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>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
36 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37 unsigned long hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 __initdata LIST_HEAD(huge_boot_pages);
45 /* for command line parsing */
46 static struct hstate * __initdata parsed_hstate;
47 static unsigned long __initdata default_hstate_max_huge_pages;
48 static unsigned long __initdata default_hstate_size;
51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 DEFINE_SPINLOCK(hugetlb_lock);
55 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59 spin_unlock(&spool->lock);
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
67 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 struct hugepage_subpool *spool;
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
75 spin_lock_init(&spool->lock);
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
83 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
88 unlock_or_release_subpool(spool);
91 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
105 spin_unlock(&spool->lock);
110 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
123 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 return HUGETLBFS_SB(inode->i_sb)->spool;
128 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 return subpool_inode(file_inode(vma->vm_file));
134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation_mutex:
142 * down_write(&mm->mmap_sem);
144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
148 struct list_head link;
153 static long region_add(struct list_head *head, long f, long t)
155 struct file_region *rg, *nrg, *trg;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg, head, link)
162 /* Round our left edge to the current segment if it encloses us. */
166 /* Check for and consume any regions we now overlap with. */
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
189 static long region_chg(struct list_head *head, long f, long t)
191 struct file_region *rg, *nrg;
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg, head, link)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg->link == head || t < rg->from) {
203 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
208 INIT_LIST_HEAD(&nrg->link);
209 list_add(&nrg->link, rg->link.prev);
214 /* Round our left edge to the current segment if it encloses us. */
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg, rg->link.prev, link) {
221 if (&rg->link == head)
226 /* We overlap with this area, if it extends further than
227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
233 chg -= rg->to - rg->from;
238 static long region_truncate(struct list_head *head, long end)
240 struct file_region *rg, *trg;
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg, head, link)
247 if (&rg->link == head)
250 /* If we are in the middle of a region then adjust it. */
251 if (end > rg->from) {
254 rg = list_entry(rg->link.next, typeof(*rg), link);
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
259 if (&rg->link == head)
261 chg += rg->to - rg->from;
268 static long region_count(struct list_head *head, long f, long t)
270 struct file_region *rg;
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg, head, link) {
283 seg_from = max(rg->from, f);
284 seg_to = min(rg->to, t);
286 chg += seg_to - seg_from;
293 * Convert the address within this vma to the page offset within
294 * the mapping, in pagecache page units; huge pages here.
296 static pgoff_t vma_hugecache_offset(struct hstate *h,
297 struct vm_area_struct *vma, unsigned long address)
299 return ((address - vma->vm_start) >> huge_page_shift(h)) +
300 (vma->vm_pgoff >> huge_page_order(h));
303 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
304 unsigned long address)
306 return vma_hugecache_offset(hstate_vma(vma), vma, address);
310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
313 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 struct hstate *hstate;
317 if (!is_vm_hugetlb_page(vma))
320 hstate = hstate_vma(vma);
322 return 1UL << huge_page_shift(hstate);
324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
332 #ifndef vma_mmu_pagesize
333 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 return vma_kernel_pagesize(vma);
340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
344 #define HPAGE_RESV_OWNER (1UL << 0)
345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
367 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 return (unsigned long)vma->vm_private_data;
372 static void set_vma_private_data(struct vm_area_struct *vma,
375 vma->vm_private_data = (void *)value;
380 struct list_head regions;
383 static struct resv_map *resv_map_alloc(void)
385 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
389 kref_init(&resv_map->refs);
390 INIT_LIST_HEAD(&resv_map->regions);
395 static void resv_map_release(struct kref *ref)
397 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map->regions, 0);
404 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma));
407 if (!(vma->vm_flags & VM_MAYSHARE))
408 return (struct resv_map *)(get_vma_private_data(vma) &
413 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma));
416 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418 set_vma_private_data(vma, (get_vma_private_data(vma) &
419 HPAGE_RESV_MASK) | (unsigned long)map);
422 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma));
425 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
430 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma));
434 return (get_vma_private_data(vma) & flag) != 0;
437 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
438 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440 VM_BUG_ON(!is_vm_hugetlb_page(vma));
441 if (!(vma->vm_flags & VM_MAYSHARE))
442 vma->vm_private_data = (void *)0;
445 /* Returns true if the VMA has associated reserve pages */
446 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
448 if (vma->vm_flags & VM_NORESERVE) {
450 * This address is already reserved by other process(chg == 0),
451 * so, we should decrement reserved count. Without decrementing,
452 * reserve count remains after releasing inode, because this
453 * allocated page will go into page cache and is regarded as
454 * coming from reserved pool in releasing step. Currently, we
455 * don't have any other solution to deal with this situation
456 * properly, so add work-around here.
458 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
464 /* Shared mappings always use reserves */
465 if (vma->vm_flags & VM_MAYSHARE)
469 * Only the process that called mmap() has reserves for
472 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
478 static void copy_gigantic_page(struct page *dst, struct page *src)
481 struct hstate *h = page_hstate(src);
482 struct page *dst_base = dst;
483 struct page *src_base = src;
485 for (i = 0; i < pages_per_huge_page(h); ) {
487 copy_highpage(dst, src);
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
495 void copy_huge_page(struct page *dst, struct page *src)
498 struct hstate *h = page_hstate(src);
500 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
501 copy_gigantic_page(dst, src);
506 for (i = 0; i < pages_per_huge_page(h); i++) {
508 copy_highpage(dst + i, src + i);
512 static void enqueue_huge_page(struct hstate *h, struct page *page)
514 int nid = page_to_nid(page);
515 list_move(&page->lru, &h->hugepage_freelists[nid]);
516 h->free_huge_pages++;
517 h->free_huge_pages_node[nid]++;
520 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524 if (list_empty(&h->hugepage_freelists[nid]))
526 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
527 list_move(&page->lru, &h->hugepage_activelist);
528 set_page_refcounted(page);
529 h->free_huge_pages--;
530 h->free_huge_pages_node[nid]--;
534 static struct page *dequeue_huge_page_vma(struct hstate *h,
535 struct vm_area_struct *vma,
536 unsigned long address, int avoid_reserve,
539 struct page *page = NULL;
540 struct mempolicy *mpol;
541 nodemask_t *nodemask;
542 struct zonelist *zonelist;
545 unsigned int cpuset_mems_cookie;
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma, chg) &&
553 h->free_huge_pages - h->resv_huge_pages == 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 cpuset_mems_cookie = get_mems_allowed();
562 zonelist = huge_zonelist(vma, address,
563 htlb_alloc_mask, &mpol, &nodemask);
565 for_each_zone_zonelist_nodemask(zone, z, zonelist,
566 MAX_NR_ZONES - 1, nodemask) {
567 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
568 page = dequeue_huge_page_node(h, zone_to_nid(zone));
572 if (!vma_has_reserves(vma, chg))
575 SetPagePrivate(page);
576 h->resv_huge_pages--;
583 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
591 static void update_and_free_page(struct hstate *h, struct page *page)
595 VM_BUG_ON(h->order >= MAX_ORDER);
598 h->nr_huge_pages_node[page_to_nid(page)]--;
599 for (i = 0; i < pages_per_huge_page(h); i++) {
600 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
601 1 << PG_referenced | 1 << PG_dirty |
602 1 << PG_active | 1 << PG_reserved |
603 1 << PG_private | 1 << PG_writeback);
605 VM_BUG_ON(hugetlb_cgroup_from_page(page));
606 set_compound_page_dtor(page, NULL);
607 set_page_refcounted(page);
608 arch_release_hugepage(page);
609 __free_pages(page, huge_page_order(h));
612 struct hstate *size_to_hstate(unsigned long size)
617 if (huge_page_size(h) == size)
623 static void free_huge_page(struct page *page)
626 * Can't pass hstate in here because it is called from the
627 * compound page destructor.
629 struct hstate *h = page_hstate(page);
630 int nid = page_to_nid(page);
631 struct hugepage_subpool *spool =
632 (struct hugepage_subpool *)page_private(page);
633 bool restore_reserve;
635 set_page_private(page, 0);
636 page->mapping = NULL;
637 BUG_ON(page_count(page));
638 BUG_ON(page_mapcount(page));
639 restore_reserve = PagePrivate(page);
641 spin_lock(&hugetlb_lock);
642 hugetlb_cgroup_uncharge_page(hstate_index(h),
643 pages_per_huge_page(h), page);
645 h->resv_huge_pages++;
647 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
648 /* remove the page from active list */
649 list_del(&page->lru);
650 update_and_free_page(h, page);
651 h->surplus_huge_pages--;
652 h->surplus_huge_pages_node[nid]--;
654 arch_clear_hugepage_flags(page);
655 enqueue_huge_page(h, page);
657 spin_unlock(&hugetlb_lock);
658 hugepage_subpool_put_pages(spool, 1);
661 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
663 INIT_LIST_HEAD(&page->lru);
664 set_compound_page_dtor(page, free_huge_page);
665 spin_lock(&hugetlb_lock);
666 set_hugetlb_cgroup(page, NULL);
668 h->nr_huge_pages_node[nid]++;
669 spin_unlock(&hugetlb_lock);
670 put_page(page); /* free it into the hugepage allocator */
673 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
676 int nr_pages = 1 << order;
677 struct page *p = page + 1;
679 /* we rely on prep_new_huge_page to set the destructor */
680 set_compound_order(page, order);
682 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
684 set_page_count(p, 0);
685 p->first_page = page;
690 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
691 * transparent huge pages. See the PageTransHuge() documentation for more
694 int PageHuge(struct page *page)
696 compound_page_dtor *dtor;
698 if (!PageCompound(page))
701 page = compound_head(page);
702 dtor = get_compound_page_dtor(page);
704 return dtor == free_huge_page;
706 EXPORT_SYMBOL_GPL(PageHuge);
708 pgoff_t __basepage_index(struct page *page)
710 struct page *page_head = compound_head(page);
711 pgoff_t index = page_index(page_head);
712 unsigned long compound_idx;
714 if (!PageHuge(page_head))
715 return page_index(page);
717 if (compound_order(page_head) >= MAX_ORDER)
718 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
720 compound_idx = page - page_head;
722 return (index << compound_order(page_head)) + compound_idx;
725 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
729 if (h->order >= MAX_ORDER)
732 page = alloc_pages_exact_node(nid,
733 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
734 __GFP_REPEAT|__GFP_NOWARN,
737 if (arch_prepare_hugepage(page)) {
738 __free_pages(page, huge_page_order(h));
741 prep_new_huge_page(h, page, nid);
748 * common helper functions for hstate_next_node_to_{alloc|free}.
749 * We may have allocated or freed a huge page based on a different
750 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
751 * be outside of *nodes_allowed. Ensure that we use an allowed
752 * node for alloc or free.
754 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
756 nid = next_node(nid, *nodes_allowed);
757 if (nid == MAX_NUMNODES)
758 nid = first_node(*nodes_allowed);
759 VM_BUG_ON(nid >= MAX_NUMNODES);
764 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
766 if (!node_isset(nid, *nodes_allowed))
767 nid = next_node_allowed(nid, nodes_allowed);
772 * returns the previously saved node ["this node"] from which to
773 * allocate a persistent huge page for the pool and advance the
774 * next node from which to allocate, handling wrap at end of node
777 static int hstate_next_node_to_alloc(struct hstate *h,
778 nodemask_t *nodes_allowed)
782 VM_BUG_ON(!nodes_allowed);
784 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
785 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
791 * helper for free_pool_huge_page() - return the previously saved
792 * node ["this node"] from which to free a huge page. Advance the
793 * next node id whether or not we find a free huge page to free so
794 * that the next attempt to free addresses the next node.
796 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
800 VM_BUG_ON(!nodes_allowed);
802 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
803 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
808 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
809 for (nr_nodes = nodes_weight(*mask); \
811 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
814 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
815 for (nr_nodes = nodes_weight(*mask); \
817 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
820 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
826 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
827 page = alloc_fresh_huge_page_node(h, node);
835 count_vm_event(HTLB_BUDDY_PGALLOC);
837 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
843 * Free huge page from pool from next node to free.
844 * Attempt to keep persistent huge pages more or less
845 * balanced over allowed nodes.
846 * Called with hugetlb_lock locked.
848 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
854 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
856 * If we're returning unused surplus pages, only examine
857 * nodes with surplus pages.
859 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
860 !list_empty(&h->hugepage_freelists[node])) {
862 list_entry(h->hugepage_freelists[node].next,
864 list_del(&page->lru);
865 h->free_huge_pages--;
866 h->free_huge_pages_node[node]--;
868 h->surplus_huge_pages--;
869 h->surplus_huge_pages_node[node]--;
871 update_and_free_page(h, page);
880 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
885 if (h->order >= MAX_ORDER)
889 * Assume we will successfully allocate the surplus page to
890 * prevent racing processes from causing the surplus to exceed
893 * This however introduces a different race, where a process B
894 * tries to grow the static hugepage pool while alloc_pages() is
895 * called by process A. B will only examine the per-node
896 * counters in determining if surplus huge pages can be
897 * converted to normal huge pages in adjust_pool_surplus(). A
898 * won't be able to increment the per-node counter, until the
899 * lock is dropped by B, but B doesn't drop hugetlb_lock until
900 * no more huge pages can be converted from surplus to normal
901 * state (and doesn't try to convert again). Thus, we have a
902 * case where a surplus huge page exists, the pool is grown, and
903 * the surplus huge page still exists after, even though it
904 * should just have been converted to a normal huge page. This
905 * does not leak memory, though, as the hugepage will be freed
906 * once it is out of use. It also does not allow the counters to
907 * go out of whack in adjust_pool_surplus() as we don't modify
908 * the node values until we've gotten the hugepage and only the
909 * per-node value is checked there.
911 spin_lock(&hugetlb_lock);
912 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
913 spin_unlock(&hugetlb_lock);
917 h->surplus_huge_pages++;
919 spin_unlock(&hugetlb_lock);
921 if (nid == NUMA_NO_NODE)
922 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
923 __GFP_REPEAT|__GFP_NOWARN,
926 page = alloc_pages_exact_node(nid,
927 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
928 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
930 if (page && arch_prepare_hugepage(page)) {
931 __free_pages(page, huge_page_order(h));
935 spin_lock(&hugetlb_lock);
937 INIT_LIST_HEAD(&page->lru);
938 r_nid = page_to_nid(page);
939 set_compound_page_dtor(page, free_huge_page);
940 set_hugetlb_cgroup(page, NULL);
942 * We incremented the global counters already
944 h->nr_huge_pages_node[r_nid]++;
945 h->surplus_huge_pages_node[r_nid]++;
946 __count_vm_event(HTLB_BUDDY_PGALLOC);
949 h->surplus_huge_pages--;
950 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
952 spin_unlock(&hugetlb_lock);
958 * This allocation function is useful in the context where vma is irrelevant.
959 * E.g. soft-offlining uses this function because it only cares physical
960 * address of error page.
962 struct page *alloc_huge_page_node(struct hstate *h, int nid)
964 struct page *page = NULL;
966 spin_lock(&hugetlb_lock);
967 if (h->free_huge_pages - h->resv_huge_pages > 0)
968 page = dequeue_huge_page_node(h, nid);
969 spin_unlock(&hugetlb_lock);
972 page = alloc_buddy_huge_page(h, nid);
978 * Increase the hugetlb pool such that it can accommodate a reservation
981 static int gather_surplus_pages(struct hstate *h, int delta)
983 struct list_head surplus_list;
984 struct page *page, *tmp;
986 int needed, allocated;
987 bool alloc_ok = true;
989 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
991 h->resv_huge_pages += delta;
996 INIT_LIST_HEAD(&surplus_list);
1000 spin_unlock(&hugetlb_lock);
1001 for (i = 0; i < needed; i++) {
1002 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1007 list_add(&page->lru, &surplus_list);
1012 * After retaking hugetlb_lock, we need to recalculate 'needed'
1013 * because either resv_huge_pages or free_huge_pages may have changed.
1015 spin_lock(&hugetlb_lock);
1016 needed = (h->resv_huge_pages + delta) -
1017 (h->free_huge_pages + allocated);
1022 * We were not able to allocate enough pages to
1023 * satisfy the entire reservation so we free what
1024 * we've allocated so far.
1029 * The surplus_list now contains _at_least_ the number of extra pages
1030 * needed to accommodate the reservation. Add the appropriate number
1031 * of pages to the hugetlb pool and free the extras back to the buddy
1032 * allocator. Commit the entire reservation here to prevent another
1033 * process from stealing the pages as they are added to the pool but
1034 * before they are reserved.
1036 needed += allocated;
1037 h->resv_huge_pages += delta;
1040 /* Free the needed pages to the hugetlb pool */
1041 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1045 * This page is now managed by the hugetlb allocator and has
1046 * no users -- drop the buddy allocator's reference.
1048 put_page_testzero(page);
1049 VM_BUG_ON(page_count(page));
1050 enqueue_huge_page(h, page);
1053 spin_unlock(&hugetlb_lock);
1055 /* Free unnecessary surplus pages to the buddy allocator */
1056 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1058 spin_lock(&hugetlb_lock);
1064 * When releasing a hugetlb pool reservation, any surplus pages that were
1065 * allocated to satisfy the reservation must be explicitly freed if they were
1067 * Called with hugetlb_lock held.
1069 static void return_unused_surplus_pages(struct hstate *h,
1070 unsigned long unused_resv_pages)
1072 unsigned long nr_pages;
1074 /* Uncommit the reservation */
1075 h->resv_huge_pages -= unused_resv_pages;
1077 /* Cannot return gigantic pages currently */
1078 if (h->order >= MAX_ORDER)
1081 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1084 * We want to release as many surplus pages as possible, spread
1085 * evenly across all nodes with memory. Iterate across these nodes
1086 * until we can no longer free unreserved surplus pages. This occurs
1087 * when the nodes with surplus pages have no free pages.
1088 * free_pool_huge_page() will balance the the freed pages across the
1089 * on-line nodes with memory and will handle the hstate accounting.
1091 while (nr_pages--) {
1092 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1098 * Determine if the huge page at addr within the vma has an associated
1099 * reservation. Where it does not we will need to logically increase
1100 * reservation and actually increase subpool usage before an allocation
1101 * can occur. Where any new reservation would be required the
1102 * reservation change is prepared, but not committed. Once the page
1103 * has been allocated from the subpool and instantiated the change should
1104 * be committed via vma_commit_reservation. No action is required on
1107 static long vma_needs_reservation(struct hstate *h,
1108 struct vm_area_struct *vma, unsigned long addr)
1110 struct address_space *mapping = vma->vm_file->f_mapping;
1111 struct inode *inode = mapping->host;
1113 if (vma->vm_flags & VM_MAYSHARE) {
1114 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1115 return region_chg(&inode->i_mapping->private_list,
1118 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1123 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1124 struct resv_map *resv = vma_resv_map(vma);
1126 err = region_chg(&resv->regions, idx, idx + 1);
1132 static void vma_commit_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 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1142 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1143 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1144 struct resv_map *resv = vma_resv_map(vma);
1146 /* Mark this page used in the map. */
1147 region_add(&resv->regions, idx, idx + 1);
1151 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1152 unsigned long addr, int avoid_reserve)
1154 struct hugepage_subpool *spool = subpool_vma(vma);
1155 struct hstate *h = hstate_vma(vma);
1159 struct hugetlb_cgroup *h_cg;
1161 idx = hstate_index(h);
1163 * Processes that did not create the mapping will have no
1164 * reserves and will not have accounted against subpool
1165 * limit. Check that the subpool limit can be made before
1166 * satisfying the allocation MAP_NORESERVE mappings may also
1167 * need pages and subpool limit allocated allocated if no reserve
1170 chg = vma_needs_reservation(h, vma, addr);
1172 return ERR_PTR(-ENOMEM);
1173 if (chg || avoid_reserve)
1174 if (hugepage_subpool_get_pages(spool, 1))
1175 return ERR_PTR(-ENOSPC);
1177 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1179 if (chg || avoid_reserve)
1180 hugepage_subpool_put_pages(spool, 1);
1181 return ERR_PTR(-ENOSPC);
1183 spin_lock(&hugetlb_lock);
1184 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1186 spin_unlock(&hugetlb_lock);
1187 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1189 hugetlb_cgroup_uncharge_cgroup(idx,
1190 pages_per_huge_page(h),
1192 if (chg || avoid_reserve)
1193 hugepage_subpool_put_pages(spool, 1);
1194 return ERR_PTR(-ENOSPC);
1196 spin_lock(&hugetlb_lock);
1197 list_move(&page->lru, &h->hugepage_activelist);
1200 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1201 spin_unlock(&hugetlb_lock);
1203 set_page_private(page, (unsigned long)spool);
1205 vma_commit_reservation(h, vma, addr);
1209 int __weak alloc_bootmem_huge_page(struct hstate *h)
1211 struct huge_bootmem_page *m;
1214 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1217 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1218 huge_page_size(h), huge_page_size(h), 0);
1222 * Use the beginning of the huge page to store the
1223 * huge_bootmem_page struct (until gather_bootmem
1224 * puts them into the mem_map).
1233 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1234 /* Put them into a private list first because mem_map is not up yet */
1235 list_add(&m->list, &huge_boot_pages);
1240 static void prep_compound_huge_page(struct page *page, int order)
1242 if (unlikely(order > (MAX_ORDER - 1)))
1243 prep_compound_gigantic_page(page, order);
1245 prep_compound_page(page, order);
1248 /* Put bootmem huge pages into the standard lists after mem_map is up */
1249 static void __init gather_bootmem_prealloc(void)
1251 struct huge_bootmem_page *m;
1253 list_for_each_entry(m, &huge_boot_pages, list) {
1254 struct hstate *h = m->hstate;
1257 #ifdef CONFIG_HIGHMEM
1258 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1259 free_bootmem_late((unsigned long)m,
1260 sizeof(struct huge_bootmem_page));
1262 page = virt_to_page(m);
1264 __ClearPageReserved(page);
1265 WARN_ON(page_count(page) != 1);
1266 prep_compound_huge_page(page, h->order);
1267 prep_new_huge_page(h, page, page_to_nid(page));
1269 * If we had gigantic hugepages allocated at boot time, we need
1270 * to restore the 'stolen' pages to totalram_pages in order to
1271 * fix confusing memory reports from free(1) and another
1272 * side-effects, like CommitLimit going negative.
1274 if (h->order > (MAX_ORDER - 1))
1275 adjust_managed_page_count(page, 1 << h->order);
1279 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1283 for (i = 0; i < h->max_huge_pages; ++i) {
1284 if (h->order >= MAX_ORDER) {
1285 if (!alloc_bootmem_huge_page(h))
1287 } else if (!alloc_fresh_huge_page(h,
1288 &node_states[N_MEMORY]))
1291 h->max_huge_pages = i;
1294 static void __init hugetlb_init_hstates(void)
1298 for_each_hstate(h) {
1299 /* oversize hugepages were init'ed in early boot */
1300 if (h->order < MAX_ORDER)
1301 hugetlb_hstate_alloc_pages(h);
1305 static char * __init memfmt(char *buf, unsigned long n)
1307 if (n >= (1UL << 30))
1308 sprintf(buf, "%lu GB", n >> 30);
1309 else if (n >= (1UL << 20))
1310 sprintf(buf, "%lu MB", n >> 20);
1312 sprintf(buf, "%lu KB", n >> 10);
1316 static void __init report_hugepages(void)
1320 for_each_hstate(h) {
1322 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1323 memfmt(buf, huge_page_size(h)),
1324 h->free_huge_pages);
1328 #ifdef CONFIG_HIGHMEM
1329 static void try_to_free_low(struct hstate *h, unsigned long count,
1330 nodemask_t *nodes_allowed)
1334 if (h->order >= MAX_ORDER)
1337 for_each_node_mask(i, *nodes_allowed) {
1338 struct page *page, *next;
1339 struct list_head *freel = &h->hugepage_freelists[i];
1340 list_for_each_entry_safe(page, next, freel, lru) {
1341 if (count >= h->nr_huge_pages)
1343 if (PageHighMem(page))
1345 list_del(&page->lru);
1346 update_and_free_page(h, page);
1347 h->free_huge_pages--;
1348 h->free_huge_pages_node[page_to_nid(page)]--;
1353 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1354 nodemask_t *nodes_allowed)
1360 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1361 * balanced by operating on them in a round-robin fashion.
1362 * Returns 1 if an adjustment was made.
1364 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1369 VM_BUG_ON(delta != -1 && delta != 1);
1372 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1373 if (h->surplus_huge_pages_node[node])
1377 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1378 if (h->surplus_huge_pages_node[node] <
1379 h->nr_huge_pages_node[node])
1386 h->surplus_huge_pages += delta;
1387 h->surplus_huge_pages_node[node] += delta;
1391 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1392 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1393 nodemask_t *nodes_allowed)
1395 unsigned long min_count, ret;
1397 if (h->order >= MAX_ORDER)
1398 return h->max_huge_pages;
1401 * Increase the pool size
1402 * First take pages out of surplus state. Then make up the
1403 * remaining difference by allocating fresh huge pages.
1405 * We might race with alloc_buddy_huge_page() here and be unable
1406 * to convert a surplus huge page to a normal huge page. That is
1407 * not critical, though, it just means the overall size of the
1408 * pool might be one hugepage larger than it needs to be, but
1409 * within all the constraints specified by the sysctls.
1411 spin_lock(&hugetlb_lock);
1412 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1413 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1417 while (count > persistent_huge_pages(h)) {
1419 * If this allocation races such that we no longer need the
1420 * page, free_huge_page will handle it by freeing the page
1421 * and reducing the surplus.
1423 spin_unlock(&hugetlb_lock);
1424 ret = alloc_fresh_huge_page(h, nodes_allowed);
1425 spin_lock(&hugetlb_lock);
1429 /* Bail for signals. Probably ctrl-c from user */
1430 if (signal_pending(current))
1435 * Decrease the pool size
1436 * First return free pages to the buddy allocator (being careful
1437 * to keep enough around to satisfy reservations). Then place
1438 * pages into surplus state as needed so the pool will shrink
1439 * to the desired size as pages become free.
1441 * By placing pages into the surplus state independent of the
1442 * overcommit value, we are allowing the surplus pool size to
1443 * exceed overcommit. There are few sane options here. Since
1444 * alloc_buddy_huge_page() is checking the global counter,
1445 * though, we'll note that we're not allowed to exceed surplus
1446 * and won't grow the pool anywhere else. Not until one of the
1447 * sysctls are changed, or the surplus pages go out of use.
1449 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1450 min_count = max(count, min_count);
1451 try_to_free_low(h, min_count, nodes_allowed);
1452 while (min_count < persistent_huge_pages(h)) {
1453 if (!free_pool_huge_page(h, nodes_allowed, 0))
1456 while (count < persistent_huge_pages(h)) {
1457 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1461 ret = persistent_huge_pages(h);
1462 spin_unlock(&hugetlb_lock);
1466 #define HSTATE_ATTR_RO(_name) \
1467 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1469 #define HSTATE_ATTR(_name) \
1470 static struct kobj_attribute _name##_attr = \
1471 __ATTR(_name, 0644, _name##_show, _name##_store)
1473 static struct kobject *hugepages_kobj;
1474 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1476 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1478 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1482 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1483 if (hstate_kobjs[i] == kobj) {
1485 *nidp = NUMA_NO_NODE;
1489 return kobj_to_node_hstate(kobj, nidp);
1492 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1493 struct kobj_attribute *attr, char *buf)
1496 unsigned long nr_huge_pages;
1499 h = kobj_to_hstate(kobj, &nid);
1500 if (nid == NUMA_NO_NODE)
1501 nr_huge_pages = h->nr_huge_pages;
1503 nr_huge_pages = h->nr_huge_pages_node[nid];
1505 return sprintf(buf, "%lu\n", nr_huge_pages);
1508 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1509 struct kobject *kobj, struct kobj_attribute *attr,
1510 const char *buf, size_t len)
1514 unsigned long count;
1516 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1518 err = kstrtoul(buf, 10, &count);
1522 h = kobj_to_hstate(kobj, &nid);
1523 if (h->order >= MAX_ORDER) {
1528 if (nid == NUMA_NO_NODE) {
1530 * global hstate attribute
1532 if (!(obey_mempolicy &&
1533 init_nodemask_of_mempolicy(nodes_allowed))) {
1534 NODEMASK_FREE(nodes_allowed);
1535 nodes_allowed = &node_states[N_MEMORY];
1537 } else if (nodes_allowed) {
1539 * per node hstate attribute: adjust count to global,
1540 * but restrict alloc/free to the specified node.
1542 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1543 init_nodemask_of_node(nodes_allowed, nid);
1545 nodes_allowed = &node_states[N_MEMORY];
1547 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1549 if (nodes_allowed != &node_states[N_MEMORY])
1550 NODEMASK_FREE(nodes_allowed);
1554 NODEMASK_FREE(nodes_allowed);
1558 static ssize_t nr_hugepages_show(struct kobject *kobj,
1559 struct kobj_attribute *attr, char *buf)
1561 return nr_hugepages_show_common(kobj, attr, buf);
1564 static ssize_t nr_hugepages_store(struct kobject *kobj,
1565 struct kobj_attribute *attr, const char *buf, size_t len)
1567 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1569 HSTATE_ATTR(nr_hugepages);
1574 * hstate attribute for optionally mempolicy-based constraint on persistent
1575 * huge page alloc/free.
1577 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1578 struct kobj_attribute *attr, char *buf)
1580 return nr_hugepages_show_common(kobj, attr, buf);
1583 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1584 struct kobj_attribute *attr, const char *buf, size_t len)
1586 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1588 HSTATE_ATTR(nr_hugepages_mempolicy);
1592 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1593 struct kobj_attribute *attr, char *buf)
1595 struct hstate *h = kobj_to_hstate(kobj, NULL);
1596 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1599 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1600 struct kobj_attribute *attr, const char *buf, size_t count)
1603 unsigned long input;
1604 struct hstate *h = kobj_to_hstate(kobj, NULL);
1606 if (h->order >= MAX_ORDER)
1609 err = kstrtoul(buf, 10, &input);
1613 spin_lock(&hugetlb_lock);
1614 h->nr_overcommit_huge_pages = input;
1615 spin_unlock(&hugetlb_lock);
1619 HSTATE_ATTR(nr_overcommit_hugepages);
1621 static ssize_t free_hugepages_show(struct kobject *kobj,
1622 struct kobj_attribute *attr, char *buf)
1625 unsigned long free_huge_pages;
1628 h = kobj_to_hstate(kobj, &nid);
1629 if (nid == NUMA_NO_NODE)
1630 free_huge_pages = h->free_huge_pages;
1632 free_huge_pages = h->free_huge_pages_node[nid];
1634 return sprintf(buf, "%lu\n", free_huge_pages);
1636 HSTATE_ATTR_RO(free_hugepages);
1638 static ssize_t resv_hugepages_show(struct kobject *kobj,
1639 struct kobj_attribute *attr, char *buf)
1641 struct hstate *h = kobj_to_hstate(kobj, NULL);
1642 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1644 HSTATE_ATTR_RO(resv_hugepages);
1646 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1647 struct kobj_attribute *attr, char *buf)
1650 unsigned long surplus_huge_pages;
1653 h = kobj_to_hstate(kobj, &nid);
1654 if (nid == NUMA_NO_NODE)
1655 surplus_huge_pages = h->surplus_huge_pages;
1657 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1659 return sprintf(buf, "%lu\n", surplus_huge_pages);
1661 HSTATE_ATTR_RO(surplus_hugepages);
1663 static struct attribute *hstate_attrs[] = {
1664 &nr_hugepages_attr.attr,
1665 &nr_overcommit_hugepages_attr.attr,
1666 &free_hugepages_attr.attr,
1667 &resv_hugepages_attr.attr,
1668 &surplus_hugepages_attr.attr,
1670 &nr_hugepages_mempolicy_attr.attr,
1675 static struct attribute_group hstate_attr_group = {
1676 .attrs = hstate_attrs,
1679 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1680 struct kobject **hstate_kobjs,
1681 struct attribute_group *hstate_attr_group)
1684 int hi = hstate_index(h);
1686 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1687 if (!hstate_kobjs[hi])
1690 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1692 kobject_put(hstate_kobjs[hi]);
1697 static void __init hugetlb_sysfs_init(void)
1702 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1703 if (!hugepages_kobj)
1706 for_each_hstate(h) {
1707 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1708 hstate_kobjs, &hstate_attr_group);
1710 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1717 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1718 * with node devices in node_devices[] using a parallel array. The array
1719 * index of a node device or _hstate == node id.
1720 * This is here to avoid any static dependency of the node device driver, in
1721 * the base kernel, on the hugetlb module.
1723 struct node_hstate {
1724 struct kobject *hugepages_kobj;
1725 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1727 struct node_hstate node_hstates[MAX_NUMNODES];
1730 * A subset of global hstate attributes for node devices
1732 static struct attribute *per_node_hstate_attrs[] = {
1733 &nr_hugepages_attr.attr,
1734 &free_hugepages_attr.attr,
1735 &surplus_hugepages_attr.attr,
1739 static struct attribute_group per_node_hstate_attr_group = {
1740 .attrs = per_node_hstate_attrs,
1744 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1745 * Returns node id via non-NULL nidp.
1747 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1751 for (nid = 0; nid < nr_node_ids; nid++) {
1752 struct node_hstate *nhs = &node_hstates[nid];
1754 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1755 if (nhs->hstate_kobjs[i] == kobj) {
1767 * Unregister hstate attributes from a single node device.
1768 * No-op if no hstate attributes attached.
1770 static void hugetlb_unregister_node(struct node *node)
1773 struct node_hstate *nhs = &node_hstates[node->dev.id];
1775 if (!nhs->hugepages_kobj)
1776 return; /* no hstate attributes */
1778 for_each_hstate(h) {
1779 int idx = hstate_index(h);
1780 if (nhs->hstate_kobjs[idx]) {
1781 kobject_put(nhs->hstate_kobjs[idx]);
1782 nhs->hstate_kobjs[idx] = NULL;
1786 kobject_put(nhs->hugepages_kobj);
1787 nhs->hugepages_kobj = NULL;
1791 * hugetlb module exit: unregister hstate attributes from node devices
1794 static void hugetlb_unregister_all_nodes(void)
1799 * disable node device registrations.
1801 register_hugetlbfs_with_node(NULL, NULL);
1804 * remove hstate attributes from any nodes that have them.
1806 for (nid = 0; nid < nr_node_ids; nid++)
1807 hugetlb_unregister_node(node_devices[nid]);
1811 * Register hstate attributes for a single node device.
1812 * No-op if attributes already registered.
1814 static void hugetlb_register_node(struct node *node)
1817 struct node_hstate *nhs = &node_hstates[node->dev.id];
1820 if (nhs->hugepages_kobj)
1821 return; /* already allocated */
1823 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1825 if (!nhs->hugepages_kobj)
1828 for_each_hstate(h) {
1829 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1831 &per_node_hstate_attr_group);
1833 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1834 h->name, node->dev.id);
1835 hugetlb_unregister_node(node);
1842 * hugetlb init time: register hstate attributes for all registered node
1843 * devices of nodes that have memory. All on-line nodes should have
1844 * registered their associated device by this time.
1846 static void hugetlb_register_all_nodes(void)
1850 for_each_node_state(nid, N_MEMORY) {
1851 struct node *node = node_devices[nid];
1852 if (node->dev.id == nid)
1853 hugetlb_register_node(node);
1857 * Let the node device driver know we're here so it can
1858 * [un]register hstate attributes on node hotplug.
1860 register_hugetlbfs_with_node(hugetlb_register_node,
1861 hugetlb_unregister_node);
1863 #else /* !CONFIG_NUMA */
1865 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1873 static void hugetlb_unregister_all_nodes(void) { }
1875 static void hugetlb_register_all_nodes(void) { }
1879 static void __exit hugetlb_exit(void)
1883 hugetlb_unregister_all_nodes();
1885 for_each_hstate(h) {
1886 kobject_put(hstate_kobjs[hstate_index(h)]);
1889 kobject_put(hugepages_kobj);
1891 module_exit(hugetlb_exit);
1893 static int __init hugetlb_init(void)
1895 /* Some platform decide whether they support huge pages at boot
1896 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1897 * there is no such support
1899 if (HPAGE_SHIFT == 0)
1902 if (!size_to_hstate(default_hstate_size)) {
1903 default_hstate_size = HPAGE_SIZE;
1904 if (!size_to_hstate(default_hstate_size))
1905 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1907 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1908 if (default_hstate_max_huge_pages)
1909 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1911 hugetlb_init_hstates();
1912 gather_bootmem_prealloc();
1915 hugetlb_sysfs_init();
1916 hugetlb_register_all_nodes();
1917 hugetlb_cgroup_file_init();
1921 module_init(hugetlb_init);
1923 /* Should be called on processing a hugepagesz=... option */
1924 void __init hugetlb_add_hstate(unsigned order)
1929 if (size_to_hstate(PAGE_SIZE << order)) {
1930 pr_warning("hugepagesz= specified twice, ignoring\n");
1933 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1935 h = &hstates[hugetlb_max_hstate++];
1937 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1938 h->nr_huge_pages = 0;
1939 h->free_huge_pages = 0;
1940 for (i = 0; i < MAX_NUMNODES; ++i)
1941 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1942 INIT_LIST_HEAD(&h->hugepage_activelist);
1943 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1944 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1945 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1946 huge_page_size(h)/1024);
1951 static int __init hugetlb_nrpages_setup(char *s)
1954 static unsigned long *last_mhp;
1957 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1958 * so this hugepages= parameter goes to the "default hstate".
1960 if (!hugetlb_max_hstate)
1961 mhp = &default_hstate_max_huge_pages;
1963 mhp = &parsed_hstate->max_huge_pages;
1965 if (mhp == last_mhp) {
1966 pr_warning("hugepages= specified twice without "
1967 "interleaving hugepagesz=, ignoring\n");
1971 if (sscanf(s, "%lu", mhp) <= 0)
1975 * Global state is always initialized later in hugetlb_init.
1976 * But we need to allocate >= MAX_ORDER hstates here early to still
1977 * use the bootmem allocator.
1979 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1980 hugetlb_hstate_alloc_pages(parsed_hstate);
1986 __setup("hugepages=", hugetlb_nrpages_setup);
1988 static int __init hugetlb_default_setup(char *s)
1990 default_hstate_size = memparse(s, &s);
1993 __setup("default_hugepagesz=", hugetlb_default_setup);
1995 static unsigned int cpuset_mems_nr(unsigned int *array)
1998 unsigned int nr = 0;
2000 for_each_node_mask(node, cpuset_current_mems_allowed)
2006 #ifdef CONFIG_SYSCTL
2007 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2008 struct ctl_table *table, int write,
2009 void __user *buffer, size_t *length, loff_t *ppos)
2011 struct hstate *h = &default_hstate;
2015 tmp = h->max_huge_pages;
2017 if (write && h->order >= MAX_ORDER)
2021 table->maxlen = sizeof(unsigned long);
2022 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2027 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2028 GFP_KERNEL | __GFP_NORETRY);
2029 if (!(obey_mempolicy &&
2030 init_nodemask_of_mempolicy(nodes_allowed))) {
2031 NODEMASK_FREE(nodes_allowed);
2032 nodes_allowed = &node_states[N_MEMORY];
2034 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2036 if (nodes_allowed != &node_states[N_MEMORY])
2037 NODEMASK_FREE(nodes_allowed);
2043 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2044 void __user *buffer, size_t *length, loff_t *ppos)
2047 return hugetlb_sysctl_handler_common(false, table, write,
2048 buffer, length, ppos);
2052 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2053 void __user *buffer, size_t *length, loff_t *ppos)
2055 return hugetlb_sysctl_handler_common(true, table, write,
2056 buffer, length, ppos);
2058 #endif /* CONFIG_NUMA */
2060 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2061 void __user *buffer,
2062 size_t *length, loff_t *ppos)
2064 proc_dointvec(table, write, buffer, length, ppos);
2065 if (hugepages_treat_as_movable)
2066 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2068 htlb_alloc_mask = GFP_HIGHUSER;
2072 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2073 void __user *buffer,
2074 size_t *length, loff_t *ppos)
2076 struct hstate *h = &default_hstate;
2080 tmp = h->nr_overcommit_huge_pages;
2082 if (write && h->order >= MAX_ORDER)
2086 table->maxlen = sizeof(unsigned long);
2087 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2092 spin_lock(&hugetlb_lock);
2093 h->nr_overcommit_huge_pages = tmp;
2094 spin_unlock(&hugetlb_lock);
2100 #endif /* CONFIG_SYSCTL */
2102 void hugetlb_report_meminfo(struct seq_file *m)
2104 struct hstate *h = &default_hstate;
2106 "HugePages_Total: %5lu\n"
2107 "HugePages_Free: %5lu\n"
2108 "HugePages_Rsvd: %5lu\n"
2109 "HugePages_Surp: %5lu\n"
2110 "Hugepagesize: %8lu kB\n",
2114 h->surplus_huge_pages,
2115 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2118 int hugetlb_report_node_meminfo(int nid, char *buf)
2120 struct hstate *h = &default_hstate;
2122 "Node %d HugePages_Total: %5u\n"
2123 "Node %d HugePages_Free: %5u\n"
2124 "Node %d HugePages_Surp: %5u\n",
2125 nid, h->nr_huge_pages_node[nid],
2126 nid, h->free_huge_pages_node[nid],
2127 nid, h->surplus_huge_pages_node[nid]);
2130 void hugetlb_show_meminfo(void)
2135 for_each_node_state(nid, N_MEMORY)
2137 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2139 h->nr_huge_pages_node[nid],
2140 h->free_huge_pages_node[nid],
2141 h->surplus_huge_pages_node[nid],
2142 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2145 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2146 unsigned long hugetlb_total_pages(void)
2149 unsigned long nr_total_pages = 0;
2152 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2153 return nr_total_pages;
2156 static int hugetlb_acct_memory(struct hstate *h, long delta)
2160 spin_lock(&hugetlb_lock);
2162 * When cpuset is configured, it breaks the strict hugetlb page
2163 * reservation as the accounting is done on a global variable. Such
2164 * reservation is completely rubbish in the presence of cpuset because
2165 * the reservation is not checked against page availability for the
2166 * current cpuset. Application can still potentially OOM'ed by kernel
2167 * with lack of free htlb page in cpuset that the task is in.
2168 * Attempt to enforce strict accounting with cpuset is almost
2169 * impossible (or too ugly) because cpuset is too fluid that
2170 * task or memory node can be dynamically moved between cpusets.
2172 * The change of semantics for shared hugetlb mapping with cpuset is
2173 * undesirable. However, in order to preserve some of the semantics,
2174 * we fall back to check against current free page availability as
2175 * a best attempt and hopefully to minimize the impact of changing
2176 * semantics that cpuset has.
2179 if (gather_surplus_pages(h, delta) < 0)
2182 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2183 return_unused_surplus_pages(h, delta);
2190 return_unused_surplus_pages(h, (unsigned long) -delta);
2193 spin_unlock(&hugetlb_lock);
2197 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2199 struct resv_map *resv = vma_resv_map(vma);
2202 * This new VMA should share its siblings reservation map if present.
2203 * The VMA will only ever have a valid reservation map pointer where
2204 * it is being copied for another still existing VMA. As that VMA
2205 * has a reference to the reservation map it cannot disappear until
2206 * after this open call completes. It is therefore safe to take a
2207 * new reference here without additional locking.
2210 kref_get(&resv->refs);
2213 static void resv_map_put(struct vm_area_struct *vma)
2215 struct resv_map *resv = vma_resv_map(vma);
2219 kref_put(&resv->refs, resv_map_release);
2222 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2224 struct hstate *h = hstate_vma(vma);
2225 struct resv_map *resv = vma_resv_map(vma);
2226 struct hugepage_subpool *spool = subpool_vma(vma);
2227 unsigned long reserve;
2228 unsigned long start;
2232 start = vma_hugecache_offset(h, vma, vma->vm_start);
2233 end = vma_hugecache_offset(h, vma, vma->vm_end);
2235 reserve = (end - start) -
2236 region_count(&resv->regions, start, end);
2241 hugetlb_acct_memory(h, -reserve);
2242 hugepage_subpool_put_pages(spool, reserve);
2248 * We cannot handle pagefaults against hugetlb pages at all. They cause
2249 * handle_mm_fault() to try to instantiate regular-sized pages in the
2250 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2253 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2259 const struct vm_operations_struct hugetlb_vm_ops = {
2260 .fault = hugetlb_vm_op_fault,
2261 .open = hugetlb_vm_op_open,
2262 .close = hugetlb_vm_op_close,
2265 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2271 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2272 vma->vm_page_prot)));
2274 entry = huge_pte_wrprotect(mk_huge_pte(page,
2275 vma->vm_page_prot));
2277 entry = pte_mkyoung(entry);
2278 entry = pte_mkhuge(entry);
2279 entry = arch_make_huge_pte(entry, vma, page, writable);
2284 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2285 unsigned long address, pte_t *ptep)
2289 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2290 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2291 update_mmu_cache(vma, address, ptep);
2295 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2296 struct vm_area_struct *vma)
2298 pte_t *src_pte, *dst_pte, entry;
2299 struct page *ptepage;
2302 struct hstate *h = hstate_vma(vma);
2303 unsigned long sz = huge_page_size(h);
2305 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2307 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2308 src_pte = huge_pte_offset(src, addr);
2311 dst_pte = huge_pte_alloc(dst, addr, sz);
2315 /* If the pagetables are shared don't copy or take references */
2316 if (dst_pte == src_pte)
2319 spin_lock(&dst->page_table_lock);
2320 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2321 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2323 huge_ptep_set_wrprotect(src, addr, src_pte);
2324 entry = huge_ptep_get(src_pte);
2325 ptepage = pte_page(entry);
2327 page_dup_rmap(ptepage);
2328 set_huge_pte_at(dst, addr, dst_pte, entry);
2330 spin_unlock(&src->page_table_lock);
2331 spin_unlock(&dst->page_table_lock);
2339 static int is_hugetlb_entry_migration(pte_t pte)
2343 if (huge_pte_none(pte) || pte_present(pte))
2345 swp = pte_to_swp_entry(pte);
2346 if (non_swap_entry(swp) && is_migration_entry(swp))
2352 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2356 if (huge_pte_none(pte) || pte_present(pte))
2358 swp = pte_to_swp_entry(pte);
2359 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2365 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2366 unsigned long start, unsigned long end,
2367 struct page *ref_page)
2369 int force_flush = 0;
2370 struct mm_struct *mm = vma->vm_mm;
2371 unsigned long address;
2375 struct hstate *h = hstate_vma(vma);
2376 unsigned long sz = huge_page_size(h);
2377 const unsigned long mmun_start = start; /* For mmu_notifiers */
2378 const unsigned long mmun_end = end; /* For mmu_notifiers */
2380 WARN_ON(!is_vm_hugetlb_page(vma));
2381 BUG_ON(start & ~huge_page_mask(h));
2382 BUG_ON(end & ~huge_page_mask(h));
2384 tlb_start_vma(tlb, vma);
2385 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2387 spin_lock(&mm->page_table_lock);
2388 for (address = start; address < end; address += sz) {
2389 ptep = huge_pte_offset(mm, address);
2393 if (huge_pmd_unshare(mm, &address, ptep))
2396 pte = huge_ptep_get(ptep);
2397 if (huge_pte_none(pte))
2401 * HWPoisoned hugepage is already unmapped and dropped reference
2403 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2404 huge_pte_clear(mm, address, ptep);
2408 page = pte_page(pte);
2410 * If a reference page is supplied, it is because a specific
2411 * page is being unmapped, not a range. Ensure the page we
2412 * are about to unmap is the actual page of interest.
2415 if (page != ref_page)
2419 * Mark the VMA as having unmapped its page so that
2420 * future faults in this VMA will fail rather than
2421 * looking like data was lost
2423 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2426 pte = huge_ptep_get_and_clear(mm, address, ptep);
2427 tlb_remove_tlb_entry(tlb, ptep, address);
2428 if (huge_pte_dirty(pte))
2429 set_page_dirty(page);
2431 page_remove_rmap(page);
2432 force_flush = !__tlb_remove_page(tlb, page);
2435 /* Bail out after unmapping reference page if supplied */
2439 spin_unlock(&mm->page_table_lock);
2441 * mmu_gather ran out of room to batch pages, we break out of
2442 * the PTE lock to avoid doing the potential expensive TLB invalidate
2443 * and page-free while holding it.
2448 if (address < end && !ref_page)
2451 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2452 tlb_end_vma(tlb, vma);
2455 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2456 struct vm_area_struct *vma, unsigned long start,
2457 unsigned long end, struct page *ref_page)
2459 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2462 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2463 * test will fail on a vma being torn down, and not grab a page table
2464 * on its way out. We're lucky that the flag has such an appropriate
2465 * name, and can in fact be safely cleared here. We could clear it
2466 * before the __unmap_hugepage_range above, but all that's necessary
2467 * is to clear it before releasing the i_mmap_mutex. This works
2468 * because in the context this is called, the VMA is about to be
2469 * destroyed and the i_mmap_mutex is held.
2471 vma->vm_flags &= ~VM_MAYSHARE;
2474 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2475 unsigned long end, struct page *ref_page)
2477 struct mm_struct *mm;
2478 struct mmu_gather tlb;
2482 tlb_gather_mmu(&tlb, mm, start, end);
2483 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2484 tlb_finish_mmu(&tlb, start, end);
2488 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2489 * mappping it owns the reserve page for. The intention is to unmap the page
2490 * from other VMAs and let the children be SIGKILLed if they are faulting the
2493 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2494 struct page *page, unsigned long address)
2496 struct hstate *h = hstate_vma(vma);
2497 struct vm_area_struct *iter_vma;
2498 struct address_space *mapping;
2502 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2503 * from page cache lookup which is in HPAGE_SIZE units.
2505 address = address & huge_page_mask(h);
2506 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2508 mapping = file_inode(vma->vm_file)->i_mapping;
2511 * Take the mapping lock for the duration of the table walk. As
2512 * this mapping should be shared between all the VMAs,
2513 * __unmap_hugepage_range() is called as the lock is already held
2515 mutex_lock(&mapping->i_mmap_mutex);
2516 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2517 /* Do not unmap the current VMA */
2518 if (iter_vma == vma)
2522 * Unmap the page from other VMAs without their own reserves.
2523 * They get marked to be SIGKILLed if they fault in these
2524 * areas. This is because a future no-page fault on this VMA
2525 * could insert a zeroed page instead of the data existing
2526 * from the time of fork. This would look like data corruption
2528 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2529 unmap_hugepage_range(iter_vma, address,
2530 address + huge_page_size(h), page);
2532 mutex_unlock(&mapping->i_mmap_mutex);
2538 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2539 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2540 * cannot race with other handlers or page migration.
2541 * Keep the pte_same checks anyway to make transition from the mutex easier.
2543 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2544 unsigned long address, pte_t *ptep, pte_t pte,
2545 struct page *pagecache_page)
2547 struct hstate *h = hstate_vma(vma);
2548 struct page *old_page, *new_page;
2549 int outside_reserve = 0;
2550 unsigned long mmun_start; /* For mmu_notifiers */
2551 unsigned long mmun_end; /* For mmu_notifiers */
2553 old_page = pte_page(pte);
2556 /* If no-one else is actually using this page, avoid the copy
2557 * and just make the page writable */
2558 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2559 page_move_anon_rmap(old_page, vma, address);
2560 set_huge_ptep_writable(vma, address, ptep);
2565 * If the process that created a MAP_PRIVATE mapping is about to
2566 * perform a COW due to a shared page count, attempt to satisfy
2567 * the allocation without using the existing reserves. The pagecache
2568 * page is used to determine if the reserve at this address was
2569 * consumed or not. If reserves were used, a partial faulted mapping
2570 * at the time of fork() could consume its reserves on COW instead
2571 * of the full address range.
2573 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2574 old_page != pagecache_page)
2575 outside_reserve = 1;
2577 page_cache_get(old_page);
2579 /* Drop page_table_lock as buddy allocator may be called */
2580 spin_unlock(&mm->page_table_lock);
2581 new_page = alloc_huge_page(vma, address, outside_reserve);
2583 if (IS_ERR(new_page)) {
2584 long err = PTR_ERR(new_page);
2585 page_cache_release(old_page);
2588 * If a process owning a MAP_PRIVATE mapping fails to COW,
2589 * it is due to references held by a child and an insufficient
2590 * huge page pool. To guarantee the original mappers
2591 * reliability, unmap the page from child processes. The child
2592 * may get SIGKILLed if it later faults.
2594 if (outside_reserve) {
2595 BUG_ON(huge_pte_none(pte));
2596 if (unmap_ref_private(mm, vma, old_page, address)) {
2597 BUG_ON(huge_pte_none(pte));
2598 spin_lock(&mm->page_table_lock);
2599 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2600 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2601 goto retry_avoidcopy;
2603 * race occurs while re-acquiring page_table_lock, and
2611 /* Caller expects lock to be held */
2612 spin_lock(&mm->page_table_lock);
2614 return VM_FAULT_OOM;
2616 return VM_FAULT_SIGBUS;
2620 * When the original hugepage is shared one, it does not have
2621 * anon_vma prepared.
2623 if (unlikely(anon_vma_prepare(vma))) {
2624 page_cache_release(new_page);
2625 page_cache_release(old_page);
2626 /* Caller expects lock to be held */
2627 spin_lock(&mm->page_table_lock);
2628 return VM_FAULT_OOM;
2631 copy_user_huge_page(new_page, old_page, address, vma,
2632 pages_per_huge_page(h));
2633 __SetPageUptodate(new_page);
2635 mmun_start = address & huge_page_mask(h);
2636 mmun_end = mmun_start + huge_page_size(h);
2637 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2639 * Retake the page_table_lock to check for racing updates
2640 * before the page tables are altered
2642 spin_lock(&mm->page_table_lock);
2643 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2644 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2645 ClearPagePrivate(new_page);
2648 huge_ptep_clear_flush(vma, address, ptep);
2649 set_huge_pte_at(mm, address, ptep,
2650 make_huge_pte(vma, new_page, 1));
2651 page_remove_rmap(old_page);
2652 hugepage_add_new_anon_rmap(new_page, vma, address);
2653 /* Make the old page be freed below */
2654 new_page = old_page;
2656 spin_unlock(&mm->page_table_lock);
2657 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2658 page_cache_release(new_page);
2659 page_cache_release(old_page);
2661 /* Caller expects lock to be held */
2662 spin_lock(&mm->page_table_lock);
2666 /* Return the pagecache page at a given address within a VMA */
2667 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2668 struct vm_area_struct *vma, unsigned long address)
2670 struct address_space *mapping;
2673 mapping = vma->vm_file->f_mapping;
2674 idx = vma_hugecache_offset(h, vma, address);
2676 return find_lock_page(mapping, idx);
2680 * Return whether there is a pagecache page to back given address within VMA.
2681 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2683 static bool hugetlbfs_pagecache_present(struct hstate *h,
2684 struct vm_area_struct *vma, unsigned long address)
2686 struct address_space *mapping;
2690 mapping = vma->vm_file->f_mapping;
2691 idx = vma_hugecache_offset(h, vma, address);
2693 page = find_get_page(mapping, idx);
2696 return page != NULL;
2699 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2700 unsigned long address, pte_t *ptep, unsigned int flags)
2702 struct hstate *h = hstate_vma(vma);
2703 int ret = VM_FAULT_SIGBUS;
2708 struct address_space *mapping;
2712 * Currently, we are forced to kill the process in the event the
2713 * original mapper has unmapped pages from the child due to a failed
2714 * COW. Warn that such a situation has occurred as it may not be obvious
2716 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2717 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2722 mapping = vma->vm_file->f_mapping;
2723 idx = vma_hugecache_offset(h, vma, address);
2726 * Use page lock to guard against racing truncation
2727 * before we get page_table_lock.
2730 page = find_lock_page(mapping, idx);
2732 size = i_size_read(mapping->host) >> huge_page_shift(h);
2735 page = alloc_huge_page(vma, address, 0);
2737 ret = PTR_ERR(page);
2741 ret = VM_FAULT_SIGBUS;
2744 clear_huge_page(page, address, pages_per_huge_page(h));
2745 __SetPageUptodate(page);
2747 if (vma->vm_flags & VM_MAYSHARE) {
2749 struct inode *inode = mapping->host;
2751 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2758 ClearPagePrivate(page);
2760 spin_lock(&inode->i_lock);
2761 inode->i_blocks += blocks_per_huge_page(h);
2762 spin_unlock(&inode->i_lock);
2765 if (unlikely(anon_vma_prepare(vma))) {
2767 goto backout_unlocked;
2773 * If memory error occurs between mmap() and fault, some process
2774 * don't have hwpoisoned swap entry for errored virtual address.
2775 * So we need to block hugepage fault by PG_hwpoison bit check.
2777 if (unlikely(PageHWPoison(page))) {
2778 ret = VM_FAULT_HWPOISON |
2779 VM_FAULT_SET_HINDEX(hstate_index(h));
2780 goto backout_unlocked;
2785 * If we are going to COW a private mapping later, we examine the
2786 * pending reservations for this page now. This will ensure that
2787 * any allocations necessary to record that reservation occur outside
2790 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2791 if (vma_needs_reservation(h, vma, address) < 0) {
2793 goto backout_unlocked;
2796 spin_lock(&mm->page_table_lock);
2797 size = i_size_read(mapping->host) >> huge_page_shift(h);
2802 if (!huge_pte_none(huge_ptep_get(ptep)))
2806 ClearPagePrivate(page);
2807 hugepage_add_new_anon_rmap(page, vma, address);
2810 page_dup_rmap(page);
2811 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2812 && (vma->vm_flags & VM_SHARED)));
2813 set_huge_pte_at(mm, address, ptep, new_pte);
2815 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2816 /* Optimization, do the COW without a second fault */
2817 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2820 spin_unlock(&mm->page_table_lock);
2826 spin_unlock(&mm->page_table_lock);
2833 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2834 unsigned long address, unsigned int flags)
2839 struct page *page = NULL;
2840 struct page *pagecache_page = NULL;
2841 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2842 struct hstate *h = hstate_vma(vma);
2844 address &= huge_page_mask(h);
2846 ptep = huge_pte_offset(mm, address);
2848 entry = huge_ptep_get(ptep);
2849 if (unlikely(is_hugetlb_entry_migration(entry))) {
2850 migration_entry_wait_huge(mm, ptep);
2852 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2853 return VM_FAULT_HWPOISON_LARGE |
2854 VM_FAULT_SET_HINDEX(hstate_index(h));
2857 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2859 return VM_FAULT_OOM;
2862 * Serialize hugepage allocation and instantiation, so that we don't
2863 * get spurious allocation failures if two CPUs race to instantiate
2864 * the same page in the page cache.
2866 mutex_lock(&hugetlb_instantiation_mutex);
2867 entry = huge_ptep_get(ptep);
2868 if (huge_pte_none(entry)) {
2869 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2876 * If we are going to COW the mapping later, we examine the pending
2877 * reservations for this page now. This will ensure that any
2878 * allocations necessary to record that reservation occur outside the
2879 * spinlock. For private mappings, we also lookup the pagecache
2880 * page now as it is used to determine if a reservation has been
2883 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2884 if (vma_needs_reservation(h, vma, address) < 0) {
2889 if (!(vma->vm_flags & VM_MAYSHARE))
2890 pagecache_page = hugetlbfs_pagecache_page(h,
2895 * hugetlb_cow() requires page locks of pte_page(entry) and
2896 * pagecache_page, so here we need take the former one
2897 * when page != pagecache_page or !pagecache_page.
2898 * Note that locking order is always pagecache_page -> page,
2899 * so no worry about deadlock.
2901 page = pte_page(entry);
2903 if (page != pagecache_page)
2906 spin_lock(&mm->page_table_lock);
2907 /* Check for a racing update before calling hugetlb_cow */
2908 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2909 goto out_page_table_lock;
2912 if (flags & FAULT_FLAG_WRITE) {
2913 if (!huge_pte_write(entry)) {
2914 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2916 goto out_page_table_lock;
2918 entry = huge_pte_mkdirty(entry);
2920 entry = pte_mkyoung(entry);
2921 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2922 flags & FAULT_FLAG_WRITE))
2923 update_mmu_cache(vma, address, ptep);
2925 out_page_table_lock:
2926 spin_unlock(&mm->page_table_lock);
2928 if (pagecache_page) {
2929 unlock_page(pagecache_page);
2930 put_page(pagecache_page);
2932 if (page != pagecache_page)
2937 mutex_unlock(&hugetlb_instantiation_mutex);
2942 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2943 struct page **pages, struct vm_area_struct **vmas,
2944 unsigned long *position, unsigned long *nr_pages,
2945 long i, unsigned int flags)
2947 unsigned long pfn_offset;
2948 unsigned long vaddr = *position;
2949 unsigned long remainder = *nr_pages;
2950 struct hstate *h = hstate_vma(vma);
2952 spin_lock(&mm->page_table_lock);
2953 while (vaddr < vma->vm_end && remainder) {
2959 * Some archs (sparc64, sh*) have multiple pte_ts to
2960 * each hugepage. We have to make sure we get the
2961 * first, for the page indexing below to work.
2963 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2964 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2967 * When coredumping, it suits get_dump_page if we just return
2968 * an error where there's an empty slot with no huge pagecache
2969 * to back it. This way, we avoid allocating a hugepage, and
2970 * the sparse dumpfile avoids allocating disk blocks, but its
2971 * huge holes still show up with zeroes where they need to be.
2973 if (absent && (flags & FOLL_DUMP) &&
2974 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2980 * We need call hugetlb_fault for both hugepages under migration
2981 * (in which case hugetlb_fault waits for the migration,) and
2982 * hwpoisoned hugepages (in which case we need to prevent the
2983 * caller from accessing to them.) In order to do this, we use
2984 * here is_swap_pte instead of is_hugetlb_entry_migration and
2985 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2986 * both cases, and because we can't follow correct pages
2987 * directly from any kind of swap entries.
2989 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2990 ((flags & FOLL_WRITE) &&
2991 !huge_pte_write(huge_ptep_get(pte)))) {
2994 spin_unlock(&mm->page_table_lock);
2995 ret = hugetlb_fault(mm, vma, vaddr,
2996 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2997 spin_lock(&mm->page_table_lock);
2998 if (!(ret & VM_FAULT_ERROR))
3005 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3006 page = pte_page(huge_ptep_get(pte));
3009 pages[i] = mem_map_offset(page, pfn_offset);
3020 if (vaddr < vma->vm_end && remainder &&
3021 pfn_offset < pages_per_huge_page(h)) {
3023 * We use pfn_offset to avoid touching the pageframes
3024 * of this compound page.
3029 spin_unlock(&mm->page_table_lock);
3030 *nr_pages = remainder;
3033 return i ? i : -EFAULT;
3036 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3037 unsigned long address, unsigned long end, pgprot_t newprot)
3039 struct mm_struct *mm = vma->vm_mm;
3040 unsigned long start = address;
3043 struct hstate *h = hstate_vma(vma);
3044 unsigned long pages = 0;
3046 BUG_ON(address >= end);
3047 flush_cache_range(vma, address, end);
3049 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3050 spin_lock(&mm->page_table_lock);
3051 for (; address < end; address += huge_page_size(h)) {
3052 ptep = huge_pte_offset(mm, address);
3055 if (huge_pmd_unshare(mm, &address, ptep)) {
3059 if (!huge_pte_none(huge_ptep_get(ptep))) {
3060 pte = huge_ptep_get_and_clear(mm, address, ptep);
3061 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3062 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3063 set_huge_pte_at(mm, address, ptep, pte);
3067 spin_unlock(&mm->page_table_lock);
3069 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3070 * may have cleared our pud entry and done put_page on the page table:
3071 * once we release i_mmap_mutex, another task can do the final put_page
3072 * and that page table be reused and filled with junk.
3074 flush_tlb_range(vma, start, end);
3075 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3077 return pages << h->order;
3080 int hugetlb_reserve_pages(struct inode *inode,
3082 struct vm_area_struct *vma,
3083 vm_flags_t vm_flags)
3086 struct hstate *h = hstate_inode(inode);
3087 struct hugepage_subpool *spool = subpool_inode(inode);
3090 * Only apply hugepage reservation if asked. At fault time, an
3091 * attempt will be made for VM_NORESERVE to allocate a page
3092 * without using reserves
3094 if (vm_flags & VM_NORESERVE)
3098 * Shared mappings base their reservation on the number of pages that
3099 * are already allocated on behalf of the file. Private mappings need
3100 * to reserve the full area even if read-only as mprotect() may be
3101 * called to make the mapping read-write. Assume !vma is a shm mapping
3103 if (!vma || vma->vm_flags & VM_MAYSHARE)
3104 chg = region_chg(&inode->i_mapping->private_list, from, to);
3106 struct resv_map *resv_map = resv_map_alloc();
3112 set_vma_resv_map(vma, resv_map);
3113 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3121 /* There must be enough pages in the subpool for the mapping */
3122 if (hugepage_subpool_get_pages(spool, chg)) {
3128 * Check enough hugepages are available for the reservation.
3129 * Hand the pages back to the subpool if there are not
3131 ret = hugetlb_acct_memory(h, chg);
3133 hugepage_subpool_put_pages(spool, chg);
3138 * Account for the reservations made. Shared mappings record regions
3139 * that have reservations as they are shared by multiple VMAs.
3140 * When the last VMA disappears, the region map says how much
3141 * the reservation was and the page cache tells how much of
3142 * the reservation was consumed. Private mappings are per-VMA and
3143 * only the consumed reservations are tracked. When the VMA
3144 * disappears, the original reservation is the VMA size and the
3145 * consumed reservations are stored in the map. Hence, nothing
3146 * else has to be done for private mappings here
3148 if (!vma || vma->vm_flags & VM_MAYSHARE)
3149 region_add(&inode->i_mapping->private_list, from, to);
3157 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3159 struct hstate *h = hstate_inode(inode);
3160 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3161 struct hugepage_subpool *spool = subpool_inode(inode);
3163 spin_lock(&inode->i_lock);
3164 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3165 spin_unlock(&inode->i_lock);
3167 hugepage_subpool_put_pages(spool, (chg - freed));
3168 hugetlb_acct_memory(h, -(chg - freed));
3171 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3172 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3173 struct vm_area_struct *vma,
3174 unsigned long addr, pgoff_t idx)
3176 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3178 unsigned long sbase = saddr & PUD_MASK;
3179 unsigned long s_end = sbase + PUD_SIZE;
3181 /* Allow segments to share if only one is marked locked */
3182 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3183 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3186 * match the virtual addresses, permission and the alignment of the
3189 if (pmd_index(addr) != pmd_index(saddr) ||
3190 vm_flags != svm_flags ||
3191 sbase < svma->vm_start || svma->vm_end < s_end)
3197 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3199 unsigned long base = addr & PUD_MASK;
3200 unsigned long end = base + PUD_SIZE;
3203 * check on proper vm_flags and page table alignment
3205 if (vma->vm_flags & VM_MAYSHARE &&
3206 vma->vm_start <= base && end <= vma->vm_end)
3212 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3213 * and returns the corresponding pte. While this is not necessary for the
3214 * !shared pmd case because we can allocate the pmd later as well, it makes the
3215 * code much cleaner. pmd allocation is essential for the shared case because
3216 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3217 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3218 * bad pmd for sharing.
3220 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3222 struct vm_area_struct *vma = find_vma(mm, addr);
3223 struct address_space *mapping = vma->vm_file->f_mapping;
3224 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3226 struct vm_area_struct *svma;
3227 unsigned long saddr;
3231 if (!vma_shareable(vma, addr))
3232 return (pte_t *)pmd_alloc(mm, pud, addr);
3234 mutex_lock(&mapping->i_mmap_mutex);
3235 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3239 saddr = page_table_shareable(svma, vma, addr, idx);
3241 spte = huge_pte_offset(svma->vm_mm, saddr);
3243 get_page(virt_to_page(spte));
3252 spin_lock(&mm->page_table_lock);
3254 pud_populate(mm, pud,
3255 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3257 put_page(virt_to_page(spte));
3258 spin_unlock(&mm->page_table_lock);
3260 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3261 mutex_unlock(&mapping->i_mmap_mutex);
3266 * unmap huge page backed by shared pte.
3268 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3269 * indicated by page_count > 1, unmap is achieved by clearing pud and
3270 * decrementing the ref count. If count == 1, the pte page is not shared.
3272 * called with vma->vm_mm->page_table_lock held.
3274 * returns: 1 successfully unmapped a shared pte page
3275 * 0 the underlying pte page is not shared, or it is the last user
3277 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3279 pgd_t *pgd = pgd_offset(mm, *addr);
3280 pud_t *pud = pud_offset(pgd, *addr);
3282 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3283 if (page_count(virt_to_page(ptep)) == 1)
3287 put_page(virt_to_page(ptep));
3288 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3291 #define want_pmd_share() (1)
3292 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3293 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3297 #define want_pmd_share() (0)
3298 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3300 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3301 pte_t *huge_pte_alloc(struct mm_struct *mm,
3302 unsigned long addr, unsigned long sz)
3308 pgd = pgd_offset(mm, addr);
3309 pud = pud_alloc(mm, pgd, addr);
3311 if (sz == PUD_SIZE) {
3314 BUG_ON(sz != PMD_SIZE);
3315 if (want_pmd_share() && pud_none(*pud))
3316 pte = huge_pmd_share(mm, addr, pud);
3318 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3321 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3326 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3332 pgd = pgd_offset(mm, addr);
3333 if (pgd_present(*pgd)) {
3334 pud = pud_offset(pgd, addr);
3335 if (pud_present(*pud)) {
3337 return (pte_t *)pud;
3338 pmd = pmd_offset(pud, addr);
3341 return (pte_t *) pmd;
3345 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3346 pmd_t *pmd, int write)
3350 page = pte_page(*(pte_t *)pmd);
3352 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3357 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3358 pud_t *pud, int write)
3362 page = pte_page(*(pte_t *)pud);
3364 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3368 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3370 /* Can be overriden by architectures */
3371 __attribute__((weak)) struct page *
3372 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3373 pud_t *pud, int write)
3379 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3381 #ifdef CONFIG_MEMORY_FAILURE
3383 /* Should be called in hugetlb_lock */
3384 static int is_hugepage_on_freelist(struct page *hpage)
3388 struct hstate *h = page_hstate(hpage);
3389 int nid = page_to_nid(hpage);
3391 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3398 * This function is called from memory failure code.
3399 * Assume the caller holds page lock of the head page.
3401 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3403 struct hstate *h = page_hstate(hpage);
3404 int nid = page_to_nid(hpage);
3407 spin_lock(&hugetlb_lock);
3408 if (is_hugepage_on_freelist(hpage)) {
3410 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3411 * but dangling hpage->lru can trigger list-debug warnings
3412 * (this happens when we call unpoison_memory() on it),
3413 * so let it point to itself with list_del_init().
3415 list_del_init(&hpage->lru);
3416 set_page_refcounted(hpage);
3417 h->free_huge_pages--;
3418 h->free_huge_pages_node[nid]--;
3421 spin_unlock(&hugetlb_lock);