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);
651 hugetlb_cgroup_uncharge_file_region(
654 /* New entry for end of split region */
658 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
660 INIT_LIST_HEAD(&nrg->link);
662 /* Original entry is trimmed */
665 list_add(&nrg->link, &rg->link);
670 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
671 del += rg->to - rg->from;
672 hugetlb_cgroup_uncharge_file_region(resv, rg,
679 if (f <= rg->from) { /* Trim beginning of region */
680 hugetlb_cgroup_uncharge_file_region(resv, rg,
685 } else { /* Trim end of region */
686 hugetlb_cgroup_uncharge_file_region(resv, rg,
694 spin_unlock(&resv->lock);
700 * A rare out of memory error was encountered which prevented removal of
701 * the reserve map region for a page. The huge page itself was free'ed
702 * and removed from the page cache. This routine will adjust the subpool
703 * usage count, and the global reserve count if needed. By incrementing
704 * these counts, the reserve map entry which could not be deleted will
705 * appear as a "reserved" entry instead of simply dangling with incorrect
708 void hugetlb_fix_reserve_counts(struct inode *inode)
710 struct hugepage_subpool *spool = subpool_inode(inode);
713 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
715 struct hstate *h = hstate_inode(inode);
717 hugetlb_acct_memory(h, 1);
722 * Count and return the number of huge pages in the reserve map
723 * that intersect with the range [f, t).
725 static long region_count(struct resv_map *resv, long f, long t)
727 struct list_head *head = &resv->regions;
728 struct file_region *rg;
731 spin_lock(&resv->lock);
732 /* Locate each segment we overlap with, and count that overlap. */
733 list_for_each_entry(rg, head, link) {
742 seg_from = max(rg->from, f);
743 seg_to = min(rg->to, t);
745 chg += seg_to - seg_from;
747 spin_unlock(&resv->lock);
753 * Convert the address within this vma to the page offset within
754 * the mapping, in pagecache page units; huge pages here.
756 static pgoff_t vma_hugecache_offset(struct hstate *h,
757 struct vm_area_struct *vma, unsigned long address)
759 return ((address - vma->vm_start) >> huge_page_shift(h)) +
760 (vma->vm_pgoff >> huge_page_order(h));
763 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
764 unsigned long address)
766 return vma_hugecache_offset(hstate_vma(vma), vma, address);
768 EXPORT_SYMBOL_GPL(linear_hugepage_index);
771 * Return the size of the pages allocated when backing a VMA. In the majority
772 * cases this will be same size as used by the page table entries.
774 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
776 if (vma->vm_ops && vma->vm_ops->pagesize)
777 return vma->vm_ops->pagesize(vma);
780 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
783 * Return the page size being used by the MMU to back a VMA. In the majority
784 * of cases, the page size used by the kernel matches the MMU size. On
785 * architectures where it differs, an architecture-specific 'strong'
786 * version of this symbol is required.
788 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
790 return vma_kernel_pagesize(vma);
794 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
795 * bits of the reservation map pointer, which are always clear due to
798 #define HPAGE_RESV_OWNER (1UL << 0)
799 #define HPAGE_RESV_UNMAPPED (1UL << 1)
800 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
803 * These helpers are used to track how many pages are reserved for
804 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
805 * is guaranteed to have their future faults succeed.
807 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
808 * the reserve counters are updated with the hugetlb_lock held. It is safe
809 * to reset the VMA at fork() time as it is not in use yet and there is no
810 * chance of the global counters getting corrupted as a result of the values.
812 * The private mapping reservation is represented in a subtly different
813 * manner to a shared mapping. A shared mapping has a region map associated
814 * with the underlying file, this region map represents the backing file
815 * pages which have ever had a reservation assigned which this persists even
816 * after the page is instantiated. A private mapping has a region map
817 * associated with the original mmap which is attached to all VMAs which
818 * reference it, this region map represents those offsets which have consumed
819 * reservation ie. where pages have been instantiated.
821 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
823 return (unsigned long)vma->vm_private_data;
826 static void set_vma_private_data(struct vm_area_struct *vma,
829 vma->vm_private_data = (void *)value;
833 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
834 struct hugetlb_cgroup *h_cg,
837 #ifdef CONFIG_CGROUP_HUGETLB
839 resv_map->reservation_counter = NULL;
840 resv_map->pages_per_hpage = 0;
841 resv_map->css = NULL;
843 resv_map->reservation_counter =
844 &h_cg->rsvd_hugepage[hstate_index(h)];
845 resv_map->pages_per_hpage = pages_per_huge_page(h);
846 resv_map->css = &h_cg->css;
851 struct resv_map *resv_map_alloc(void)
853 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
854 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
856 if (!resv_map || !rg) {
862 kref_init(&resv_map->refs);
863 spin_lock_init(&resv_map->lock);
864 INIT_LIST_HEAD(&resv_map->regions);
866 resv_map->adds_in_progress = 0;
868 * Initialize these to 0. On shared mappings, 0's here indicate these
869 * fields don't do cgroup accounting. On private mappings, these will be
870 * re-initialized to the proper values, to indicate that hugetlb cgroup
871 * reservations are to be un-charged from here.
873 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
875 INIT_LIST_HEAD(&resv_map->region_cache);
876 list_add(&rg->link, &resv_map->region_cache);
877 resv_map->region_cache_count = 1;
882 void resv_map_release(struct kref *ref)
884 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
885 struct list_head *head = &resv_map->region_cache;
886 struct file_region *rg, *trg;
888 /* Clear out any active regions before we release the map. */
889 region_del(resv_map, 0, LONG_MAX);
891 /* ... and any entries left in the cache */
892 list_for_each_entry_safe(rg, trg, head, link) {
897 VM_BUG_ON(resv_map->adds_in_progress);
902 static inline struct resv_map *inode_resv_map(struct inode *inode)
905 * At inode evict time, i_mapping may not point to the original
906 * address space within the inode. This original address space
907 * contains the pointer to the resv_map. So, always use the
908 * address space embedded within the inode.
909 * The VERY common case is inode->mapping == &inode->i_data but,
910 * this may not be true for device special inodes.
912 return (struct resv_map *)(&inode->i_data)->private_data;
915 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
917 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
918 if (vma->vm_flags & VM_MAYSHARE) {
919 struct address_space *mapping = vma->vm_file->f_mapping;
920 struct inode *inode = mapping->host;
922 return inode_resv_map(inode);
925 return (struct resv_map *)(get_vma_private_data(vma) &
930 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
932 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
933 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
935 set_vma_private_data(vma, (get_vma_private_data(vma) &
936 HPAGE_RESV_MASK) | (unsigned long)map);
939 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
941 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
942 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
944 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
947 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
949 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
951 return (get_vma_private_data(vma) & flag) != 0;
954 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
955 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
957 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958 if (!(vma->vm_flags & VM_MAYSHARE))
959 vma->vm_private_data = (void *)0;
962 /* Returns true if the VMA has associated reserve pages */
963 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
965 if (vma->vm_flags & VM_NORESERVE) {
967 * This address is already reserved by other process(chg == 0),
968 * so, we should decrement reserved count. Without decrementing,
969 * reserve count remains after releasing inode, because this
970 * allocated page will go into page cache and is regarded as
971 * coming from reserved pool in releasing step. Currently, we
972 * don't have any other solution to deal with this situation
973 * properly, so add work-around here.
975 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
981 /* Shared mappings always use reserves */
982 if (vma->vm_flags & VM_MAYSHARE) {
984 * We know VM_NORESERVE is not set. Therefore, there SHOULD
985 * be a region map for all pages. The only situation where
986 * there is no region map is if a hole was punched via
987 * fallocate. In this case, there really are no reserves to
988 * use. This situation is indicated if chg != 0.
997 * Only the process that called mmap() has reserves for
1000 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1002 * Like the shared case above, a hole punch or truncate
1003 * could have been performed on the private mapping.
1004 * Examine the value of chg to determine if reserves
1005 * actually exist or were previously consumed.
1006 * Very Subtle - The value of chg comes from a previous
1007 * call to vma_needs_reserves(). The reserve map for
1008 * private mappings has different (opposite) semantics
1009 * than that of shared mappings. vma_needs_reserves()
1010 * has already taken this difference in semantics into
1011 * account. Therefore, the meaning of chg is the same
1012 * as in the shared case above. Code could easily be
1013 * combined, but keeping it separate draws attention to
1014 * subtle differences.
1025 static void enqueue_huge_page(struct hstate *h, struct page *page)
1027 int nid = page_to_nid(page);
1028 list_move(&page->lru, &h->hugepage_freelists[nid]);
1029 h->free_huge_pages++;
1030 h->free_huge_pages_node[nid]++;
1033 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1036 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1038 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1039 if (nocma && is_migrate_cma_page(page))
1042 if (PageHWPoison(page))
1045 list_move(&page->lru, &h->hugepage_activelist);
1046 set_page_refcounted(page);
1047 h->free_huge_pages--;
1048 h->free_huge_pages_node[nid]--;
1055 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1058 unsigned int cpuset_mems_cookie;
1059 struct zonelist *zonelist;
1062 int node = NUMA_NO_NODE;
1064 zonelist = node_zonelist(nid, gfp_mask);
1067 cpuset_mems_cookie = read_mems_allowed_begin();
1068 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1071 if (!cpuset_zone_allowed(zone, gfp_mask))
1074 * no need to ask again on the same node. Pool is node rather than
1077 if (zone_to_nid(zone) == node)
1079 node = zone_to_nid(zone);
1081 page = dequeue_huge_page_node_exact(h, node);
1085 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1091 static struct page *dequeue_huge_page_vma(struct hstate *h,
1092 struct vm_area_struct *vma,
1093 unsigned long address, int avoid_reserve,
1097 struct mempolicy *mpol;
1099 nodemask_t *nodemask;
1103 * A child process with MAP_PRIVATE mappings created by their parent
1104 * have no page reserves. This check ensures that reservations are
1105 * not "stolen". The child may still get SIGKILLed
1107 if (!vma_has_reserves(vma, chg) &&
1108 h->free_huge_pages - h->resv_huge_pages == 0)
1111 /* If reserves cannot be used, ensure enough pages are in the pool */
1112 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1115 gfp_mask = htlb_alloc_mask(h);
1116 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1117 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1118 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1119 SetPagePrivate(page);
1120 h->resv_huge_pages--;
1123 mpol_cond_put(mpol);
1131 * common helper functions for hstate_next_node_to_{alloc|free}.
1132 * We may have allocated or freed a huge page based on a different
1133 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1134 * be outside of *nodes_allowed. Ensure that we use an allowed
1135 * node for alloc or free.
1137 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1139 nid = next_node_in(nid, *nodes_allowed);
1140 VM_BUG_ON(nid >= MAX_NUMNODES);
1145 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1147 if (!node_isset(nid, *nodes_allowed))
1148 nid = next_node_allowed(nid, nodes_allowed);
1153 * returns the previously saved node ["this node"] from which to
1154 * allocate a persistent huge page for the pool and advance the
1155 * next node from which to allocate, handling wrap at end of node
1158 static int hstate_next_node_to_alloc(struct hstate *h,
1159 nodemask_t *nodes_allowed)
1163 VM_BUG_ON(!nodes_allowed);
1165 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1166 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1172 * helper for free_pool_huge_page() - return the previously saved
1173 * node ["this node"] from which to free a huge page. Advance the
1174 * next node id whether or not we find a free huge page to free so
1175 * that the next attempt to free addresses the next node.
1177 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1181 VM_BUG_ON(!nodes_allowed);
1183 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1184 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1189 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1190 for (nr_nodes = nodes_weight(*mask); \
1192 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1195 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1196 for (nr_nodes = nodes_weight(*mask); \
1198 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1201 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1202 static void destroy_compound_gigantic_page(struct page *page,
1206 int nr_pages = 1 << order;
1207 struct page *p = page + 1;
1209 atomic_set(compound_mapcount_ptr(page), 0);
1210 if (hpage_pincount_available(page))
1211 atomic_set(compound_pincount_ptr(page), 0);
1213 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1214 clear_compound_head(p);
1215 set_page_refcounted(p);
1218 set_compound_order(page, 0);
1219 __ClearPageHead(page);
1222 static void free_gigantic_page(struct page *page, unsigned int order)
1225 * If the page isn't allocated using the cma allocator,
1226 * cma_release() returns false.
1229 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1233 free_contig_range(page_to_pfn(page), 1 << order);
1236 #ifdef CONFIG_CONTIG_ALLOC
1237 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1238 int nid, nodemask_t *nodemask)
1240 unsigned long nr_pages = 1UL << huge_page_order(h);
1241 if (nid == NUMA_NO_NODE)
1242 nid = numa_mem_id();
1249 if (hugetlb_cma[nid]) {
1250 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1251 huge_page_order(h), true);
1256 if (!(gfp_mask & __GFP_THISNODE)) {
1257 for_each_node_mask(node, *nodemask) {
1258 if (node == nid || !hugetlb_cma[node])
1261 page = cma_alloc(hugetlb_cma[node], nr_pages,
1262 huge_page_order(h), true);
1270 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1273 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1274 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1275 #else /* !CONFIG_CONTIG_ALLOC */
1276 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1277 int nid, nodemask_t *nodemask)
1281 #endif /* CONFIG_CONTIG_ALLOC */
1283 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1284 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1285 int nid, nodemask_t *nodemask)
1289 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1290 static inline void destroy_compound_gigantic_page(struct page *page,
1291 unsigned int order) { }
1294 static void update_and_free_page(struct hstate *h, struct page *page)
1298 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1302 h->nr_huge_pages_node[page_to_nid(page)]--;
1303 for (i = 0; i < pages_per_huge_page(h); i++) {
1304 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1305 1 << PG_referenced | 1 << PG_dirty |
1306 1 << PG_active | 1 << PG_private |
1309 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1310 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1311 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1312 set_page_refcounted(page);
1313 if (hstate_is_gigantic(h)) {
1315 * Temporarily drop the hugetlb_lock, because
1316 * we might block in free_gigantic_page().
1318 spin_unlock(&hugetlb_lock);
1319 destroy_compound_gigantic_page(page, huge_page_order(h));
1320 free_gigantic_page(page, huge_page_order(h));
1321 spin_lock(&hugetlb_lock);
1323 __free_pages(page, huge_page_order(h));
1327 struct hstate *size_to_hstate(unsigned long size)
1331 for_each_hstate(h) {
1332 if (huge_page_size(h) == size)
1339 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1340 * to hstate->hugepage_activelist.)
1342 * This function can be called for tail pages, but never returns true for them.
1344 bool page_huge_active(struct page *page)
1346 VM_BUG_ON_PAGE(!PageHuge(page), page);
1347 return PageHead(page) && PagePrivate(&page[1]);
1350 /* never called for tail page */
1351 static void set_page_huge_active(struct page *page)
1353 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1354 SetPagePrivate(&page[1]);
1357 static void clear_page_huge_active(struct page *page)
1359 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1360 ClearPagePrivate(&page[1]);
1364 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1367 static inline bool PageHugeTemporary(struct page *page)
1369 if (!PageHuge(page))
1372 return (unsigned long)page[2].mapping == -1U;
1375 static inline void SetPageHugeTemporary(struct page *page)
1377 page[2].mapping = (void *)-1U;
1380 static inline void ClearPageHugeTemporary(struct page *page)
1382 page[2].mapping = NULL;
1385 static void __free_huge_page(struct page *page)
1388 * Can't pass hstate in here because it is called from the
1389 * compound page destructor.
1391 struct hstate *h = page_hstate(page);
1392 int nid = page_to_nid(page);
1393 struct hugepage_subpool *spool =
1394 (struct hugepage_subpool *)page_private(page);
1395 bool restore_reserve;
1397 VM_BUG_ON_PAGE(page_count(page), page);
1398 VM_BUG_ON_PAGE(page_mapcount(page), page);
1400 set_page_private(page, 0);
1401 page->mapping = NULL;
1402 restore_reserve = PagePrivate(page);
1403 ClearPagePrivate(page);
1406 * If PagePrivate() was set on page, page allocation consumed a
1407 * reservation. If the page was associated with a subpool, there
1408 * would have been a page reserved in the subpool before allocation
1409 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1410 * reservtion, do not call hugepage_subpool_put_pages() as this will
1411 * remove the reserved page from the subpool.
1413 if (!restore_reserve) {
1415 * A return code of zero implies that the subpool will be
1416 * under its minimum size if the reservation is not restored
1417 * after page is free. Therefore, force restore_reserve
1420 if (hugepage_subpool_put_pages(spool, 1) == 0)
1421 restore_reserve = true;
1424 spin_lock(&hugetlb_lock);
1425 clear_page_huge_active(page);
1426 hugetlb_cgroup_uncharge_page(hstate_index(h),
1427 pages_per_huge_page(h), page);
1428 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1429 pages_per_huge_page(h), page);
1430 if (restore_reserve)
1431 h->resv_huge_pages++;
1433 if (PageHugeTemporary(page)) {
1434 list_del(&page->lru);
1435 ClearPageHugeTemporary(page);
1436 update_and_free_page(h, page);
1437 } else if (h->surplus_huge_pages_node[nid]) {
1438 /* remove the page from active list */
1439 list_del(&page->lru);
1440 update_and_free_page(h, page);
1441 h->surplus_huge_pages--;
1442 h->surplus_huge_pages_node[nid]--;
1444 arch_clear_hugepage_flags(page);
1445 enqueue_huge_page(h, page);
1447 spin_unlock(&hugetlb_lock);
1451 * As free_huge_page() can be called from a non-task context, we have
1452 * to defer the actual freeing in a workqueue to prevent potential
1453 * hugetlb_lock deadlock.
1455 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1456 * be freed and frees them one-by-one. As the page->mapping pointer is
1457 * going to be cleared in __free_huge_page() anyway, it is reused as the
1458 * llist_node structure of a lockless linked list of huge pages to be freed.
1460 static LLIST_HEAD(hpage_freelist);
1462 static void free_hpage_workfn(struct work_struct *work)
1464 struct llist_node *node;
1467 node = llist_del_all(&hpage_freelist);
1470 page = container_of((struct address_space **)node,
1471 struct page, mapping);
1473 __free_huge_page(page);
1476 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1478 void free_huge_page(struct page *page)
1481 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1485 * Only call schedule_work() if hpage_freelist is previously
1486 * empty. Otherwise, schedule_work() had been called but the
1487 * workfn hasn't retrieved the list yet.
1489 if (llist_add((struct llist_node *)&page->mapping,
1491 schedule_work(&free_hpage_work);
1495 __free_huge_page(page);
1498 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1500 INIT_LIST_HEAD(&page->lru);
1501 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1502 set_hugetlb_cgroup(page, NULL);
1503 set_hugetlb_cgroup_rsvd(page, NULL);
1504 spin_lock(&hugetlb_lock);
1506 h->nr_huge_pages_node[nid]++;
1507 spin_unlock(&hugetlb_lock);
1510 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1513 int nr_pages = 1 << order;
1514 struct page *p = page + 1;
1516 /* we rely on prep_new_huge_page to set the destructor */
1517 set_compound_order(page, order);
1518 __ClearPageReserved(page);
1519 __SetPageHead(page);
1520 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1522 * For gigantic hugepages allocated through bootmem at
1523 * boot, it's safer to be consistent with the not-gigantic
1524 * hugepages and clear the PG_reserved bit from all tail pages
1525 * too. Otherwise drivers using get_user_pages() to access tail
1526 * pages may get the reference counting wrong if they see
1527 * PG_reserved set on a tail page (despite the head page not
1528 * having PG_reserved set). Enforcing this consistency between
1529 * head and tail pages allows drivers to optimize away a check
1530 * on the head page when they need know if put_page() is needed
1531 * after get_user_pages().
1533 __ClearPageReserved(p);
1534 set_page_count(p, 0);
1535 set_compound_head(p, page);
1537 atomic_set(compound_mapcount_ptr(page), -1);
1539 if (hpage_pincount_available(page))
1540 atomic_set(compound_pincount_ptr(page), 0);
1544 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1545 * transparent huge pages. See the PageTransHuge() documentation for more
1548 int PageHuge(struct page *page)
1550 if (!PageCompound(page))
1553 page = compound_head(page);
1554 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1556 EXPORT_SYMBOL_GPL(PageHuge);
1559 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1560 * normal or transparent huge pages.
1562 int PageHeadHuge(struct page *page_head)
1564 if (!PageHead(page_head))
1567 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1571 * Find address_space associated with hugetlbfs page.
1572 * Upon entry page is locked and page 'was' mapped although mapped state
1573 * could change. If necessary, use anon_vma to find vma and associated
1574 * address space. The returned mapping may be stale, but it can not be
1575 * invalid as page lock (which is held) is required to destroy mapping.
1577 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1579 struct anon_vma *anon_vma;
1580 pgoff_t pgoff_start, pgoff_end;
1581 struct anon_vma_chain *avc;
1582 struct address_space *mapping = page_mapping(hpage);
1584 /* Simple file based mapping */
1589 * Even anonymous hugetlbfs mappings are associated with an
1590 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1591 * code). Find a vma associated with the anonymous vma, and
1592 * use the file pointer to get address_space.
1594 anon_vma = page_lock_anon_vma_read(hpage);
1596 return mapping; /* NULL */
1598 /* Use first found vma */
1599 pgoff_start = page_to_pgoff(hpage);
1600 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
1601 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1602 pgoff_start, pgoff_end) {
1603 struct vm_area_struct *vma = avc->vma;
1605 mapping = vma->vm_file->f_mapping;
1609 anon_vma_unlock_read(anon_vma);
1614 * Find and lock address space (mapping) in write mode.
1616 * Upon entry, the page is locked which allows us to find the mapping
1617 * even in the case of an anon page. However, locking order dictates
1618 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1619 * specific. So, we first try to lock the sema while still holding the
1620 * page lock. If this works, great! If not, then we need to drop the
1621 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1622 * course, need to revalidate state along the way.
1624 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1626 struct address_space *mapping, *mapping2;
1628 mapping = _get_hugetlb_page_mapping(hpage);
1634 * If no contention, take lock and return
1636 if (i_mmap_trylock_write(mapping))
1640 * Must drop page lock and wait on mapping sema.
1641 * Note: Once page lock is dropped, mapping could become invalid.
1642 * As a hack, increase map count until we lock page again.
1644 atomic_inc(&hpage->_mapcount);
1646 i_mmap_lock_write(mapping);
1648 atomic_add_negative(-1, &hpage->_mapcount);
1650 /* verify page is still mapped */
1651 if (!page_mapped(hpage)) {
1652 i_mmap_unlock_write(mapping);
1657 * Get address space again and verify it is the same one
1658 * we locked. If not, drop lock and retry.
1660 mapping2 = _get_hugetlb_page_mapping(hpage);
1661 if (mapping2 != mapping) {
1662 i_mmap_unlock_write(mapping);
1670 pgoff_t __basepage_index(struct page *page)
1672 struct page *page_head = compound_head(page);
1673 pgoff_t index = page_index(page_head);
1674 unsigned long compound_idx;
1676 if (!PageHuge(page_head))
1677 return page_index(page);
1679 if (compound_order(page_head) >= MAX_ORDER)
1680 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1682 compound_idx = page - page_head;
1684 return (index << compound_order(page_head)) + compound_idx;
1687 static struct page *alloc_buddy_huge_page(struct hstate *h,
1688 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1689 nodemask_t *node_alloc_noretry)
1691 int order = huge_page_order(h);
1693 bool alloc_try_hard = true;
1696 * By default we always try hard to allocate the page with
1697 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1698 * a loop (to adjust global huge page counts) and previous allocation
1699 * failed, do not continue to try hard on the same node. Use the
1700 * node_alloc_noretry bitmap to manage this state information.
1702 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1703 alloc_try_hard = false;
1704 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1706 gfp_mask |= __GFP_RETRY_MAYFAIL;
1707 if (nid == NUMA_NO_NODE)
1708 nid = numa_mem_id();
1709 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1711 __count_vm_event(HTLB_BUDDY_PGALLOC);
1713 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1716 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1717 * indicates an overall state change. Clear bit so that we resume
1718 * normal 'try hard' allocations.
1720 if (node_alloc_noretry && page && !alloc_try_hard)
1721 node_clear(nid, *node_alloc_noretry);
1724 * If we tried hard to get a page but failed, set bit so that
1725 * subsequent attempts will not try as hard until there is an
1726 * overall state change.
1728 if (node_alloc_noretry && !page && alloc_try_hard)
1729 node_set(nid, *node_alloc_noretry);
1735 * Common helper to allocate a fresh hugetlb page. All specific allocators
1736 * should use this function to get new hugetlb pages
1738 static struct page *alloc_fresh_huge_page(struct hstate *h,
1739 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1740 nodemask_t *node_alloc_noretry)
1744 if (hstate_is_gigantic(h))
1745 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1747 page = alloc_buddy_huge_page(h, gfp_mask,
1748 nid, nmask, node_alloc_noretry);
1752 if (hstate_is_gigantic(h))
1753 prep_compound_gigantic_page(page, huge_page_order(h));
1754 prep_new_huge_page(h, page, page_to_nid(page));
1760 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1763 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1764 nodemask_t *node_alloc_noretry)
1768 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1770 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1771 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1772 node_alloc_noretry);
1780 put_page(page); /* free it into the hugepage allocator */
1786 * Free huge page from pool from next node to free.
1787 * Attempt to keep persistent huge pages more or less
1788 * balanced over allowed nodes.
1789 * Called with hugetlb_lock locked.
1791 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1797 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1799 * If we're returning unused surplus pages, only examine
1800 * nodes with surplus pages.
1802 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1803 !list_empty(&h->hugepage_freelists[node])) {
1805 list_entry(h->hugepage_freelists[node].next,
1807 list_del(&page->lru);
1808 h->free_huge_pages--;
1809 h->free_huge_pages_node[node]--;
1811 h->surplus_huge_pages--;
1812 h->surplus_huge_pages_node[node]--;
1814 update_and_free_page(h, page);
1824 * Dissolve a given free hugepage into free buddy pages. This function does
1825 * nothing for in-use hugepages and non-hugepages.
1826 * This function returns values like below:
1828 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1829 * (allocated or reserved.)
1830 * 0: successfully dissolved free hugepages or the page is not a
1831 * hugepage (considered as already dissolved)
1833 int dissolve_free_huge_page(struct page *page)
1837 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1838 if (!PageHuge(page))
1841 spin_lock(&hugetlb_lock);
1842 if (!PageHuge(page)) {
1847 if (!page_count(page)) {
1848 struct page *head = compound_head(page);
1849 struct hstate *h = page_hstate(head);
1850 int nid = page_to_nid(head);
1851 if (h->free_huge_pages - h->resv_huge_pages == 0)
1854 * Move PageHWPoison flag from head page to the raw error page,
1855 * which makes any subpages rather than the error page reusable.
1857 if (PageHWPoison(head) && page != head) {
1858 SetPageHWPoison(page);
1859 ClearPageHWPoison(head);
1861 list_del(&head->lru);
1862 h->free_huge_pages--;
1863 h->free_huge_pages_node[nid]--;
1864 h->max_huge_pages--;
1865 update_and_free_page(h, head);
1869 spin_unlock(&hugetlb_lock);
1874 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1875 * make specified memory blocks removable from the system.
1876 * Note that this will dissolve a free gigantic hugepage completely, if any
1877 * part of it lies within the given range.
1878 * Also note that if dissolve_free_huge_page() returns with an error, all
1879 * free hugepages that were dissolved before that error are lost.
1881 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1887 if (!hugepages_supported())
1890 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1891 page = pfn_to_page(pfn);
1892 rc = dissolve_free_huge_page(page);
1901 * Allocates a fresh surplus page from the page allocator.
1903 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1904 int nid, nodemask_t *nmask)
1906 struct page *page = NULL;
1908 if (hstate_is_gigantic(h))
1911 spin_lock(&hugetlb_lock);
1912 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1914 spin_unlock(&hugetlb_lock);
1916 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1920 spin_lock(&hugetlb_lock);
1922 * We could have raced with the pool size change.
1923 * Double check that and simply deallocate the new page
1924 * if we would end up overcommiting the surpluses. Abuse
1925 * temporary page to workaround the nasty free_huge_page
1928 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1929 SetPageHugeTemporary(page);
1930 spin_unlock(&hugetlb_lock);
1934 h->surplus_huge_pages++;
1935 h->surplus_huge_pages_node[page_to_nid(page)]++;
1939 spin_unlock(&hugetlb_lock);
1944 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1945 int nid, nodemask_t *nmask)
1949 if (hstate_is_gigantic(h))
1952 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1957 * We do not account these pages as surplus because they are only
1958 * temporary and will be released properly on the last reference
1960 SetPageHugeTemporary(page);
1966 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1969 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1970 struct vm_area_struct *vma, unsigned long addr)
1973 struct mempolicy *mpol;
1974 gfp_t gfp_mask = htlb_alloc_mask(h);
1976 nodemask_t *nodemask;
1978 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1979 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1980 mpol_cond_put(mpol);
1985 /* page migration callback function */
1986 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1987 nodemask_t *nmask, gfp_t gfp_mask)
1989 spin_lock(&hugetlb_lock);
1990 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1993 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1995 spin_unlock(&hugetlb_lock);
1999 spin_unlock(&hugetlb_lock);
2001 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2004 /* mempolicy aware migration callback */
2005 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2006 unsigned long address)
2008 struct mempolicy *mpol;
2009 nodemask_t *nodemask;
2014 gfp_mask = htlb_alloc_mask(h);
2015 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2016 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2017 mpol_cond_put(mpol);
2023 * Increase the hugetlb pool such that it can accommodate a reservation
2026 static int gather_surplus_pages(struct hstate *h, int delta)
2027 __must_hold(&hugetlb_lock)
2029 struct list_head surplus_list;
2030 struct page *page, *tmp;
2032 int needed, allocated;
2033 bool alloc_ok = true;
2035 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2037 h->resv_huge_pages += delta;
2042 INIT_LIST_HEAD(&surplus_list);
2046 spin_unlock(&hugetlb_lock);
2047 for (i = 0; i < needed; i++) {
2048 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2049 NUMA_NO_NODE, NULL);
2054 list_add(&page->lru, &surplus_list);
2060 * After retaking hugetlb_lock, we need to recalculate 'needed'
2061 * because either resv_huge_pages or free_huge_pages may have changed.
2063 spin_lock(&hugetlb_lock);
2064 needed = (h->resv_huge_pages + delta) -
2065 (h->free_huge_pages + allocated);
2070 * We were not able to allocate enough pages to
2071 * satisfy the entire reservation so we free what
2072 * we've allocated so far.
2077 * The surplus_list now contains _at_least_ the number of extra pages
2078 * needed to accommodate the reservation. Add the appropriate number
2079 * of pages to the hugetlb pool and free the extras back to the buddy
2080 * allocator. Commit the entire reservation here to prevent another
2081 * process from stealing the pages as they are added to the pool but
2082 * before they are reserved.
2084 needed += allocated;
2085 h->resv_huge_pages += delta;
2088 /* Free the needed pages to the hugetlb pool */
2089 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2093 * This page is now managed by the hugetlb allocator and has
2094 * no users -- drop the buddy allocator's reference.
2096 put_page_testzero(page);
2097 VM_BUG_ON_PAGE(page_count(page), page);
2098 enqueue_huge_page(h, page);
2101 spin_unlock(&hugetlb_lock);
2103 /* Free unnecessary surplus pages to the buddy allocator */
2104 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2106 spin_lock(&hugetlb_lock);
2112 * This routine has two main purposes:
2113 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2114 * in unused_resv_pages. This corresponds to the prior adjustments made
2115 * to the associated reservation map.
2116 * 2) Free any unused surplus pages that may have been allocated to satisfy
2117 * the reservation. As many as unused_resv_pages may be freed.
2119 * Called with hugetlb_lock held. However, the lock could be dropped (and
2120 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2121 * we must make sure nobody else can claim pages we are in the process of
2122 * freeing. Do this by ensuring resv_huge_page always is greater than the
2123 * number of huge pages we plan to free when dropping the lock.
2125 static void return_unused_surplus_pages(struct hstate *h,
2126 unsigned long unused_resv_pages)
2128 unsigned long nr_pages;
2130 /* Cannot return gigantic pages currently */
2131 if (hstate_is_gigantic(h))
2135 * Part (or even all) of the reservation could have been backed
2136 * by pre-allocated pages. Only free surplus pages.
2138 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2141 * We want to release as many surplus pages as possible, spread
2142 * evenly across all nodes with memory. Iterate across these nodes
2143 * until we can no longer free unreserved surplus pages. This occurs
2144 * when the nodes with surplus pages have no free pages.
2145 * free_pool_huge_page() will balance the freed pages across the
2146 * on-line nodes with memory and will handle the hstate accounting.
2148 * Note that we decrement resv_huge_pages as we free the pages. If
2149 * we drop the lock, resv_huge_pages will still be sufficiently large
2150 * to cover subsequent pages we may free.
2152 while (nr_pages--) {
2153 h->resv_huge_pages--;
2154 unused_resv_pages--;
2155 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2157 cond_resched_lock(&hugetlb_lock);
2161 /* Fully uncommit the reservation */
2162 h->resv_huge_pages -= unused_resv_pages;
2167 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2168 * are used by the huge page allocation routines to manage reservations.
2170 * vma_needs_reservation is called to determine if the huge page at addr
2171 * within the vma has an associated reservation. If a reservation is
2172 * needed, the value 1 is returned. The caller is then responsible for
2173 * managing the global reservation and subpool usage counts. After
2174 * the huge page has been allocated, vma_commit_reservation is called
2175 * to add the page to the reservation map. If the page allocation fails,
2176 * the reservation must be ended instead of committed. vma_end_reservation
2177 * is called in such cases.
2179 * In the normal case, vma_commit_reservation returns the same value
2180 * as the preceding vma_needs_reservation call. The only time this
2181 * is not the case is if a reserve map was changed between calls. It
2182 * is the responsibility of the caller to notice the difference and
2183 * take appropriate action.
2185 * vma_add_reservation is used in error paths where a reservation must
2186 * be restored when a newly allocated huge page must be freed. It is
2187 * to be called after calling vma_needs_reservation to determine if a
2188 * reservation exists.
2190 enum vma_resv_mode {
2196 static long __vma_reservation_common(struct hstate *h,
2197 struct vm_area_struct *vma, unsigned long addr,
2198 enum vma_resv_mode mode)
2200 struct resv_map *resv;
2203 long dummy_out_regions_needed;
2205 resv = vma_resv_map(vma);
2209 idx = vma_hugecache_offset(h, vma, addr);
2211 case VMA_NEEDS_RESV:
2212 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2213 /* We assume that vma_reservation_* routines always operate on
2214 * 1 page, and that adding to resv map a 1 page entry can only
2215 * ever require 1 region.
2217 VM_BUG_ON(dummy_out_regions_needed != 1);
2219 case VMA_COMMIT_RESV:
2220 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2221 /* region_add calls of range 1 should never fail. */
2225 region_abort(resv, idx, idx + 1, 1);
2229 if (vma->vm_flags & VM_MAYSHARE) {
2230 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2231 /* region_add calls of range 1 should never fail. */
2234 region_abort(resv, idx, idx + 1, 1);
2235 ret = region_del(resv, idx, idx + 1);
2242 if (vma->vm_flags & VM_MAYSHARE)
2244 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2246 * In most cases, reserves always exist for private mappings.
2247 * However, a file associated with mapping could have been
2248 * hole punched or truncated after reserves were consumed.
2249 * As subsequent fault on such a range will not use reserves.
2250 * Subtle - The reserve map for private mappings has the
2251 * opposite meaning than that of shared mappings. If NO
2252 * entry is in the reserve map, it means a reservation exists.
2253 * If an entry exists in the reserve map, it means the
2254 * reservation has already been consumed. As a result, the
2255 * return value of this routine is the opposite of the
2256 * value returned from reserve map manipulation routines above.
2264 return ret < 0 ? ret : 0;
2267 static long vma_needs_reservation(struct hstate *h,
2268 struct vm_area_struct *vma, unsigned long addr)
2270 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2273 static long vma_commit_reservation(struct hstate *h,
2274 struct vm_area_struct *vma, unsigned long addr)
2276 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2279 static void vma_end_reservation(struct hstate *h,
2280 struct vm_area_struct *vma, unsigned long addr)
2282 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2285 static long vma_add_reservation(struct hstate *h,
2286 struct vm_area_struct *vma, unsigned long addr)
2288 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2292 * This routine is called to restore a reservation on error paths. In the
2293 * specific error paths, a huge page was allocated (via alloc_huge_page)
2294 * and is about to be freed. If a reservation for the page existed,
2295 * alloc_huge_page would have consumed the reservation and set PagePrivate
2296 * in the newly allocated page. When the page is freed via free_huge_page,
2297 * the global reservation count will be incremented if PagePrivate is set.
2298 * However, free_huge_page can not adjust the reserve map. Adjust the
2299 * reserve map here to be consistent with global reserve count adjustments
2300 * to be made by free_huge_page.
2302 static void restore_reserve_on_error(struct hstate *h,
2303 struct vm_area_struct *vma, unsigned long address,
2306 if (unlikely(PagePrivate(page))) {
2307 long rc = vma_needs_reservation(h, vma, address);
2309 if (unlikely(rc < 0)) {
2311 * Rare out of memory condition in reserve map
2312 * manipulation. Clear PagePrivate so that
2313 * global reserve count will not be incremented
2314 * by free_huge_page. This will make it appear
2315 * as though the reservation for this page was
2316 * consumed. This may prevent the task from
2317 * faulting in the page at a later time. This
2318 * is better than inconsistent global huge page
2319 * accounting of reserve counts.
2321 ClearPagePrivate(page);
2323 rc = vma_add_reservation(h, vma, address);
2324 if (unlikely(rc < 0))
2326 * See above comment about rare out of
2329 ClearPagePrivate(page);
2331 vma_end_reservation(h, vma, address);
2335 struct page *alloc_huge_page(struct vm_area_struct *vma,
2336 unsigned long addr, int avoid_reserve)
2338 struct hugepage_subpool *spool = subpool_vma(vma);
2339 struct hstate *h = hstate_vma(vma);
2341 long map_chg, map_commit;
2344 struct hugetlb_cgroup *h_cg;
2345 bool deferred_reserve;
2347 idx = hstate_index(h);
2349 * Examine the region/reserve map to determine if the process
2350 * has a reservation for the page to be allocated. A return
2351 * code of zero indicates a reservation exists (no change).
2353 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2355 return ERR_PTR(-ENOMEM);
2358 * Processes that did not create the mapping will have no
2359 * reserves as indicated by the region/reserve map. Check
2360 * that the allocation will not exceed the subpool limit.
2361 * Allocations for MAP_NORESERVE mappings also need to be
2362 * checked against any subpool limit.
2364 if (map_chg || avoid_reserve) {
2365 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2367 vma_end_reservation(h, vma, addr);
2368 return ERR_PTR(-ENOSPC);
2372 * Even though there was no reservation in the region/reserve
2373 * map, there could be reservations associated with the
2374 * subpool that can be used. This would be indicated if the
2375 * return value of hugepage_subpool_get_pages() is zero.
2376 * However, if avoid_reserve is specified we still avoid even
2377 * the subpool reservations.
2383 /* If this allocation is not consuming a reservation, charge it now.
2385 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2386 if (deferred_reserve) {
2387 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2388 idx, pages_per_huge_page(h), &h_cg);
2390 goto out_subpool_put;
2393 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2395 goto out_uncharge_cgroup_reservation;
2397 spin_lock(&hugetlb_lock);
2399 * glb_chg is passed to indicate whether or not a page must be taken
2400 * from the global free pool (global change). gbl_chg == 0 indicates
2401 * a reservation exists for the allocation.
2403 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2405 spin_unlock(&hugetlb_lock);
2406 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2408 goto out_uncharge_cgroup;
2409 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2410 SetPagePrivate(page);
2411 h->resv_huge_pages--;
2413 spin_lock(&hugetlb_lock);
2414 list_add(&page->lru, &h->hugepage_activelist);
2417 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2418 /* If allocation is not consuming a reservation, also store the
2419 * hugetlb_cgroup pointer on the page.
2421 if (deferred_reserve) {
2422 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2426 spin_unlock(&hugetlb_lock);
2428 set_page_private(page, (unsigned long)spool);
2430 map_commit = vma_commit_reservation(h, vma, addr);
2431 if (unlikely(map_chg > map_commit)) {
2433 * The page was added to the reservation map between
2434 * vma_needs_reservation and vma_commit_reservation.
2435 * This indicates a race with hugetlb_reserve_pages.
2436 * Adjust for the subpool count incremented above AND
2437 * in hugetlb_reserve_pages for the same page. Also,
2438 * the reservation count added in hugetlb_reserve_pages
2439 * no longer applies.
2443 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2444 hugetlb_acct_memory(h, -rsv_adjust);
2445 if (deferred_reserve)
2446 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2447 pages_per_huge_page(h), page);
2451 out_uncharge_cgroup:
2452 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2453 out_uncharge_cgroup_reservation:
2454 if (deferred_reserve)
2455 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2458 if (map_chg || avoid_reserve)
2459 hugepage_subpool_put_pages(spool, 1);
2460 vma_end_reservation(h, vma, addr);
2461 return ERR_PTR(-ENOSPC);
2464 int alloc_bootmem_huge_page(struct hstate *h)
2465 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2466 int __alloc_bootmem_huge_page(struct hstate *h)
2468 struct huge_bootmem_page *m;
2471 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2474 addr = memblock_alloc_try_nid_raw(
2475 huge_page_size(h), huge_page_size(h),
2476 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2479 * Use the beginning of the huge page to store the
2480 * huge_bootmem_page struct (until gather_bootmem
2481 * puts them into the mem_map).
2490 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2491 /* Put them into a private list first because mem_map is not up yet */
2492 INIT_LIST_HEAD(&m->list);
2493 list_add(&m->list, &huge_boot_pages);
2498 static void __init prep_compound_huge_page(struct page *page,
2501 if (unlikely(order > (MAX_ORDER - 1)))
2502 prep_compound_gigantic_page(page, order);
2504 prep_compound_page(page, order);
2507 /* Put bootmem huge pages into the standard lists after mem_map is up */
2508 static void __init gather_bootmem_prealloc(void)
2510 struct huge_bootmem_page *m;
2512 list_for_each_entry(m, &huge_boot_pages, list) {
2513 struct page *page = virt_to_page(m);
2514 struct hstate *h = m->hstate;
2516 WARN_ON(page_count(page) != 1);
2517 prep_compound_huge_page(page, h->order);
2518 WARN_ON(PageReserved(page));
2519 prep_new_huge_page(h, page, page_to_nid(page));
2520 put_page(page); /* free it into the hugepage allocator */
2523 * If we had gigantic hugepages allocated at boot time, we need
2524 * to restore the 'stolen' pages to totalram_pages in order to
2525 * fix confusing memory reports from free(1) and another
2526 * side-effects, like CommitLimit going negative.
2528 if (hstate_is_gigantic(h))
2529 adjust_managed_page_count(page, 1 << h->order);
2534 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2537 nodemask_t *node_alloc_noretry;
2539 if (!hstate_is_gigantic(h)) {
2541 * Bit mask controlling how hard we retry per-node allocations.
2542 * Ignore errors as lower level routines can deal with
2543 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2544 * time, we are likely in bigger trouble.
2546 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2549 /* allocations done at boot time */
2550 node_alloc_noretry = NULL;
2553 /* bit mask controlling how hard we retry per-node allocations */
2554 if (node_alloc_noretry)
2555 nodes_clear(*node_alloc_noretry);
2557 for (i = 0; i < h->max_huge_pages; ++i) {
2558 if (hstate_is_gigantic(h)) {
2559 if (hugetlb_cma_size) {
2560 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2563 if (!alloc_bootmem_huge_page(h))
2565 } else if (!alloc_pool_huge_page(h,
2566 &node_states[N_MEMORY],
2567 node_alloc_noretry))
2571 if (i < h->max_huge_pages) {
2574 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2575 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2576 h->max_huge_pages, buf, i);
2577 h->max_huge_pages = i;
2580 kfree(node_alloc_noretry);
2583 static void __init hugetlb_init_hstates(void)
2587 for_each_hstate(h) {
2588 if (minimum_order > huge_page_order(h))
2589 minimum_order = huge_page_order(h);
2591 /* oversize hugepages were init'ed in early boot */
2592 if (!hstate_is_gigantic(h))
2593 hugetlb_hstate_alloc_pages(h);
2595 VM_BUG_ON(minimum_order == UINT_MAX);
2598 static void __init report_hugepages(void)
2602 for_each_hstate(h) {
2605 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2606 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2607 buf, h->free_huge_pages);
2611 #ifdef CONFIG_HIGHMEM
2612 static void try_to_free_low(struct hstate *h, unsigned long count,
2613 nodemask_t *nodes_allowed)
2617 if (hstate_is_gigantic(h))
2620 for_each_node_mask(i, *nodes_allowed) {
2621 struct page *page, *next;
2622 struct list_head *freel = &h->hugepage_freelists[i];
2623 list_for_each_entry_safe(page, next, freel, lru) {
2624 if (count >= h->nr_huge_pages)
2626 if (PageHighMem(page))
2628 list_del(&page->lru);
2629 update_and_free_page(h, page);
2630 h->free_huge_pages--;
2631 h->free_huge_pages_node[page_to_nid(page)]--;
2636 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2637 nodemask_t *nodes_allowed)
2643 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2644 * balanced by operating on them in a round-robin fashion.
2645 * Returns 1 if an adjustment was made.
2647 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2652 VM_BUG_ON(delta != -1 && delta != 1);
2655 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2656 if (h->surplus_huge_pages_node[node])
2660 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2661 if (h->surplus_huge_pages_node[node] <
2662 h->nr_huge_pages_node[node])
2669 h->surplus_huge_pages += delta;
2670 h->surplus_huge_pages_node[node] += delta;
2674 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2675 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2676 nodemask_t *nodes_allowed)
2678 unsigned long min_count, ret;
2679 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2682 * Bit mask controlling how hard we retry per-node allocations.
2683 * If we can not allocate the bit mask, do not attempt to allocate
2684 * the requested huge pages.
2686 if (node_alloc_noretry)
2687 nodes_clear(*node_alloc_noretry);
2691 spin_lock(&hugetlb_lock);
2694 * Check for a node specific request.
2695 * Changing node specific huge page count may require a corresponding
2696 * change to the global count. In any case, the passed node mask
2697 * (nodes_allowed) will restrict alloc/free to the specified node.
2699 if (nid != NUMA_NO_NODE) {
2700 unsigned long old_count = count;
2702 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2704 * User may have specified a large count value which caused the
2705 * above calculation to overflow. In this case, they wanted
2706 * to allocate as many huge pages as possible. Set count to
2707 * largest possible value to align with their intention.
2709 if (count < old_count)
2714 * Gigantic pages runtime allocation depend on the capability for large
2715 * page range allocation.
2716 * If the system does not provide this feature, return an error when
2717 * the user tries to allocate gigantic pages but let the user free the
2718 * boottime allocated gigantic pages.
2720 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2721 if (count > persistent_huge_pages(h)) {
2722 spin_unlock(&hugetlb_lock);
2723 NODEMASK_FREE(node_alloc_noretry);
2726 /* Fall through to decrease pool */
2730 * Increase the pool size
2731 * First take pages out of surplus state. Then make up the
2732 * remaining difference by allocating fresh huge pages.
2734 * We might race with alloc_surplus_huge_page() here and be unable
2735 * to convert a surplus huge page to a normal huge page. That is
2736 * not critical, though, it just means the overall size of the
2737 * pool might be one hugepage larger than it needs to be, but
2738 * within all the constraints specified by the sysctls.
2740 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2741 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2745 while (count > persistent_huge_pages(h)) {
2747 * If this allocation races such that we no longer need the
2748 * page, free_huge_page will handle it by freeing the page
2749 * and reducing the surplus.
2751 spin_unlock(&hugetlb_lock);
2753 /* yield cpu to avoid soft lockup */
2756 ret = alloc_pool_huge_page(h, nodes_allowed,
2757 node_alloc_noretry);
2758 spin_lock(&hugetlb_lock);
2762 /* Bail for signals. Probably ctrl-c from user */
2763 if (signal_pending(current))
2768 * Decrease the pool size
2769 * First return free pages to the buddy allocator (being careful
2770 * to keep enough around to satisfy reservations). Then place
2771 * pages into surplus state as needed so the pool will shrink
2772 * to the desired size as pages become free.
2774 * By placing pages into the surplus state independent of the
2775 * overcommit value, we are allowing the surplus pool size to
2776 * exceed overcommit. There are few sane options here. Since
2777 * alloc_surplus_huge_page() is checking the global counter,
2778 * though, we'll note that we're not allowed to exceed surplus
2779 * and won't grow the pool anywhere else. Not until one of the
2780 * sysctls are changed, or the surplus pages go out of use.
2782 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2783 min_count = max(count, min_count);
2784 try_to_free_low(h, min_count, nodes_allowed);
2785 while (min_count < persistent_huge_pages(h)) {
2786 if (!free_pool_huge_page(h, nodes_allowed, 0))
2788 cond_resched_lock(&hugetlb_lock);
2790 while (count < persistent_huge_pages(h)) {
2791 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2795 h->max_huge_pages = persistent_huge_pages(h);
2796 spin_unlock(&hugetlb_lock);
2798 NODEMASK_FREE(node_alloc_noretry);
2803 #define HSTATE_ATTR_RO(_name) \
2804 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2806 #define HSTATE_ATTR(_name) \
2807 static struct kobj_attribute _name##_attr = \
2808 __ATTR(_name, 0644, _name##_show, _name##_store)
2810 static struct kobject *hugepages_kobj;
2811 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2813 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2815 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2819 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2820 if (hstate_kobjs[i] == kobj) {
2822 *nidp = NUMA_NO_NODE;
2826 return kobj_to_node_hstate(kobj, nidp);
2829 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2830 struct kobj_attribute *attr, char *buf)
2833 unsigned long nr_huge_pages;
2836 h = kobj_to_hstate(kobj, &nid);
2837 if (nid == NUMA_NO_NODE)
2838 nr_huge_pages = h->nr_huge_pages;
2840 nr_huge_pages = h->nr_huge_pages_node[nid];
2842 return sprintf(buf, "%lu\n", nr_huge_pages);
2845 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2846 struct hstate *h, int nid,
2847 unsigned long count, size_t len)
2850 nodemask_t nodes_allowed, *n_mask;
2852 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2855 if (nid == NUMA_NO_NODE) {
2857 * global hstate attribute
2859 if (!(obey_mempolicy &&
2860 init_nodemask_of_mempolicy(&nodes_allowed)))
2861 n_mask = &node_states[N_MEMORY];
2863 n_mask = &nodes_allowed;
2866 * Node specific request. count adjustment happens in
2867 * set_max_huge_pages() after acquiring hugetlb_lock.
2869 init_nodemask_of_node(&nodes_allowed, nid);
2870 n_mask = &nodes_allowed;
2873 err = set_max_huge_pages(h, count, nid, n_mask);
2875 return err ? err : len;
2878 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2879 struct kobject *kobj, const char *buf,
2883 unsigned long count;
2887 err = kstrtoul(buf, 10, &count);
2891 h = kobj_to_hstate(kobj, &nid);
2892 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2895 static ssize_t nr_hugepages_show(struct kobject *kobj,
2896 struct kobj_attribute *attr, char *buf)
2898 return nr_hugepages_show_common(kobj, attr, buf);
2901 static ssize_t nr_hugepages_store(struct kobject *kobj,
2902 struct kobj_attribute *attr, const char *buf, size_t len)
2904 return nr_hugepages_store_common(false, kobj, buf, len);
2906 HSTATE_ATTR(nr_hugepages);
2911 * hstate attribute for optionally mempolicy-based constraint on persistent
2912 * huge page alloc/free.
2914 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2915 struct kobj_attribute *attr, char *buf)
2917 return nr_hugepages_show_common(kobj, attr, buf);
2920 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2921 struct kobj_attribute *attr, const char *buf, size_t len)
2923 return nr_hugepages_store_common(true, kobj, buf, len);
2925 HSTATE_ATTR(nr_hugepages_mempolicy);
2929 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2930 struct kobj_attribute *attr, char *buf)
2932 struct hstate *h = kobj_to_hstate(kobj, NULL);
2933 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2936 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2937 struct kobj_attribute *attr, const char *buf, size_t count)
2940 unsigned long input;
2941 struct hstate *h = kobj_to_hstate(kobj, NULL);
2943 if (hstate_is_gigantic(h))
2946 err = kstrtoul(buf, 10, &input);
2950 spin_lock(&hugetlb_lock);
2951 h->nr_overcommit_huge_pages = input;
2952 spin_unlock(&hugetlb_lock);
2956 HSTATE_ATTR(nr_overcommit_hugepages);
2958 static ssize_t free_hugepages_show(struct kobject *kobj,
2959 struct kobj_attribute *attr, char *buf)
2962 unsigned long free_huge_pages;
2965 h = kobj_to_hstate(kobj, &nid);
2966 if (nid == NUMA_NO_NODE)
2967 free_huge_pages = h->free_huge_pages;
2969 free_huge_pages = h->free_huge_pages_node[nid];
2971 return sprintf(buf, "%lu\n", free_huge_pages);
2973 HSTATE_ATTR_RO(free_hugepages);
2975 static ssize_t resv_hugepages_show(struct kobject *kobj,
2976 struct kobj_attribute *attr, char *buf)
2978 struct hstate *h = kobj_to_hstate(kobj, NULL);
2979 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2981 HSTATE_ATTR_RO(resv_hugepages);
2983 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2984 struct kobj_attribute *attr, char *buf)
2987 unsigned long surplus_huge_pages;
2990 h = kobj_to_hstate(kobj, &nid);
2991 if (nid == NUMA_NO_NODE)
2992 surplus_huge_pages = h->surplus_huge_pages;
2994 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2996 return sprintf(buf, "%lu\n", surplus_huge_pages);
2998 HSTATE_ATTR_RO(surplus_hugepages);
3000 static struct attribute *hstate_attrs[] = {
3001 &nr_hugepages_attr.attr,
3002 &nr_overcommit_hugepages_attr.attr,
3003 &free_hugepages_attr.attr,
3004 &resv_hugepages_attr.attr,
3005 &surplus_hugepages_attr.attr,
3007 &nr_hugepages_mempolicy_attr.attr,
3012 static const struct attribute_group hstate_attr_group = {
3013 .attrs = hstate_attrs,
3016 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3017 struct kobject **hstate_kobjs,
3018 const struct attribute_group *hstate_attr_group)
3021 int hi = hstate_index(h);
3023 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3024 if (!hstate_kobjs[hi])
3027 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3029 kobject_put(hstate_kobjs[hi]);
3034 static void __init hugetlb_sysfs_init(void)
3039 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3040 if (!hugepages_kobj)
3043 for_each_hstate(h) {
3044 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3045 hstate_kobjs, &hstate_attr_group);
3047 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3054 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3055 * with node devices in node_devices[] using a parallel array. The array
3056 * index of a node device or _hstate == node id.
3057 * This is here to avoid any static dependency of the node device driver, in
3058 * the base kernel, on the hugetlb module.
3060 struct node_hstate {
3061 struct kobject *hugepages_kobj;
3062 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3064 static struct node_hstate node_hstates[MAX_NUMNODES];
3067 * A subset of global hstate attributes for node devices
3069 static struct attribute *per_node_hstate_attrs[] = {
3070 &nr_hugepages_attr.attr,
3071 &free_hugepages_attr.attr,
3072 &surplus_hugepages_attr.attr,
3076 static const struct attribute_group per_node_hstate_attr_group = {
3077 .attrs = per_node_hstate_attrs,
3081 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3082 * Returns node id via non-NULL nidp.
3084 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3088 for (nid = 0; nid < nr_node_ids; nid++) {
3089 struct node_hstate *nhs = &node_hstates[nid];
3091 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3092 if (nhs->hstate_kobjs[i] == kobj) {
3104 * Unregister hstate attributes from a single node device.
3105 * No-op if no hstate attributes attached.
3107 static void hugetlb_unregister_node(struct node *node)
3110 struct node_hstate *nhs = &node_hstates[node->dev.id];
3112 if (!nhs->hugepages_kobj)
3113 return; /* no hstate attributes */
3115 for_each_hstate(h) {
3116 int idx = hstate_index(h);
3117 if (nhs->hstate_kobjs[idx]) {
3118 kobject_put(nhs->hstate_kobjs[idx]);
3119 nhs->hstate_kobjs[idx] = NULL;
3123 kobject_put(nhs->hugepages_kobj);
3124 nhs->hugepages_kobj = NULL;
3129 * Register hstate attributes for a single node device.
3130 * No-op if attributes already registered.
3132 static void hugetlb_register_node(struct node *node)
3135 struct node_hstate *nhs = &node_hstates[node->dev.id];
3138 if (nhs->hugepages_kobj)
3139 return; /* already allocated */
3141 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3143 if (!nhs->hugepages_kobj)
3146 for_each_hstate(h) {
3147 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3149 &per_node_hstate_attr_group);
3151 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3152 h->name, node->dev.id);
3153 hugetlb_unregister_node(node);
3160 * hugetlb init time: register hstate attributes for all registered node
3161 * devices of nodes that have memory. All on-line nodes should have
3162 * registered their associated device by this time.
3164 static void __init hugetlb_register_all_nodes(void)
3168 for_each_node_state(nid, N_MEMORY) {
3169 struct node *node = node_devices[nid];
3170 if (node->dev.id == nid)
3171 hugetlb_register_node(node);
3175 * Let the node device driver know we're here so it can
3176 * [un]register hstate attributes on node hotplug.
3178 register_hugetlbfs_with_node(hugetlb_register_node,
3179 hugetlb_unregister_node);
3181 #else /* !CONFIG_NUMA */
3183 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3191 static void hugetlb_register_all_nodes(void) { }
3195 static int __init hugetlb_init(void)
3199 if (!hugepages_supported()) {
3200 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3201 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3206 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3207 * architectures depend on setup being done here.
3209 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3210 if (!parsed_default_hugepagesz) {
3212 * If we did not parse a default huge page size, set
3213 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3214 * number of huge pages for this default size was implicitly
3215 * specified, set that here as well.
3216 * Note that the implicit setting will overwrite an explicit
3217 * setting. A warning will be printed in this case.
3219 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3220 if (default_hstate_max_huge_pages) {
3221 if (default_hstate.max_huge_pages) {
3224 string_get_size(huge_page_size(&default_hstate),
3225 1, STRING_UNITS_2, buf, 32);
3226 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3227 default_hstate.max_huge_pages, buf);
3228 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3229 default_hstate_max_huge_pages);
3231 default_hstate.max_huge_pages =
3232 default_hstate_max_huge_pages;
3236 hugetlb_cma_check();
3237 hugetlb_init_hstates();
3238 gather_bootmem_prealloc();
3241 hugetlb_sysfs_init();
3242 hugetlb_register_all_nodes();
3243 hugetlb_cgroup_file_init();
3246 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3248 num_fault_mutexes = 1;
3250 hugetlb_fault_mutex_table =
3251 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3253 BUG_ON(!hugetlb_fault_mutex_table);
3255 for (i = 0; i < num_fault_mutexes; i++)
3256 mutex_init(&hugetlb_fault_mutex_table[i]);
3259 subsys_initcall(hugetlb_init);
3261 /* Overwritten by architectures with more huge page sizes */
3262 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3264 return size == HPAGE_SIZE;
3267 void __init hugetlb_add_hstate(unsigned int order)
3272 if (size_to_hstate(PAGE_SIZE << order)) {
3275 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3277 h = &hstates[hugetlb_max_hstate++];
3279 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3280 h->nr_huge_pages = 0;
3281 h->free_huge_pages = 0;
3282 for (i = 0; i < MAX_NUMNODES; ++i)
3283 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3284 INIT_LIST_HEAD(&h->hugepage_activelist);
3285 h->next_nid_to_alloc = first_memory_node;
3286 h->next_nid_to_free = first_memory_node;
3287 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3288 huge_page_size(h)/1024);
3294 * hugepages command line processing
3295 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3296 * specification. If not, ignore the hugepages value. hugepages can also
3297 * be the first huge page command line option in which case it implicitly
3298 * specifies the number of huge pages for the default size.
3300 static int __init hugepages_setup(char *s)
3303 static unsigned long *last_mhp;
3305 if (!parsed_valid_hugepagesz) {
3306 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3307 parsed_valid_hugepagesz = true;
3312 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3313 * yet, so this hugepages= parameter goes to the "default hstate".
3314 * Otherwise, it goes with the previously parsed hugepagesz or
3315 * default_hugepagesz.
3317 else if (!hugetlb_max_hstate)
3318 mhp = &default_hstate_max_huge_pages;
3320 mhp = &parsed_hstate->max_huge_pages;
3322 if (mhp == last_mhp) {
3323 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3327 if (sscanf(s, "%lu", mhp) <= 0)
3331 * Global state is always initialized later in hugetlb_init.
3332 * But we need to allocate >= MAX_ORDER hstates here early to still
3333 * use the bootmem allocator.
3335 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3336 hugetlb_hstate_alloc_pages(parsed_hstate);
3342 __setup("hugepages=", hugepages_setup);
3345 * hugepagesz command line processing
3346 * A specific huge page size can only be specified once with hugepagesz.
3347 * hugepagesz is followed by hugepages on the command line. The global
3348 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3349 * hugepagesz argument was valid.
3351 static int __init hugepagesz_setup(char *s)
3356 parsed_valid_hugepagesz = false;
3357 size = (unsigned long)memparse(s, NULL);
3359 if (!arch_hugetlb_valid_size(size)) {
3360 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3364 h = size_to_hstate(size);
3367 * hstate for this size already exists. This is normally
3368 * an error, but is allowed if the existing hstate is the
3369 * default hstate. More specifically, it is only allowed if
3370 * the number of huge pages for the default hstate was not
3371 * previously specified.
3373 if (!parsed_default_hugepagesz || h != &default_hstate ||
3374 default_hstate.max_huge_pages) {
3375 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3380 * No need to call hugetlb_add_hstate() as hstate already
3381 * exists. But, do set parsed_hstate so that a following
3382 * hugepages= parameter will be applied to this hstate.
3385 parsed_valid_hugepagesz = true;
3389 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3390 parsed_valid_hugepagesz = true;
3393 __setup("hugepagesz=", hugepagesz_setup);
3396 * default_hugepagesz command line input
3397 * Only one instance of default_hugepagesz allowed on command line.
3399 static int __init default_hugepagesz_setup(char *s)
3403 parsed_valid_hugepagesz = false;
3404 if (parsed_default_hugepagesz) {
3405 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3409 size = (unsigned long)memparse(s, NULL);
3411 if (!arch_hugetlb_valid_size(size)) {
3412 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3416 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3417 parsed_valid_hugepagesz = true;
3418 parsed_default_hugepagesz = true;
3419 default_hstate_idx = hstate_index(size_to_hstate(size));
3422 * The number of default huge pages (for this size) could have been
3423 * specified as the first hugetlb parameter: hugepages=X. If so,
3424 * then default_hstate_max_huge_pages is set. If the default huge
3425 * page size is gigantic (>= MAX_ORDER), then the pages must be
3426 * allocated here from bootmem allocator.
3428 if (default_hstate_max_huge_pages) {
3429 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3430 if (hstate_is_gigantic(&default_hstate))
3431 hugetlb_hstate_alloc_pages(&default_hstate);
3432 default_hstate_max_huge_pages = 0;
3437 __setup("default_hugepagesz=", default_hugepagesz_setup);
3439 static unsigned int allowed_mems_nr(struct hstate *h)
3442 unsigned int nr = 0;
3443 nodemask_t *mpol_allowed;
3444 unsigned int *array = h->free_huge_pages_node;
3445 gfp_t gfp_mask = htlb_alloc_mask(h);
3447 mpol_allowed = policy_nodemask_current(gfp_mask);
3449 for_each_node_mask(node, cpuset_current_mems_allowed) {
3450 if (!mpol_allowed ||
3451 (mpol_allowed && node_isset(node, *mpol_allowed)))
3458 #ifdef CONFIG_SYSCTL
3459 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3460 void *buffer, size_t *length,
3461 loff_t *ppos, unsigned long *out)
3463 struct ctl_table dup_table;
3466 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3467 * can duplicate the @table and alter the duplicate of it.
3470 dup_table.data = out;
3472 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3475 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3476 struct ctl_table *table, int write,
3477 void *buffer, size_t *length, loff_t *ppos)
3479 struct hstate *h = &default_hstate;
3480 unsigned long tmp = h->max_huge_pages;
3483 if (!hugepages_supported())
3486 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3492 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3493 NUMA_NO_NODE, tmp, *length);
3498 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3499 void *buffer, size_t *length, loff_t *ppos)
3502 return hugetlb_sysctl_handler_common(false, table, write,
3503 buffer, length, ppos);
3507 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3508 void *buffer, size_t *length, loff_t *ppos)
3510 return hugetlb_sysctl_handler_common(true, table, write,
3511 buffer, length, ppos);
3513 #endif /* CONFIG_NUMA */
3515 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3516 void *buffer, size_t *length, loff_t *ppos)
3518 struct hstate *h = &default_hstate;
3522 if (!hugepages_supported())
3525 tmp = h->nr_overcommit_huge_pages;
3527 if (write && hstate_is_gigantic(h))
3530 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3536 spin_lock(&hugetlb_lock);
3537 h->nr_overcommit_huge_pages = tmp;
3538 spin_unlock(&hugetlb_lock);
3544 #endif /* CONFIG_SYSCTL */
3546 void hugetlb_report_meminfo(struct seq_file *m)
3549 unsigned long total = 0;
3551 if (!hugepages_supported())
3554 for_each_hstate(h) {
3555 unsigned long count = h->nr_huge_pages;
3557 total += (PAGE_SIZE << huge_page_order(h)) * count;
3559 if (h == &default_hstate)
3561 "HugePages_Total: %5lu\n"
3562 "HugePages_Free: %5lu\n"
3563 "HugePages_Rsvd: %5lu\n"
3564 "HugePages_Surp: %5lu\n"
3565 "Hugepagesize: %8lu kB\n",
3569 h->surplus_huge_pages,
3570 (PAGE_SIZE << huge_page_order(h)) / 1024);
3573 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3576 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3578 struct hstate *h = &default_hstate;
3580 if (!hugepages_supported())
3583 return sysfs_emit_at(buf, len,
3584 "Node %d HugePages_Total: %5u\n"
3585 "Node %d HugePages_Free: %5u\n"
3586 "Node %d HugePages_Surp: %5u\n",
3587 nid, h->nr_huge_pages_node[nid],
3588 nid, h->free_huge_pages_node[nid],
3589 nid, h->surplus_huge_pages_node[nid]);
3592 void hugetlb_show_meminfo(void)
3597 if (!hugepages_supported())
3600 for_each_node_state(nid, N_MEMORY)
3602 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3604 h->nr_huge_pages_node[nid],
3605 h->free_huge_pages_node[nid],
3606 h->surplus_huge_pages_node[nid],
3607 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3610 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3612 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3613 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3616 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3617 unsigned long hugetlb_total_pages(void)
3620 unsigned long nr_total_pages = 0;
3623 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3624 return nr_total_pages;
3627 static int hugetlb_acct_memory(struct hstate *h, long delta)
3631 spin_lock(&hugetlb_lock);
3633 * When cpuset is configured, it breaks the strict hugetlb page
3634 * reservation as the accounting is done on a global variable. Such
3635 * reservation is completely rubbish in the presence of cpuset because
3636 * the reservation is not checked against page availability for the
3637 * current cpuset. Application can still potentially OOM'ed by kernel
3638 * with lack of free htlb page in cpuset that the task is in.
3639 * Attempt to enforce strict accounting with cpuset is almost
3640 * impossible (or too ugly) because cpuset is too fluid that
3641 * task or memory node can be dynamically moved between cpusets.
3643 * The change of semantics for shared hugetlb mapping with cpuset is
3644 * undesirable. However, in order to preserve some of the semantics,
3645 * we fall back to check against current free page availability as
3646 * a best attempt and hopefully to minimize the impact of changing
3647 * semantics that cpuset has.
3649 * Apart from cpuset, we also have memory policy mechanism that
3650 * also determines from which node the kernel will allocate memory
3651 * in a NUMA system. So similar to cpuset, we also should consider
3652 * the memory policy of the current task. Similar to the description
3656 if (gather_surplus_pages(h, delta) < 0)
3659 if (delta > allowed_mems_nr(h)) {
3660 return_unused_surplus_pages(h, delta);
3667 return_unused_surplus_pages(h, (unsigned long) -delta);
3670 spin_unlock(&hugetlb_lock);
3674 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3676 struct resv_map *resv = vma_resv_map(vma);
3679 * This new VMA should share its siblings reservation map if present.
3680 * The VMA will only ever have a valid reservation map pointer where
3681 * it is being copied for another still existing VMA. As that VMA
3682 * has a reference to the reservation map it cannot disappear until
3683 * after this open call completes. It is therefore safe to take a
3684 * new reference here without additional locking.
3686 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3687 kref_get(&resv->refs);
3690 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3692 struct hstate *h = hstate_vma(vma);
3693 struct resv_map *resv = vma_resv_map(vma);
3694 struct hugepage_subpool *spool = subpool_vma(vma);
3695 unsigned long reserve, start, end;
3698 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3701 start = vma_hugecache_offset(h, vma, vma->vm_start);
3702 end = vma_hugecache_offset(h, vma, vma->vm_end);
3704 reserve = (end - start) - region_count(resv, start, end);
3705 hugetlb_cgroup_uncharge_counter(resv, start, end);
3708 * Decrement reserve counts. The global reserve count may be
3709 * adjusted if the subpool has a minimum size.
3711 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3712 hugetlb_acct_memory(h, -gbl_reserve);
3715 kref_put(&resv->refs, resv_map_release);
3718 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3720 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3725 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3727 struct hstate *hstate = hstate_vma(vma);
3729 return 1UL << huge_page_shift(hstate);
3733 * We cannot handle pagefaults against hugetlb pages at all. They cause
3734 * handle_mm_fault() to try to instantiate regular-sized pages in the
3735 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3738 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3745 * When a new function is introduced to vm_operations_struct and added
3746 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3747 * This is because under System V memory model, mappings created via
3748 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3749 * their original vm_ops are overwritten with shm_vm_ops.
3751 const struct vm_operations_struct hugetlb_vm_ops = {
3752 .fault = hugetlb_vm_op_fault,
3753 .open = hugetlb_vm_op_open,
3754 .close = hugetlb_vm_op_close,
3755 .split = hugetlb_vm_op_split,
3756 .pagesize = hugetlb_vm_op_pagesize,
3759 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3765 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3766 vma->vm_page_prot)));
3768 entry = huge_pte_wrprotect(mk_huge_pte(page,
3769 vma->vm_page_prot));
3771 entry = pte_mkyoung(entry);
3772 entry = pte_mkhuge(entry);
3773 entry = arch_make_huge_pte(entry, vma, page, writable);
3778 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3779 unsigned long address, pte_t *ptep)
3783 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3784 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3785 update_mmu_cache(vma, address, ptep);
3788 bool is_hugetlb_entry_migration(pte_t pte)
3792 if (huge_pte_none(pte) || pte_present(pte))
3794 swp = pte_to_swp_entry(pte);
3795 if (is_migration_entry(swp))
3801 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3805 if (huge_pte_none(pte) || pte_present(pte))
3807 swp = pte_to_swp_entry(pte);
3808 if (is_hwpoison_entry(swp))
3814 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3815 struct vm_area_struct *vma)
3817 pte_t *src_pte, *dst_pte, entry, dst_entry;
3818 struct page *ptepage;
3821 struct hstate *h = hstate_vma(vma);
3822 unsigned long sz = huge_page_size(h);
3823 struct address_space *mapping = vma->vm_file->f_mapping;
3824 struct mmu_notifier_range range;
3827 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3830 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3833 mmu_notifier_invalidate_range_start(&range);
3836 * For shared mappings i_mmap_rwsem must be held to call
3837 * huge_pte_alloc, otherwise the returned ptep could go
3838 * away if part of a shared pmd and another thread calls
3841 i_mmap_lock_read(mapping);
3844 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3845 spinlock_t *src_ptl, *dst_ptl;
3846 src_pte = huge_pte_offset(src, addr, sz);
3849 dst_pte = huge_pte_alloc(dst, addr, sz);
3856 * If the pagetables are shared don't copy or take references.
3857 * dst_pte == src_pte is the common case of src/dest sharing.
3859 * However, src could have 'unshared' and dst shares with
3860 * another vma. If dst_pte !none, this implies sharing.
3861 * Check here before taking page table lock, and once again
3862 * after taking the lock below.
3864 dst_entry = huge_ptep_get(dst_pte);
3865 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3868 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3869 src_ptl = huge_pte_lockptr(h, src, src_pte);
3870 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3871 entry = huge_ptep_get(src_pte);
3872 dst_entry = huge_ptep_get(dst_pte);
3873 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3875 * Skip if src entry none. Also, skip in the
3876 * unlikely case dst entry !none as this implies
3877 * sharing with another vma.
3880 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3881 is_hugetlb_entry_hwpoisoned(entry))) {
3882 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3884 if (is_write_migration_entry(swp_entry) && cow) {
3886 * COW mappings require pages in both
3887 * parent and child to be set to read.
3889 make_migration_entry_read(&swp_entry);
3890 entry = swp_entry_to_pte(swp_entry);
3891 set_huge_swap_pte_at(src, addr, src_pte,
3894 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3898 * No need to notify as we are downgrading page
3899 * table protection not changing it to point
3902 * See Documentation/vm/mmu_notifier.rst
3904 huge_ptep_set_wrprotect(src, addr, src_pte);
3906 entry = huge_ptep_get(src_pte);
3907 ptepage = pte_page(entry);
3909 page_dup_rmap(ptepage, true);
3910 set_huge_pte_at(dst, addr, dst_pte, entry);
3911 hugetlb_count_add(pages_per_huge_page(h), dst);
3913 spin_unlock(src_ptl);
3914 spin_unlock(dst_ptl);
3918 mmu_notifier_invalidate_range_end(&range);
3920 i_mmap_unlock_read(mapping);
3925 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3926 unsigned long start, unsigned long end,
3927 struct page *ref_page)
3929 struct mm_struct *mm = vma->vm_mm;
3930 unsigned long address;
3935 struct hstate *h = hstate_vma(vma);
3936 unsigned long sz = huge_page_size(h);
3937 struct mmu_notifier_range range;
3939 WARN_ON(!is_vm_hugetlb_page(vma));
3940 BUG_ON(start & ~huge_page_mask(h));
3941 BUG_ON(end & ~huge_page_mask(h));
3944 * This is a hugetlb vma, all the pte entries should point
3947 tlb_change_page_size(tlb, sz);
3948 tlb_start_vma(tlb, vma);
3951 * If sharing possible, alert mmu notifiers of worst case.
3953 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3955 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3956 mmu_notifier_invalidate_range_start(&range);
3958 for (; address < end; address += sz) {
3959 ptep = huge_pte_offset(mm, address, sz);
3963 ptl = huge_pte_lock(h, mm, ptep);
3964 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3967 * We just unmapped a page of PMDs by clearing a PUD.
3968 * The caller's TLB flush range should cover this area.
3973 pte = huge_ptep_get(ptep);
3974 if (huge_pte_none(pte)) {
3980 * Migrating hugepage or HWPoisoned hugepage is already
3981 * unmapped and its refcount is dropped, so just clear pte here.
3983 if (unlikely(!pte_present(pte))) {
3984 huge_pte_clear(mm, address, ptep, sz);
3989 page = pte_page(pte);
3991 * If a reference page is supplied, it is because a specific
3992 * page is being unmapped, not a range. Ensure the page we
3993 * are about to unmap is the actual page of interest.
3996 if (page != ref_page) {
4001 * Mark the VMA as having unmapped its page so that
4002 * future faults in this VMA will fail rather than
4003 * looking like data was lost
4005 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4008 pte = huge_ptep_get_and_clear(mm, address, ptep);
4009 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4010 if (huge_pte_dirty(pte))
4011 set_page_dirty(page);
4013 hugetlb_count_sub(pages_per_huge_page(h), mm);
4014 page_remove_rmap(page, true);
4017 tlb_remove_page_size(tlb, page, huge_page_size(h));
4019 * Bail out after unmapping reference page if supplied
4024 mmu_notifier_invalidate_range_end(&range);
4025 tlb_end_vma(tlb, vma);
4028 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4029 struct vm_area_struct *vma, unsigned long start,
4030 unsigned long end, struct page *ref_page)
4032 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4035 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4036 * test will fail on a vma being torn down, and not grab a page table
4037 * on its way out. We're lucky that the flag has such an appropriate
4038 * name, and can in fact be safely cleared here. We could clear it
4039 * before the __unmap_hugepage_range above, but all that's necessary
4040 * is to clear it before releasing the i_mmap_rwsem. This works
4041 * because in the context this is called, the VMA is about to be
4042 * destroyed and the i_mmap_rwsem is held.
4044 vma->vm_flags &= ~VM_MAYSHARE;
4047 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4048 unsigned long end, struct page *ref_page)
4050 struct mm_struct *mm;
4051 struct mmu_gather tlb;
4052 unsigned long tlb_start = start;
4053 unsigned long tlb_end = end;
4056 * If shared PMDs were possibly used within this vma range, adjust
4057 * start/end for worst case tlb flushing.
4058 * Note that we can not be sure if PMDs are shared until we try to
4059 * unmap pages. However, we want to make sure TLB flushing covers
4060 * the largest possible range.
4062 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4066 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4067 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4068 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4072 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4073 * mappping it owns the reserve page for. The intention is to unmap the page
4074 * from other VMAs and let the children be SIGKILLed if they are faulting the
4077 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4078 struct page *page, unsigned long address)
4080 struct hstate *h = hstate_vma(vma);
4081 struct vm_area_struct *iter_vma;
4082 struct address_space *mapping;
4086 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4087 * from page cache lookup which is in HPAGE_SIZE units.
4089 address = address & huge_page_mask(h);
4090 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4092 mapping = vma->vm_file->f_mapping;
4095 * Take the mapping lock for the duration of the table walk. As
4096 * this mapping should be shared between all the VMAs,
4097 * __unmap_hugepage_range() is called as the lock is already held
4099 i_mmap_lock_write(mapping);
4100 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4101 /* Do not unmap the current VMA */
4102 if (iter_vma == vma)
4106 * Shared VMAs have their own reserves and do not affect
4107 * MAP_PRIVATE accounting but it is possible that a shared
4108 * VMA is using the same page so check and skip such VMAs.
4110 if (iter_vma->vm_flags & VM_MAYSHARE)
4114 * Unmap the page from other VMAs without their own reserves.
4115 * They get marked to be SIGKILLed if they fault in these
4116 * areas. This is because a future no-page fault on this VMA
4117 * could insert a zeroed page instead of the data existing
4118 * from the time of fork. This would look like data corruption
4120 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4121 unmap_hugepage_range(iter_vma, address,
4122 address + huge_page_size(h), page);
4124 i_mmap_unlock_write(mapping);
4128 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4129 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4130 * cannot race with other handlers or page migration.
4131 * Keep the pte_same checks anyway to make transition from the mutex easier.
4133 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4134 unsigned long address, pte_t *ptep,
4135 struct page *pagecache_page, spinlock_t *ptl)
4138 struct hstate *h = hstate_vma(vma);
4139 struct page *old_page, *new_page;
4140 int outside_reserve = 0;
4142 unsigned long haddr = address & huge_page_mask(h);
4143 struct mmu_notifier_range range;
4145 pte = huge_ptep_get(ptep);
4146 old_page = pte_page(pte);
4149 /* If no-one else is actually using this page, avoid the copy
4150 * and just make the page writable */
4151 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4152 page_move_anon_rmap(old_page, vma);
4153 set_huge_ptep_writable(vma, haddr, ptep);
4158 * If the process that created a MAP_PRIVATE mapping is about to
4159 * perform a COW due to a shared page count, attempt to satisfy
4160 * the allocation without using the existing reserves. The pagecache
4161 * page is used to determine if the reserve at this address was
4162 * consumed or not. If reserves were used, a partial faulted mapping
4163 * at the time of fork() could consume its reserves on COW instead
4164 * of the full address range.
4166 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4167 old_page != pagecache_page)
4168 outside_reserve = 1;
4173 * Drop page table lock as buddy allocator may be called. It will
4174 * be acquired again before returning to the caller, as expected.
4177 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4179 if (IS_ERR(new_page)) {
4181 * If a process owning a MAP_PRIVATE mapping fails to COW,
4182 * it is due to references held by a child and an insufficient
4183 * huge page pool. To guarantee the original mappers
4184 * reliability, unmap the page from child processes. The child
4185 * may get SIGKILLed if it later faults.
4187 if (outside_reserve) {
4189 BUG_ON(huge_pte_none(pte));
4190 unmap_ref_private(mm, vma, old_page, haddr);
4191 BUG_ON(huge_pte_none(pte));
4193 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4195 pte_same(huge_ptep_get(ptep), pte)))
4196 goto retry_avoidcopy;
4198 * race occurs while re-acquiring page table
4199 * lock, and our job is done.
4204 ret = vmf_error(PTR_ERR(new_page));
4205 goto out_release_old;
4209 * When the original hugepage is shared one, it does not have
4210 * anon_vma prepared.
4212 if (unlikely(anon_vma_prepare(vma))) {
4214 goto out_release_all;
4217 copy_user_huge_page(new_page, old_page, address, vma,
4218 pages_per_huge_page(h));
4219 __SetPageUptodate(new_page);
4221 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4222 haddr + huge_page_size(h));
4223 mmu_notifier_invalidate_range_start(&range);
4226 * Retake the page table lock to check for racing updates
4227 * before the page tables are altered
4230 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4231 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4232 ClearPagePrivate(new_page);
4235 huge_ptep_clear_flush(vma, haddr, ptep);
4236 mmu_notifier_invalidate_range(mm, range.start, range.end);
4237 set_huge_pte_at(mm, haddr, ptep,
4238 make_huge_pte(vma, new_page, 1));
4239 page_remove_rmap(old_page, true);
4240 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4241 set_page_huge_active(new_page);
4242 /* Make the old page be freed below */
4243 new_page = old_page;
4246 mmu_notifier_invalidate_range_end(&range);
4248 restore_reserve_on_error(h, vma, haddr, new_page);
4253 spin_lock(ptl); /* Caller expects lock to be held */
4257 /* Return the pagecache page at a given address within a VMA */
4258 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4259 struct vm_area_struct *vma, unsigned long address)
4261 struct address_space *mapping;
4264 mapping = vma->vm_file->f_mapping;
4265 idx = vma_hugecache_offset(h, vma, address);
4267 return find_lock_page(mapping, idx);
4271 * Return whether there is a pagecache page to back given address within VMA.
4272 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4274 static bool hugetlbfs_pagecache_present(struct hstate *h,
4275 struct vm_area_struct *vma, unsigned long address)
4277 struct address_space *mapping;
4281 mapping = vma->vm_file->f_mapping;
4282 idx = vma_hugecache_offset(h, vma, address);
4284 page = find_get_page(mapping, idx);
4287 return page != NULL;
4290 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4293 struct inode *inode = mapping->host;
4294 struct hstate *h = hstate_inode(inode);
4295 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4299 ClearPagePrivate(page);
4302 * set page dirty so that it will not be removed from cache/file
4303 * by non-hugetlbfs specific code paths.
4305 set_page_dirty(page);
4307 spin_lock(&inode->i_lock);
4308 inode->i_blocks += blocks_per_huge_page(h);
4309 spin_unlock(&inode->i_lock);
4313 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4314 struct vm_area_struct *vma,
4315 struct address_space *mapping, pgoff_t idx,
4316 unsigned long address, pte_t *ptep, unsigned int flags)
4318 struct hstate *h = hstate_vma(vma);
4319 vm_fault_t ret = VM_FAULT_SIGBUS;
4325 unsigned long haddr = address & huge_page_mask(h);
4326 bool new_page = false;
4329 * Currently, we are forced to kill the process in the event the
4330 * original mapper has unmapped pages from the child due to a failed
4331 * COW. Warn that such a situation has occurred as it may not be obvious
4333 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4334 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4340 * We can not race with truncation due to holding i_mmap_rwsem.
4341 * i_size is modified when holding i_mmap_rwsem, so check here
4342 * once for faults beyond end of file.
4344 size = i_size_read(mapping->host) >> huge_page_shift(h);
4349 page = find_lock_page(mapping, idx);
4352 * Check for page in userfault range
4354 if (userfaultfd_missing(vma)) {
4356 struct vm_fault vmf = {
4361 * Hard to debug if it ends up being
4362 * used by a callee that assumes
4363 * something about the other
4364 * uninitialized fields... same as in
4370 * hugetlb_fault_mutex and i_mmap_rwsem must be
4371 * dropped before handling userfault. Reacquire
4372 * after handling fault to make calling code simpler.
4374 hash = hugetlb_fault_mutex_hash(mapping, idx);
4375 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4376 i_mmap_unlock_read(mapping);
4377 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4378 i_mmap_lock_read(mapping);
4379 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4383 page = alloc_huge_page(vma, haddr, 0);
4386 * Returning error will result in faulting task being
4387 * sent SIGBUS. The hugetlb fault mutex prevents two
4388 * tasks from racing to fault in the same page which
4389 * could result in false unable to allocate errors.
4390 * Page migration does not take the fault mutex, but
4391 * does a clear then write of pte's under page table
4392 * lock. Page fault code could race with migration,
4393 * notice the clear pte and try to allocate a page
4394 * here. Before returning error, get ptl and make
4395 * sure there really is no pte entry.
4397 ptl = huge_pte_lock(h, mm, ptep);
4398 if (!huge_pte_none(huge_ptep_get(ptep))) {
4404 ret = vmf_error(PTR_ERR(page));
4407 clear_huge_page(page, address, pages_per_huge_page(h));
4408 __SetPageUptodate(page);
4411 if (vma->vm_flags & VM_MAYSHARE) {
4412 int err = huge_add_to_page_cache(page, mapping, idx);
4421 if (unlikely(anon_vma_prepare(vma))) {
4423 goto backout_unlocked;
4429 * If memory error occurs between mmap() and fault, some process
4430 * don't have hwpoisoned swap entry for errored virtual address.
4431 * So we need to block hugepage fault by PG_hwpoison bit check.
4433 if (unlikely(PageHWPoison(page))) {
4434 ret = VM_FAULT_HWPOISON |
4435 VM_FAULT_SET_HINDEX(hstate_index(h));
4436 goto backout_unlocked;
4441 * If we are going to COW a private mapping later, we examine the
4442 * pending reservations for this page now. This will ensure that
4443 * any allocations necessary to record that reservation occur outside
4446 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4447 if (vma_needs_reservation(h, vma, haddr) < 0) {
4449 goto backout_unlocked;
4451 /* Just decrements count, does not deallocate */
4452 vma_end_reservation(h, vma, haddr);
4455 ptl = huge_pte_lock(h, mm, ptep);
4457 if (!huge_pte_none(huge_ptep_get(ptep)))
4461 ClearPagePrivate(page);
4462 hugepage_add_new_anon_rmap(page, vma, haddr);
4464 page_dup_rmap(page, true);
4465 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4466 && (vma->vm_flags & VM_SHARED)));
4467 set_huge_pte_at(mm, haddr, ptep, new_pte);
4469 hugetlb_count_add(pages_per_huge_page(h), mm);
4470 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4471 /* Optimization, do the COW without a second fault */
4472 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4478 * Only make newly allocated pages active. Existing pages found
4479 * in the pagecache could be !page_huge_active() if they have been
4480 * isolated for migration.
4483 set_page_huge_active(page);
4493 restore_reserve_on_error(h, vma, haddr, page);
4499 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4501 unsigned long key[2];
4504 key[0] = (unsigned long) mapping;
4507 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4509 return hash & (num_fault_mutexes - 1);
4513 * For uniprocesor systems we always use a single mutex, so just
4514 * return 0 and avoid the hashing overhead.
4516 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4522 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4523 unsigned long address, unsigned int flags)
4530 struct page *page = NULL;
4531 struct page *pagecache_page = NULL;
4532 struct hstate *h = hstate_vma(vma);
4533 struct address_space *mapping;
4534 int need_wait_lock = 0;
4535 unsigned long haddr = address & huge_page_mask(h);
4537 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4540 * Since we hold no locks, ptep could be stale. That is
4541 * OK as we are only making decisions based on content and
4542 * not actually modifying content here.
4544 entry = huge_ptep_get(ptep);
4545 if (unlikely(is_hugetlb_entry_migration(entry))) {
4546 migration_entry_wait_huge(vma, mm, ptep);
4548 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4549 return VM_FAULT_HWPOISON_LARGE |
4550 VM_FAULT_SET_HINDEX(hstate_index(h));
4554 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4555 * until finished with ptep. This serves two purposes:
4556 * 1) It prevents huge_pmd_unshare from being called elsewhere
4557 * and making the ptep no longer valid.
4558 * 2) It synchronizes us with i_size modifications during truncation.
4560 * ptep could have already be assigned via huge_pte_offset. That
4561 * is OK, as huge_pte_alloc will return the same value unless
4562 * something has changed.
4564 mapping = vma->vm_file->f_mapping;
4565 i_mmap_lock_read(mapping);
4566 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4568 i_mmap_unlock_read(mapping);
4569 return VM_FAULT_OOM;
4573 * Serialize hugepage allocation and instantiation, so that we don't
4574 * get spurious allocation failures if two CPUs race to instantiate
4575 * the same page in the page cache.
4577 idx = vma_hugecache_offset(h, vma, haddr);
4578 hash = hugetlb_fault_mutex_hash(mapping, idx);
4579 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4581 entry = huge_ptep_get(ptep);
4582 if (huge_pte_none(entry)) {
4583 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4590 * entry could be a migration/hwpoison entry at this point, so this
4591 * check prevents the kernel from going below assuming that we have
4592 * an active hugepage in pagecache. This goto expects the 2nd page
4593 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4594 * properly handle it.
4596 if (!pte_present(entry))
4600 * If we are going to COW the mapping later, we examine the pending
4601 * reservations for this page now. This will ensure that any
4602 * allocations necessary to record that reservation occur outside the
4603 * spinlock. For private mappings, we also lookup the pagecache
4604 * page now as it is used to determine if a reservation has been
4607 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4608 if (vma_needs_reservation(h, vma, haddr) < 0) {
4612 /* Just decrements count, does not deallocate */
4613 vma_end_reservation(h, vma, haddr);
4615 if (!(vma->vm_flags & VM_MAYSHARE))
4616 pagecache_page = hugetlbfs_pagecache_page(h,
4620 ptl = huge_pte_lock(h, mm, ptep);
4622 /* Check for a racing update before calling hugetlb_cow */
4623 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4627 * hugetlb_cow() requires page locks of pte_page(entry) and
4628 * pagecache_page, so here we need take the former one
4629 * when page != pagecache_page or !pagecache_page.
4631 page = pte_page(entry);
4632 if (page != pagecache_page)
4633 if (!trylock_page(page)) {
4640 if (flags & FAULT_FLAG_WRITE) {
4641 if (!huge_pte_write(entry)) {
4642 ret = hugetlb_cow(mm, vma, address, ptep,
4643 pagecache_page, ptl);
4646 entry = huge_pte_mkdirty(entry);
4648 entry = pte_mkyoung(entry);
4649 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4650 flags & FAULT_FLAG_WRITE))
4651 update_mmu_cache(vma, haddr, ptep);
4653 if (page != pagecache_page)
4659 if (pagecache_page) {
4660 unlock_page(pagecache_page);
4661 put_page(pagecache_page);
4664 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4665 i_mmap_unlock_read(mapping);
4667 * Generally it's safe to hold refcount during waiting page lock. But
4668 * here we just wait to defer the next page fault to avoid busy loop and
4669 * the page is not used after unlocked before returning from the current
4670 * page fault. So we are safe from accessing freed page, even if we wait
4671 * here without taking refcount.
4674 wait_on_page_locked(page);
4679 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4680 * modifications for huge pages.
4682 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4684 struct vm_area_struct *dst_vma,
4685 unsigned long dst_addr,
4686 unsigned long src_addr,
4687 struct page **pagep)
4689 struct address_space *mapping;
4692 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4693 struct hstate *h = hstate_vma(dst_vma);
4701 page = alloc_huge_page(dst_vma, dst_addr, 0);
4705 ret = copy_huge_page_from_user(page,
4706 (const void __user *) src_addr,
4707 pages_per_huge_page(h), false);
4709 /* fallback to copy_from_user outside mmap_lock */
4710 if (unlikely(ret)) {
4713 /* don't free the page */
4722 * The memory barrier inside __SetPageUptodate makes sure that
4723 * preceding stores to the page contents become visible before
4724 * the set_pte_at() write.
4726 __SetPageUptodate(page);
4728 mapping = dst_vma->vm_file->f_mapping;
4729 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4732 * If shared, add to page cache
4735 size = i_size_read(mapping->host) >> huge_page_shift(h);
4738 goto out_release_nounlock;
4741 * Serialization between remove_inode_hugepages() and
4742 * huge_add_to_page_cache() below happens through the
4743 * hugetlb_fault_mutex_table that here must be hold by
4746 ret = huge_add_to_page_cache(page, mapping, idx);
4748 goto out_release_nounlock;
4751 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4755 * Recheck the i_size after holding PT lock to make sure not
4756 * to leave any page mapped (as page_mapped()) beyond the end
4757 * of the i_size (remove_inode_hugepages() is strict about
4758 * enforcing that). If we bail out here, we'll also leave a
4759 * page in the radix tree in the vm_shared case beyond the end
4760 * of the i_size, but remove_inode_hugepages() will take care
4761 * of it as soon as we drop the hugetlb_fault_mutex_table.
4763 size = i_size_read(mapping->host) >> huge_page_shift(h);
4766 goto out_release_unlock;
4769 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4770 goto out_release_unlock;
4773 page_dup_rmap(page, true);
4775 ClearPagePrivate(page);
4776 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4779 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4780 if (dst_vma->vm_flags & VM_WRITE)
4781 _dst_pte = huge_pte_mkdirty(_dst_pte);
4782 _dst_pte = pte_mkyoung(_dst_pte);
4784 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4786 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4787 dst_vma->vm_flags & VM_WRITE);
4788 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4790 /* No need to invalidate - it was non-present before */
4791 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4794 set_page_huge_active(page);
4804 out_release_nounlock:
4809 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4810 struct page **pages, struct vm_area_struct **vmas,
4811 unsigned long *position, unsigned long *nr_pages,
4812 long i, unsigned int flags, int *locked)
4814 unsigned long pfn_offset;
4815 unsigned long vaddr = *position;
4816 unsigned long remainder = *nr_pages;
4817 struct hstate *h = hstate_vma(vma);
4820 while (vaddr < vma->vm_end && remainder) {
4822 spinlock_t *ptl = NULL;
4827 * If we have a pending SIGKILL, don't keep faulting pages and
4828 * potentially allocating memory.
4830 if (fatal_signal_pending(current)) {
4836 * Some archs (sparc64, sh*) have multiple pte_ts to
4837 * each hugepage. We have to make sure we get the
4838 * first, for the page indexing below to work.
4840 * Note that page table lock is not held when pte is null.
4842 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4845 ptl = huge_pte_lock(h, mm, pte);
4846 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4849 * When coredumping, it suits get_dump_page if we just return
4850 * an error where there's an empty slot with no huge pagecache
4851 * to back it. This way, we avoid allocating a hugepage, and
4852 * the sparse dumpfile avoids allocating disk blocks, but its
4853 * huge holes still show up with zeroes where they need to be.
4855 if (absent && (flags & FOLL_DUMP) &&
4856 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4864 * We need call hugetlb_fault for both hugepages under migration
4865 * (in which case hugetlb_fault waits for the migration,) and
4866 * hwpoisoned hugepages (in which case we need to prevent the
4867 * caller from accessing to them.) In order to do this, we use
4868 * here is_swap_pte instead of is_hugetlb_entry_migration and
4869 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4870 * both cases, and because we can't follow correct pages
4871 * directly from any kind of swap entries.
4873 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4874 ((flags & FOLL_WRITE) &&
4875 !huge_pte_write(huge_ptep_get(pte)))) {
4877 unsigned int fault_flags = 0;
4881 if (flags & FOLL_WRITE)
4882 fault_flags |= FAULT_FLAG_WRITE;
4884 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4885 FAULT_FLAG_KILLABLE;
4886 if (flags & FOLL_NOWAIT)
4887 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4888 FAULT_FLAG_RETRY_NOWAIT;
4889 if (flags & FOLL_TRIED) {
4891 * Note: FAULT_FLAG_ALLOW_RETRY and
4892 * FAULT_FLAG_TRIED can co-exist
4894 fault_flags |= FAULT_FLAG_TRIED;
4896 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4897 if (ret & VM_FAULT_ERROR) {
4898 err = vm_fault_to_errno(ret, flags);
4902 if (ret & VM_FAULT_RETRY) {
4904 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4908 * VM_FAULT_RETRY must not return an
4909 * error, it will return zero
4912 * No need to update "position" as the
4913 * caller will not check it after
4914 * *nr_pages is set to 0.
4921 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4922 page = pte_page(huge_ptep_get(pte));
4925 * If subpage information not requested, update counters
4926 * and skip the same_page loop below.
4928 if (!pages && !vmas && !pfn_offset &&
4929 (vaddr + huge_page_size(h) < vma->vm_end) &&
4930 (remainder >= pages_per_huge_page(h))) {
4931 vaddr += huge_page_size(h);
4932 remainder -= pages_per_huge_page(h);
4933 i += pages_per_huge_page(h);
4940 pages[i] = mem_map_offset(page, pfn_offset);
4942 * try_grab_page() should always succeed here, because:
4943 * a) we hold the ptl lock, and b) we've just checked
4944 * that the huge page is present in the page tables. If
4945 * the huge page is present, then the tail pages must
4946 * also be present. The ptl prevents the head page and
4947 * tail pages from being rearranged in any way. So this
4948 * page must be available at this point, unless the page
4949 * refcount overflowed:
4951 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4966 if (vaddr < vma->vm_end && remainder &&
4967 pfn_offset < pages_per_huge_page(h)) {
4969 * We use pfn_offset to avoid touching the pageframes
4970 * of this compound page.
4976 *nr_pages = remainder;
4978 * setting position is actually required only if remainder is
4979 * not zero but it's faster not to add a "if (remainder)"
4987 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4989 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4992 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4995 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4996 unsigned long address, unsigned long end, pgprot_t newprot)
4998 struct mm_struct *mm = vma->vm_mm;
4999 unsigned long start = address;
5002 struct hstate *h = hstate_vma(vma);
5003 unsigned long pages = 0;
5004 bool shared_pmd = false;
5005 struct mmu_notifier_range range;
5008 * In the case of shared PMDs, the area to flush could be beyond
5009 * start/end. Set range.start/range.end to cover the maximum possible
5010 * range if PMD sharing is possible.
5012 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5013 0, vma, mm, start, end);
5014 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5016 BUG_ON(address >= end);
5017 flush_cache_range(vma, range.start, range.end);
5019 mmu_notifier_invalidate_range_start(&range);
5020 i_mmap_lock_write(vma->vm_file->f_mapping);
5021 for (; address < end; address += huge_page_size(h)) {
5023 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5026 ptl = huge_pte_lock(h, mm, ptep);
5027 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5033 pte = huge_ptep_get(ptep);
5034 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5038 if (unlikely(is_hugetlb_entry_migration(pte))) {
5039 swp_entry_t entry = pte_to_swp_entry(pte);
5041 if (is_write_migration_entry(entry)) {
5044 make_migration_entry_read(&entry);
5045 newpte = swp_entry_to_pte(entry);
5046 set_huge_swap_pte_at(mm, address, ptep,
5047 newpte, huge_page_size(h));
5053 if (!huge_pte_none(pte)) {
5056 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5057 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5058 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5059 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5065 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5066 * may have cleared our pud entry and done put_page on the page table:
5067 * once we release i_mmap_rwsem, another task can do the final put_page
5068 * and that page table be reused and filled with junk. If we actually
5069 * did unshare a page of pmds, flush the range corresponding to the pud.
5072 flush_hugetlb_tlb_range(vma, range.start, range.end);
5074 flush_hugetlb_tlb_range(vma, start, end);
5076 * No need to call mmu_notifier_invalidate_range() we are downgrading
5077 * page table protection not changing it to point to a new page.
5079 * See Documentation/vm/mmu_notifier.rst
5081 i_mmap_unlock_write(vma->vm_file->f_mapping);
5082 mmu_notifier_invalidate_range_end(&range);
5084 return pages << h->order;
5087 int hugetlb_reserve_pages(struct inode *inode,
5089 struct vm_area_struct *vma,
5090 vm_flags_t vm_flags)
5092 long ret, chg, add = -1;
5093 struct hstate *h = hstate_inode(inode);
5094 struct hugepage_subpool *spool = subpool_inode(inode);
5095 struct resv_map *resv_map;
5096 struct hugetlb_cgroup *h_cg = NULL;
5097 long gbl_reserve, regions_needed = 0;
5099 /* This should never happen */
5101 VM_WARN(1, "%s called with a negative range\n", __func__);
5106 * Only apply hugepage reservation if asked. At fault time, an
5107 * attempt will be made for VM_NORESERVE to allocate a page
5108 * without using reserves
5110 if (vm_flags & VM_NORESERVE)
5114 * Shared mappings base their reservation on the number of pages that
5115 * are already allocated on behalf of the file. Private mappings need
5116 * to reserve the full area even if read-only as mprotect() may be
5117 * called to make the mapping read-write. Assume !vma is a shm mapping
5119 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5121 * resv_map can not be NULL as hugetlb_reserve_pages is only
5122 * called for inodes for which resv_maps were created (see
5123 * hugetlbfs_get_inode).
5125 resv_map = inode_resv_map(inode);
5127 chg = region_chg(resv_map, from, to, ®ions_needed);
5130 /* Private mapping. */
5131 resv_map = resv_map_alloc();
5137 set_vma_resv_map(vma, resv_map);
5138 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5146 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5147 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5154 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5155 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5158 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5162 * There must be enough pages in the subpool for the mapping. If
5163 * the subpool has a minimum size, there may be some global
5164 * reservations already in place (gbl_reserve).
5166 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5167 if (gbl_reserve < 0) {
5169 goto out_uncharge_cgroup;
5173 * Check enough hugepages are available for the reservation.
5174 * Hand the pages back to the subpool if there are not
5176 ret = hugetlb_acct_memory(h, gbl_reserve);
5182 * Account for the reservations made. Shared mappings record regions
5183 * that have reservations as they are shared by multiple VMAs.
5184 * When the last VMA disappears, the region map says how much
5185 * the reservation was and the page cache tells how much of
5186 * the reservation was consumed. Private mappings are per-VMA and
5187 * only the consumed reservations are tracked. When the VMA
5188 * disappears, the original reservation is the VMA size and the
5189 * consumed reservations are stored in the map. Hence, nothing
5190 * else has to be done for private mappings here
5192 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5193 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5195 if (unlikely(add < 0)) {
5196 hugetlb_acct_memory(h, -gbl_reserve);
5198 } else if (unlikely(chg > add)) {
5200 * pages in this range were added to the reserve
5201 * map between region_chg and region_add. This
5202 * indicates a race with alloc_huge_page. Adjust
5203 * the subpool and reserve counts modified above
5204 * based on the difference.
5208 hugetlb_cgroup_uncharge_cgroup_rsvd(
5210 (chg - add) * pages_per_huge_page(h), h_cg);
5212 rsv_adjust = hugepage_subpool_put_pages(spool,
5214 hugetlb_acct_memory(h, -rsv_adjust);
5219 /* put back original number of pages, chg */
5220 (void)hugepage_subpool_put_pages(spool, chg);
5221 out_uncharge_cgroup:
5222 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5223 chg * pages_per_huge_page(h), h_cg);
5225 if (!vma || vma->vm_flags & VM_MAYSHARE)
5226 /* Only call region_abort if the region_chg succeeded but the
5227 * region_add failed or didn't run.
5229 if (chg >= 0 && add < 0)
5230 region_abort(resv_map, from, to, regions_needed);
5231 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5232 kref_put(&resv_map->refs, resv_map_release);
5236 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5239 struct hstate *h = hstate_inode(inode);
5240 struct resv_map *resv_map = inode_resv_map(inode);
5242 struct hugepage_subpool *spool = subpool_inode(inode);
5246 * Since this routine can be called in the evict inode path for all
5247 * hugetlbfs inodes, resv_map could be NULL.
5250 chg = region_del(resv_map, start, end);
5252 * region_del() can fail in the rare case where a region
5253 * must be split and another region descriptor can not be
5254 * allocated. If end == LONG_MAX, it will not fail.
5260 spin_lock(&inode->i_lock);
5261 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5262 spin_unlock(&inode->i_lock);
5265 * If the subpool has a minimum size, the number of global
5266 * reservations to be released may be adjusted.
5268 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5269 hugetlb_acct_memory(h, -gbl_reserve);
5274 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5275 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5276 struct vm_area_struct *vma,
5277 unsigned long addr, pgoff_t idx)
5279 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5281 unsigned long sbase = saddr & PUD_MASK;
5282 unsigned long s_end = sbase + PUD_SIZE;
5284 /* Allow segments to share if only one is marked locked */
5285 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5286 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5289 * match the virtual addresses, permission and the alignment of the
5292 if (pmd_index(addr) != pmd_index(saddr) ||
5293 vm_flags != svm_flags ||
5294 sbase < svma->vm_start || svma->vm_end < s_end)
5300 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5302 unsigned long base = addr & PUD_MASK;
5303 unsigned long end = base + PUD_SIZE;
5306 * check on proper vm_flags and page table alignment
5308 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5314 * Determine if start,end range within vma could be mapped by shared pmd.
5315 * If yes, adjust start and end to cover range associated with possible
5316 * shared pmd mappings.
5318 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5319 unsigned long *start, unsigned long *end)
5321 unsigned long a_start, a_end;
5323 if (!(vma->vm_flags & VM_MAYSHARE))
5326 /* Extend the range to be PUD aligned for a worst case scenario */
5327 a_start = ALIGN_DOWN(*start, PUD_SIZE);
5328 a_end = ALIGN(*end, PUD_SIZE);
5331 * Intersect the range with the vma range, since pmd sharing won't be
5332 * across vma after all
5334 *start = max(vma->vm_start, a_start);
5335 *end = min(vma->vm_end, a_end);
5339 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5340 * and returns the corresponding pte. While this is not necessary for the
5341 * !shared pmd case because we can allocate the pmd later as well, it makes the
5342 * code much cleaner.
5344 * This routine must be called with i_mmap_rwsem held in at least read mode if
5345 * sharing is possible. For hugetlbfs, this prevents removal of any page
5346 * table entries associated with the address space. This is important as we
5347 * are setting up sharing based on existing page table entries (mappings).
5349 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5350 * huge_pte_alloc know that sharing is not possible and do not take
5351 * i_mmap_rwsem as a performance optimization. This is handled by the
5352 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5353 * only required for subsequent processing.
5355 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5357 struct vm_area_struct *vma = find_vma(mm, addr);
5358 struct address_space *mapping = vma->vm_file->f_mapping;
5359 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5361 struct vm_area_struct *svma;
5362 unsigned long saddr;
5367 if (!vma_shareable(vma, addr))
5368 return (pte_t *)pmd_alloc(mm, pud, addr);
5370 i_mmap_assert_locked(mapping);
5371 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5375 saddr = page_table_shareable(svma, vma, addr, idx);
5377 spte = huge_pte_offset(svma->vm_mm, saddr,
5378 vma_mmu_pagesize(svma));
5380 get_page(virt_to_page(spte));
5389 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5390 if (pud_none(*pud)) {
5391 pud_populate(mm, pud,
5392 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5395 put_page(virt_to_page(spte));
5399 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5404 * unmap huge page backed by shared pte.
5406 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5407 * indicated by page_count > 1, unmap is achieved by clearing pud and
5408 * decrementing the ref count. If count == 1, the pte page is not shared.
5410 * Called with page table lock held and i_mmap_rwsem held in write mode.
5412 * returns: 1 successfully unmapped a shared pte page
5413 * 0 the underlying pte page is not shared, or it is the last user
5415 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5416 unsigned long *addr, pte_t *ptep)
5418 pgd_t *pgd = pgd_offset(mm, *addr);
5419 p4d_t *p4d = p4d_offset(pgd, *addr);
5420 pud_t *pud = pud_offset(p4d, *addr);
5422 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5423 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5424 if (page_count(virt_to_page(ptep)) == 1)
5428 put_page(virt_to_page(ptep));
5430 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5433 #define want_pmd_share() (1)
5434 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5435 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5440 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5441 unsigned long *addr, pte_t *ptep)
5446 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5447 unsigned long *start, unsigned long *end)
5450 #define want_pmd_share() (0)
5451 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5453 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5454 pte_t *huge_pte_alloc(struct mm_struct *mm,
5455 unsigned long addr, unsigned long sz)
5462 pgd = pgd_offset(mm, addr);
5463 p4d = p4d_alloc(mm, pgd, addr);
5466 pud = pud_alloc(mm, p4d, addr);
5468 if (sz == PUD_SIZE) {
5471 BUG_ON(sz != PMD_SIZE);
5472 if (want_pmd_share() && pud_none(*pud))
5473 pte = huge_pmd_share(mm, addr, pud);
5475 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5478 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5484 * huge_pte_offset() - Walk the page table to resolve the hugepage
5485 * entry at address @addr
5487 * Return: Pointer to page table entry (PUD or PMD) for
5488 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5489 * size @sz doesn't match the hugepage size at this level of the page
5492 pte_t *huge_pte_offset(struct mm_struct *mm,
5493 unsigned long addr, unsigned long sz)
5500 pgd = pgd_offset(mm, addr);
5501 if (!pgd_present(*pgd))
5503 p4d = p4d_offset(pgd, addr);
5504 if (!p4d_present(*p4d))
5507 pud = pud_offset(p4d, addr);
5509 /* must be pud huge, non-present or none */
5510 return (pte_t *)pud;
5511 if (!pud_present(*pud))
5513 /* must have a valid entry and size to go further */
5515 pmd = pmd_offset(pud, addr);
5516 /* must be pmd huge, non-present or none */
5517 return (pte_t *)pmd;
5520 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5523 * These functions are overwritable if your architecture needs its own
5526 struct page * __weak
5527 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5530 return ERR_PTR(-EINVAL);
5533 struct page * __weak
5534 follow_huge_pd(struct vm_area_struct *vma,
5535 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5537 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5541 struct page * __weak
5542 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5543 pmd_t *pmd, int flags)
5545 struct page *page = NULL;
5549 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5550 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5551 (FOLL_PIN | FOLL_GET)))
5555 ptl = pmd_lockptr(mm, pmd);
5558 * make sure that the address range covered by this pmd is not
5559 * unmapped from other threads.
5561 if (!pmd_huge(*pmd))
5563 pte = huge_ptep_get((pte_t *)pmd);
5564 if (pte_present(pte)) {
5565 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5567 * try_grab_page() should always succeed here, because: a) we
5568 * hold the pmd (ptl) lock, and b) we've just checked that the
5569 * huge pmd (head) page is present in the page tables. The ptl
5570 * prevents the head page and tail pages from being rearranged
5571 * in any way. So this page must be available at this point,
5572 * unless the page refcount overflowed:
5574 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5579 if (is_hugetlb_entry_migration(pte)) {
5581 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5585 * hwpoisoned entry is treated as no_page_table in
5586 * follow_page_mask().
5594 struct page * __weak
5595 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5596 pud_t *pud, int flags)
5598 if (flags & (FOLL_GET | FOLL_PIN))
5601 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5604 struct page * __weak
5605 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5607 if (flags & (FOLL_GET | FOLL_PIN))
5610 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5613 bool isolate_huge_page(struct page *page, struct list_head *list)
5617 VM_BUG_ON_PAGE(!PageHead(page), page);
5618 spin_lock(&hugetlb_lock);
5619 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5623 clear_page_huge_active(page);
5624 list_move_tail(&page->lru, list);
5626 spin_unlock(&hugetlb_lock);
5630 void putback_active_hugepage(struct page *page)
5632 VM_BUG_ON_PAGE(!PageHead(page), page);
5633 spin_lock(&hugetlb_lock);
5634 set_page_huge_active(page);
5635 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5636 spin_unlock(&hugetlb_lock);
5640 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5642 struct hstate *h = page_hstate(oldpage);
5644 hugetlb_cgroup_migrate(oldpage, newpage);
5645 set_page_owner_migrate_reason(newpage, reason);
5648 * transfer temporary state of the new huge page. This is
5649 * reverse to other transitions because the newpage is going to
5650 * be final while the old one will be freed so it takes over
5651 * the temporary status.
5653 * Also note that we have to transfer the per-node surplus state
5654 * here as well otherwise the global surplus count will not match
5657 if (PageHugeTemporary(newpage)) {
5658 int old_nid = page_to_nid(oldpage);
5659 int new_nid = page_to_nid(newpage);
5661 SetPageHugeTemporary(oldpage);
5662 ClearPageHugeTemporary(newpage);
5664 spin_lock(&hugetlb_lock);
5665 if (h->surplus_huge_pages_node[old_nid]) {
5666 h->surplus_huge_pages_node[old_nid]--;
5667 h->surplus_huge_pages_node[new_nid]++;
5669 spin_unlock(&hugetlb_lock);
5674 static bool cma_reserve_called __initdata;
5676 static int __init cmdline_parse_hugetlb_cma(char *p)
5678 hugetlb_cma_size = memparse(p, &p);
5682 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5684 void __init hugetlb_cma_reserve(int order)
5686 unsigned long size, reserved, per_node;
5689 cma_reserve_called = true;
5691 if (!hugetlb_cma_size)
5694 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5695 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5696 (PAGE_SIZE << order) / SZ_1M);
5701 * If 3 GB area is requested on a machine with 4 numa nodes,
5702 * let's allocate 1 GB on first three nodes and ignore the last one.
5704 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5705 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5706 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5709 for_each_node_state(nid, N_ONLINE) {
5711 char name[CMA_MAX_NAME];
5713 size = min(per_node, hugetlb_cma_size - reserved);
5714 size = round_up(size, PAGE_SIZE << order);
5716 snprintf(name, sizeof(name), "hugetlb%d", nid);
5717 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5719 &hugetlb_cma[nid], nid);
5721 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5727 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5730 if (reserved >= hugetlb_cma_size)
5735 void __init hugetlb_cma_check(void)
5737 if (!hugetlb_cma_size || cma_reserve_called)
5740 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5743 #endif /* CONFIG_CMA */