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 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
441 VM_BUG_ON(!is_vm_hugetlb_page(vma));
442 if (!(vma->vm_flags & VM_MAYSHARE))
443 vma->vm_private_data = (void *)0;
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
449 if (vma->vm_flags & VM_NORESERVE) {
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
459 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
465 /* Shared mappings always use reserves */
466 if (vma->vm_flags & VM_MAYSHARE)
470 * Only the process that called mmap() has reserves for
473 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
479 static void copy_gigantic_page(struct page *dst, struct page *src)
482 struct hstate *h = page_hstate(src);
483 struct page *dst_base = dst;
484 struct page *src_base = src;
486 for (i = 0; i < pages_per_huge_page(h); ) {
488 copy_highpage(dst, src);
491 dst = mem_map_next(dst, dst_base, i);
492 src = mem_map_next(src, src_base, i);
496 void copy_huge_page(struct page *dst, struct page *src)
499 struct hstate *h = page_hstate(src);
501 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
502 copy_gigantic_page(dst, src);
507 for (i = 0; i < pages_per_huge_page(h); i++) {
509 copy_highpage(dst + i, src + i);
513 static void enqueue_huge_page(struct hstate *h, struct page *page)
515 int nid = page_to_nid(page);
516 list_move(&page->lru, &h->hugepage_freelists[nid]);
517 h->free_huge_pages++;
518 h->free_huge_pages_node[nid]++;
521 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
525 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
526 if (!is_migrate_isolate_page(page))
529 * if 'non-isolated free hugepage' not found on the list,
530 * the allocation fails.
532 if (&h->hugepage_freelists[nid] == &page->lru)
534 list_move(&page->lru, &h->hugepage_activelist);
535 set_page_refcounted(page);
536 h->free_huge_pages--;
537 h->free_huge_pages_node[nid]--;
541 static struct page *dequeue_huge_page_vma(struct hstate *h,
542 struct vm_area_struct *vma,
543 unsigned long address, int avoid_reserve,
546 struct page *page = NULL;
547 struct mempolicy *mpol;
548 nodemask_t *nodemask;
549 struct zonelist *zonelist;
552 unsigned int cpuset_mems_cookie;
555 cpuset_mems_cookie = get_mems_allowed();
556 zonelist = huge_zonelist(vma, address,
557 htlb_alloc_mask, &mpol, &nodemask);
559 * A child process with MAP_PRIVATE mappings created by their parent
560 * have no page reserves. This check ensures that reservations are
561 * not "stolen". The child may still get SIGKILLed
563 if (!vma_has_reserves(vma, chg) &&
564 h->free_huge_pages - h->resv_huge_pages == 0)
567 /* If reserves cannot be used, ensure enough pages are in the pool */
568 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
571 for_each_zone_zonelist_nodemask(zone, z, zonelist,
572 MAX_NR_ZONES - 1, nodemask) {
573 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
574 page = dequeue_huge_page_node(h, zone_to_nid(zone));
578 if (!vma_has_reserves(vma, chg))
581 SetPagePrivate(page);
582 h->resv_huge_pages--;
589 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
598 static void update_and_free_page(struct hstate *h, struct page *page)
602 VM_BUG_ON(h->order >= MAX_ORDER);
605 h->nr_huge_pages_node[page_to_nid(page)]--;
606 for (i = 0; i < pages_per_huge_page(h); i++) {
607 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
608 1 << PG_referenced | 1 << PG_dirty |
609 1 << PG_active | 1 << PG_reserved |
610 1 << PG_private | 1 << PG_writeback);
612 VM_BUG_ON(hugetlb_cgroup_from_page(page));
613 set_compound_page_dtor(page, NULL);
614 set_page_refcounted(page);
615 arch_release_hugepage(page);
616 __free_pages(page, huge_page_order(h));
619 struct hstate *size_to_hstate(unsigned long size)
624 if (huge_page_size(h) == size)
630 static void free_huge_page(struct page *page)
633 * Can't pass hstate in here because it is called from the
634 * compound page destructor.
636 struct hstate *h = page_hstate(page);
637 int nid = page_to_nid(page);
638 struct hugepage_subpool *spool =
639 (struct hugepage_subpool *)page_private(page);
640 bool restore_reserve;
642 set_page_private(page, 0);
643 page->mapping = NULL;
644 BUG_ON(page_count(page));
645 BUG_ON(page_mapcount(page));
646 restore_reserve = PagePrivate(page);
648 spin_lock(&hugetlb_lock);
649 hugetlb_cgroup_uncharge_page(hstate_index(h),
650 pages_per_huge_page(h), page);
652 h->resv_huge_pages++;
654 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
655 /* remove the page from active list */
656 list_del(&page->lru);
657 update_and_free_page(h, page);
658 h->surplus_huge_pages--;
659 h->surplus_huge_pages_node[nid]--;
661 arch_clear_hugepage_flags(page);
662 enqueue_huge_page(h, page);
664 spin_unlock(&hugetlb_lock);
665 hugepage_subpool_put_pages(spool, 1);
668 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
670 INIT_LIST_HEAD(&page->lru);
671 set_compound_page_dtor(page, free_huge_page);
672 spin_lock(&hugetlb_lock);
673 set_hugetlb_cgroup(page, NULL);
675 h->nr_huge_pages_node[nid]++;
676 spin_unlock(&hugetlb_lock);
677 put_page(page); /* free it into the hugepage allocator */
680 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
683 int nr_pages = 1 << order;
684 struct page *p = page + 1;
686 /* we rely on prep_new_huge_page to set the destructor */
687 set_compound_order(page, order);
689 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
691 set_page_count(p, 0);
692 p->first_page = page;
697 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
698 * transparent huge pages. See the PageTransHuge() documentation for more
701 int PageHuge(struct page *page)
703 compound_page_dtor *dtor;
705 if (!PageCompound(page))
708 page = compound_head(page);
709 dtor = get_compound_page_dtor(page);
711 return dtor == free_huge_page;
713 EXPORT_SYMBOL_GPL(PageHuge);
716 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
717 * normal or transparent huge pages.
719 int PageHeadHuge(struct page *page_head)
721 compound_page_dtor *dtor;
723 if (!PageHead(page_head))
726 dtor = get_compound_page_dtor(page_head);
728 return dtor == free_huge_page;
730 EXPORT_SYMBOL_GPL(PageHeadHuge);
732 pgoff_t __basepage_index(struct page *page)
734 struct page *page_head = compound_head(page);
735 pgoff_t index = page_index(page_head);
736 unsigned long compound_idx;
738 if (!PageHuge(page_head))
739 return page_index(page);
741 if (compound_order(page_head) >= MAX_ORDER)
742 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
744 compound_idx = page - page_head;
746 return (index << compound_order(page_head)) + compound_idx;
749 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
753 if (h->order >= MAX_ORDER)
756 page = alloc_pages_exact_node(nid,
757 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
758 __GFP_REPEAT|__GFP_NOWARN,
761 if (arch_prepare_hugepage(page)) {
762 __free_pages(page, huge_page_order(h));
765 prep_new_huge_page(h, page, nid);
772 * common helper functions for hstate_next_node_to_{alloc|free}.
773 * We may have allocated or freed a huge page based on a different
774 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
775 * be outside of *nodes_allowed. Ensure that we use an allowed
776 * node for alloc or free.
778 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
780 nid = next_node(nid, *nodes_allowed);
781 if (nid == MAX_NUMNODES)
782 nid = first_node(*nodes_allowed);
783 VM_BUG_ON(nid >= MAX_NUMNODES);
788 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
790 if (!node_isset(nid, *nodes_allowed))
791 nid = next_node_allowed(nid, nodes_allowed);
796 * returns the previously saved node ["this node"] from which to
797 * allocate a persistent huge page for the pool and advance the
798 * next node from which to allocate, handling wrap at end of node
801 static int hstate_next_node_to_alloc(struct hstate *h,
802 nodemask_t *nodes_allowed)
806 VM_BUG_ON(!nodes_allowed);
808 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
809 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
815 * helper for free_pool_huge_page() - return the previously saved
816 * node ["this node"] from which to free a huge page. Advance the
817 * next node id whether or not we find a free huge page to free so
818 * that the next attempt to free addresses the next node.
820 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
824 VM_BUG_ON(!nodes_allowed);
826 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
827 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
832 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
833 for (nr_nodes = nodes_weight(*mask); \
835 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
838 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
839 for (nr_nodes = nodes_weight(*mask); \
841 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
844 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
850 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
851 page = alloc_fresh_huge_page_node(h, node);
859 count_vm_event(HTLB_BUDDY_PGALLOC);
861 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
867 * Free huge page from pool from next node to free.
868 * Attempt to keep persistent huge pages more or less
869 * balanced over allowed nodes.
870 * Called with hugetlb_lock locked.
872 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
878 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
880 * If we're returning unused surplus pages, only examine
881 * nodes with surplus pages.
883 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
884 !list_empty(&h->hugepage_freelists[node])) {
886 list_entry(h->hugepage_freelists[node].next,
888 list_del(&page->lru);
889 h->free_huge_pages--;
890 h->free_huge_pages_node[node]--;
892 h->surplus_huge_pages--;
893 h->surplus_huge_pages_node[node]--;
895 update_and_free_page(h, page);
904 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
909 if (h->order >= MAX_ORDER)
913 * Assume we will successfully allocate the surplus page to
914 * prevent racing processes from causing the surplus to exceed
917 * This however introduces a different race, where a process B
918 * tries to grow the static hugepage pool while alloc_pages() is
919 * called by process A. B will only examine the per-node
920 * counters in determining if surplus huge pages can be
921 * converted to normal huge pages in adjust_pool_surplus(). A
922 * won't be able to increment the per-node counter, until the
923 * lock is dropped by B, but B doesn't drop hugetlb_lock until
924 * no more huge pages can be converted from surplus to normal
925 * state (and doesn't try to convert again). Thus, we have a
926 * case where a surplus huge page exists, the pool is grown, and
927 * the surplus huge page still exists after, even though it
928 * should just have been converted to a normal huge page. This
929 * does not leak memory, though, as the hugepage will be freed
930 * once it is out of use. It also does not allow the counters to
931 * go out of whack in adjust_pool_surplus() as we don't modify
932 * the node values until we've gotten the hugepage and only the
933 * per-node value is checked there.
935 spin_lock(&hugetlb_lock);
936 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
937 spin_unlock(&hugetlb_lock);
941 h->surplus_huge_pages++;
943 spin_unlock(&hugetlb_lock);
945 if (nid == NUMA_NO_NODE)
946 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
947 __GFP_REPEAT|__GFP_NOWARN,
950 page = alloc_pages_exact_node(nid,
951 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
952 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
954 if (page && arch_prepare_hugepage(page)) {
955 __free_pages(page, huge_page_order(h));
959 spin_lock(&hugetlb_lock);
961 INIT_LIST_HEAD(&page->lru);
962 r_nid = page_to_nid(page);
963 set_compound_page_dtor(page, free_huge_page);
964 set_hugetlb_cgroup(page, NULL);
966 * We incremented the global counters already
968 h->nr_huge_pages_node[r_nid]++;
969 h->surplus_huge_pages_node[r_nid]++;
970 __count_vm_event(HTLB_BUDDY_PGALLOC);
973 h->surplus_huge_pages--;
974 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
976 spin_unlock(&hugetlb_lock);
982 * This allocation function is useful in the context where vma is irrelevant.
983 * E.g. soft-offlining uses this function because it only cares physical
984 * address of error page.
986 struct page *alloc_huge_page_node(struct hstate *h, int nid)
988 struct page *page = NULL;
990 spin_lock(&hugetlb_lock);
991 if (h->free_huge_pages - h->resv_huge_pages > 0)
992 page = dequeue_huge_page_node(h, nid);
993 spin_unlock(&hugetlb_lock);
996 page = alloc_buddy_huge_page(h, nid);
1002 * Increase the hugetlb pool such that it can accommodate a reservation
1005 static int gather_surplus_pages(struct hstate *h, int delta)
1007 struct list_head surplus_list;
1008 struct page *page, *tmp;
1010 int needed, allocated;
1011 bool alloc_ok = true;
1013 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1015 h->resv_huge_pages += delta;
1020 INIT_LIST_HEAD(&surplus_list);
1024 spin_unlock(&hugetlb_lock);
1025 for (i = 0; i < needed; i++) {
1026 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1031 list_add(&page->lru, &surplus_list);
1036 * After retaking hugetlb_lock, we need to recalculate 'needed'
1037 * because either resv_huge_pages or free_huge_pages may have changed.
1039 spin_lock(&hugetlb_lock);
1040 needed = (h->resv_huge_pages + delta) -
1041 (h->free_huge_pages + allocated);
1046 * We were not able to allocate enough pages to
1047 * satisfy the entire reservation so we free what
1048 * we've allocated so far.
1053 * The surplus_list now contains _at_least_ the number of extra pages
1054 * needed to accommodate the reservation. Add the appropriate number
1055 * of pages to the hugetlb pool and free the extras back to the buddy
1056 * allocator. Commit the entire reservation here to prevent another
1057 * process from stealing the pages as they are added to the pool but
1058 * before they are reserved.
1060 needed += allocated;
1061 h->resv_huge_pages += delta;
1064 /* Free the needed pages to the hugetlb pool */
1065 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1069 * This page is now managed by the hugetlb allocator and has
1070 * no users -- drop the buddy allocator's reference.
1072 put_page_testzero(page);
1073 VM_BUG_ON(page_count(page));
1074 enqueue_huge_page(h, page);
1077 spin_unlock(&hugetlb_lock);
1079 /* Free unnecessary surplus pages to the buddy allocator */
1080 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1082 spin_lock(&hugetlb_lock);
1088 * When releasing a hugetlb pool reservation, any surplus pages that were
1089 * allocated to satisfy the reservation must be explicitly freed if they were
1091 * Called with hugetlb_lock held.
1093 static void return_unused_surplus_pages(struct hstate *h,
1094 unsigned long unused_resv_pages)
1096 unsigned long nr_pages;
1098 /* Uncommit the reservation */
1099 h->resv_huge_pages -= unused_resv_pages;
1101 /* Cannot return gigantic pages currently */
1102 if (h->order >= MAX_ORDER)
1105 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1108 * We want to release as many surplus pages as possible, spread
1109 * evenly across all nodes with memory. Iterate across these nodes
1110 * until we can no longer free unreserved surplus pages. This occurs
1111 * when the nodes with surplus pages have no free pages.
1112 * free_pool_huge_page() will balance the the freed pages across the
1113 * on-line nodes with memory and will handle the hstate accounting.
1115 while (nr_pages--) {
1116 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1118 cond_resched_lock(&hugetlb_lock);
1123 * Determine if the huge page at addr within the vma has an associated
1124 * reservation. Where it does not we will need to logically increase
1125 * reservation and actually increase subpool usage before an allocation
1126 * can occur. Where any new reservation would be required the
1127 * reservation change is prepared, but not committed. Once the page
1128 * has been allocated from the subpool and instantiated the change should
1129 * be committed via vma_commit_reservation. No action is required on
1132 static long vma_needs_reservation(struct hstate *h,
1133 struct vm_area_struct *vma, unsigned long addr)
1135 struct address_space *mapping = vma->vm_file->f_mapping;
1136 struct inode *inode = mapping->host;
1138 if (vma->vm_flags & VM_MAYSHARE) {
1139 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1140 return region_chg(&inode->i_mapping->private_list,
1143 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1148 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1149 struct resv_map *resv = vma_resv_map(vma);
1151 err = region_chg(&resv->regions, idx, idx + 1);
1157 static void vma_commit_reservation(struct hstate *h,
1158 struct vm_area_struct *vma, unsigned long addr)
1160 struct address_space *mapping = vma->vm_file->f_mapping;
1161 struct inode *inode = mapping->host;
1163 if (vma->vm_flags & VM_MAYSHARE) {
1164 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1165 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1167 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1168 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1169 struct resv_map *resv = vma_resv_map(vma);
1171 /* Mark this page used in the map. */
1172 region_add(&resv->regions, idx, idx + 1);
1176 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1177 unsigned long addr, int avoid_reserve)
1179 struct hugepage_subpool *spool = subpool_vma(vma);
1180 struct hstate *h = hstate_vma(vma);
1184 struct hugetlb_cgroup *h_cg;
1186 idx = hstate_index(h);
1188 * Processes that did not create the mapping will have no
1189 * reserves and will not have accounted against subpool
1190 * limit. Check that the subpool limit can be made before
1191 * satisfying the allocation MAP_NORESERVE mappings may also
1192 * need pages and subpool limit allocated allocated if no reserve
1195 chg = vma_needs_reservation(h, vma, addr);
1197 return ERR_PTR(-ENOMEM);
1198 if (chg || avoid_reserve)
1199 if (hugepage_subpool_get_pages(spool, 1))
1200 return ERR_PTR(-ENOSPC);
1202 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1204 if (chg || avoid_reserve)
1205 hugepage_subpool_put_pages(spool, 1);
1206 return ERR_PTR(-ENOSPC);
1208 spin_lock(&hugetlb_lock);
1209 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1211 spin_unlock(&hugetlb_lock);
1212 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1214 hugetlb_cgroup_uncharge_cgroup(idx,
1215 pages_per_huge_page(h),
1217 if (chg || avoid_reserve)
1218 hugepage_subpool_put_pages(spool, 1);
1219 return ERR_PTR(-ENOSPC);
1221 spin_lock(&hugetlb_lock);
1222 list_move(&page->lru, &h->hugepage_activelist);
1225 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1226 spin_unlock(&hugetlb_lock);
1228 set_page_private(page, (unsigned long)spool);
1230 vma_commit_reservation(h, vma, addr);
1234 int __weak alloc_bootmem_huge_page(struct hstate *h)
1236 struct huge_bootmem_page *m;
1239 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1242 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1243 huge_page_size(h), huge_page_size(h), 0);
1247 * Use the beginning of the huge page to store the
1248 * huge_bootmem_page struct (until gather_bootmem
1249 * puts them into the mem_map).
1258 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1259 /* Put them into a private list first because mem_map is not up yet */
1260 list_add(&m->list, &huge_boot_pages);
1265 static void prep_compound_huge_page(struct page *page, int order)
1267 if (unlikely(order > (MAX_ORDER - 1)))
1268 prep_compound_gigantic_page(page, order);
1270 prep_compound_page(page, order);
1273 /* Put bootmem huge pages into the standard lists after mem_map is up */
1274 static void __init gather_bootmem_prealloc(void)
1276 struct huge_bootmem_page *m;
1278 list_for_each_entry(m, &huge_boot_pages, list) {
1279 struct hstate *h = m->hstate;
1282 #ifdef CONFIG_HIGHMEM
1283 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1284 free_bootmem_late((unsigned long)m,
1285 sizeof(struct huge_bootmem_page));
1287 page = virt_to_page(m);
1289 __ClearPageReserved(page);
1290 WARN_ON(page_count(page) != 1);
1291 prep_compound_huge_page(page, h->order);
1292 prep_new_huge_page(h, page, page_to_nid(page));
1294 * If we had gigantic hugepages allocated at boot time, we need
1295 * to restore the 'stolen' pages to totalram_pages in order to
1296 * fix confusing memory reports from free(1) and another
1297 * side-effects, like CommitLimit going negative.
1299 if (h->order > (MAX_ORDER - 1))
1300 totalram_pages += 1 << h->order;
1304 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1308 for (i = 0; i < h->max_huge_pages; ++i) {
1309 if (h->order >= MAX_ORDER) {
1310 if (!alloc_bootmem_huge_page(h))
1312 } else if (!alloc_fresh_huge_page(h,
1313 &node_states[N_MEMORY]))
1316 h->max_huge_pages = i;
1319 static void __init hugetlb_init_hstates(void)
1323 for_each_hstate(h) {
1324 /* oversize hugepages were init'ed in early boot */
1325 if (h->order < MAX_ORDER)
1326 hugetlb_hstate_alloc_pages(h);
1330 static char * __init memfmt(char *buf, unsigned long n)
1332 if (n >= (1UL << 30))
1333 sprintf(buf, "%lu GB", n >> 30);
1334 else if (n >= (1UL << 20))
1335 sprintf(buf, "%lu MB", n >> 20);
1337 sprintf(buf, "%lu KB", n >> 10);
1341 static void __init report_hugepages(void)
1345 for_each_hstate(h) {
1347 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1348 memfmt(buf, huge_page_size(h)),
1349 h->free_huge_pages);
1353 #ifdef CONFIG_HIGHMEM
1354 static void try_to_free_low(struct hstate *h, unsigned long count,
1355 nodemask_t *nodes_allowed)
1359 if (h->order >= MAX_ORDER)
1362 for_each_node_mask(i, *nodes_allowed) {
1363 struct page *page, *next;
1364 struct list_head *freel = &h->hugepage_freelists[i];
1365 list_for_each_entry_safe(page, next, freel, lru) {
1366 if (count >= h->nr_huge_pages)
1368 if (PageHighMem(page))
1370 list_del(&page->lru);
1371 update_and_free_page(h, page);
1372 h->free_huge_pages--;
1373 h->free_huge_pages_node[page_to_nid(page)]--;
1378 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1379 nodemask_t *nodes_allowed)
1385 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1386 * balanced by operating on them in a round-robin fashion.
1387 * Returns 1 if an adjustment was made.
1389 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1394 VM_BUG_ON(delta != -1 && delta != 1);
1397 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1398 if (h->surplus_huge_pages_node[node])
1402 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1403 if (h->surplus_huge_pages_node[node] <
1404 h->nr_huge_pages_node[node])
1411 h->surplus_huge_pages += delta;
1412 h->surplus_huge_pages_node[node] += delta;
1416 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1417 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1418 nodemask_t *nodes_allowed)
1420 unsigned long min_count, ret;
1422 if (h->order >= MAX_ORDER)
1423 return h->max_huge_pages;
1426 * Increase the pool size
1427 * First take pages out of surplus state. Then make up the
1428 * remaining difference by allocating fresh huge pages.
1430 * We might race with alloc_buddy_huge_page() here and be unable
1431 * to convert a surplus huge page to a normal huge page. That is
1432 * not critical, though, it just means the overall size of the
1433 * pool might be one hugepage larger than it needs to be, but
1434 * within all the constraints specified by the sysctls.
1436 spin_lock(&hugetlb_lock);
1437 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1438 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1442 while (count > persistent_huge_pages(h)) {
1444 * If this allocation races such that we no longer need the
1445 * page, free_huge_page will handle it by freeing the page
1446 * and reducing the surplus.
1448 spin_unlock(&hugetlb_lock);
1449 ret = alloc_fresh_huge_page(h, nodes_allowed);
1450 spin_lock(&hugetlb_lock);
1454 /* Bail for signals. Probably ctrl-c from user */
1455 if (signal_pending(current))
1460 * Decrease the pool size
1461 * First return free pages to the buddy allocator (being careful
1462 * to keep enough around to satisfy reservations). Then place
1463 * pages into surplus state as needed so the pool will shrink
1464 * to the desired size as pages become free.
1466 * By placing pages into the surplus state independent of the
1467 * overcommit value, we are allowing the surplus pool size to
1468 * exceed overcommit. There are few sane options here. Since
1469 * alloc_buddy_huge_page() is checking the global counter,
1470 * though, we'll note that we're not allowed to exceed surplus
1471 * and won't grow the pool anywhere else. Not until one of the
1472 * sysctls are changed, or the surplus pages go out of use.
1474 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1475 min_count = max(count, min_count);
1476 try_to_free_low(h, min_count, nodes_allowed);
1477 while (min_count < persistent_huge_pages(h)) {
1478 if (!free_pool_huge_page(h, nodes_allowed, 0))
1480 cond_resched_lock(&hugetlb_lock);
1482 while (count < persistent_huge_pages(h)) {
1483 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1487 ret = persistent_huge_pages(h);
1488 spin_unlock(&hugetlb_lock);
1492 #define HSTATE_ATTR_RO(_name) \
1493 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1495 #define HSTATE_ATTR(_name) \
1496 static struct kobj_attribute _name##_attr = \
1497 __ATTR(_name, 0644, _name##_show, _name##_store)
1499 static struct kobject *hugepages_kobj;
1500 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1502 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1504 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1508 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1509 if (hstate_kobjs[i] == kobj) {
1511 *nidp = NUMA_NO_NODE;
1515 return kobj_to_node_hstate(kobj, nidp);
1518 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1519 struct kobj_attribute *attr, char *buf)
1522 unsigned long nr_huge_pages;
1525 h = kobj_to_hstate(kobj, &nid);
1526 if (nid == NUMA_NO_NODE)
1527 nr_huge_pages = h->nr_huge_pages;
1529 nr_huge_pages = h->nr_huge_pages_node[nid];
1531 return sprintf(buf, "%lu\n", nr_huge_pages);
1534 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1535 struct kobject *kobj, struct kobj_attribute *attr,
1536 const char *buf, size_t len)
1540 unsigned long count;
1542 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1544 err = strict_strtoul(buf, 10, &count);
1548 h = kobj_to_hstate(kobj, &nid);
1549 if (h->order >= MAX_ORDER) {
1554 if (nid == NUMA_NO_NODE) {
1556 * global hstate attribute
1558 if (!(obey_mempolicy &&
1559 init_nodemask_of_mempolicy(nodes_allowed))) {
1560 NODEMASK_FREE(nodes_allowed);
1561 nodes_allowed = &node_states[N_MEMORY];
1563 } else if (nodes_allowed) {
1565 * per node hstate attribute: adjust count to global,
1566 * but restrict alloc/free to the specified node.
1568 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1569 init_nodemask_of_node(nodes_allowed, nid);
1571 nodes_allowed = &node_states[N_MEMORY];
1573 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1575 if (nodes_allowed != &node_states[N_MEMORY])
1576 NODEMASK_FREE(nodes_allowed);
1580 NODEMASK_FREE(nodes_allowed);
1584 static ssize_t nr_hugepages_show(struct kobject *kobj,
1585 struct kobj_attribute *attr, char *buf)
1587 return nr_hugepages_show_common(kobj, attr, buf);
1590 static ssize_t nr_hugepages_store(struct kobject *kobj,
1591 struct kobj_attribute *attr, const char *buf, size_t len)
1593 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1595 HSTATE_ATTR(nr_hugepages);
1600 * hstate attribute for optionally mempolicy-based constraint on persistent
1601 * huge page alloc/free.
1603 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1606 return nr_hugepages_show_common(kobj, attr, buf);
1609 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1610 struct kobj_attribute *attr, const char *buf, size_t len)
1612 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1614 HSTATE_ATTR(nr_hugepages_mempolicy);
1618 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1619 struct kobj_attribute *attr, char *buf)
1621 struct hstate *h = kobj_to_hstate(kobj, NULL);
1622 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1625 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1626 struct kobj_attribute *attr, const char *buf, size_t count)
1629 unsigned long input;
1630 struct hstate *h = kobj_to_hstate(kobj, NULL);
1632 if (h->order >= MAX_ORDER)
1635 err = strict_strtoul(buf, 10, &input);
1639 spin_lock(&hugetlb_lock);
1640 h->nr_overcommit_huge_pages = input;
1641 spin_unlock(&hugetlb_lock);
1645 HSTATE_ATTR(nr_overcommit_hugepages);
1647 static ssize_t free_hugepages_show(struct kobject *kobj,
1648 struct kobj_attribute *attr, char *buf)
1651 unsigned long free_huge_pages;
1654 h = kobj_to_hstate(kobj, &nid);
1655 if (nid == NUMA_NO_NODE)
1656 free_huge_pages = h->free_huge_pages;
1658 free_huge_pages = h->free_huge_pages_node[nid];
1660 return sprintf(buf, "%lu\n", free_huge_pages);
1662 HSTATE_ATTR_RO(free_hugepages);
1664 static ssize_t resv_hugepages_show(struct kobject *kobj,
1665 struct kobj_attribute *attr, char *buf)
1667 struct hstate *h = kobj_to_hstate(kobj, NULL);
1668 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1670 HSTATE_ATTR_RO(resv_hugepages);
1672 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1673 struct kobj_attribute *attr, char *buf)
1676 unsigned long surplus_huge_pages;
1679 h = kobj_to_hstate(kobj, &nid);
1680 if (nid == NUMA_NO_NODE)
1681 surplus_huge_pages = h->surplus_huge_pages;
1683 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1685 return sprintf(buf, "%lu\n", surplus_huge_pages);
1687 HSTATE_ATTR_RO(surplus_hugepages);
1689 static struct attribute *hstate_attrs[] = {
1690 &nr_hugepages_attr.attr,
1691 &nr_overcommit_hugepages_attr.attr,
1692 &free_hugepages_attr.attr,
1693 &resv_hugepages_attr.attr,
1694 &surplus_hugepages_attr.attr,
1696 &nr_hugepages_mempolicy_attr.attr,
1701 static struct attribute_group hstate_attr_group = {
1702 .attrs = hstate_attrs,
1705 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1706 struct kobject **hstate_kobjs,
1707 struct attribute_group *hstate_attr_group)
1710 int hi = hstate_index(h);
1712 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1713 if (!hstate_kobjs[hi])
1716 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1718 kobject_put(hstate_kobjs[hi]);
1723 static void __init hugetlb_sysfs_init(void)
1728 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1729 if (!hugepages_kobj)
1732 for_each_hstate(h) {
1733 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1734 hstate_kobjs, &hstate_attr_group);
1736 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1743 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1744 * with node devices in node_devices[] using a parallel array. The array
1745 * index of a node device or _hstate == node id.
1746 * This is here to avoid any static dependency of the node device driver, in
1747 * the base kernel, on the hugetlb module.
1749 struct node_hstate {
1750 struct kobject *hugepages_kobj;
1751 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1753 struct node_hstate node_hstates[MAX_NUMNODES];
1756 * A subset of global hstate attributes for node devices
1758 static struct attribute *per_node_hstate_attrs[] = {
1759 &nr_hugepages_attr.attr,
1760 &free_hugepages_attr.attr,
1761 &surplus_hugepages_attr.attr,
1765 static struct attribute_group per_node_hstate_attr_group = {
1766 .attrs = per_node_hstate_attrs,
1770 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1771 * Returns node id via non-NULL nidp.
1773 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1777 for (nid = 0; nid < nr_node_ids; nid++) {
1778 struct node_hstate *nhs = &node_hstates[nid];
1780 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1781 if (nhs->hstate_kobjs[i] == kobj) {
1793 * Unregister hstate attributes from a single node device.
1794 * No-op if no hstate attributes attached.
1796 static void hugetlb_unregister_node(struct node *node)
1799 struct node_hstate *nhs = &node_hstates[node->dev.id];
1801 if (!nhs->hugepages_kobj)
1802 return; /* no hstate attributes */
1804 for_each_hstate(h) {
1805 int idx = hstate_index(h);
1806 if (nhs->hstate_kobjs[idx]) {
1807 kobject_put(nhs->hstate_kobjs[idx]);
1808 nhs->hstate_kobjs[idx] = NULL;
1812 kobject_put(nhs->hugepages_kobj);
1813 nhs->hugepages_kobj = NULL;
1817 * hugetlb module exit: unregister hstate attributes from node devices
1820 static void hugetlb_unregister_all_nodes(void)
1825 * disable node device registrations.
1827 register_hugetlbfs_with_node(NULL, NULL);
1830 * remove hstate attributes from any nodes that have them.
1832 for (nid = 0; nid < nr_node_ids; nid++)
1833 hugetlb_unregister_node(node_devices[nid]);
1837 * Register hstate attributes for a single node device.
1838 * No-op if attributes already registered.
1840 static void hugetlb_register_node(struct node *node)
1843 struct node_hstate *nhs = &node_hstates[node->dev.id];
1846 if (nhs->hugepages_kobj)
1847 return; /* already allocated */
1849 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1851 if (!nhs->hugepages_kobj)
1854 for_each_hstate(h) {
1855 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1857 &per_node_hstate_attr_group);
1859 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1860 h->name, node->dev.id);
1861 hugetlb_unregister_node(node);
1868 * hugetlb init time: register hstate attributes for all registered node
1869 * devices of nodes that have memory. All on-line nodes should have
1870 * registered their associated device by this time.
1872 static void hugetlb_register_all_nodes(void)
1876 for_each_node_state(nid, N_MEMORY) {
1877 struct node *node = node_devices[nid];
1878 if (node->dev.id == nid)
1879 hugetlb_register_node(node);
1883 * Let the node device driver know we're here so it can
1884 * [un]register hstate attributes on node hotplug.
1886 register_hugetlbfs_with_node(hugetlb_register_node,
1887 hugetlb_unregister_node);
1889 #else /* !CONFIG_NUMA */
1891 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1899 static void hugetlb_unregister_all_nodes(void) { }
1901 static void hugetlb_register_all_nodes(void) { }
1905 static void __exit hugetlb_exit(void)
1909 hugetlb_unregister_all_nodes();
1911 for_each_hstate(h) {
1912 kobject_put(hstate_kobjs[hstate_index(h)]);
1915 kobject_put(hugepages_kobj);
1917 module_exit(hugetlb_exit);
1919 static int __init hugetlb_init(void)
1921 /* Some platform decide whether they support huge pages at boot
1922 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1923 * there is no such support
1925 if (HPAGE_SHIFT == 0)
1928 if (!size_to_hstate(default_hstate_size)) {
1929 default_hstate_size = HPAGE_SIZE;
1930 if (!size_to_hstate(default_hstate_size))
1931 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1933 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1934 if (default_hstate_max_huge_pages)
1935 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1937 hugetlb_init_hstates();
1938 gather_bootmem_prealloc();
1941 hugetlb_sysfs_init();
1942 hugetlb_register_all_nodes();
1943 hugetlb_cgroup_file_init();
1947 module_init(hugetlb_init);
1949 /* Should be called on processing a hugepagesz=... option */
1950 void __init hugetlb_add_hstate(unsigned order)
1955 if (size_to_hstate(PAGE_SIZE << order)) {
1956 pr_warning("hugepagesz= specified twice, ignoring\n");
1959 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1961 h = &hstates[hugetlb_max_hstate++];
1963 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1964 h->nr_huge_pages = 0;
1965 h->free_huge_pages = 0;
1966 for (i = 0; i < MAX_NUMNODES; ++i)
1967 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1968 INIT_LIST_HEAD(&h->hugepage_activelist);
1969 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1970 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1971 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1972 huge_page_size(h)/1024);
1977 static int __init hugetlb_nrpages_setup(char *s)
1980 static unsigned long *last_mhp;
1983 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1984 * so this hugepages= parameter goes to the "default hstate".
1986 if (!hugetlb_max_hstate)
1987 mhp = &default_hstate_max_huge_pages;
1989 mhp = &parsed_hstate->max_huge_pages;
1991 if (mhp == last_mhp) {
1992 pr_warning("hugepages= specified twice without "
1993 "interleaving hugepagesz=, ignoring\n");
1997 if (sscanf(s, "%lu", mhp) <= 0)
2001 * Global state is always initialized later in hugetlb_init.
2002 * But we need to allocate >= MAX_ORDER hstates here early to still
2003 * use the bootmem allocator.
2005 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2006 hugetlb_hstate_alloc_pages(parsed_hstate);
2012 __setup("hugepages=", hugetlb_nrpages_setup);
2014 static int __init hugetlb_default_setup(char *s)
2016 default_hstate_size = memparse(s, &s);
2019 __setup("default_hugepagesz=", hugetlb_default_setup);
2021 static unsigned int cpuset_mems_nr(unsigned int *array)
2024 unsigned int nr = 0;
2026 for_each_node_mask(node, cpuset_current_mems_allowed)
2032 #ifdef CONFIG_SYSCTL
2033 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2034 struct ctl_table *table, int write,
2035 void __user *buffer, size_t *length, loff_t *ppos)
2037 struct hstate *h = &default_hstate;
2041 tmp = h->max_huge_pages;
2043 if (write && h->order >= MAX_ORDER)
2047 table->maxlen = sizeof(unsigned long);
2048 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2053 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2054 GFP_KERNEL | __GFP_NORETRY);
2055 if (!(obey_mempolicy &&
2056 init_nodemask_of_mempolicy(nodes_allowed))) {
2057 NODEMASK_FREE(nodes_allowed);
2058 nodes_allowed = &node_states[N_MEMORY];
2060 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2062 if (nodes_allowed != &node_states[N_MEMORY])
2063 NODEMASK_FREE(nodes_allowed);
2069 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2070 void __user *buffer, size_t *length, loff_t *ppos)
2073 return hugetlb_sysctl_handler_common(false, table, write,
2074 buffer, length, ppos);
2078 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2079 void __user *buffer, size_t *length, loff_t *ppos)
2081 return hugetlb_sysctl_handler_common(true, table, write,
2082 buffer, length, ppos);
2084 #endif /* CONFIG_NUMA */
2086 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2087 void __user *buffer,
2088 size_t *length, loff_t *ppos)
2090 proc_dointvec(table, write, buffer, length, ppos);
2091 if (hugepages_treat_as_movable)
2092 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2094 htlb_alloc_mask = GFP_HIGHUSER;
2098 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2099 void __user *buffer,
2100 size_t *length, loff_t *ppos)
2102 struct hstate *h = &default_hstate;
2106 tmp = h->nr_overcommit_huge_pages;
2108 if (write && h->order >= MAX_ORDER)
2112 table->maxlen = sizeof(unsigned long);
2113 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2118 spin_lock(&hugetlb_lock);
2119 h->nr_overcommit_huge_pages = tmp;
2120 spin_unlock(&hugetlb_lock);
2126 #endif /* CONFIG_SYSCTL */
2128 void hugetlb_report_meminfo(struct seq_file *m)
2130 struct hstate *h = &default_hstate;
2132 "HugePages_Total: %5lu\n"
2133 "HugePages_Free: %5lu\n"
2134 "HugePages_Rsvd: %5lu\n"
2135 "HugePages_Surp: %5lu\n"
2136 "Hugepagesize: %8lu kB\n",
2140 h->surplus_huge_pages,
2141 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2144 int hugetlb_report_node_meminfo(int nid, char *buf)
2146 struct hstate *h = &default_hstate;
2148 "Node %d HugePages_Total: %5u\n"
2149 "Node %d HugePages_Free: %5u\n"
2150 "Node %d HugePages_Surp: %5u\n",
2151 nid, h->nr_huge_pages_node[nid],
2152 nid, h->free_huge_pages_node[nid],
2153 nid, h->surplus_huge_pages_node[nid]);
2156 void hugetlb_show_meminfo(void)
2161 for_each_node_state(nid, N_MEMORY)
2163 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2165 h->nr_huge_pages_node[nid],
2166 h->free_huge_pages_node[nid],
2167 h->surplus_huge_pages_node[nid],
2168 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2171 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2172 unsigned long hugetlb_total_pages(void)
2175 unsigned long nr_total_pages = 0;
2178 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2179 return nr_total_pages;
2182 static int hugetlb_acct_memory(struct hstate *h, long delta)
2186 spin_lock(&hugetlb_lock);
2188 * When cpuset is configured, it breaks the strict hugetlb page
2189 * reservation as the accounting is done on a global variable. Such
2190 * reservation is completely rubbish in the presence of cpuset because
2191 * the reservation is not checked against page availability for the
2192 * current cpuset. Application can still potentially OOM'ed by kernel
2193 * with lack of free htlb page in cpuset that the task is in.
2194 * Attempt to enforce strict accounting with cpuset is almost
2195 * impossible (or too ugly) because cpuset is too fluid that
2196 * task or memory node can be dynamically moved between cpusets.
2198 * The change of semantics for shared hugetlb mapping with cpuset is
2199 * undesirable. However, in order to preserve some of the semantics,
2200 * we fall back to check against current free page availability as
2201 * a best attempt and hopefully to minimize the impact of changing
2202 * semantics that cpuset has.
2205 if (gather_surplus_pages(h, delta) < 0)
2208 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2209 return_unused_surplus_pages(h, delta);
2216 return_unused_surplus_pages(h, (unsigned long) -delta);
2219 spin_unlock(&hugetlb_lock);
2223 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2225 struct resv_map *resv = vma_resv_map(vma);
2228 * This new VMA should share its siblings reservation map if present.
2229 * The VMA will only ever have a valid reservation map pointer where
2230 * it is being copied for another still existing VMA. As that VMA
2231 * has a reference to the reservation map it cannot disappear until
2232 * after this open call completes. It is therefore safe to take a
2233 * new reference here without additional locking.
2236 kref_get(&resv->refs);
2239 static void resv_map_put(struct vm_area_struct *vma)
2241 struct resv_map *resv = vma_resv_map(vma);
2245 kref_put(&resv->refs, resv_map_release);
2248 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2250 struct hstate *h = hstate_vma(vma);
2251 struct resv_map *resv = vma_resv_map(vma);
2252 struct hugepage_subpool *spool = subpool_vma(vma);
2253 unsigned long reserve;
2254 unsigned long start;
2258 start = vma_hugecache_offset(h, vma, vma->vm_start);
2259 end = vma_hugecache_offset(h, vma, vma->vm_end);
2261 reserve = (end - start) -
2262 region_count(&resv->regions, start, end);
2267 hugetlb_acct_memory(h, -reserve);
2268 hugepage_subpool_put_pages(spool, reserve);
2274 * We cannot handle pagefaults against hugetlb pages at all. They cause
2275 * handle_mm_fault() to try to instantiate regular-sized pages in the
2276 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2279 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2285 const struct vm_operations_struct hugetlb_vm_ops = {
2286 .fault = hugetlb_vm_op_fault,
2287 .open = hugetlb_vm_op_open,
2288 .close = hugetlb_vm_op_close,
2291 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2297 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2298 vma->vm_page_prot)));
2300 entry = huge_pte_wrprotect(mk_huge_pte(page,
2301 vma->vm_page_prot));
2303 entry = pte_mkyoung(entry);
2304 entry = pte_mkhuge(entry);
2305 entry = arch_make_huge_pte(entry, vma, page, writable);
2310 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2311 unsigned long address, pte_t *ptep)
2315 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2316 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2317 update_mmu_cache(vma, address, ptep);
2320 static int is_hugetlb_entry_migration(pte_t pte)
2324 if (huge_pte_none(pte) || pte_present(pte))
2326 swp = pte_to_swp_entry(pte);
2327 if (non_swap_entry(swp) && is_migration_entry(swp))
2333 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2337 if (huge_pte_none(pte) || pte_present(pte))
2339 swp = pte_to_swp_entry(pte);
2340 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2346 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2347 struct vm_area_struct *vma)
2349 pte_t *src_pte, *dst_pte, entry;
2350 struct page *ptepage;
2353 struct hstate *h = hstate_vma(vma);
2354 unsigned long sz = huge_page_size(h);
2355 unsigned long mmun_start; /* For mmu_notifiers */
2356 unsigned long mmun_end; /* For mmu_notifiers */
2359 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2361 mmun_start = vma->vm_start;
2362 mmun_end = vma->vm_end;
2364 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2366 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2367 src_pte = huge_pte_offset(src, addr);
2370 dst_pte = huge_pte_alloc(dst, addr, sz);
2376 /* If the pagetables are shared don't copy or take references */
2377 if (dst_pte == src_pte)
2380 spin_lock(&dst->page_table_lock);
2381 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2382 entry = huge_ptep_get(src_pte);
2383 if (huge_pte_none(entry)) { /* skip none entry */
2385 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2386 is_hugetlb_entry_hwpoisoned(entry))) {
2387 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2389 if (is_write_migration_entry(swp_entry) && cow) {
2391 * COW mappings require pages in both
2392 * parent and child to be set to read.
2394 make_migration_entry_read(&swp_entry);
2395 entry = swp_entry_to_pte(swp_entry);
2396 set_huge_pte_at(src, addr, src_pte, entry);
2398 set_huge_pte_at(dst, addr, dst_pte, entry);
2401 huge_ptep_set_wrprotect(src, addr, src_pte);
2402 entry = huge_ptep_get(src_pte);
2403 ptepage = pte_page(entry);
2405 page_dup_rmap(ptepage);
2406 set_huge_pte_at(dst, addr, dst_pte, entry);
2408 spin_unlock(&src->page_table_lock);
2409 spin_unlock(&dst->page_table_lock);
2413 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2418 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2419 unsigned long start, unsigned long end,
2420 struct page *ref_page)
2422 int force_flush = 0;
2423 struct mm_struct *mm = vma->vm_mm;
2424 unsigned long address;
2428 struct hstate *h = hstate_vma(vma);
2429 unsigned long sz = huge_page_size(h);
2430 const unsigned long mmun_start = start; /* For mmu_notifiers */
2431 const unsigned long mmun_end = end; /* For mmu_notifiers */
2433 WARN_ON(!is_vm_hugetlb_page(vma));
2434 BUG_ON(start & ~huge_page_mask(h));
2435 BUG_ON(end & ~huge_page_mask(h));
2437 tlb_start_vma(tlb, vma);
2438 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2440 spin_lock(&mm->page_table_lock);
2441 for (address = start; address < end; address += sz) {
2442 ptep = huge_pte_offset(mm, address);
2446 if (huge_pmd_unshare(mm, &address, ptep))
2449 pte = huge_ptep_get(ptep);
2450 if (huge_pte_none(pte))
2454 * HWPoisoned hugepage is already unmapped and dropped reference
2456 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2457 huge_pte_clear(mm, address, ptep);
2461 page = pte_page(pte);
2463 * If a reference page is supplied, it is because a specific
2464 * page is being unmapped, not a range. Ensure the page we
2465 * are about to unmap is the actual page of interest.
2468 if (page != ref_page)
2472 * Mark the VMA as having unmapped its page so that
2473 * future faults in this VMA will fail rather than
2474 * looking like data was lost
2476 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2479 pte = huge_ptep_get_and_clear(mm, address, ptep);
2480 tlb_remove_tlb_entry(tlb, ptep, address);
2481 if (huge_pte_dirty(pte))
2482 set_page_dirty(page);
2484 page_remove_rmap(page);
2485 force_flush = !__tlb_remove_page(tlb, page);
2488 /* Bail out after unmapping reference page if supplied */
2492 spin_unlock(&mm->page_table_lock);
2494 * mmu_gather ran out of room to batch pages, we break out of
2495 * the PTE lock to avoid doing the potential expensive TLB invalidate
2496 * and page-free while holding it.
2501 if (address < end && !ref_page)
2504 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2505 tlb_end_vma(tlb, vma);
2508 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2509 struct vm_area_struct *vma, unsigned long start,
2510 unsigned long end, struct page *ref_page)
2512 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2515 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2516 * test will fail on a vma being torn down, and not grab a page table
2517 * on its way out. We're lucky that the flag has such an appropriate
2518 * name, and can in fact be safely cleared here. We could clear it
2519 * before the __unmap_hugepage_range above, but all that's necessary
2520 * is to clear it before releasing the i_mmap_mutex. This works
2521 * because in the context this is called, the VMA is about to be
2522 * destroyed and the i_mmap_mutex is held.
2524 vma->vm_flags &= ~VM_MAYSHARE;
2527 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2528 unsigned long end, struct page *ref_page)
2530 struct mm_struct *mm;
2531 struct mmu_gather tlb;
2535 tlb_gather_mmu(&tlb, mm, start, end);
2536 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2537 tlb_finish_mmu(&tlb, start, end);
2541 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2542 * mappping it owns the reserve page for. The intention is to unmap the page
2543 * from other VMAs and let the children be SIGKILLed if they are faulting the
2546 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2547 struct page *page, unsigned long address)
2549 struct hstate *h = hstate_vma(vma);
2550 struct vm_area_struct *iter_vma;
2551 struct address_space *mapping;
2555 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2556 * from page cache lookup which is in HPAGE_SIZE units.
2558 address = address & huge_page_mask(h);
2559 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2561 mapping = file_inode(vma->vm_file)->i_mapping;
2564 * Take the mapping lock for the duration of the table walk. As
2565 * this mapping should be shared between all the VMAs,
2566 * __unmap_hugepage_range() is called as the lock is already held
2568 mutex_lock(&mapping->i_mmap_mutex);
2569 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2570 /* Do not unmap the current VMA */
2571 if (iter_vma == vma)
2575 * Unmap the page from other VMAs without their own reserves.
2576 * They get marked to be SIGKILLed if they fault in these
2577 * areas. This is because a future no-page fault on this VMA
2578 * could insert a zeroed page instead of the data existing
2579 * from the time of fork. This would look like data corruption
2581 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2582 unmap_hugepage_range(iter_vma, address,
2583 address + huge_page_size(h), page);
2585 mutex_unlock(&mapping->i_mmap_mutex);
2591 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2592 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2593 * cannot race with other handlers or page migration.
2594 * Keep the pte_same checks anyway to make transition from the mutex easier.
2596 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2597 unsigned long address, pte_t *ptep, pte_t pte,
2598 struct page *pagecache_page)
2600 struct hstate *h = hstate_vma(vma);
2601 struct page *old_page, *new_page;
2602 int outside_reserve = 0;
2603 unsigned long mmun_start; /* For mmu_notifiers */
2604 unsigned long mmun_end; /* For mmu_notifiers */
2606 old_page = pte_page(pte);
2609 /* If no-one else is actually using this page, avoid the copy
2610 * and just make the page writable */
2611 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2612 page_move_anon_rmap(old_page, vma, address);
2613 set_huge_ptep_writable(vma, address, ptep);
2618 * If the process that created a MAP_PRIVATE mapping is about to
2619 * perform a COW due to a shared page count, attempt to satisfy
2620 * the allocation without using the existing reserves. The pagecache
2621 * page is used to determine if the reserve at this address was
2622 * consumed or not. If reserves were used, a partial faulted mapping
2623 * at the time of fork() could consume its reserves on COW instead
2624 * of the full address range.
2626 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2627 old_page != pagecache_page)
2628 outside_reserve = 1;
2630 page_cache_get(old_page);
2632 /* Drop page_table_lock as buddy allocator may be called */
2633 spin_unlock(&mm->page_table_lock);
2634 new_page = alloc_huge_page(vma, address, outside_reserve);
2636 if (IS_ERR(new_page)) {
2637 long err = PTR_ERR(new_page);
2638 page_cache_release(old_page);
2641 * If a process owning a MAP_PRIVATE mapping fails to COW,
2642 * it is due to references held by a child and an insufficient
2643 * huge page pool. To guarantee the original mappers
2644 * reliability, unmap the page from child processes. The child
2645 * may get SIGKILLed if it later faults.
2647 if (outside_reserve) {
2648 BUG_ON(huge_pte_none(pte));
2649 if (unmap_ref_private(mm, vma, old_page, address)) {
2650 BUG_ON(huge_pte_none(pte));
2651 spin_lock(&mm->page_table_lock);
2652 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2653 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2654 goto retry_avoidcopy;
2656 * race occurs while re-acquiring page_table_lock, and
2664 /* Caller expects lock to be held */
2665 spin_lock(&mm->page_table_lock);
2667 return VM_FAULT_OOM;
2669 return VM_FAULT_SIGBUS;
2673 * When the original hugepage is shared one, it does not have
2674 * anon_vma prepared.
2676 if (unlikely(anon_vma_prepare(vma))) {
2677 page_cache_release(new_page);
2678 page_cache_release(old_page);
2679 /* Caller expects lock to be held */
2680 spin_lock(&mm->page_table_lock);
2681 return VM_FAULT_OOM;
2684 copy_user_huge_page(new_page, old_page, address, vma,
2685 pages_per_huge_page(h));
2686 __SetPageUptodate(new_page);
2688 mmun_start = address & huge_page_mask(h);
2689 mmun_end = mmun_start + huge_page_size(h);
2690 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2692 * Retake the page_table_lock to check for racing updates
2693 * before the page tables are altered
2695 spin_lock(&mm->page_table_lock);
2696 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2697 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2698 ClearPagePrivate(new_page);
2701 huge_ptep_clear_flush(vma, address, ptep);
2702 set_huge_pte_at(mm, address, ptep,
2703 make_huge_pte(vma, new_page, 1));
2704 page_remove_rmap(old_page);
2705 hugepage_add_new_anon_rmap(new_page, vma, address);
2706 /* Make the old page be freed below */
2707 new_page = old_page;
2709 spin_unlock(&mm->page_table_lock);
2710 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2711 page_cache_release(new_page);
2712 page_cache_release(old_page);
2714 /* Caller expects lock to be held */
2715 spin_lock(&mm->page_table_lock);
2719 /* Return the pagecache page at a given address within a VMA */
2720 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2721 struct vm_area_struct *vma, unsigned long address)
2723 struct address_space *mapping;
2726 mapping = vma->vm_file->f_mapping;
2727 idx = vma_hugecache_offset(h, vma, address);
2729 return find_lock_page(mapping, idx);
2733 * Return whether there is a pagecache page to back given address within VMA.
2734 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2736 static bool hugetlbfs_pagecache_present(struct hstate *h,
2737 struct vm_area_struct *vma, unsigned long address)
2739 struct address_space *mapping;
2743 mapping = vma->vm_file->f_mapping;
2744 idx = vma_hugecache_offset(h, vma, address);
2746 page = find_get_page(mapping, idx);
2749 return page != NULL;
2752 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2753 unsigned long address, pte_t *ptep, unsigned int flags)
2755 struct hstate *h = hstate_vma(vma);
2756 int ret = VM_FAULT_SIGBUS;
2761 struct address_space *mapping;
2765 * Currently, we are forced to kill the process in the event the
2766 * original mapper has unmapped pages from the child due to a failed
2767 * COW. Warn that such a situation has occurred as it may not be obvious
2769 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2770 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2775 mapping = vma->vm_file->f_mapping;
2776 idx = vma_hugecache_offset(h, vma, address);
2779 * Use page lock to guard against racing truncation
2780 * before we get page_table_lock.
2783 page = find_lock_page(mapping, idx);
2785 size = i_size_read(mapping->host) >> huge_page_shift(h);
2788 page = alloc_huge_page(vma, address, 0);
2790 ret = PTR_ERR(page);
2794 ret = VM_FAULT_SIGBUS;
2797 clear_huge_page(page, address, pages_per_huge_page(h));
2798 __SetPageUptodate(page);
2800 if (vma->vm_flags & VM_MAYSHARE) {
2802 struct inode *inode = mapping->host;
2804 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2811 ClearPagePrivate(page);
2813 spin_lock(&inode->i_lock);
2814 inode->i_blocks += blocks_per_huge_page(h);
2815 spin_unlock(&inode->i_lock);
2818 if (unlikely(anon_vma_prepare(vma))) {
2820 goto backout_unlocked;
2826 * If memory error occurs between mmap() and fault, some process
2827 * don't have hwpoisoned swap entry for errored virtual address.
2828 * So we need to block hugepage fault by PG_hwpoison bit check.
2830 if (unlikely(PageHWPoison(page))) {
2831 ret = VM_FAULT_HWPOISON |
2832 VM_FAULT_SET_HINDEX(hstate_index(h));
2833 goto backout_unlocked;
2838 * If we are going to COW a private mapping later, we examine the
2839 * pending reservations for this page now. This will ensure that
2840 * any allocations necessary to record that reservation occur outside
2843 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2844 if (vma_needs_reservation(h, vma, address) < 0) {
2846 goto backout_unlocked;
2849 spin_lock(&mm->page_table_lock);
2850 size = i_size_read(mapping->host) >> huge_page_shift(h);
2855 if (!huge_pte_none(huge_ptep_get(ptep)))
2859 ClearPagePrivate(page);
2860 hugepage_add_new_anon_rmap(page, vma, address);
2863 page_dup_rmap(page);
2864 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2865 && (vma->vm_flags & VM_SHARED)));
2866 set_huge_pte_at(mm, address, ptep, new_pte);
2868 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2869 /* Optimization, do the COW without a second fault */
2870 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2873 spin_unlock(&mm->page_table_lock);
2879 spin_unlock(&mm->page_table_lock);
2886 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2887 unsigned long address, unsigned int flags)
2892 struct page *page = NULL;
2893 struct page *pagecache_page = NULL;
2894 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2895 struct hstate *h = hstate_vma(vma);
2897 address &= huge_page_mask(h);
2899 ptep = huge_pte_offset(mm, address);
2901 entry = huge_ptep_get(ptep);
2902 if (unlikely(is_hugetlb_entry_migration(entry))) {
2903 migration_entry_wait_huge(mm, ptep);
2905 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2906 return VM_FAULT_HWPOISON_LARGE |
2907 VM_FAULT_SET_HINDEX(hstate_index(h));
2910 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2912 return VM_FAULT_OOM;
2915 * Serialize hugepage allocation and instantiation, so that we don't
2916 * get spurious allocation failures if two CPUs race to instantiate
2917 * the same page in the page cache.
2919 mutex_lock(&hugetlb_instantiation_mutex);
2920 entry = huge_ptep_get(ptep);
2921 if (huge_pte_none(entry)) {
2922 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2929 * If we are going to COW the mapping later, we examine the pending
2930 * reservations for this page now. This will ensure that any
2931 * allocations necessary to record that reservation occur outside the
2932 * spinlock. For private mappings, we also lookup the pagecache
2933 * page now as it is used to determine if a reservation has been
2936 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2937 if (vma_needs_reservation(h, vma, address) < 0) {
2942 if (!(vma->vm_flags & VM_MAYSHARE))
2943 pagecache_page = hugetlbfs_pagecache_page(h,
2948 * hugetlb_cow() requires page locks of pte_page(entry) and
2949 * pagecache_page, so here we need take the former one
2950 * when page != pagecache_page or !pagecache_page.
2951 * Note that locking order is always pagecache_page -> page,
2952 * so no worry about deadlock.
2954 page = pte_page(entry);
2956 if (page != pagecache_page)
2959 spin_lock(&mm->page_table_lock);
2960 /* Check for a racing update before calling hugetlb_cow */
2961 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2962 goto out_page_table_lock;
2965 if (flags & FAULT_FLAG_WRITE) {
2966 if (!huge_pte_write(entry)) {
2967 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2969 goto out_page_table_lock;
2971 entry = huge_pte_mkdirty(entry);
2973 entry = pte_mkyoung(entry);
2974 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2975 flags & FAULT_FLAG_WRITE))
2976 update_mmu_cache(vma, address, ptep);
2978 out_page_table_lock:
2979 spin_unlock(&mm->page_table_lock);
2981 if (pagecache_page) {
2982 unlock_page(pagecache_page);
2983 put_page(pagecache_page);
2985 if (page != pagecache_page)
2990 mutex_unlock(&hugetlb_instantiation_mutex);
2995 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2996 struct page **pages, struct vm_area_struct **vmas,
2997 unsigned long *position, unsigned long *nr_pages,
2998 long i, unsigned int flags)
3000 unsigned long pfn_offset;
3001 unsigned long vaddr = *position;
3002 unsigned long remainder = *nr_pages;
3003 struct hstate *h = hstate_vma(vma);
3005 spin_lock(&mm->page_table_lock);
3006 while (vaddr < vma->vm_end && remainder) {
3012 * Some archs (sparc64, sh*) have multiple pte_ts to
3013 * each hugepage. We have to make sure we get the
3014 * first, for the page indexing below to work.
3016 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3017 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3020 * When coredumping, it suits get_dump_page if we just return
3021 * an error where there's an empty slot with no huge pagecache
3022 * to back it. This way, we avoid allocating a hugepage, and
3023 * the sparse dumpfile avoids allocating disk blocks, but its
3024 * huge holes still show up with zeroes where they need to be.
3026 if (absent && (flags & FOLL_DUMP) &&
3027 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3033 * We need call hugetlb_fault for both hugepages under migration
3034 * (in which case hugetlb_fault waits for the migration,) and
3035 * hwpoisoned hugepages (in which case we need to prevent the
3036 * caller from accessing to them.) In order to do this, we use
3037 * here is_swap_pte instead of is_hugetlb_entry_migration and
3038 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3039 * both cases, and because we can't follow correct pages
3040 * directly from any kind of swap entries.
3042 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3043 ((flags & FOLL_WRITE) &&
3044 !huge_pte_write(huge_ptep_get(pte)))) {
3047 spin_unlock(&mm->page_table_lock);
3048 ret = hugetlb_fault(mm, vma, vaddr,
3049 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3050 spin_lock(&mm->page_table_lock);
3051 if (!(ret & VM_FAULT_ERROR))
3058 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3059 page = pte_page(huge_ptep_get(pte));
3062 pages[i] = mem_map_offset(page, pfn_offset);
3073 if (vaddr < vma->vm_end && remainder &&
3074 pfn_offset < pages_per_huge_page(h)) {
3076 * We use pfn_offset to avoid touching the pageframes
3077 * of this compound page.
3082 spin_unlock(&mm->page_table_lock);
3083 *nr_pages = remainder;
3086 return i ? i : -EFAULT;
3089 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3090 unsigned long address, unsigned long end, pgprot_t newprot)
3092 struct mm_struct *mm = vma->vm_mm;
3093 unsigned long start = address;
3096 struct hstate *h = hstate_vma(vma);
3097 unsigned long pages = 0;
3099 BUG_ON(address >= end);
3100 flush_cache_range(vma, address, end);
3102 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3103 spin_lock(&mm->page_table_lock);
3104 for (; address < end; address += huge_page_size(h)) {
3105 ptep = huge_pte_offset(mm, address);
3108 if (huge_pmd_unshare(mm, &address, ptep)) {
3112 if (!huge_pte_none(huge_ptep_get(ptep))) {
3113 pte = huge_ptep_get_and_clear(mm, address, ptep);
3114 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3115 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3116 set_huge_pte_at(mm, address, ptep, pte);
3120 spin_unlock(&mm->page_table_lock);
3122 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3123 * may have cleared our pud entry and done put_page on the page table:
3124 * once we release i_mmap_mutex, another task can do the final put_page
3125 * and that page table be reused and filled with junk.
3127 flush_tlb_range(vma, start, end);
3128 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3130 return pages << h->order;
3133 int hugetlb_reserve_pages(struct inode *inode,
3135 struct vm_area_struct *vma,
3136 vm_flags_t vm_flags)
3139 struct hstate *h = hstate_inode(inode);
3140 struct hugepage_subpool *spool = subpool_inode(inode);
3143 * Only apply hugepage reservation if asked. At fault time, an
3144 * attempt will be made for VM_NORESERVE to allocate a page
3145 * without using reserves
3147 if (vm_flags & VM_NORESERVE)
3151 * Shared mappings base their reservation on the number of pages that
3152 * are already allocated on behalf of the file. Private mappings need
3153 * to reserve the full area even if read-only as mprotect() may be
3154 * called to make the mapping read-write. Assume !vma is a shm mapping
3156 if (!vma || vma->vm_flags & VM_MAYSHARE)
3157 chg = region_chg(&inode->i_mapping->private_list, from, to);
3159 struct resv_map *resv_map = resv_map_alloc();
3165 set_vma_resv_map(vma, resv_map);
3166 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3174 /* There must be enough pages in the subpool for the mapping */
3175 if (hugepage_subpool_get_pages(spool, chg)) {
3181 * Check enough hugepages are available for the reservation.
3182 * Hand the pages back to the subpool if there are not
3184 ret = hugetlb_acct_memory(h, chg);
3186 hugepage_subpool_put_pages(spool, chg);
3191 * Account for the reservations made. Shared mappings record regions
3192 * that have reservations as they are shared by multiple VMAs.
3193 * When the last VMA disappears, the region map says how much
3194 * the reservation was and the page cache tells how much of
3195 * the reservation was consumed. Private mappings are per-VMA and
3196 * only the consumed reservations are tracked. When the VMA
3197 * disappears, the original reservation is the VMA size and the
3198 * consumed reservations are stored in the map. Hence, nothing
3199 * else has to be done for private mappings here
3201 if (!vma || vma->vm_flags & VM_MAYSHARE)
3202 region_add(&inode->i_mapping->private_list, from, to);
3210 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3212 struct hstate *h = hstate_inode(inode);
3213 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3214 struct hugepage_subpool *spool = subpool_inode(inode);
3216 spin_lock(&inode->i_lock);
3217 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3218 spin_unlock(&inode->i_lock);
3220 hugepage_subpool_put_pages(spool, (chg - freed));
3221 hugetlb_acct_memory(h, -(chg - freed));
3224 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3225 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3226 struct vm_area_struct *vma,
3227 unsigned long addr, pgoff_t idx)
3229 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3231 unsigned long sbase = saddr & PUD_MASK;
3232 unsigned long s_end = sbase + PUD_SIZE;
3234 /* Allow segments to share if only one is marked locked */
3235 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3236 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3239 * match the virtual addresses, permission and the alignment of the
3242 if (pmd_index(addr) != pmd_index(saddr) ||
3243 vm_flags != svm_flags ||
3244 sbase < svma->vm_start || svma->vm_end < s_end)
3250 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3252 unsigned long base = addr & PUD_MASK;
3253 unsigned long end = base + PUD_SIZE;
3256 * check on proper vm_flags and page table alignment
3258 if (vma->vm_flags & VM_MAYSHARE &&
3259 vma->vm_start <= base && end <= vma->vm_end)
3265 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3266 * and returns the corresponding pte. While this is not necessary for the
3267 * !shared pmd case because we can allocate the pmd later as well, it makes the
3268 * code much cleaner. pmd allocation is essential for the shared case because
3269 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3270 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3271 * bad pmd for sharing.
3273 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3275 struct vm_area_struct *vma = find_vma(mm, addr);
3276 struct address_space *mapping = vma->vm_file->f_mapping;
3277 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3279 struct vm_area_struct *svma;
3280 unsigned long saddr;
3284 if (!vma_shareable(vma, addr))
3285 return (pte_t *)pmd_alloc(mm, pud, addr);
3287 mutex_lock(&mapping->i_mmap_mutex);
3288 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3292 saddr = page_table_shareable(svma, vma, addr, idx);
3294 spte = huge_pte_offset(svma->vm_mm, saddr);
3296 get_page(virt_to_page(spte));
3305 spin_lock(&mm->page_table_lock);
3307 pud_populate(mm, pud,
3308 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3310 put_page(virt_to_page(spte));
3311 spin_unlock(&mm->page_table_lock);
3313 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3314 mutex_unlock(&mapping->i_mmap_mutex);
3319 * unmap huge page backed by shared pte.
3321 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3322 * indicated by page_count > 1, unmap is achieved by clearing pud and
3323 * decrementing the ref count. If count == 1, the pte page is not shared.
3325 * called with vma->vm_mm->page_table_lock held.
3327 * returns: 1 successfully unmapped a shared pte page
3328 * 0 the underlying pte page is not shared, or it is the last user
3330 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3332 pgd_t *pgd = pgd_offset(mm, *addr);
3333 pud_t *pud = pud_offset(pgd, *addr);
3335 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3336 if (page_count(virt_to_page(ptep)) == 1)
3340 put_page(virt_to_page(ptep));
3341 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3344 #define want_pmd_share() (1)
3345 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3346 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3350 #define want_pmd_share() (0)
3351 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3353 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3354 pte_t *huge_pte_alloc(struct mm_struct *mm,
3355 unsigned long addr, unsigned long sz)
3361 pgd = pgd_offset(mm, addr);
3362 pud = pud_alloc(mm, pgd, addr);
3364 if (sz == PUD_SIZE) {
3367 BUG_ON(sz != PMD_SIZE);
3368 if (want_pmd_share() && pud_none(*pud))
3369 pte = huge_pmd_share(mm, addr, pud);
3371 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3374 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3379 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3385 pgd = pgd_offset(mm, addr);
3386 if (pgd_present(*pgd)) {
3387 pud = pud_offset(pgd, addr);
3388 if (pud_present(*pud)) {
3390 return (pte_t *)pud;
3391 pmd = pmd_offset(pud, addr);
3394 return (pte_t *) pmd;
3398 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3399 pmd_t *pmd, int write)
3403 page = pte_page(*(pte_t *)pmd);
3405 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3410 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3411 pud_t *pud, int write)
3415 page = pte_page(*(pte_t *)pud);
3417 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3421 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3423 /* Can be overriden by architectures */
3424 __attribute__((weak)) struct page *
3425 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3426 pud_t *pud, int write)
3432 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3434 #ifdef CONFIG_MEMORY_FAILURE
3436 /* Should be called in hugetlb_lock */
3437 static int is_hugepage_on_freelist(struct page *hpage)
3441 struct hstate *h = page_hstate(hpage);
3442 int nid = page_to_nid(hpage);
3444 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3451 * This function is called from memory failure code.
3452 * Assume the caller holds page lock of the head page.
3454 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3456 struct hstate *h = page_hstate(hpage);
3457 int nid = page_to_nid(hpage);
3460 spin_lock(&hugetlb_lock);
3461 if (is_hugepage_on_freelist(hpage)) {
3463 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3464 * but dangling hpage->lru can trigger list-debug warnings
3465 * (this happens when we call unpoison_memory() on it),
3466 * so let it point to itself with list_del_init().
3468 list_del_init(&hpage->lru);
3469 set_page_refcounted(hpage);
3470 h->free_huge_pages--;
3471 h->free_huge_pages_node[nid]--;
3474 spin_unlock(&hugetlb_lock);