(mm, haddr, ptep,
4234 make_huge_pte(vma, new_page, 1));
4235 page_remove_rmap(old_page, true);
4236 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4237 set_page_huge_active(new_page);
4238 /* Make the old page be freed below */
4239 new_page = old_page;
4242 mmu_notifier_invalidate_range_end(&range);
4244 restore_reserve_on_error(h, vma, haddr, new_page);
4249 spin_lock(ptl); /* Caller expects lock to be held */
4253 /* Return the pagecache page at a given address within a VMA */
4254 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4255 struct vm_area_struct *vma, unsigned long address)
4257 struct address_space *mapping;
4260 mapping = vma->vm_file->f_mapping;
4261 idx = vma_hugecache_offset(h, vma, address);
4263 return find_lock_page(mapping, idx);
4267 * Return whether there is a pagecache page to back given address within VMA.
4268 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4270 static bool hugetlbfs_pagecache_present(struct hstate *h,
4271 struct vm_area_struct *vma, unsigned long address)
4273 struct address_space *mapping;
4277 mapping = vma->vm_file->f_mapping;
4278 idx = vma_hugecache_offset(h, vma, address);
4280 page = find_get_page(mapping, idx);
4283 return page != NULL;
4286 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4289 struct inode *inode = mapping->host;
4290 struct hstate *h = hstate_inode(inode);
4291 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4295 ClearPagePrivate(page);
4298 * set page dirty so that it will not be removed from cache/file
4299 * by non-hugetlbfs specific code paths.
4301 set_page_dirty(page);
4303 spin_lock(&inode->i_lock);
4304 inode->i_blocks += blocks_per_huge_page(h);
4305 spin_unlock(&inode->i_lock);
4309 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4310 struct vm_area_struct *vma,
4311 struct address_space *mapping, pgoff_t idx,
4312 unsigned long address, pte_t *ptep, unsigned int flags)
4314 struct hstate *h = hstate_vma(vma);
4315 vm_fault_t ret = VM_FAULT_SIGBUS;
4321 unsigned long haddr = address & huge_page_mask(h);
4322 bool new_page = false;
4325 * Currently, we are forced to kill the process in the event the
4326 * original mapper has unmapped pages from the child due to a failed
4327 * COW. Warn that such a situation has occurred as it may not be obvious
4329 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4330 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4336 * We can not race with truncation due to holding i_mmap_rwsem.
4337 * i_size is modified when holding i_mmap_rwsem, so check here
4338 * once for faults beyond end of file.
4340 size = i_size_read(mapping->host) >> huge_page_shift(h);
4345 page = find_lock_page(mapping, idx);
4348 * Check for page in userfault range
4350 if (userfaultfd_missing(vma)) {
4352 struct vm_fault vmf = {
4357 * Hard to debug if it ends up being
4358 * used by a callee that assumes
4359 * something about the other
4360 * uninitialized fields... same as in
4366 * hugetlb_fault_mutex and i_mmap_rwsem must be
4367 * dropped before handling userfault. Reacquire
4368 * after handling fault to make calling code simpler.
4370 hash = hugetlb_fault_mutex_hash(mapping, idx);
4371 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4372 i_mmap_unlock_read(mapping);
4373 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4374 i_mmap_lock_read(mapping);
4375 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4379 page = alloc_huge_page(vma, haddr, 0);
4382 * Returning error will result in faulting task being
4383 * sent SIGBUS. The hugetlb fault mutex prevents two
4384 * tasks from racing to fault in the same page which
4385 * could result in false unable to allocate errors.
4386 * Page migration does not take the fault mutex, but
4387 * does a clear then write of pte's under page table
4388 * lock. Page fault code could race with migration,
4389 * notice the clear pte and try to allocate a page
4390 * here. Before returning error, get ptl and make
4391 * sure there really is no pte entry.
4393 ptl = huge_pte_lock(h, mm, ptep);
4394 if (!huge_pte_none(huge_ptep_get(ptep))) {
4400 ret = vmf_error(PTR_ERR(page));
4403 clear_huge_page(page, address, pages_per_huge_page(h));
4404 __SetPageUptodate(page);
4407 if (vma->vm_flags & VM_MAYSHARE) {
4408 int err = huge_add_to_page_cache(page, mapping, idx);
4417 if (unlikely(anon_vma_prepare(vma))) {
4419 goto backout_unlocked;
4425 * If memory error occurs between mmap() and fault, some process
4426 * don't have hwpoisoned swap entry for errored virtual address.
4427 * So we need to block hugepage fault by PG_hwpoison bit check.
4429 if (unlikely(PageHWPoison(page))) {
4430 ret = VM_FAULT_HWPOISON |
4431 VM_FAULT_SET_HINDEX(hstate_index(h));
4432 goto backout_unlocked;
4437 * If we are going to COW a private mapping later, we examine the
4438 * pending reservations for this page now. This will ensure that
4439 * any allocations necessary to record that reservation occur outside
4442 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4443 if (vma_needs_reservation(h, vma, haddr) < 0) {
4445 goto backout_unlocked;
4447 /* Just decrements count, does not deallocate */
4448 vma_end_reservation(h, vma, haddr);
4451 ptl = huge_pte_lock(h, mm, ptep);
4453 if (!huge_pte_none(huge_ptep_get(ptep)))
4457 ClearPagePrivate(page);
4458 hugepage_add_new_anon_rmap(page, vma, haddr);
4460 page_dup_rmap(page, true);
4461 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4462 && (vma->vm_flags & VM_SHARED)));
4463 set_huge_pte_at(mm, haddr, ptep, new_pte);
4465 hugetlb_count_add(pages_per_huge_page(h), mm);
4466 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4467 /* Optimization, do the COW without a second fault */
4468 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4474 * Only make newly allocated pages active. Existing pages found
4475 * in the pagecache could be !page_huge_active() if they have been
4476 * isolated for migration.
4479 set_page_huge_active(page);
4489 restore_reserve_on_error(h, vma, haddr, page);
4495 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4497 unsigned long key[2];
4500 key[0] = (unsigned long) mapping;
4503 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4505 return hash & (num_fault_mutexes - 1);
4509 * For uniprocesor systems we always use a single mutex, so just
4510 * return 0 and avoid the hashing overhead.
4512 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4518 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4519 unsigned long address, unsigned int flags)
4526 struct page *page = NULL;
4527 struct page *pagecache_page = NULL;
4528 struct hstate *h = hstate_vma(vma);
4529 struct address_space *mapping;
4530 int need_wait_lock = 0;
4531 unsigned long haddr = address & huge_page_mask(h);
4533 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4536 * Since we hold no locks, ptep could be stale. That is
4537 * OK as we are only making decisions based on content and
4538 * not actually modifying content here.
4540 entry = huge_ptep_get(ptep);
4541 if (unlikely(is_hugetlb_entry_migration(entry))) {
4542 migration_entry_wait_huge(vma, mm, ptep);
4544 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4545 return VM_FAULT_HWPOISON_LARGE |
4546 VM_FAULT_SET_HINDEX(hstate_index(h));
4550 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4551 * until finished with ptep. This serves two purposes:
4552 * 1) It prevents huge_pmd_unshare from being called elsewhere
4553 * and making the ptep no longer valid.
4554 * 2) It synchronizes us with i_size modifications during truncation.
4556 * ptep could have already be assigned via huge_pte_offset. That
4557 * is OK, as huge_pte_alloc will return the same value unless
4558 * something has changed.
4560 mapping = vma->vm_file->f_mapping;
4561 i_mmap_lock_read(mapping);
4562 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4564 i_mmap_unlock_read(mapping);
4565 return VM_FAULT_OOM;
4569 * Serialize hugepage allocation and instantiation, so that we don't
4570 * get spurious allocation failures if two CPUs race to instantiate
4571 * the same page in the page cache.
4573 idx = vma_hugecache_offset(h, vma, haddr);
4574 hash = hugetlb_fault_mutex_hash(mapping, idx);
4575 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4577 entry = huge_ptep_get(ptep);
4578 if (huge_pte_none(entry)) {
4579 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4586 * entry could be a migration/hwpoison entry at this point, so this
4587 * check prevents the kernel from going below assuming that we have
4588 * an active hugepage in pagecache. This goto expects the 2nd page
4589 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4590 * properly handle it.
4592 if (!pte_present(entry))
4596 * If we are going to COW the mapping later, we examine the pending
4597 * reservations for this page now. This will ensure that any
4598 * allocations necessary to record that reservation occur outside the
4599 * spinlock. For private mappings, we also lookup the pagecache
4600 * page now as it is used to determine if a reservation has been
4603 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4604 if (vma_needs_reservation(h, vma, haddr) < 0) {
4608 /* Just decrements count, does not deallocate */
4609 vma_end_reservation(h, vma, haddr);
4611 if (!(vma->vm_flags & VM_MAYSHARE))
4612 pagecache_page = hugetlbfs_pagecache_page(h,
4616 ptl = huge_pte_lock(h, mm, ptep);
4618 /* Check for a racing update before calling hugetlb_cow */
4619 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4623 * hugetlb_cow() requires page locks of pte_page(entry) and
4624 * pagecache_page, so here we need take the former one
4625 * when page != pagecache_page or !pagecache_page.
4627 page = pte_page(entry);
4628 if (page != pagecache_page)
4629 if (!trylock_page(page)) {
4636 if (flags & FAULT_FLAG_WRITE) {
4637 if (!huge_pte_write(entry)) {
4638 ret = hugetlb_cow(mm, vma, address, ptep,
4639 pagecache_page, ptl);
4642 entry = huge_pte_mkdirty(entry);
4644 entry = pte_mkyoung(entry);
4645 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4646 flags & FAULT_FLAG_WRITE))
4647 update_mmu_cache(vma, haddr, ptep);
4649 if (page != pagecache_page)
4655 if (pagecache_page) {
4656 unlock_page(pagecache_page);
4657 put_page(pagecache_page);
4660 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4661 i_mmap_unlock_read(mapping);
4663 * Generally it's safe to hold refcount during waiting page lock. But
4664 * here we just wait to defer the next page fault to avoid busy loop and
4665 * the page is not used after unlocked before returning from the current
4666 * page fault. So we are safe from accessing freed page, even if we wait
4667 * here without taking refcount.
4670 wait_on_page_locked(page);
4675 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4676 * modifications for huge pages.
4678 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4680 struct vm_area_struct *dst_vma,
4681 unsigned long dst_addr,
4682 unsigned long src_addr,
4683 struct page **pagep)
4685 struct address_space *mapping;
4688 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4689 struct hstate *h = hstate_vma(dst_vma);
4697 page = alloc_huge_page(dst_vma, dst_addr, 0);
4701 ret = copy_huge_page_from_user(page,
4702 (const void __user *) src_addr,
4703 pages_per_huge_page(h), false);
4705 /* fallback to copy_from_user outside mmap_lock */
4706 if (unlikely(ret)) {
4709 /* don't free the page */
4718 * The memory barrier inside __SetPageUptodate makes sure that
4719 * preceding stores to the page contents become visible before
4720 * the set_pte_at() write.
4722 __SetPageUptodate(page);
4724 mapping = dst_vma->vm_file->f_mapping;
4725 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4728 * If shared, add to page cache
4731 size = i_size_read(mapping->host) >> huge_page_shift(h);
4734 goto out_release_nounlock;
4737 * Serialization between remove_inode_hugepages() and
4738 * huge_add_to_page_cache() below happens through the
4739 * hugetlb_fault_mutex_table that here must be hold by
4742 ret = huge_add_to_page_cache(page, mapping, idx);
4744 goto out_release_nounlock;
4747 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4751 * Recheck the i_size after holding PT lock to make sure not
4752 * to leave any page mapped (as page_mapped()) beyond the end
4753 * of the i_size (remove_inode_hugepages() is strict about
4754 * enforcing that). If we bail out here, we'll also leave a
4755 * page in the radix tree in the vm_shared case beyond the end
4756 * of the i_size, but remove_inode_hugepages() will take care
4757 * of it as soon as we drop the hugetlb_fault_mutex_table.
4759 size = i_size_read(mapping->host) >> huge_page_shift(h);
4762 goto out_release_unlock;
4765 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4766 goto out_release_unlock;
4769 page_dup_rmap(page, true);
4771 ClearPagePrivate(page);
4772 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4775 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4776 if (dst_vma->vm_flags & VM_WRITE)
4777 _dst_pte = huge_pte_mkdirty(_dst_pte);
4778 _dst_pte = pte_mkyoung(_dst_pte);
4780 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4782 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4783 dst_vma->vm_flags & VM_WRITE);
4784 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4786 /* No need to invalidate - it was non-present before */
4787 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4790 set_page_huge_active(page);
4800 out_release_nounlock:
4805 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4806 struct page **pages, struct vm_area_struct **vmas,
4807 unsigned long *position, unsigned long *nr_pages,
4808 long i, unsigned int flags, int *locked)
4810 unsigned long pfn_offset;
4811 unsigned long vaddr = *position;
4812 unsigned long remainder = *nr_pages;
4813 struct hstate *h = hstate_vma(vma);
4816 while (vaddr < vma->vm_end && remainder) {
4818 spinlock_t *ptl = NULL;
4823 * If we have a pending SIGKILL, don't keep faulting pages and
4824 * potentially allocating memory.
4826 if (fatal_signal_pending(current)) {
4832 * Some archs (sparc64, sh*) have multiple pte_ts to
4833 * each hugepage. We have to make sure we get the
4834 * first, for the page indexing below to work.
4836 * Note that page table lock is not held when pte is null.
4838 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4841 ptl = huge_pte_lock(h, mm, pte);
4842 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4845 * When coredumping, it suits get_dump_page if we just return
4846 * an error where there's an empty slot with no huge pagecache
4847 * to back it. This way, we avoid allocating a hugepage, and
4848 * the sparse dumpfile avoids allocating disk blocks, but its
4849 * huge holes still show up with zeroes where they need to be.
4851 if (absent && (flags & FOLL_DUMP) &&
4852 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4860 * We need call hugetlb_fault for both hugepages under migration
4861 * (in which case hugetlb_fault waits for the migration,) and
4862 * hwpoisoned hugepages (in which case we need to prevent the
4863 * caller from accessing to them.) In order to do this, we use
4864 * here is_swap_pte instead of is_hugetlb_entry_migration and
4865 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4866 * both cases, and because we can't follow correct pages
4867 * directly from any kind of swap entries.
4869 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4870 ((flags & FOLL_WRITE) &&
4871 !huge_pte_write(huge_ptep_get(pte)))) {
4873 unsigned int fault_flags = 0;
4877 if (flags & FOLL_WRITE)
4878 fault_flags |= FAULT_FLAG_WRITE;
4880 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4881 FAULT_FLAG_KILLABLE;
4882 if (flags & FOLL_NOWAIT)
4883 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4884 FAULT_FLAG_RETRY_NOWAIT;
4885 if (flags & FOLL_TRIED) {
4887 * Note: FAULT_FLAG_ALLOW_RETRY and
4888 * FAULT_FLAG_TRIED can co-exist
4890 fault_flags |= FAULT_FLAG_TRIED;
4892 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4893 if (ret & VM_FAULT_ERROR) {
4894 err = vm_fault_to_errno(ret, flags);
4898 if (ret & VM_FAULT_RETRY) {
4900 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4904 * VM_FAULT_RETRY must not return an
4905 * error, it will return zero
4908 * No need to update "position" as the
4909 * caller will not check it after
4910 * *nr_pages is set to 0.
4917 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4918 page = pte_page(huge_ptep_get(pte));
4921 * If subpage information not requested, update counters
4922 * and skip the same_page loop below.
4924 if (!pages && !vmas && !pfn_offset &&
4925 (vaddr + huge_page_size(h) < vma->vm_end) &&
4926 (remainder >= pages_per_huge_page(h))) {
4927 vaddr += huge_page_size(h);
4928 remainder -= pages_per_huge_page(h);
4929 i += pages_per_huge_page(h);
4936 pages[i] = mem_map_offset(page, pfn_offset);
4938 * try_grab_page() should always succeed here, because:
4939 * a) we hold the ptl lock, and b) we've just checked
4940 * that the huge page is present in the page tables. If
4941 * the huge page is present, then the tail pages must
4942 * also be present. The ptl prevents the head page and
4943 * tail pages from being rearranged in any way. So this
4944 * page must be available at this point, unless the page
4945 * refcount overflowed:
4947 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4962 if (vaddr < vma->vm_end && remainder &&
4963 pfn_offset < pages_per_huge_page(h)) {
4965 * We use pfn_offset to avoid touching the pageframes
4966 * of this compound page.
4972 *nr_pages = remainder;
4974 * setting position is actually required only if remainder is
4975 * not zero but it's faster not to add a "if (remainder)"
4983 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4985 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4988 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4991 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4992 unsigned long address, unsigned long end, pgprot_t newprot)
4994 struct mm_struct *mm = vma->vm_mm;
4995 unsigned long start = address;
4998 struct hstate *h = hstate_vma(vma);
4999 unsigned long pages = 0;
5000 bool shared_pmd = false;
5001 struct mmu_notifier_range range;
5004 * In the case of shared PMDs, the area to flush could be beyond
5005 * start/end. Set range.start/range.end to cover the maximum possible
5006 * range if PMD sharing is possible.
5008 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5009 0, vma, mm, start, end);
5010 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5012 BUG_ON(address >= end);
5013 flush_cache_range(vma, range.start, range.end);
5015 mmu_notifier_invalidate_range_start(&range);
5016 i_mmap_lock_write(vma->vm_file->f_mapping);
5017 for (; address < end; address += huge_page_size(h)) {
5019 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5022 ptl = huge_pte_lock(h, mm, ptep);
5023 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5029 pte = huge_ptep_get(ptep);
5030 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5034 if (unlikely(is_hugetlb_entry_migration(pte))) {
5035 swp_entry_t entry = pte_to_swp_entry(pte);
5037 if (is_write_migration_entry(entry)) {
5040 make_migration_entry_read(&entry);
5041 newpte = swp_entry_to_pte(entry);
5042 set_huge_swap_pte_at(mm, address, ptep,
5043 newpte, huge_page_size(h));
5049 if (!huge_pte_none(pte)) {
5052 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5053 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5054 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5055 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5061 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5062 * may have cleared our pud entry and done put_page on the page table:
5063 * once we release i_mmap_rwsem, another task can do the final put_page
5064 * and that page table be reused and filled with junk. If we actually
5065 * did unshare a page of pmds, flush the range corresponding to the pud.
5068 flush_hugetlb_tlb_range(vma, range.start, range.end);
5070 flush_hugetlb_tlb_range(vma, start, end);
5072 * No need to call mmu_notifier_invalidate_range() we are downgrading
5073 * page table protection not changing it to point to a new page.
5075 * See Documentation/vm/mmu_notifier.rst
5077 i_mmap_unlock_write(vma->vm_file->f_mapping);
5078 mmu_notifier_invalidate_range_end(&range);
5080 return pages << h->order;
5083 int hugetlb_reserve_pages(struct inode *inode,
5085 struct vm_area_struct *vma,
5086 vm_flags_t vm_flags)
5088 long ret, chg, add = -1;
5089 struct hstate *h = hstate_inode(inode);
5090 struct hugepage_subpool *spool = subpool_inode(inode);
5091 struct resv_map *resv_map;
5092 struct hugetlb_cgroup *h_cg = NULL;
5093 long gbl_reserve, regions_needed = 0;
5095 /* This should never happen */
5097 VM_WARN(1, "%s called with a negative range\n", __func__);
5102 * Only apply hugepage reservation if asked. At fault time, an
5103 * attempt will be made for VM_NORESERVE to allocate a page
5104 * without using reserves
5106 if (vm_flags & VM_NORESERVE)
5110 * Shared mappings base their reservation on the number of pages that
5111 * are already allocated on behalf of the file. Private mappings need
5112 * to reserve the full area even if read-only as mprotect() may be
5113 * called to make the mapping read-write. Assume !vma is a shm mapping
5115 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5117 * resv_map can not be NULL as hugetlb_reserve_pages is only
5118 * called for inodes for which resv_maps were created (see
5119 * hugetlbfs_get_inode).
5121 resv_map = inode_resv_map(inode);
5123 chg = region_chg(resv_map, from, to, ®ions_needed);
5126 /* Private mapping. */
5127 resv_map = resv_map_alloc();
5133 set_vma_resv_map(vma, resv_map);
5134 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5142 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5143 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5150 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5151 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5154 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5158 * There must be enough pages in the subpool for the mapping. If
5159 * the subpool has a minimum size, there may be some global
5160 * reservations already in place (gbl_reserve).
5162 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5163 if (gbl_reserve < 0) {
5165 goto out_uncharge_cgroup;
5169 * Check enough hugepages are available for the reservation.
5170 * Hand the pages back to the subpool if there are not
5172 ret = hugetlb_acct_memory(h, gbl_reserve);
5178 * Account for the reservations made. Shared mappings record regions
5179 * that have reservations as they are shared by multiple VMAs.
5180 * When the last VMA disappears, the region map says how much
5181 * the reservation was and the page cache tells how much of
5182 * the reservation was consumed. Private mappings are per-VMA and
5183 * only the consumed reservations are tracked. When the VMA
5184 * disappears, the original reservation is the VMA size and the
5185 * consumed reservations are stored in the map. Hence, nothing
5186 * else has to be done for private mappings here
5188 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5189 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5191 if (unlikely(add < 0)) {
5192 hugetlb_acct_memory(h, -gbl_reserve);
5194 } else if (unlikely(chg > add)) {
5196 * pages in this range were added to the reserve
5197 * map between region_chg and region_add. This
5198 * indicates a race with alloc_huge_page. Adjust
5199 * the subpool and reserve counts modified above
5200 * based on the difference.
5204 hugetlb_cgroup_uncharge_cgroup_rsvd(
5206 (chg - add) * pages_per_huge_page(h), h_cg);
5208 rsv_adjust = hugepage_subpool_put_pages(spool,
5210 hugetlb_acct_memory(h, -rsv_adjust);
5215 /* put back original number of pages, chg */
5216 (void)hugepage_subpool_put_pages(spool, chg);
5217 out_uncharge_cgroup:
5218 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5219 chg * pages_per_huge_page(h), h_cg);
5221 if (!vma || vma->vm_flags & VM_MAYSHARE)
5222 /* Only call region_abort if the region_chg succeeded but the
5223 * region_add failed or didn't run.
5225 if (chg >= 0 && add < 0)
5226 region_abort(resv_map, from, to, regions_needed);
5227 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5228 kref_put(&resv_map->refs, resv_map_release);
5232 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5235 struct hstate *h = hstate_inode(inode);
5236 struct resv_map *resv_map = inode_resv_map(inode);
5238 struct hugepage_subpool *spool = subpool_inode(inode);
5242 * Since this routine can be called in the evict inode path for all
5243 * hugetlbfs inodes, resv_map could be NULL.
5246 chg = region_del(resv_map, start, end);
5248 * region_del() can fail in the rare case where a region
5249 * must be split and another region descriptor can not be
5250 * allocated. If end == LONG_MAX, it will not fail.
5256 spin_lock(&inode->i_lock);
5257 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5258 spin_unlock(&inode->i_lock);
5261 * If the subpool has a minimum size, the number of global
5262 * reservations to be released may be adjusted.
5264 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5265 hugetlb_acct_memory(h, -gbl_reserve);
5270 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5271 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5272 struct vm_area_struct *vma,
5273 unsigned long addr, pgoff_t idx)
5275 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5277 unsigned long sbase = saddr & PUD_MASK;
5278 unsigned long s_end = sbase + PUD_SIZE;
5280 /* Allow segments to share if only one is marked locked */
5281 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5282 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5285 * match the virtual addresses, permission and the alignment of the
5288 if (pmd_index(addr) != pmd_index(saddr) ||
5289 vm_flags != svm_flags ||
5290 sbase < svma->vm_start || svma->vm_end < s_end)
5296 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5298 unsigned long base = addr & PUD_MASK;
5299 unsigned long end = base + PUD_SIZE;
5302 * check on proper vm_flags and page table alignment
5304 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5310 * Determine if start,end range within vma could be mapped by shared pmd.
5311 * If yes, adjust start and end to cover range associated with possible
5312 * shared pmd mappings.
5314 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5315 unsigned long *start, unsigned long *end)
5317 unsigned long a_start, a_end;
5319 if (!(vma->vm_flags & VM_MAYSHARE))
5322 /* Extend the range to be PUD aligned for a worst case scenario */
5323 a_start = ALIGN_DOWN(*start, PUD_SIZE);
5324 a_end = ALIGN(*end, PUD_SIZE);
5327 * Intersect the range with the vma range, since pmd sharing won't be
5328 * across vma after all
5330 *start = max(vma->vm_start, a_start);
5331 *end = min(vma->vm_end, a_end);
5335 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5336 * and returns the corresponding pte. While this is not necessary for the
5337 * !shared pmd case because we can allocate the pmd later as well, it makes the
5338 * code much cleaner.
5340 * This routine must be called with i_mmap_rwsem held in at least read mode if
5341 * sharing is possible. For hugetlbfs, this prevents removal of any page
5342 * table entries associated with the address space. This is important as we
5343 * are setting up sharing based on existing page table entries (mappings).
5345 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5346 * huge_pte_alloc know that sharing is not possible and do not take
5347 * i_mmap_rwsem as a performance optimization. This is handled by the
5348 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5349 * only required for subsequent processing.
5351 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5353 struct vm_area_struct *vma = find_vma(mm, addr);
5354 struct address_space *mapping = vma->vm_file->f_mapping;
5355 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5357 struct vm_area_struct *svma;
5358 unsigned long saddr;
5363 if (!vma_shareable(vma, addr))
5364 return (pte_t *)pmd_alloc(mm, pud, addr);
5366 i_mmap_assert_locked(mapping);
5367 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5371 saddr = page_table_shareable(svma, vma, addr, idx);
5373 spte = huge_pte_offset(svma->vm_mm, saddr,
5374 vma_mmu_pagesize(svma));
5376 get_page(virt_to_page(spte));
5385 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5386 if (pud_none(*pud)) {
5387 pud_populate(mm, pud,
5388 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5391 put_page(virt_to_page(spte));
5395 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5400 * unmap huge page backed by shared pte.
5402 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5403 * indicated by page_count > 1, unmap is achieved by clearing pud and
5404 * decrementing the ref count. If count == 1, the pte page is not shared.
5406 * Called with page table lock held and i_mmap_rwsem held in write mode.
5408 * returns: 1 successfully unmapped a shared pte page
5409 * 0 the underlying pte page is not shared, or it is the last user
5411 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5412 unsigned long *addr, pte_t *ptep)
5414 pgd_t *pgd = pgd_offset(mm, *addr);
5415 p4d_t *p4d = p4d_offset(pgd, *addr);
5416 pud_t *pud = pud_offset(p4d, *addr);
5418 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5419 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5420 if (page_count(virt_to_page(ptep)) == 1)
5424 put_page(virt_to_page(ptep));
5426 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5429 #define want_pmd_share() (1)
5430 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5431 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5436 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5437 unsigned long *addr, pte_t *ptep)
5442 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5443 unsigned long *start, unsigned long *end)
5446 #define want_pmd_share() (0)
5447 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5449 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5450 pte_t *huge_pte_alloc(struct mm_struct *mm,
5451 unsigned long addr, unsigned long sz)
5458 pgd = pgd_offset(mm, addr);
5459 p4d = p4d_alloc(mm, pgd, addr);
5462 pud = pud_alloc(mm, p4d, addr);
5464 if (sz == PUD_SIZE) {
5467 BUG_ON(sz != PMD_SIZE);
5468 if (want_pmd_share() && pud_none(*pud))
5469 pte = huge_pmd_share(mm, addr, pud);
5471 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5474 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5480 * huge_pte_offset() - Walk the page table to resolve the hugepage
5481 * entry at address @addr
5483 * Return: Pointer to page table entry (PUD or PMD) for
5484 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5485 * size @sz doesn't match the hugepage size at this level of the page
5488 pte_t *huge_pte_offset(struct mm_struct *mm,
5489 unsigned long addr, unsigned long sz)
5496 pgd = pgd_offset(mm, addr);
5497 if (!pgd_present(*pgd))
5499 p4d = p4d_offset(pgd, addr);
5500 if (!p4d_present(*p4d))
5503 pud = pud_offset(p4d, addr);
5505 /* must be pud huge, non-present or none */
5506 return (pte_t *)pud;
5507 if (!pud_present(*pud))
5509 /* must have a valid entry and size to go further */
5511 pmd = pmd_offset(pud, addr);
5512 /* must be pmd huge, non-present or none */
5513 return (pte_t *)pmd;
5516 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5519 * These functions are overwritable if your architecture needs its own
5522 struct page * __weak
5523 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5526 return ERR_PTR(-EINVAL);
5529 struct page * __weak
5530 follow_huge_pd(struct vm_area_struct *vma,
5531 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5533 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5537 struct page * __weak
5538 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5539 pmd_t *pmd, int flags)
5541 struct page *page = NULL;
5545 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5546 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5547 (FOLL_PIN | FOLL_GET)))
5551 ptl = pmd_lockptr(mm, pmd);
5554 * make sure that the address range covered by this pmd is not
5555 * unmapped from other threads.
5557 if (!pmd_huge(*pmd))
5559 pte = huge_ptep_get((pte_t *)pmd);
5560 if (pte_present(pte)) {
5561 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5563 * try_grab_page() should always succeed here, because: a) we
5564 * hold the pmd (ptl) lock, and b) we've just checked that the
5565 * huge pmd (head) page is present in the page tables. The ptl
5566 * prevents the head page and tail pages from being rearranged
5567 * in any way. So this page must be available at this point,
5568 * unless the page refcount overflowed:
5570 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5575 if (is_hugetlb_entry_migration(pte)) {
5577 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5581 * hwpoisoned entry is treated as no_page_table in
5582 * follow_page_mask().
5590 struct page * __weak
5591 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5592 pud_t *pud, int flags)
5594 if (flags & (FOLL_GET | FOLL_PIN))
5597 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5600 struct page * __weak
5601 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5603 if (flags & (FOLL_GET | FOLL_PIN))
5606 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5609 bool isolate_huge_page(struct page *page, struct list_head *list)
5613 VM_BUG_ON_PAGE(!PageHead(page), page);
5614 spin_lock(&hugetlb_lock);
5615 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5619 clear_page_huge_active(page);
5620 list_move_tail(&page->lru, list);
5622 spin_unlock(&hugetlb_lock);
5626 void putback_active_hugepage(struct page *page)
5628 VM_BUG_ON_PAGE(!PageHead(page), page);
5629 spin_lock(&hugetlb_lock);
5630 set_page_huge_active(page);
5631 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5632 spin_unlock(&hugetlb_lock);
5636 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5638 struct hstate *h = page_hstate(oldpage);
5640 hugetlb_cgroup_migrate(oldpage, newpage);
5641 set_page_owner_migrate_reason(newpage, reason);
5644 * transfer temporary state of the new huge page. This is
5645 * reverse to other transitions because the newpage is going to
5646 * be final while the old one will be freed so it takes over
5647 * the temporary status.
5649 * Also note that we have to transfer the per-node surplus state
5650 * here as well otherwise the global surplus count will not match
5653 if (PageHugeTemporary(newpage)) {
5654 int old_nid = page_to_nid(oldpage);
5655 int new_nid = page_to_nid(newpage);
5657 SetPageHugeTemporary(oldpage);
5658 ClearPageHugeTemporary(newpage);
5660 spin_lock(&hugetlb_lock);
5661 if (h->surplus_huge_pages_node[old_nid]) {
5662 h->surplus_huge_pages_node[old_nid]--;
5663 h->surplus_huge_pages_node[new_nid]++;
5665 spin_unlock(&hugetlb_lock);
5670 static bool cma_reserve_called __initdata;
5672 static int __init cmdline_parse_hugetlb_cma(char *p)
5674 hugetlb_cma_size = memparse(p, &p);
5678 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5680 void __init hugetlb_cma_reserve(int order)
5682 unsigned long size, reserved, per_node;
5685 cma_reserve_called = true;
5687 if (!hugetlb_cma_size)
5690 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5691 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5692 (PAGE_SIZE << order) / SZ_1M);
5697 * If 3 GB area is requested on a machine with 4 numa nodes,
5698 * let's allocate 1 GB on first three nodes and ignore the last one.
5700 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5701 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5702 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5705 for_each_node_state(nid, N_ONLINE) {
5709 size = min(per_node, hugetlb_cma_size - reserved);
5710 size = round_up(size, PAGE_SIZE << order);
5712 snprintf(name, 20, "hugetlb%d", nid);
5713 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5715 &hugetlb_cma[nid], nid);
5717 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5723 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5726 if (reserved >= hugetlb_cma_size)
5731 void __init hugetlb_cma_check(void)
5733 if (!hugetlb_cma_size || cma_reserve_called)
5736 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5739 #endif /* CONFIG_CMA */