2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.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/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
27 #include <asm/pgtable.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
37 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
38 unsigned long hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 __initdata LIST_HEAD(huge_boot_pages);
46 /* for command line parsing */
47 static struct hstate * __initdata parsed_hstate;
48 static unsigned long __initdata default_hstate_max_huge_pages;
49 static unsigned long __initdata default_hstate_size;
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 DEFINE_SPINLOCK(hugetlb_lock);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
60 spin_unlock(&spool->lock);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
70 struct hugepage_subpool *spool;
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
76 spin_lock_init(&spool->lock);
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
89 unlock_or_release_subpool(spool);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
106 spin_unlock(&spool->lock);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
126 return HUGETLBFS_SB(inode->i_sb)->spool;
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
131 return subpool_inode(file_inode(vma->vm_file));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link;
154 static long region_add(struct list_head *head, long f, long t)
156 struct file_region *rg, *nrg, *trg;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head *head, long f, long t)
192 struct file_region *rg, *nrg;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg -= rg->to - rg->from;
239 static long region_truncate(struct list_head *head, long end)
241 struct file_region *rg, *trg;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
248 if (&rg->link == head)
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
255 rg = list_entry(rg->link.next, typeof(*rg), link);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
262 chg += rg->to - rg->from;
269 static long region_count(struct list_head *head, long f, long t)
271 struct file_region *rg;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
287 chg += seg_to - seg_from;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
316 struct hstate *hstate;
318 if (!is_vm_hugetlb_page(vma))
321 hstate = hstate_vma(vma);
323 return 1UL << (hstate->order + PAGE_SHIFT);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
336 return vma_kernel_pagesize(vma);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
370 return (unsigned long)vma->vm_private_data;
373 static void set_vma_private_data(struct vm_area_struct *vma,
376 vma->vm_private_data = (void *)value;
381 struct list_head regions;
384 static struct resv_map *resv_map_alloc(void)
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
396 static void resv_map_release(struct kref *ref)
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
435 return (get_vma_private_data(vma) & flag) != 0;
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
442 if (vma->vm_flags & VM_NORESERVE)
445 if (vma->vm_flags & VM_MAYSHARE) {
446 /* Shared mappings always use reserves */
447 h->resv_huge_pages--;
448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
450 * Only the process that called mmap() has reserves for
453 h->resv_huge_pages--;
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
461 if (!(vma->vm_flags & VM_MAYSHARE))
462 vma->vm_private_data = (void *)0;
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct *vma)
468 if (vma->vm_flags & VM_MAYSHARE)
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
475 static void copy_gigantic_page(struct page *dst, struct page *src)
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
482 for (i = 0; i < pages_per_huge_page(h); ) {
484 copy_highpage(dst, src);
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
492 void copy_huge_page(struct page *dst, struct page *src)
495 struct hstate *h = page_hstate(src);
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
503 for (i = 0; i < pages_per_huge_page(h); i++) {
505 copy_highpage(dst + i, src + i);
509 static void enqueue_huge_page(struct hstate *h, struct page *page)
511 int nid = page_to_nid(page);
512 list_move(&page->lru, &h->hugepage_freelists[nid]);
513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
517 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
521 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
522 if (!is_migrate_isolate_page(page))
525 * if 'non-isolated free hugepage' not found on the list,
526 * the allocation fails.
528 if (&h->hugepage_freelists[nid] == &page->lru)
530 list_move(&page->lru, &h->hugepage_activelist);
531 set_page_refcounted(page);
532 h->free_huge_pages--;
533 h->free_huge_pages_node[nid]--;
537 static struct page *dequeue_huge_page_vma(struct hstate *h,
538 struct vm_area_struct *vma,
539 unsigned long address, int avoid_reserve)
541 struct page *page = NULL;
542 struct mempolicy *mpol;
543 nodemask_t *nodemask;
544 struct zonelist *zonelist;
547 unsigned int cpuset_mems_cookie;
550 cpuset_mems_cookie = get_mems_allowed();
551 zonelist = huge_zonelist(vma, address,
552 htlb_alloc_mask, &mpol, &nodemask);
554 * A child process with MAP_PRIVATE mappings created by their parent
555 * have no page reserves. This check ensures that reservations are
556 * not "stolen". The child may still get SIGKILLed
558 if (!vma_has_reserves(vma) &&
559 h->free_huge_pages - h->resv_huge_pages == 0)
562 /* If reserves cannot be used, ensure enough pages are in the pool */
563 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
566 for_each_zone_zonelist_nodemask(zone, z, zonelist,
567 MAX_NR_ZONES - 1, nodemask) {
568 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
569 page = dequeue_huge_page_node(h, zone_to_nid(zone));
572 decrement_hugepage_resv_vma(h, vma);
579 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
588 static void update_and_free_page(struct hstate *h, struct page *page)
592 VM_BUG_ON(h->order >= MAX_ORDER);
595 h->nr_huge_pages_node[page_to_nid(page)]--;
596 for (i = 0; i < pages_per_huge_page(h); i++) {
597 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
598 1 << PG_referenced | 1 << PG_dirty |
599 1 << PG_active | 1 << PG_reserved |
600 1 << PG_private | 1 << PG_writeback);
602 VM_BUG_ON(hugetlb_cgroup_from_page(page));
603 set_compound_page_dtor(page, NULL);
604 set_page_refcounted(page);
605 arch_release_hugepage(page);
606 __free_pages(page, huge_page_order(h));
609 struct hstate *size_to_hstate(unsigned long size)
614 if (huge_page_size(h) == size)
620 static void free_huge_page(struct page *page)
623 * Can't pass hstate in here because it is called from the
624 * compound page destructor.
626 struct hstate *h = page_hstate(page);
627 int nid = page_to_nid(page);
628 struct hugepage_subpool *spool =
629 (struct hugepage_subpool *)page_private(page);
631 set_page_private(page, 0);
632 page->mapping = NULL;
633 BUG_ON(page_count(page));
634 BUG_ON(page_mapcount(page));
636 spin_lock(&hugetlb_lock);
637 hugetlb_cgroup_uncharge_page(hstate_index(h),
638 pages_per_huge_page(h), page);
639 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
640 /* remove the page from active list */
641 list_del(&page->lru);
642 update_and_free_page(h, page);
643 h->surplus_huge_pages--;
644 h->surplus_huge_pages_node[nid]--;
646 arch_clear_hugepage_flags(page);
647 enqueue_huge_page(h, page);
649 spin_unlock(&hugetlb_lock);
650 hugepage_subpool_put_pages(spool, 1);
653 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
655 INIT_LIST_HEAD(&page->lru);
656 set_compound_page_dtor(page, free_huge_page);
657 spin_lock(&hugetlb_lock);
658 set_hugetlb_cgroup(page, NULL);
660 h->nr_huge_pages_node[nid]++;
661 spin_unlock(&hugetlb_lock);
662 put_page(page); /* free it into the hugepage allocator */
665 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
668 int nr_pages = 1 << order;
669 struct page *p = page + 1;
671 /* we rely on prep_new_huge_page to set the destructor */
672 set_compound_order(page, order);
674 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
676 set_page_count(p, 0);
677 p->first_page = page;
682 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
683 * transparent huge pages. See the PageTransHuge() documentation for more
686 int PageHuge(struct page *page)
688 compound_page_dtor *dtor;
690 if (!PageCompound(page))
693 page = compound_head(page);
694 dtor = get_compound_page_dtor(page);
696 return dtor == free_huge_page;
698 EXPORT_SYMBOL_GPL(PageHuge);
701 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
702 * normal or transparent huge pages.
704 int PageHeadHuge(struct page *page_head)
706 compound_page_dtor *dtor;
708 if (!PageHead(page_head))
711 dtor = get_compound_page_dtor(page_head);
713 return dtor == free_huge_page;
715 EXPORT_SYMBOL_GPL(PageHeadHuge);
717 pgoff_t __basepage_index(struct page *page)
719 struct page *page_head = compound_head(page);
720 pgoff_t index = page_index(page_head);
721 unsigned long compound_idx;
723 if (!PageHuge(page_head))
724 return page_index(page);
726 if (compound_order(page_head) >= MAX_ORDER)
727 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
729 compound_idx = page - page_head;
731 return (index << compound_order(page_head)) + compound_idx;
734 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
738 if (h->order >= MAX_ORDER)
741 page = alloc_pages_exact_node(nid,
742 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
743 __GFP_REPEAT|__GFP_NOWARN,
746 if (arch_prepare_hugepage(page)) {
747 __free_pages(page, huge_page_order(h));
750 prep_new_huge_page(h, page, nid);
757 * common helper functions for hstate_next_node_to_{alloc|free}.
758 * We may have allocated or freed a huge page based on a different
759 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
760 * be outside of *nodes_allowed. Ensure that we use an allowed
761 * node for alloc or free.
763 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
765 nid = next_node(nid, *nodes_allowed);
766 if (nid == MAX_NUMNODES)
767 nid = first_node(*nodes_allowed);
768 VM_BUG_ON(nid >= MAX_NUMNODES);
773 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
775 if (!node_isset(nid, *nodes_allowed))
776 nid = next_node_allowed(nid, nodes_allowed);
781 * returns the previously saved node ["this node"] from which to
782 * allocate a persistent huge page for the pool and advance the
783 * next node from which to allocate, handling wrap at end of node
786 static int hstate_next_node_to_alloc(struct hstate *h,
787 nodemask_t *nodes_allowed)
791 VM_BUG_ON(!nodes_allowed);
793 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
794 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
799 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
806 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
807 next_nid = start_nid;
810 page = alloc_fresh_huge_page_node(h, next_nid);
815 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
816 } while (next_nid != start_nid);
819 count_vm_event(HTLB_BUDDY_PGALLOC);
821 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
827 * helper for free_pool_huge_page() - return the previously saved
828 * node ["this node"] from which to free a huge page. Advance the
829 * next node id whether or not we find a free huge page to free so
830 * that the next attempt to free addresses the next node.
832 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
836 VM_BUG_ON(!nodes_allowed);
838 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
839 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
845 * Free huge page from pool from next node to free.
846 * Attempt to keep persistent huge pages more or less
847 * balanced over allowed nodes.
848 * Called with hugetlb_lock locked.
850 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
857 start_nid = hstate_next_node_to_free(h, nodes_allowed);
858 next_nid = start_nid;
862 * If we're returning unused surplus pages, only examine
863 * nodes with surplus pages.
865 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
866 !list_empty(&h->hugepage_freelists[next_nid])) {
868 list_entry(h->hugepage_freelists[next_nid].next,
870 list_del(&page->lru);
871 h->free_huge_pages--;
872 h->free_huge_pages_node[next_nid]--;
874 h->surplus_huge_pages--;
875 h->surplus_huge_pages_node[next_nid]--;
877 update_and_free_page(h, page);
881 next_nid = hstate_next_node_to_free(h, nodes_allowed);
882 } while (next_nid != start_nid);
887 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
892 if (h->order >= MAX_ORDER)
896 * Assume we will successfully allocate the surplus page to
897 * prevent racing processes from causing the surplus to exceed
900 * This however introduces a different race, where a process B
901 * tries to grow the static hugepage pool while alloc_pages() is
902 * called by process A. B will only examine the per-node
903 * counters in determining if surplus huge pages can be
904 * converted to normal huge pages in adjust_pool_surplus(). A
905 * won't be able to increment the per-node counter, until the
906 * lock is dropped by B, but B doesn't drop hugetlb_lock until
907 * no more huge pages can be converted from surplus to normal
908 * state (and doesn't try to convert again). Thus, we have a
909 * case where a surplus huge page exists, the pool is grown, and
910 * the surplus huge page still exists after, even though it
911 * should just have been converted to a normal huge page. This
912 * does not leak memory, though, as the hugepage will be freed
913 * once it is out of use. It also does not allow the counters to
914 * go out of whack in adjust_pool_surplus() as we don't modify
915 * the node values until we've gotten the hugepage and only the
916 * per-node value is checked there.
918 spin_lock(&hugetlb_lock);
919 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
920 spin_unlock(&hugetlb_lock);
924 h->surplus_huge_pages++;
926 spin_unlock(&hugetlb_lock);
928 if (nid == NUMA_NO_NODE)
929 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
930 __GFP_REPEAT|__GFP_NOWARN,
933 page = alloc_pages_exact_node(nid,
934 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
935 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
937 if (page && arch_prepare_hugepage(page)) {
938 __free_pages(page, huge_page_order(h));
942 spin_lock(&hugetlb_lock);
944 INIT_LIST_HEAD(&page->lru);
945 r_nid = page_to_nid(page);
946 set_compound_page_dtor(page, free_huge_page);
947 set_hugetlb_cgroup(page, NULL);
949 * We incremented the global counters already
951 h->nr_huge_pages_node[r_nid]++;
952 h->surplus_huge_pages_node[r_nid]++;
953 __count_vm_event(HTLB_BUDDY_PGALLOC);
956 h->surplus_huge_pages--;
957 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
959 spin_unlock(&hugetlb_lock);
965 * This allocation function is useful in the context where vma is irrelevant.
966 * E.g. soft-offlining uses this function because it only cares physical
967 * address of error page.
969 struct page *alloc_huge_page_node(struct hstate *h, int nid)
973 spin_lock(&hugetlb_lock);
974 page = dequeue_huge_page_node(h, nid);
975 spin_unlock(&hugetlb_lock);
978 page = alloc_buddy_huge_page(h, nid);
984 * Increase the hugetlb pool such that it can accommodate a reservation
987 static int gather_surplus_pages(struct hstate *h, int delta)
989 struct list_head surplus_list;
990 struct page *page, *tmp;
992 int needed, allocated;
993 bool alloc_ok = true;
995 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
997 h->resv_huge_pages += delta;
1002 INIT_LIST_HEAD(&surplus_list);
1006 spin_unlock(&hugetlb_lock);
1007 for (i = 0; i < needed; i++) {
1008 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1013 list_add(&page->lru, &surplus_list);
1018 * After retaking hugetlb_lock, we need to recalculate 'needed'
1019 * because either resv_huge_pages or free_huge_pages may have changed.
1021 spin_lock(&hugetlb_lock);
1022 needed = (h->resv_huge_pages + delta) -
1023 (h->free_huge_pages + allocated);
1028 * We were not able to allocate enough pages to
1029 * satisfy the entire reservation so we free what
1030 * we've allocated so far.
1035 * The surplus_list now contains _at_least_ the number of extra pages
1036 * needed to accommodate the reservation. Add the appropriate number
1037 * of pages to the hugetlb pool and free the extras back to the buddy
1038 * allocator. Commit the entire reservation here to prevent another
1039 * process from stealing the pages as they are added to the pool but
1040 * before they are reserved.
1042 needed += allocated;
1043 h->resv_huge_pages += delta;
1046 /* Free the needed pages to the hugetlb pool */
1047 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1051 * This page is now managed by the hugetlb allocator and has
1052 * no users -- drop the buddy allocator's reference.
1054 put_page_testzero(page);
1055 VM_BUG_ON(page_count(page));
1056 enqueue_huge_page(h, page);
1059 spin_unlock(&hugetlb_lock);
1061 /* Free unnecessary surplus pages to the buddy allocator */
1062 if (!list_empty(&surplus_list)) {
1063 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1067 spin_lock(&hugetlb_lock);
1073 * When releasing a hugetlb pool reservation, any surplus pages that were
1074 * allocated to satisfy the reservation must be explicitly freed if they were
1076 * Called with hugetlb_lock held.
1078 static void return_unused_surplus_pages(struct hstate *h,
1079 unsigned long unused_resv_pages)
1081 unsigned long nr_pages;
1083 /* Uncommit the reservation */
1084 h->resv_huge_pages -= unused_resv_pages;
1086 /* Cannot return gigantic pages currently */
1087 if (h->order >= MAX_ORDER)
1090 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1093 * We want to release as many surplus pages as possible, spread
1094 * evenly across all nodes with memory. Iterate across these nodes
1095 * until we can no longer free unreserved surplus pages. This occurs
1096 * when the nodes with surplus pages have no free pages.
1097 * free_pool_huge_page() will balance the the freed pages across the
1098 * on-line nodes with memory and will handle the hstate accounting.
1100 while (nr_pages--) {
1101 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1103 cond_resched_lock(&hugetlb_lock);
1108 * Determine if the huge page at addr within the vma has an associated
1109 * reservation. Where it does not we will need to logically increase
1110 * reservation and actually increase subpool usage before an allocation
1111 * can occur. Where any new reservation would be required the
1112 * reservation change is prepared, but not committed. Once the page
1113 * has been allocated from the subpool and instantiated the change should
1114 * be committed via vma_commit_reservation. No action is required on
1117 static long vma_needs_reservation(struct hstate *h,
1118 struct vm_area_struct *vma, unsigned long addr)
1120 struct address_space *mapping = vma->vm_file->f_mapping;
1121 struct inode *inode = mapping->host;
1123 if (vma->vm_flags & VM_MAYSHARE) {
1124 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125 return region_chg(&inode->i_mapping->private_list,
1128 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1133 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1134 struct resv_map *reservations = vma_resv_map(vma);
1136 err = region_chg(&reservations->regions, idx, idx + 1);
1142 static void vma_commit_reservation(struct hstate *h,
1143 struct vm_area_struct *vma, unsigned long addr)
1145 struct address_space *mapping = vma->vm_file->f_mapping;
1146 struct inode *inode = mapping->host;
1148 if (vma->vm_flags & VM_MAYSHARE) {
1149 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1150 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1152 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1153 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1154 struct resv_map *reservations = vma_resv_map(vma);
1156 /* Mark this page used in the map. */
1157 region_add(&reservations->regions, idx, idx + 1);
1161 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1162 unsigned long addr, int avoid_reserve)
1164 struct hugepage_subpool *spool = subpool_vma(vma);
1165 struct hstate *h = hstate_vma(vma);
1169 struct hugetlb_cgroup *h_cg;
1171 idx = hstate_index(h);
1173 * Processes that did not create the mapping will have no
1174 * reserves and will not have accounted against subpool
1175 * limit. Check that the subpool limit can be made before
1176 * satisfying the allocation MAP_NORESERVE mappings may also
1177 * need pages and subpool limit allocated allocated if no reserve
1180 chg = vma_needs_reservation(h, vma, addr);
1182 return ERR_PTR(-ENOMEM);
1184 if (hugepage_subpool_get_pages(spool, chg))
1185 return ERR_PTR(-ENOSPC);
1187 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1189 hugepage_subpool_put_pages(spool, chg);
1190 return ERR_PTR(-ENOSPC);
1192 spin_lock(&hugetlb_lock);
1193 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1195 /* update page cgroup details */
1196 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1198 spin_unlock(&hugetlb_lock);
1200 spin_unlock(&hugetlb_lock);
1201 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1203 hugetlb_cgroup_uncharge_cgroup(idx,
1204 pages_per_huge_page(h),
1206 hugepage_subpool_put_pages(spool, chg);
1207 return ERR_PTR(-ENOSPC);
1209 spin_lock(&hugetlb_lock);
1210 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1212 list_move(&page->lru, &h->hugepage_activelist);
1213 spin_unlock(&hugetlb_lock);
1216 set_page_private(page, (unsigned long)spool);
1218 vma_commit_reservation(h, vma, addr);
1222 int __weak alloc_bootmem_huge_page(struct hstate *h)
1224 struct huge_bootmem_page *m;
1225 int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1230 addr = __alloc_bootmem_node_nopanic(
1231 NODE_DATA(hstate_next_node_to_alloc(h,
1232 &node_states[N_MEMORY])),
1233 huge_page_size(h), huge_page_size(h), 0);
1237 * Use the beginning of the huge page to store the
1238 * huge_bootmem_page struct (until gather_bootmem
1239 * puts them into the mem_map).
1249 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1250 /* Put them into a private list first because mem_map is not up yet */
1251 list_add(&m->list, &huge_boot_pages);
1256 static void prep_compound_huge_page(struct page *page, int order)
1258 if (unlikely(order > (MAX_ORDER - 1)))
1259 prep_compound_gigantic_page(page, order);
1261 prep_compound_page(page, order);
1264 /* Put bootmem huge pages into the standard lists after mem_map is up */
1265 static void __init gather_bootmem_prealloc(void)
1267 struct huge_bootmem_page *m;
1269 list_for_each_entry(m, &huge_boot_pages, list) {
1270 struct hstate *h = m->hstate;
1273 #ifdef CONFIG_HIGHMEM
1274 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1275 free_bootmem_late((unsigned long)m,
1276 sizeof(struct huge_bootmem_page));
1278 page = virt_to_page(m);
1280 __ClearPageReserved(page);
1281 WARN_ON(page_count(page) != 1);
1282 prep_compound_huge_page(page, h->order);
1283 prep_new_huge_page(h, page, page_to_nid(page));
1285 * If we had gigantic hugepages allocated at boot time, we need
1286 * to restore the 'stolen' pages to totalram_pages in order to
1287 * fix confusing memory reports from free(1) and another
1288 * side-effects, like CommitLimit going negative.
1290 if (h->order > (MAX_ORDER - 1))
1291 totalram_pages += 1 << h->order;
1295 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1299 for (i = 0; i < h->max_huge_pages; ++i) {
1300 if (h->order >= MAX_ORDER) {
1301 if (!alloc_bootmem_huge_page(h))
1303 } else if (!alloc_fresh_huge_page(h,
1304 &node_states[N_MEMORY]))
1307 h->max_huge_pages = i;
1310 static void __init hugetlb_init_hstates(void)
1314 for_each_hstate(h) {
1315 /* oversize hugepages were init'ed in early boot */
1316 if (h->order < MAX_ORDER)
1317 hugetlb_hstate_alloc_pages(h);
1321 static char * __init memfmt(char *buf, unsigned long n)
1323 if (n >= (1UL << 30))
1324 sprintf(buf, "%lu GB", n >> 30);
1325 else if (n >= (1UL << 20))
1326 sprintf(buf, "%lu MB", n >> 20);
1328 sprintf(buf, "%lu KB", n >> 10);
1332 static void __init report_hugepages(void)
1336 for_each_hstate(h) {
1338 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1339 memfmt(buf, huge_page_size(h)),
1340 h->free_huge_pages);
1344 #ifdef CONFIG_HIGHMEM
1345 static void try_to_free_low(struct hstate *h, unsigned long count,
1346 nodemask_t *nodes_allowed)
1350 if (h->order >= MAX_ORDER)
1353 for_each_node_mask(i, *nodes_allowed) {
1354 struct page *page, *next;
1355 struct list_head *freel = &h->hugepage_freelists[i];
1356 list_for_each_entry_safe(page, next, freel, lru) {
1357 if (count >= h->nr_huge_pages)
1359 if (PageHighMem(page))
1361 list_del(&page->lru);
1362 update_and_free_page(h, page);
1363 h->free_huge_pages--;
1364 h->free_huge_pages_node[page_to_nid(page)]--;
1369 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1370 nodemask_t *nodes_allowed)
1376 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1377 * balanced by operating on them in a round-robin fashion.
1378 * Returns 1 if an adjustment was made.
1380 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1383 int start_nid, next_nid;
1386 VM_BUG_ON(delta != -1 && delta != 1);
1389 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1391 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1392 next_nid = start_nid;
1398 * To shrink on this node, there must be a surplus page
1400 if (!h->surplus_huge_pages_node[nid]) {
1401 next_nid = hstate_next_node_to_alloc(h,
1408 * Surplus cannot exceed the total number of pages
1410 if (h->surplus_huge_pages_node[nid] >=
1411 h->nr_huge_pages_node[nid]) {
1412 next_nid = hstate_next_node_to_free(h,
1418 h->surplus_huge_pages += delta;
1419 h->surplus_huge_pages_node[nid] += delta;
1422 } while (next_nid != start_nid);
1427 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1428 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1429 nodemask_t *nodes_allowed)
1431 unsigned long min_count, ret;
1433 if (h->order >= MAX_ORDER)
1434 return h->max_huge_pages;
1437 * Increase the pool size
1438 * First take pages out of surplus state. Then make up the
1439 * remaining difference by allocating fresh huge pages.
1441 * We might race with alloc_buddy_huge_page() here and be unable
1442 * to convert a surplus huge page to a normal huge page. That is
1443 * not critical, though, it just means the overall size of the
1444 * pool might be one hugepage larger than it needs to be, but
1445 * within all the constraints specified by the sysctls.
1447 spin_lock(&hugetlb_lock);
1448 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1449 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1453 while (count > persistent_huge_pages(h)) {
1455 * If this allocation races such that we no longer need the
1456 * page, free_huge_page will handle it by freeing the page
1457 * and reducing the surplus.
1459 spin_unlock(&hugetlb_lock);
1460 ret = alloc_fresh_huge_page(h, nodes_allowed);
1461 spin_lock(&hugetlb_lock);
1465 /* Bail for signals. Probably ctrl-c from user */
1466 if (signal_pending(current))
1471 * Decrease the pool size
1472 * First return free pages to the buddy allocator (being careful
1473 * to keep enough around to satisfy reservations). Then place
1474 * pages into surplus state as needed so the pool will shrink
1475 * to the desired size as pages become free.
1477 * By placing pages into the surplus state independent of the
1478 * overcommit value, we are allowing the surplus pool size to
1479 * exceed overcommit. There are few sane options here. Since
1480 * alloc_buddy_huge_page() is checking the global counter,
1481 * though, we'll note that we're not allowed to exceed surplus
1482 * and won't grow the pool anywhere else. Not until one of the
1483 * sysctls are changed, or the surplus pages go out of use.
1485 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1486 min_count = max(count, min_count);
1487 try_to_free_low(h, min_count, nodes_allowed);
1488 while (min_count < persistent_huge_pages(h)) {
1489 if (!free_pool_huge_page(h, nodes_allowed, 0))
1491 cond_resched_lock(&hugetlb_lock);
1493 while (count < persistent_huge_pages(h)) {
1494 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1498 ret = persistent_huge_pages(h);
1499 spin_unlock(&hugetlb_lock);
1503 #define HSTATE_ATTR_RO(_name) \
1504 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1506 #define HSTATE_ATTR(_name) \
1507 static struct kobj_attribute _name##_attr = \
1508 __ATTR(_name, 0644, _name##_show, _name##_store)
1510 static struct kobject *hugepages_kobj;
1511 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1513 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1515 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1519 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1520 if (hstate_kobjs[i] == kobj) {
1522 *nidp = NUMA_NO_NODE;
1526 return kobj_to_node_hstate(kobj, nidp);
1529 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1530 struct kobj_attribute *attr, char *buf)
1533 unsigned long nr_huge_pages;
1536 h = kobj_to_hstate(kobj, &nid);
1537 if (nid == NUMA_NO_NODE)
1538 nr_huge_pages = h->nr_huge_pages;
1540 nr_huge_pages = h->nr_huge_pages_node[nid];
1542 return sprintf(buf, "%lu\n", nr_huge_pages);
1545 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1546 struct kobject *kobj, struct kobj_attribute *attr,
1547 const char *buf, size_t len)
1551 unsigned long count;
1553 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1555 err = strict_strtoul(buf, 10, &count);
1559 h = kobj_to_hstate(kobj, &nid);
1560 if (h->order >= MAX_ORDER) {
1565 if (nid == NUMA_NO_NODE) {
1567 * global hstate attribute
1569 if (!(obey_mempolicy &&
1570 init_nodemask_of_mempolicy(nodes_allowed))) {
1571 NODEMASK_FREE(nodes_allowed);
1572 nodes_allowed = &node_states[N_MEMORY];
1574 } else if (nodes_allowed) {
1576 * per node hstate attribute: adjust count to global,
1577 * but restrict alloc/free to the specified node.
1579 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1580 init_nodemask_of_node(nodes_allowed, nid);
1582 nodes_allowed = &node_states[N_MEMORY];
1584 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1586 if (nodes_allowed != &node_states[N_MEMORY])
1587 NODEMASK_FREE(nodes_allowed);
1591 NODEMASK_FREE(nodes_allowed);
1595 static ssize_t nr_hugepages_show(struct kobject *kobj,
1596 struct kobj_attribute *attr, char *buf)
1598 return nr_hugepages_show_common(kobj, attr, buf);
1601 static ssize_t nr_hugepages_store(struct kobject *kobj,
1602 struct kobj_attribute *attr, const char *buf, size_t len)
1604 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1606 HSTATE_ATTR(nr_hugepages);
1611 * hstate attribute for optionally mempolicy-based constraint on persistent
1612 * huge page alloc/free.
1614 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1615 struct kobj_attribute *attr, char *buf)
1617 return nr_hugepages_show_common(kobj, attr, buf);
1620 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1621 struct kobj_attribute *attr, const char *buf, size_t len)
1623 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1625 HSTATE_ATTR(nr_hugepages_mempolicy);
1629 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1630 struct kobj_attribute *attr, char *buf)
1632 struct hstate *h = kobj_to_hstate(kobj, NULL);
1633 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1636 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1637 struct kobj_attribute *attr, const char *buf, size_t count)
1640 unsigned long input;
1641 struct hstate *h = kobj_to_hstate(kobj, NULL);
1643 if (h->order >= MAX_ORDER)
1646 err = strict_strtoul(buf, 10, &input);
1650 spin_lock(&hugetlb_lock);
1651 h->nr_overcommit_huge_pages = input;
1652 spin_unlock(&hugetlb_lock);
1656 HSTATE_ATTR(nr_overcommit_hugepages);
1658 static ssize_t free_hugepages_show(struct kobject *kobj,
1659 struct kobj_attribute *attr, char *buf)
1662 unsigned long free_huge_pages;
1665 h = kobj_to_hstate(kobj, &nid);
1666 if (nid == NUMA_NO_NODE)
1667 free_huge_pages = h->free_huge_pages;
1669 free_huge_pages = h->free_huge_pages_node[nid];
1671 return sprintf(buf, "%lu\n", free_huge_pages);
1673 HSTATE_ATTR_RO(free_hugepages);
1675 static ssize_t resv_hugepages_show(struct kobject *kobj,
1676 struct kobj_attribute *attr, char *buf)
1678 struct hstate *h = kobj_to_hstate(kobj, NULL);
1679 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1681 HSTATE_ATTR_RO(resv_hugepages);
1683 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1684 struct kobj_attribute *attr, char *buf)
1687 unsigned long surplus_huge_pages;
1690 h = kobj_to_hstate(kobj, &nid);
1691 if (nid == NUMA_NO_NODE)
1692 surplus_huge_pages = h->surplus_huge_pages;
1694 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1696 return sprintf(buf, "%lu\n", surplus_huge_pages);
1698 HSTATE_ATTR_RO(surplus_hugepages);
1700 static struct attribute *hstate_attrs[] = {
1701 &nr_hugepages_attr.attr,
1702 &nr_overcommit_hugepages_attr.attr,
1703 &free_hugepages_attr.attr,
1704 &resv_hugepages_attr.attr,
1705 &surplus_hugepages_attr.attr,
1707 &nr_hugepages_mempolicy_attr.attr,
1712 static struct attribute_group hstate_attr_group = {
1713 .attrs = hstate_attrs,
1716 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1717 struct kobject **hstate_kobjs,
1718 struct attribute_group *hstate_attr_group)
1721 int hi = hstate_index(h);
1723 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1724 if (!hstate_kobjs[hi])
1727 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1729 kobject_put(hstate_kobjs[hi]);
1734 static void __init hugetlb_sysfs_init(void)
1739 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1740 if (!hugepages_kobj)
1743 for_each_hstate(h) {
1744 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1745 hstate_kobjs, &hstate_attr_group);
1747 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1754 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1755 * with node devices in node_devices[] using a parallel array. The array
1756 * index of a node device or _hstate == node id.
1757 * This is here to avoid any static dependency of the node device driver, in
1758 * the base kernel, on the hugetlb module.
1760 struct node_hstate {
1761 struct kobject *hugepages_kobj;
1762 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1764 struct node_hstate node_hstates[MAX_NUMNODES];
1767 * A subset of global hstate attributes for node devices
1769 static struct attribute *per_node_hstate_attrs[] = {
1770 &nr_hugepages_attr.attr,
1771 &free_hugepages_attr.attr,
1772 &surplus_hugepages_attr.attr,
1776 static struct attribute_group per_node_hstate_attr_group = {
1777 .attrs = per_node_hstate_attrs,
1781 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1782 * Returns node id via non-NULL nidp.
1784 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1788 for (nid = 0; nid < nr_node_ids; nid++) {
1789 struct node_hstate *nhs = &node_hstates[nid];
1791 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1792 if (nhs->hstate_kobjs[i] == kobj) {
1804 * Unregister hstate attributes from a single node device.
1805 * No-op if no hstate attributes attached.
1807 static void hugetlb_unregister_node(struct node *node)
1810 struct node_hstate *nhs = &node_hstates[node->dev.id];
1812 if (!nhs->hugepages_kobj)
1813 return; /* no hstate attributes */
1815 for_each_hstate(h) {
1816 int idx = hstate_index(h);
1817 if (nhs->hstate_kobjs[idx]) {
1818 kobject_put(nhs->hstate_kobjs[idx]);
1819 nhs->hstate_kobjs[idx] = NULL;
1823 kobject_put(nhs->hugepages_kobj);
1824 nhs->hugepages_kobj = NULL;
1828 * hugetlb module exit: unregister hstate attributes from node devices
1831 static void hugetlb_unregister_all_nodes(void)
1836 * disable node device registrations.
1838 register_hugetlbfs_with_node(NULL, NULL);
1841 * remove hstate attributes from any nodes that have them.
1843 for (nid = 0; nid < nr_node_ids; nid++)
1844 hugetlb_unregister_node(node_devices[nid]);
1848 * Register hstate attributes for a single node device.
1849 * No-op if attributes already registered.
1851 static void hugetlb_register_node(struct node *node)
1854 struct node_hstate *nhs = &node_hstates[node->dev.id];
1857 if (nhs->hugepages_kobj)
1858 return; /* already allocated */
1860 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1862 if (!nhs->hugepages_kobj)
1865 for_each_hstate(h) {
1866 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1868 &per_node_hstate_attr_group);
1870 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1871 h->name, node->dev.id);
1872 hugetlb_unregister_node(node);
1879 * hugetlb init time: register hstate attributes for all registered node
1880 * devices of nodes that have memory. All on-line nodes should have
1881 * registered their associated device by this time.
1883 static void hugetlb_register_all_nodes(void)
1887 for_each_node_state(nid, N_MEMORY) {
1888 struct node *node = node_devices[nid];
1889 if (node->dev.id == nid)
1890 hugetlb_register_node(node);
1894 * Let the node device driver know we're here so it can
1895 * [un]register hstate attributes on node hotplug.
1897 register_hugetlbfs_with_node(hugetlb_register_node,
1898 hugetlb_unregister_node);
1900 #else /* !CONFIG_NUMA */
1902 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1910 static void hugetlb_unregister_all_nodes(void) { }
1912 static void hugetlb_register_all_nodes(void) { }
1916 static void __exit hugetlb_exit(void)
1920 hugetlb_unregister_all_nodes();
1922 for_each_hstate(h) {
1923 kobject_put(hstate_kobjs[hstate_index(h)]);
1926 kobject_put(hugepages_kobj);
1928 module_exit(hugetlb_exit);
1930 static int __init hugetlb_init(void)
1932 /* Some platform decide whether they support huge pages at boot
1933 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1934 * there is no such support
1936 if (HPAGE_SHIFT == 0)
1939 if (!size_to_hstate(default_hstate_size)) {
1940 default_hstate_size = HPAGE_SIZE;
1941 if (!size_to_hstate(default_hstate_size))
1942 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1944 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1945 if (default_hstate_max_huge_pages)
1946 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1948 hugetlb_init_hstates();
1949 gather_bootmem_prealloc();
1952 hugetlb_sysfs_init();
1953 hugetlb_register_all_nodes();
1954 hugetlb_cgroup_file_init();
1958 module_init(hugetlb_init);
1960 /* Should be called on processing a hugepagesz=... option */
1961 void __init hugetlb_add_hstate(unsigned order)
1966 if (size_to_hstate(PAGE_SIZE << order)) {
1967 pr_warning("hugepagesz= specified twice, ignoring\n");
1970 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1972 h = &hstates[hugetlb_max_hstate++];
1974 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1975 h->nr_huge_pages = 0;
1976 h->free_huge_pages = 0;
1977 for (i = 0; i < MAX_NUMNODES; ++i)
1978 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1979 INIT_LIST_HEAD(&h->hugepage_activelist);
1980 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1981 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1982 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1983 huge_page_size(h)/1024);
1988 static int __init hugetlb_nrpages_setup(char *s)
1991 static unsigned long *last_mhp;
1994 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1995 * so this hugepages= parameter goes to the "default hstate".
1997 if (!hugetlb_max_hstate)
1998 mhp = &default_hstate_max_huge_pages;
2000 mhp = &parsed_hstate->max_huge_pages;
2002 if (mhp == last_mhp) {
2003 pr_warning("hugepages= specified twice without "
2004 "interleaving hugepagesz=, ignoring\n");
2008 if (sscanf(s, "%lu", mhp) <= 0)
2012 * Global state is always initialized later in hugetlb_init.
2013 * But we need to allocate >= MAX_ORDER hstates here early to still
2014 * use the bootmem allocator.
2016 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2017 hugetlb_hstate_alloc_pages(parsed_hstate);
2023 __setup("hugepages=", hugetlb_nrpages_setup);
2025 static int __init hugetlb_default_setup(char *s)
2027 default_hstate_size = memparse(s, &s);
2030 __setup("default_hugepagesz=", hugetlb_default_setup);
2032 static unsigned int cpuset_mems_nr(unsigned int *array)
2035 unsigned int nr = 0;
2037 for_each_node_mask(node, cpuset_current_mems_allowed)
2043 #ifdef CONFIG_SYSCTL
2044 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2045 struct ctl_table *table, int write,
2046 void __user *buffer, size_t *length, loff_t *ppos)
2048 struct hstate *h = &default_hstate;
2052 tmp = h->max_huge_pages;
2054 if (write && h->order >= MAX_ORDER)
2058 table->maxlen = sizeof(unsigned long);
2059 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2064 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2065 GFP_KERNEL | __GFP_NORETRY);
2066 if (!(obey_mempolicy &&
2067 init_nodemask_of_mempolicy(nodes_allowed))) {
2068 NODEMASK_FREE(nodes_allowed);
2069 nodes_allowed = &node_states[N_MEMORY];
2071 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2073 if (nodes_allowed != &node_states[N_MEMORY])
2074 NODEMASK_FREE(nodes_allowed);
2080 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2081 void __user *buffer, size_t *length, loff_t *ppos)
2084 return hugetlb_sysctl_handler_common(false, table, write,
2085 buffer, length, ppos);
2089 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2090 void __user *buffer, size_t *length, loff_t *ppos)
2092 return hugetlb_sysctl_handler_common(true, table, write,
2093 buffer, length, ppos);
2095 #endif /* CONFIG_NUMA */
2097 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2098 void __user *buffer,
2099 size_t *length, loff_t *ppos)
2101 proc_dointvec(table, write, buffer, length, ppos);
2102 if (hugepages_treat_as_movable)
2103 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2105 htlb_alloc_mask = GFP_HIGHUSER;
2109 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2110 void __user *buffer,
2111 size_t *length, loff_t *ppos)
2113 struct hstate *h = &default_hstate;
2117 tmp = h->nr_overcommit_huge_pages;
2119 if (write && h->order >= MAX_ORDER)
2123 table->maxlen = sizeof(unsigned long);
2124 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2129 spin_lock(&hugetlb_lock);
2130 h->nr_overcommit_huge_pages = tmp;
2131 spin_unlock(&hugetlb_lock);
2137 #endif /* CONFIG_SYSCTL */
2139 void hugetlb_report_meminfo(struct seq_file *m)
2141 struct hstate *h = &default_hstate;
2143 "HugePages_Total: %5lu\n"
2144 "HugePages_Free: %5lu\n"
2145 "HugePages_Rsvd: %5lu\n"
2146 "HugePages_Surp: %5lu\n"
2147 "Hugepagesize: %8lu kB\n",
2151 h->surplus_huge_pages,
2152 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2155 int hugetlb_report_node_meminfo(int nid, char *buf)
2157 struct hstate *h = &default_hstate;
2159 "Node %d HugePages_Total: %5u\n"
2160 "Node %d HugePages_Free: %5u\n"
2161 "Node %d HugePages_Surp: %5u\n",
2162 nid, h->nr_huge_pages_node[nid],
2163 nid, h->free_huge_pages_node[nid],
2164 nid, h->surplus_huge_pages_node[nid]);
2167 void hugetlb_show_meminfo(void)
2172 for_each_node_state(nid, N_MEMORY)
2174 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2176 h->nr_huge_pages_node[nid],
2177 h->free_huge_pages_node[nid],
2178 h->surplus_huge_pages_node[nid],
2179 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2182 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2183 unsigned long hugetlb_total_pages(void)
2186 unsigned long nr_total_pages = 0;
2189 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2190 return nr_total_pages;
2193 static int hugetlb_acct_memory(struct hstate *h, long delta)
2197 spin_lock(&hugetlb_lock);
2199 * When cpuset is configured, it breaks the strict hugetlb page
2200 * reservation as the accounting is done on a global variable. Such
2201 * reservation is completely rubbish in the presence of cpuset because
2202 * the reservation is not checked against page availability for the
2203 * current cpuset. Application can still potentially OOM'ed by kernel
2204 * with lack of free htlb page in cpuset that the task is in.
2205 * Attempt to enforce strict accounting with cpuset is almost
2206 * impossible (or too ugly) because cpuset is too fluid that
2207 * task or memory node can be dynamically moved between cpusets.
2209 * The change of semantics for shared hugetlb mapping with cpuset is
2210 * undesirable. However, in order to preserve some of the semantics,
2211 * we fall back to check against current free page availability as
2212 * a best attempt and hopefully to minimize the impact of changing
2213 * semantics that cpuset has.
2216 if (gather_surplus_pages(h, delta) < 0)
2219 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2220 return_unused_surplus_pages(h, delta);
2227 return_unused_surplus_pages(h, (unsigned long) -delta);
2230 spin_unlock(&hugetlb_lock);
2234 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2236 struct resv_map *reservations = vma_resv_map(vma);
2239 * This new VMA should share its siblings reservation map if present.
2240 * The VMA will only ever have a valid reservation map pointer where
2241 * it is being copied for another still existing VMA. As that VMA
2242 * has a reference to the reservation map it cannot disappear until
2243 * after this open call completes. It is therefore safe to take a
2244 * new reference here without additional locking.
2247 kref_get(&reservations->refs);
2250 static void resv_map_put(struct vm_area_struct *vma)
2252 struct resv_map *reservations = vma_resv_map(vma);
2256 kref_put(&reservations->refs, resv_map_release);
2259 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2261 struct hstate *h = hstate_vma(vma);
2262 struct resv_map *reservations = vma_resv_map(vma);
2263 struct hugepage_subpool *spool = subpool_vma(vma);
2264 unsigned long reserve;
2265 unsigned long start;
2269 start = vma_hugecache_offset(h, vma, vma->vm_start);
2270 end = vma_hugecache_offset(h, vma, vma->vm_end);
2272 reserve = (end - start) -
2273 region_count(&reservations->regions, start, end);
2278 hugetlb_acct_memory(h, -reserve);
2279 hugepage_subpool_put_pages(spool, reserve);
2285 * We cannot handle pagefaults against hugetlb pages at all. They cause
2286 * handle_mm_fault() to try to instantiate regular-sized pages in the
2287 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2290 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2296 const struct vm_operations_struct hugetlb_vm_ops = {
2297 .fault = hugetlb_vm_op_fault,
2298 .open = hugetlb_vm_op_open,
2299 .close = hugetlb_vm_op_close,
2302 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2308 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2309 vma->vm_page_prot)));
2311 entry = huge_pte_wrprotect(mk_huge_pte(page,
2312 vma->vm_page_prot));
2314 entry = pte_mkyoung(entry);
2315 entry = pte_mkhuge(entry);
2316 entry = arch_make_huge_pte(entry, vma, page, writable);
2321 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2322 unsigned long address, pte_t *ptep)
2326 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2327 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2328 update_mmu_cache(vma, address, ptep);
2332 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2333 struct vm_area_struct *vma)
2335 pte_t *src_pte, *dst_pte, entry;
2336 struct page *ptepage;
2339 struct hstate *h = hstate_vma(vma);
2340 unsigned long sz = huge_page_size(h);
2342 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2344 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2345 src_pte = huge_pte_offset(src, addr);
2348 dst_pte = huge_pte_alloc(dst, addr, sz);
2352 /* If the pagetables are shared don't copy or take references */
2353 if (dst_pte == src_pte)
2356 spin_lock(&dst->page_table_lock);
2357 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2358 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2360 huge_ptep_set_wrprotect(src, addr, src_pte);
2361 entry = huge_ptep_get(src_pte);
2362 ptepage = pte_page(entry);
2364 page_dup_rmap(ptepage);
2365 set_huge_pte_at(dst, addr, dst_pte, entry);
2367 spin_unlock(&src->page_table_lock);
2368 spin_unlock(&dst->page_table_lock);
2376 static int is_hugetlb_entry_migration(pte_t pte)
2380 if (huge_pte_none(pte) || pte_present(pte))
2382 swp = pte_to_swp_entry(pte);
2383 if (non_swap_entry(swp) && is_migration_entry(swp))
2389 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2393 if (huge_pte_none(pte) || pte_present(pte))
2395 swp = pte_to_swp_entry(pte);
2396 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2402 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2403 unsigned long start, unsigned long end,
2404 struct page *ref_page)
2406 int force_flush = 0;
2407 struct mm_struct *mm = vma->vm_mm;
2408 unsigned long address;
2412 struct hstate *h = hstate_vma(vma);
2413 unsigned long sz = huge_page_size(h);
2414 const unsigned long mmun_start = start; /* For mmu_notifiers */
2415 const unsigned long mmun_end = end; /* For mmu_notifiers */
2417 WARN_ON(!is_vm_hugetlb_page(vma));
2418 BUG_ON(start & ~huge_page_mask(h));
2419 BUG_ON(end & ~huge_page_mask(h));
2421 tlb_start_vma(tlb, vma);
2422 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2424 spin_lock(&mm->page_table_lock);
2425 for (address = start; address < end; address += sz) {
2426 ptep = huge_pte_offset(mm, address);
2430 if (huge_pmd_unshare(mm, &address, ptep))
2433 pte = huge_ptep_get(ptep);
2434 if (huge_pte_none(pte))
2438 * HWPoisoned hugepage is already unmapped and dropped reference
2440 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2441 huge_pte_clear(mm, address, ptep);
2445 page = pte_page(pte);
2447 * If a reference page is supplied, it is because a specific
2448 * page is being unmapped, not a range. Ensure the page we
2449 * are about to unmap is the actual page of interest.
2452 if (page != ref_page)
2456 * Mark the VMA as having unmapped its page so that
2457 * future faults in this VMA will fail rather than
2458 * looking like data was lost
2460 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2463 pte = huge_ptep_get_and_clear(mm, address, ptep);
2464 tlb_remove_tlb_entry(tlb, ptep, address);
2465 if (huge_pte_dirty(pte))
2466 set_page_dirty(page);
2468 page_remove_rmap(page);
2469 force_flush = !__tlb_remove_page(tlb, page);
2472 /* Bail out after unmapping reference page if supplied */
2476 spin_unlock(&mm->page_table_lock);
2478 * mmu_gather ran out of room to batch pages, we break out of
2479 * the PTE lock to avoid doing the potential expensive TLB invalidate
2480 * and page-free while holding it.
2485 if (address < end && !ref_page)
2488 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2489 tlb_end_vma(tlb, vma);
2492 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2493 struct vm_area_struct *vma, unsigned long start,
2494 unsigned long end, struct page *ref_page)
2496 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2499 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2500 * test will fail on a vma being torn down, and not grab a page table
2501 * on its way out. We're lucky that the flag has such an appropriate
2502 * name, and can in fact be safely cleared here. We could clear it
2503 * before the __unmap_hugepage_range above, but all that's necessary
2504 * is to clear it before releasing the i_mmap_mutex. This works
2505 * because in the context this is called, the VMA is about to be
2506 * destroyed and the i_mmap_mutex is held.
2508 vma->vm_flags &= ~VM_MAYSHARE;
2511 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2512 unsigned long end, struct page *ref_page)
2514 struct mm_struct *mm;
2515 struct mmu_gather tlb;
2519 tlb_gather_mmu(&tlb, mm, start, end);
2520 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2521 tlb_finish_mmu(&tlb, start, end);
2525 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2526 * mappping it owns the reserve page for. The intention is to unmap the page
2527 * from other VMAs and let the children be SIGKILLed if they are faulting the
2530 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2531 struct page *page, unsigned long address)
2533 struct hstate *h = hstate_vma(vma);
2534 struct vm_area_struct *iter_vma;
2535 struct address_space *mapping;
2539 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2540 * from page cache lookup which is in HPAGE_SIZE units.
2542 address = address & huge_page_mask(h);
2543 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2545 mapping = file_inode(vma->vm_file)->i_mapping;
2548 * Take the mapping lock for the duration of the table walk. As
2549 * this mapping should be shared between all the VMAs,
2550 * __unmap_hugepage_range() is called as the lock is already held
2552 mutex_lock(&mapping->i_mmap_mutex);
2553 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2554 /* Do not unmap the current VMA */
2555 if (iter_vma == vma)
2559 * Unmap the page from other VMAs without their own reserves.
2560 * They get marked to be SIGKILLed if they fault in these
2561 * areas. This is because a future no-page fault on this VMA
2562 * could insert a zeroed page instead of the data existing
2563 * from the time of fork. This would look like data corruption
2565 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2566 unmap_hugepage_range(iter_vma, address,
2567 address + huge_page_size(h), page);
2569 mutex_unlock(&mapping->i_mmap_mutex);
2575 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2576 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2577 * cannot race with other handlers or page migration.
2578 * Keep the pte_same checks anyway to make transition from the mutex easier.
2580 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2581 unsigned long address, pte_t *ptep, pte_t pte,
2582 struct page *pagecache_page)
2584 struct hstate *h = hstate_vma(vma);
2585 struct page *old_page, *new_page;
2587 int outside_reserve = 0;
2588 unsigned long mmun_start; /* For mmu_notifiers */
2589 unsigned long mmun_end; /* For mmu_notifiers */
2591 old_page = pte_page(pte);
2594 /* If no-one else is actually using this page, avoid the copy
2595 * and just make the page writable */
2596 avoidcopy = (page_mapcount(old_page) == 1);
2598 if (PageAnon(old_page))
2599 page_move_anon_rmap(old_page, vma, address);
2600 set_huge_ptep_writable(vma, address, ptep);
2605 * If the process that created a MAP_PRIVATE mapping is about to
2606 * perform a COW due to a shared page count, attempt to satisfy
2607 * the allocation without using the existing reserves. The pagecache
2608 * page is used to determine if the reserve at this address was
2609 * consumed or not. If reserves were used, a partial faulted mapping
2610 * at the time of fork() could consume its reserves on COW instead
2611 * of the full address range.
2613 if (!(vma->vm_flags & VM_MAYSHARE) &&
2614 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2615 old_page != pagecache_page)
2616 outside_reserve = 1;
2618 page_cache_get(old_page);
2620 /* Drop page_table_lock as buddy allocator may be called */
2621 spin_unlock(&mm->page_table_lock);
2622 new_page = alloc_huge_page(vma, address, outside_reserve);
2624 if (IS_ERR(new_page)) {
2625 long err = PTR_ERR(new_page);
2626 page_cache_release(old_page);
2629 * If a process owning a MAP_PRIVATE mapping fails to COW,
2630 * it is due to references held by a child and an insufficient
2631 * huge page pool. To guarantee the original mappers
2632 * reliability, unmap the page from child processes. The child
2633 * may get SIGKILLed if it later faults.
2635 if (outside_reserve) {
2636 BUG_ON(huge_pte_none(pte));
2637 if (unmap_ref_private(mm, vma, old_page, address)) {
2638 BUG_ON(huge_pte_none(pte));
2639 spin_lock(&mm->page_table_lock);
2640 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2641 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2642 goto retry_avoidcopy;
2644 * race occurs while re-acquiring page_table_lock, and
2652 /* Caller expects lock to be held */
2653 spin_lock(&mm->page_table_lock);
2655 return VM_FAULT_OOM;
2657 return VM_FAULT_SIGBUS;
2661 * When the original hugepage is shared one, it does not have
2662 * anon_vma prepared.
2664 if (unlikely(anon_vma_prepare(vma))) {
2665 page_cache_release(new_page);
2666 page_cache_release(old_page);
2667 /* Caller expects lock to be held */
2668 spin_lock(&mm->page_table_lock);
2669 return VM_FAULT_OOM;
2672 copy_user_huge_page(new_page, old_page, address, vma,
2673 pages_per_huge_page(h));
2674 __SetPageUptodate(new_page);
2676 mmun_start = address & huge_page_mask(h);
2677 mmun_end = mmun_start + huge_page_size(h);
2678 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2680 * Retake the page_table_lock to check for racing updates
2681 * before the page tables are altered
2683 spin_lock(&mm->page_table_lock);
2684 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2685 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2687 huge_ptep_clear_flush(vma, address, ptep);
2688 set_huge_pte_at(mm, address, ptep,
2689 make_huge_pte(vma, new_page, 1));
2690 page_remove_rmap(old_page);
2691 hugepage_add_new_anon_rmap(new_page, vma, address);
2692 /* Make the old page be freed below */
2693 new_page = old_page;
2695 spin_unlock(&mm->page_table_lock);
2696 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2697 /* Caller expects lock to be held */
2698 spin_lock(&mm->page_table_lock);
2699 page_cache_release(new_page);
2700 page_cache_release(old_page);
2704 /* Return the pagecache page at a given address within a VMA */
2705 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2706 struct vm_area_struct *vma, unsigned long address)
2708 struct address_space *mapping;
2711 mapping = vma->vm_file->f_mapping;
2712 idx = vma_hugecache_offset(h, vma, address);
2714 return find_lock_page(mapping, idx);
2718 * Return whether there is a pagecache page to back given address within VMA.
2719 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2721 static bool hugetlbfs_pagecache_present(struct hstate *h,
2722 struct vm_area_struct *vma, unsigned long address)
2724 struct address_space *mapping;
2728 mapping = vma->vm_file->f_mapping;
2729 idx = vma_hugecache_offset(h, vma, address);
2731 page = find_get_page(mapping, idx);
2734 return page != NULL;
2737 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2738 unsigned long address, pte_t *ptep, unsigned int flags)
2740 struct hstate *h = hstate_vma(vma);
2741 int ret = VM_FAULT_SIGBUS;
2746 struct address_space *mapping;
2750 * Currently, we are forced to kill the process in the event the
2751 * original mapper has unmapped pages from the child due to a failed
2752 * COW. Warn that such a situation has occurred as it may not be obvious
2754 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2755 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2760 mapping = vma->vm_file->f_mapping;
2761 idx = vma_hugecache_offset(h, vma, address);
2764 * Use page lock to guard against racing truncation
2765 * before we get page_table_lock.
2768 page = find_lock_page(mapping, idx);
2770 size = i_size_read(mapping->host) >> huge_page_shift(h);
2773 page = alloc_huge_page(vma, address, 0);
2775 ret = PTR_ERR(page);
2779 ret = VM_FAULT_SIGBUS;
2782 clear_huge_page(page, address, pages_per_huge_page(h));
2783 __SetPageUptodate(page);
2785 if (vma->vm_flags & VM_MAYSHARE) {
2787 struct inode *inode = mapping->host;
2789 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2797 spin_lock(&inode->i_lock);
2798 inode->i_blocks += blocks_per_huge_page(h);
2799 spin_unlock(&inode->i_lock);
2802 if (unlikely(anon_vma_prepare(vma))) {
2804 goto backout_unlocked;
2810 * If memory error occurs between mmap() and fault, some process
2811 * don't have hwpoisoned swap entry for errored virtual address.
2812 * So we need to block hugepage fault by PG_hwpoison bit check.
2814 if (unlikely(PageHWPoison(page))) {
2815 ret = VM_FAULT_HWPOISON |
2816 VM_FAULT_SET_HINDEX(hstate_index(h));
2817 goto backout_unlocked;
2822 * If we are going to COW a private mapping later, we examine the
2823 * pending reservations for this page now. This will ensure that
2824 * any allocations necessary to record that reservation occur outside
2827 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2828 if (vma_needs_reservation(h, vma, address) < 0) {
2830 goto backout_unlocked;
2833 spin_lock(&mm->page_table_lock);
2834 size = i_size_read(mapping->host) >> huge_page_shift(h);
2839 if (!huge_pte_none(huge_ptep_get(ptep)))
2843 hugepage_add_new_anon_rmap(page, vma, address);
2845 page_dup_rmap(page);
2846 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2847 && (vma->vm_flags & VM_SHARED)));
2848 set_huge_pte_at(mm, address, ptep, new_pte);
2850 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2851 /* Optimization, do the COW without a second fault */
2852 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2855 spin_unlock(&mm->page_table_lock);
2861 spin_unlock(&mm->page_table_lock);
2868 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2869 unsigned long address, unsigned int flags)
2874 struct page *page = NULL;
2875 struct page *pagecache_page = NULL;
2876 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2877 struct hstate *h = hstate_vma(vma);
2879 address &= huge_page_mask(h);
2881 ptep = huge_pte_offset(mm, address);
2883 entry = huge_ptep_get(ptep);
2884 if (unlikely(is_hugetlb_entry_migration(entry))) {
2885 migration_entry_wait_huge(mm, ptep);
2887 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2888 return VM_FAULT_HWPOISON_LARGE |
2889 VM_FAULT_SET_HINDEX(hstate_index(h));
2892 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2894 return VM_FAULT_OOM;
2897 * Serialize hugepage allocation and instantiation, so that we don't
2898 * get spurious allocation failures if two CPUs race to instantiate
2899 * the same page in the page cache.
2901 mutex_lock(&hugetlb_instantiation_mutex);
2902 entry = huge_ptep_get(ptep);
2903 if (huge_pte_none(entry)) {
2904 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2911 * If we are going to COW the mapping later, we examine the pending
2912 * reservations for this page now. This will ensure that any
2913 * allocations necessary to record that reservation occur outside the
2914 * spinlock. For private mappings, we also lookup the pagecache
2915 * page now as it is used to determine if a reservation has been
2918 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2919 if (vma_needs_reservation(h, vma, address) < 0) {
2924 if (!(vma->vm_flags & VM_MAYSHARE))
2925 pagecache_page = hugetlbfs_pagecache_page(h,
2930 * hugetlb_cow() requires page locks of pte_page(entry) and
2931 * pagecache_page, so here we need take the former one
2932 * when page != pagecache_page or !pagecache_page.
2933 * Note that locking order is always pagecache_page -> page,
2934 * so no worry about deadlock.
2936 page = pte_page(entry);
2938 if (page != pagecache_page)
2941 spin_lock(&mm->page_table_lock);
2942 /* Check for a racing update before calling hugetlb_cow */
2943 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2944 goto out_page_table_lock;
2947 if (flags & FAULT_FLAG_WRITE) {
2948 if (!huge_pte_write(entry)) {
2949 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2951 goto out_page_table_lock;
2953 entry = huge_pte_mkdirty(entry);
2955 entry = pte_mkyoung(entry);
2956 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2957 flags & FAULT_FLAG_WRITE))
2958 update_mmu_cache(vma, address, ptep);
2960 out_page_table_lock:
2961 spin_unlock(&mm->page_table_lock);
2963 if (pagecache_page) {
2964 unlock_page(pagecache_page);
2965 put_page(pagecache_page);
2967 if (page != pagecache_page)
2972 mutex_unlock(&hugetlb_instantiation_mutex);
2977 /* Can be overriden by architectures */
2978 __attribute__((weak)) struct page *
2979 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2980 pud_t *pud, int write)
2986 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2987 struct page **pages, struct vm_area_struct **vmas,
2988 unsigned long *position, unsigned long *nr_pages,
2989 long i, unsigned int flags)
2991 unsigned long pfn_offset;
2992 unsigned long vaddr = *position;
2993 unsigned long remainder = *nr_pages;
2994 struct hstate *h = hstate_vma(vma);
2996 spin_lock(&mm->page_table_lock);
2997 while (vaddr < vma->vm_end && remainder) {
3003 * Some archs (sparc64, sh*) have multiple pte_ts to
3004 * each hugepage. We have to make sure we get the
3005 * first, for the page indexing below to work.
3007 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3008 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3011 * When coredumping, it suits get_dump_page if we just return
3012 * an error where there's an empty slot with no huge pagecache
3013 * to back it. This way, we avoid allocating a hugepage, and
3014 * the sparse dumpfile avoids allocating disk blocks, but its
3015 * huge holes still show up with zeroes where they need to be.
3017 if (absent && (flags & FOLL_DUMP) &&
3018 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3024 * We need call hugetlb_fault for both hugepages under migration
3025 * (in which case hugetlb_fault waits for the migration,) and
3026 * hwpoisoned hugepages (in which case we need to prevent the
3027 * caller from accessing to them.) In order to do this, we use
3028 * here is_swap_pte instead of is_hugetlb_entry_migration and
3029 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3030 * both cases, and because we can't follow correct pages
3031 * directly from any kind of swap entries.
3033 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3034 ((flags & FOLL_WRITE) &&
3035 !huge_pte_write(huge_ptep_get(pte)))) {
3038 spin_unlock(&mm->page_table_lock);
3039 ret = hugetlb_fault(mm, vma, vaddr,
3040 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3041 spin_lock(&mm->page_table_lock);
3042 if (!(ret & VM_FAULT_ERROR))
3049 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3050 page = pte_page(huge_ptep_get(pte));
3053 pages[i] = mem_map_offset(page, pfn_offset);
3064 if (vaddr < vma->vm_end && remainder &&
3065 pfn_offset < pages_per_huge_page(h)) {
3067 * We use pfn_offset to avoid touching the pageframes
3068 * of this compound page.
3073 spin_unlock(&mm->page_table_lock);
3074 *nr_pages = remainder;
3077 return i ? i : -EFAULT;
3080 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3081 unsigned long address, unsigned long end, pgprot_t newprot)
3083 struct mm_struct *mm = vma->vm_mm;
3084 unsigned long start = address;
3087 struct hstate *h = hstate_vma(vma);
3088 unsigned long pages = 0;
3090 BUG_ON(address >= end);
3091 flush_cache_range(vma, address, end);
3093 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3094 spin_lock(&mm->page_table_lock);
3095 for (; address < end; address += huge_page_size(h)) {
3096 ptep = huge_pte_offset(mm, address);
3099 if (huge_pmd_unshare(mm, &address, ptep)) {
3103 if (!huge_pte_none(huge_ptep_get(ptep))) {
3104 pte = huge_ptep_get_and_clear(mm, address, ptep);
3105 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3106 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3107 set_huge_pte_at(mm, address, ptep, pte);
3111 spin_unlock(&mm->page_table_lock);
3113 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3114 * may have cleared our pud entry and done put_page on the page table:
3115 * once we release i_mmap_mutex, another task can do the final put_page
3116 * and that page table be reused and filled with junk.
3118 flush_tlb_range(vma, start, end);
3119 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3121 return pages << h->order;
3124 int hugetlb_reserve_pages(struct inode *inode,
3126 struct vm_area_struct *vma,
3127 vm_flags_t vm_flags)
3130 struct hstate *h = hstate_inode(inode);
3131 struct hugepage_subpool *spool = subpool_inode(inode);
3134 * Only apply hugepage reservation if asked. At fault time, an
3135 * attempt will be made for VM_NORESERVE to allocate a page
3136 * without using reserves
3138 if (vm_flags & VM_NORESERVE)
3142 * Shared mappings base their reservation on the number of pages that
3143 * are already allocated on behalf of the file. Private mappings need
3144 * to reserve the full area even if read-only as mprotect() may be
3145 * called to make the mapping read-write. Assume !vma is a shm mapping
3147 if (!vma || vma->vm_flags & VM_MAYSHARE)
3148 chg = region_chg(&inode->i_mapping->private_list, from, to);
3150 struct resv_map *resv_map = resv_map_alloc();
3156 set_vma_resv_map(vma, resv_map);
3157 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3165 /* There must be enough pages in the subpool for the mapping */
3166 if (hugepage_subpool_get_pages(spool, chg)) {
3172 * Check enough hugepages are available for the reservation.
3173 * Hand the pages back to the subpool if there are not
3175 ret = hugetlb_acct_memory(h, chg);
3177 hugepage_subpool_put_pages(spool, chg);
3182 * Account for the reservations made. Shared mappings record regions
3183 * that have reservations as they are shared by multiple VMAs.
3184 * When the last VMA disappears, the region map says how much
3185 * the reservation was and the page cache tells how much of
3186 * the reservation was consumed. Private mappings are per-VMA and
3187 * only the consumed reservations are tracked. When the VMA
3188 * disappears, the original reservation is the VMA size and the
3189 * consumed reservations are stored in the map. Hence, nothing
3190 * else has to be done for private mappings here
3192 if (!vma || vma->vm_flags & VM_MAYSHARE)
3193 region_add(&inode->i_mapping->private_list, from, to);
3201 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3203 struct hstate *h = hstate_inode(inode);
3204 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3205 struct hugepage_subpool *spool = subpool_inode(inode);
3207 spin_lock(&inode->i_lock);
3208 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3209 spin_unlock(&inode->i_lock);
3211 hugepage_subpool_put_pages(spool, (chg - freed));
3212 hugetlb_acct_memory(h, -(chg - freed));
3215 #ifdef CONFIG_MEMORY_FAILURE
3217 /* Should be called in hugetlb_lock */
3218 static int is_hugepage_on_freelist(struct page *hpage)
3222 struct hstate *h = page_hstate(hpage);
3223 int nid = page_to_nid(hpage);
3225 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3232 * This function is called from memory failure code.
3233 * Assume the caller holds page lock of the head page.
3235 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3237 struct hstate *h = page_hstate(hpage);
3238 int nid = page_to_nid(hpage);
3241 spin_lock(&hugetlb_lock);
3242 if (is_hugepage_on_freelist(hpage)) {
3244 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3245 * but dangling hpage->lru can trigger list-debug warnings
3246 * (this happens when we call unpoison_memory() on it),
3247 * so let it point to itself with list_del_init().
3249 list_del_init(&hpage->lru);
3250 set_page_refcounted(hpage);
3251 h->free_huge_pages--;
3252 h->free_huge_pages_node[nid]--;
3255 spin_unlock(&hugetlb_lock);