2 * Generic hugetlb support.
3 * (C) William Irwin, 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/node.h>
34 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
35 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
36 unsigned long hugepages_treat_as_movable;
38 int hugetlb_max_hstate __read_mostly;
39 unsigned int default_hstate_idx;
40 struct hstate hstates[HUGE_MAX_HSTATE];
42 __initdata LIST_HEAD(huge_boot_pages);
44 /* for command line parsing */
45 static struct hstate * __initdata parsed_hstate;
46 static unsigned long __initdata default_hstate_max_huge_pages;
47 static unsigned long __initdata default_hstate_size;
50 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
52 DEFINE_SPINLOCK(hugetlb_lock);
54 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
56 bool free = (spool->count == 0) && (spool->used_hpages == 0);
58 spin_unlock(&spool->lock);
60 /* If no pages are used, and no other handles to the subpool
61 * remain, free the subpool the subpool remain */
66 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
68 struct hugepage_subpool *spool;
70 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
74 spin_lock_init(&spool->lock);
76 spool->max_hpages = nr_blocks;
77 spool->used_hpages = 0;
82 void hugepage_put_subpool(struct hugepage_subpool *spool)
84 spin_lock(&spool->lock);
85 BUG_ON(!spool->count);
87 unlock_or_release_subpool(spool);
90 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
98 spin_lock(&spool->lock);
99 if ((spool->used_hpages + delta) <= spool->max_hpages) {
100 spool->used_hpages += delta;
104 spin_unlock(&spool->lock);
109 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
115 spin_lock(&spool->lock);
116 spool->used_hpages -= delta;
117 /* If hugetlbfs_put_super couldn't free spool due to
118 * an outstanding quota reference, free it now. */
119 unlock_or_release_subpool(spool);
122 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
124 return HUGETLBFS_SB(inode->i_sb)->spool;
127 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
129 return subpool_inode(vma->vm_file->f_dentry->d_inode);
133 * Region tracking -- allows tracking of reservations and instantiated pages
134 * across the pages in a mapping.
136 * The region data structures are protected by a combination of the mmap_sem
137 * and the hugetlb_instantion_mutex. To access or modify a region the caller
138 * must either hold the mmap_sem for write, or the mmap_sem for read and
139 * the hugetlb_instantiation mutex:
141 * down_write(&mm->mmap_sem);
143 * down_read(&mm->mmap_sem);
144 * mutex_lock(&hugetlb_instantiation_mutex);
147 struct list_head link;
152 static long region_add(struct list_head *head, long f, long t)
154 struct file_region *rg, *nrg, *trg;
156 /* Locate the region we are either in or before. */
157 list_for_each_entry(rg, head, link)
161 /* Round our left edge to the current segment if it encloses us. */
165 /* Check for and consume any regions we now overlap with. */
167 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
168 if (&rg->link == head)
173 /* If this area reaches higher then extend our area to
174 * include it completely. If this is not the first area
175 * which we intend to reuse, free it. */
188 static long region_chg(struct list_head *head, long f, long t)
190 struct file_region *rg, *nrg;
193 /* Locate the region we are before or in. */
194 list_for_each_entry(rg, head, link)
198 /* If we are below the current region then a new region is required.
199 * Subtle, allocate a new region at the position but make it zero
200 * size such that we can guarantee to record the reservation. */
201 if (&rg->link == head || t < rg->from) {
202 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
207 INIT_LIST_HEAD(&nrg->link);
208 list_add(&nrg->link, rg->link.prev);
213 /* Round our left edge to the current segment if it encloses us. */
218 /* Check for and consume any regions we now overlap with. */
219 list_for_each_entry(rg, rg->link.prev, link) {
220 if (&rg->link == head)
225 /* We overlap with this area, if it extends further than
226 * us then we must extend ourselves. Account for its
227 * existing reservation. */
232 chg -= rg->to - rg->from;
237 static long region_truncate(struct list_head *head, long end)
239 struct file_region *rg, *trg;
242 /* Locate the region we are either in or before. */
243 list_for_each_entry(rg, head, link)
246 if (&rg->link == head)
249 /* If we are in the middle of a region then adjust it. */
250 if (end > rg->from) {
253 rg = list_entry(rg->link.next, typeof(*rg), link);
256 /* Drop any remaining regions. */
257 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
258 if (&rg->link == head)
260 chg += rg->to - rg->from;
267 static long region_count(struct list_head *head, long f, long t)
269 struct file_region *rg;
272 /* Locate each segment we overlap with, and count that overlap. */
273 list_for_each_entry(rg, head, link) {
282 seg_from = max(rg->from, f);
283 seg_to = min(rg->to, t);
285 chg += seg_to - seg_from;
292 * Convert the address within this vma to the page offset within
293 * the mapping, in pagecache page units; huge pages here.
295 static pgoff_t vma_hugecache_offset(struct hstate *h,
296 struct vm_area_struct *vma, unsigned long address)
298 return ((address - vma->vm_start) >> huge_page_shift(h)) +
299 (vma->vm_pgoff >> huge_page_order(h));
302 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
303 unsigned long address)
305 return vma_hugecache_offset(hstate_vma(vma), vma, address);
309 * Return the size of the pages allocated when backing a VMA. In the majority
310 * cases this will be same size as used by the page table entries.
312 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
314 struct hstate *hstate;
316 if (!is_vm_hugetlb_page(vma))
319 hstate = hstate_vma(vma);
321 return 1UL << (hstate->order + PAGE_SHIFT);
323 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
326 * Return the page size being used by the MMU to back a VMA. In the majority
327 * of cases, the page size used by the kernel matches the MMU size. On
328 * architectures where it differs, an architecture-specific version of this
329 * function is required.
331 #ifndef vma_mmu_pagesize
332 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
334 return vma_kernel_pagesize(vma);
339 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
340 * bits of the reservation map pointer, which are always clear due to
343 #define HPAGE_RESV_OWNER (1UL << 0)
344 #define HPAGE_RESV_UNMAPPED (1UL << 1)
345 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
348 * These helpers are used to track how many pages are reserved for
349 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
350 * is guaranteed to have their future faults succeed.
352 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
353 * the reserve counters are updated with the hugetlb_lock held. It is safe
354 * to reset the VMA at fork() time as it is not in use yet and there is no
355 * chance of the global counters getting corrupted as a result of the values.
357 * The private mapping reservation is represented in a subtly different
358 * manner to a shared mapping. A shared mapping has a region map associated
359 * with the underlying file, this region map represents the backing file
360 * pages which have ever had a reservation assigned which this persists even
361 * after the page is instantiated. A private mapping has a region map
362 * associated with the original mmap which is attached to all VMAs which
363 * reference it, this region map represents those offsets which have consumed
364 * reservation ie. where pages have been instantiated.
366 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
368 return (unsigned long)vma->vm_private_data;
371 static void set_vma_private_data(struct vm_area_struct *vma,
374 vma->vm_private_data = (void *)value;
379 struct list_head regions;
382 static struct resv_map *resv_map_alloc(void)
384 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
388 kref_init(&resv_map->refs);
389 INIT_LIST_HEAD(&resv_map->regions);
394 static void resv_map_release(struct kref *ref)
396 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
398 /* Clear out any active regions before we release the map. */
399 region_truncate(&resv_map->regions, 0);
403 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
405 VM_BUG_ON(!is_vm_hugetlb_page(vma));
406 if (!(vma->vm_flags & VM_MAYSHARE))
407 return (struct resv_map *)(get_vma_private_data(vma) &
412 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
414 VM_BUG_ON(!is_vm_hugetlb_page(vma));
415 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
417 set_vma_private_data(vma, (get_vma_private_data(vma) &
418 HPAGE_RESV_MASK) | (unsigned long)map);
421 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
423 VM_BUG_ON(!is_vm_hugetlb_page(vma));
424 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
426 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
429 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
431 VM_BUG_ON(!is_vm_hugetlb_page(vma));
433 return (get_vma_private_data(vma) & flag) != 0;
436 /* Decrement the reserved pages in the hugepage pool by one */
437 static void decrement_hugepage_resv_vma(struct hstate *h,
438 struct vm_area_struct *vma)
440 if (vma->vm_flags & VM_NORESERVE)
443 if (vma->vm_flags & VM_MAYSHARE) {
444 /* Shared mappings always use reserves */
445 h->resv_huge_pages--;
446 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
448 * Only the process that called mmap() has reserves for
451 h->resv_huge_pages--;
455 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
456 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
458 VM_BUG_ON(!is_vm_hugetlb_page(vma));
459 if (!(vma->vm_flags & VM_MAYSHARE))
460 vma->vm_private_data = (void *)0;
463 /* Returns true if the VMA has associated reserve pages */
464 static int vma_has_reserves(struct vm_area_struct *vma)
466 if (vma->vm_flags & VM_MAYSHARE)
468 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
473 static void copy_gigantic_page(struct page *dst, struct page *src)
476 struct hstate *h = page_hstate(src);
477 struct page *dst_base = dst;
478 struct page *src_base = src;
480 for (i = 0; i < pages_per_huge_page(h); ) {
482 copy_highpage(dst, src);
485 dst = mem_map_next(dst, dst_base, i);
486 src = mem_map_next(src, src_base, i);
490 void copy_huge_page(struct page *dst, struct page *src)
493 struct hstate *h = page_hstate(src);
495 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
496 copy_gigantic_page(dst, src);
501 for (i = 0; i < pages_per_huge_page(h); i++) {
503 copy_highpage(dst + i, src + i);
507 static void enqueue_huge_page(struct hstate *h, struct page *page)
509 int nid = page_to_nid(page);
510 list_move(&page->lru, &h->hugepage_freelists[nid]);
511 h->free_huge_pages++;
512 h->free_huge_pages_node[nid]++;
515 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
519 if (list_empty(&h->hugepage_freelists[nid]))
521 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
522 list_move(&page->lru, &h->hugepage_activelist);
523 set_page_refcounted(page);
524 h->free_huge_pages--;
525 h->free_huge_pages_node[nid]--;
529 static struct page *dequeue_huge_page_vma(struct hstate *h,
530 struct vm_area_struct *vma,
531 unsigned long address, int avoid_reserve)
533 struct page *page = NULL;
534 struct mempolicy *mpol;
535 nodemask_t *nodemask;
536 struct zonelist *zonelist;
539 unsigned int cpuset_mems_cookie;
542 cpuset_mems_cookie = get_mems_allowed();
543 zonelist = huge_zonelist(vma, address,
544 htlb_alloc_mask, &mpol, &nodemask);
546 * A child process with MAP_PRIVATE mappings created by their parent
547 * have no page reserves. This check ensures that reservations are
548 * not "stolen". The child may still get SIGKILLed
550 if (!vma_has_reserves(vma) &&
551 h->free_huge_pages - h->resv_huge_pages == 0)
554 /* If reserves cannot be used, ensure enough pages are in the pool */
555 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
558 for_each_zone_zonelist_nodemask(zone, z, zonelist,
559 MAX_NR_ZONES - 1, nodemask) {
560 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
561 page = dequeue_huge_page_node(h, zone_to_nid(zone));
564 decrement_hugepage_resv_vma(h, vma);
571 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
580 static void update_and_free_page(struct hstate *h, struct page *page)
584 VM_BUG_ON(h->order >= MAX_ORDER);
587 h->nr_huge_pages_node[page_to_nid(page)]--;
588 for (i = 0; i < pages_per_huge_page(h); i++) {
589 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
590 1 << PG_referenced | 1 << PG_dirty |
591 1 << PG_active | 1 << PG_reserved |
592 1 << PG_private | 1 << PG_writeback);
594 set_compound_page_dtor(page, NULL);
595 set_page_refcounted(page);
596 arch_release_hugepage(page);
597 __free_pages(page, huge_page_order(h));
600 struct hstate *size_to_hstate(unsigned long size)
605 if (huge_page_size(h) == size)
611 static void free_huge_page(struct page *page)
614 * Can't pass hstate in here because it is called from the
615 * compound page destructor.
617 struct hstate *h = page_hstate(page);
618 int nid = page_to_nid(page);
619 struct hugepage_subpool *spool =
620 (struct hugepage_subpool *)page_private(page);
622 set_page_private(page, 0);
623 page->mapping = NULL;
624 BUG_ON(page_count(page));
625 BUG_ON(page_mapcount(page));
627 spin_lock(&hugetlb_lock);
628 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
629 /* remove the page from active list */
630 list_del(&page->lru);
631 update_and_free_page(h, page);
632 h->surplus_huge_pages--;
633 h->surplus_huge_pages_node[nid]--;
635 enqueue_huge_page(h, page);
637 spin_unlock(&hugetlb_lock);
638 hugepage_subpool_put_pages(spool, 1);
641 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
643 INIT_LIST_HEAD(&page->lru);
644 set_compound_page_dtor(page, free_huge_page);
645 spin_lock(&hugetlb_lock);
647 h->nr_huge_pages_node[nid]++;
648 spin_unlock(&hugetlb_lock);
649 put_page(page); /* free it into the hugepage allocator */
652 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
655 int nr_pages = 1 << order;
656 struct page *p = page + 1;
658 /* we rely on prep_new_huge_page to set the destructor */
659 set_compound_order(page, order);
661 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
663 set_page_count(p, 0);
664 p->first_page = page;
668 int PageHuge(struct page *page)
670 compound_page_dtor *dtor;
672 if (!PageCompound(page))
675 page = compound_head(page);
676 dtor = get_compound_page_dtor(page);
678 return dtor == free_huge_page;
680 EXPORT_SYMBOL_GPL(PageHuge);
682 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
686 if (h->order >= MAX_ORDER)
689 page = alloc_pages_exact_node(nid,
690 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
691 __GFP_REPEAT|__GFP_NOWARN,
694 if (arch_prepare_hugepage(page)) {
695 __free_pages(page, huge_page_order(h));
698 prep_new_huge_page(h, page, nid);
705 * common helper functions for hstate_next_node_to_{alloc|free}.
706 * We may have allocated or freed a huge page based on a different
707 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
708 * be outside of *nodes_allowed. Ensure that we use an allowed
709 * node for alloc or free.
711 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
713 nid = next_node(nid, *nodes_allowed);
714 if (nid == MAX_NUMNODES)
715 nid = first_node(*nodes_allowed);
716 VM_BUG_ON(nid >= MAX_NUMNODES);
721 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
723 if (!node_isset(nid, *nodes_allowed))
724 nid = next_node_allowed(nid, nodes_allowed);
729 * returns the previously saved node ["this node"] from which to
730 * allocate a persistent huge page for the pool and advance the
731 * next node from which to allocate, handling wrap at end of node
734 static int hstate_next_node_to_alloc(struct hstate *h,
735 nodemask_t *nodes_allowed)
739 VM_BUG_ON(!nodes_allowed);
741 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
742 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
747 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
754 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
755 next_nid = start_nid;
758 page = alloc_fresh_huge_page_node(h, next_nid);
763 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
764 } while (next_nid != start_nid);
767 count_vm_event(HTLB_BUDDY_PGALLOC);
769 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
775 * helper for free_pool_huge_page() - return the previously saved
776 * node ["this node"] from which to free a huge page. Advance the
777 * next node id whether or not we find a free huge page to free so
778 * that the next attempt to free addresses the next node.
780 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
784 VM_BUG_ON(!nodes_allowed);
786 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
787 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
793 * Free huge page from pool from next node to free.
794 * Attempt to keep persistent huge pages more or less
795 * balanced over allowed nodes.
796 * Called with hugetlb_lock locked.
798 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
805 start_nid = hstate_next_node_to_free(h, nodes_allowed);
806 next_nid = start_nid;
810 * If we're returning unused surplus pages, only examine
811 * nodes with surplus pages.
813 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
814 !list_empty(&h->hugepage_freelists[next_nid])) {
816 list_entry(h->hugepage_freelists[next_nid].next,
818 list_del(&page->lru);
819 h->free_huge_pages--;
820 h->free_huge_pages_node[next_nid]--;
822 h->surplus_huge_pages--;
823 h->surplus_huge_pages_node[next_nid]--;
825 update_and_free_page(h, page);
829 next_nid = hstate_next_node_to_free(h, nodes_allowed);
830 } while (next_nid != start_nid);
835 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
840 if (h->order >= MAX_ORDER)
844 * Assume we will successfully allocate the surplus page to
845 * prevent racing processes from causing the surplus to exceed
848 * This however introduces a different race, where a process B
849 * tries to grow the static hugepage pool while alloc_pages() is
850 * called by process A. B will only examine the per-node
851 * counters in determining if surplus huge pages can be
852 * converted to normal huge pages in adjust_pool_surplus(). A
853 * won't be able to increment the per-node counter, until the
854 * lock is dropped by B, but B doesn't drop hugetlb_lock until
855 * no more huge pages can be converted from surplus to normal
856 * state (and doesn't try to convert again). Thus, we have a
857 * case where a surplus huge page exists, the pool is grown, and
858 * the surplus huge page still exists after, even though it
859 * should just have been converted to a normal huge page. This
860 * does not leak memory, though, as the hugepage will be freed
861 * once it is out of use. It also does not allow the counters to
862 * go out of whack in adjust_pool_surplus() as we don't modify
863 * the node values until we've gotten the hugepage and only the
864 * per-node value is checked there.
866 spin_lock(&hugetlb_lock);
867 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
868 spin_unlock(&hugetlb_lock);
872 h->surplus_huge_pages++;
874 spin_unlock(&hugetlb_lock);
876 if (nid == NUMA_NO_NODE)
877 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
878 __GFP_REPEAT|__GFP_NOWARN,
881 page = alloc_pages_exact_node(nid,
882 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
883 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
885 if (page && arch_prepare_hugepage(page)) {
886 __free_pages(page, huge_page_order(h));
890 spin_lock(&hugetlb_lock);
892 INIT_LIST_HEAD(&page->lru);
893 r_nid = page_to_nid(page);
894 set_compound_page_dtor(page, free_huge_page);
896 * We incremented the global counters already
898 h->nr_huge_pages_node[r_nid]++;
899 h->surplus_huge_pages_node[r_nid]++;
900 __count_vm_event(HTLB_BUDDY_PGALLOC);
903 h->surplus_huge_pages--;
904 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
906 spin_unlock(&hugetlb_lock);
912 * This allocation function is useful in the context where vma is irrelevant.
913 * E.g. soft-offlining uses this function because it only cares physical
914 * address of error page.
916 struct page *alloc_huge_page_node(struct hstate *h, int nid)
920 spin_lock(&hugetlb_lock);
921 page = dequeue_huge_page_node(h, nid);
922 spin_unlock(&hugetlb_lock);
925 page = alloc_buddy_huge_page(h, nid);
931 * Increase the hugetlb pool such that it can accommodate a reservation
934 static int gather_surplus_pages(struct hstate *h, int delta)
936 struct list_head surplus_list;
937 struct page *page, *tmp;
939 int needed, allocated;
940 bool alloc_ok = true;
942 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
944 h->resv_huge_pages += delta;
949 INIT_LIST_HEAD(&surplus_list);
953 spin_unlock(&hugetlb_lock);
954 for (i = 0; i < needed; i++) {
955 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
960 list_add(&page->lru, &surplus_list);
965 * After retaking hugetlb_lock, we need to recalculate 'needed'
966 * because either resv_huge_pages or free_huge_pages may have changed.
968 spin_lock(&hugetlb_lock);
969 needed = (h->resv_huge_pages + delta) -
970 (h->free_huge_pages + allocated);
975 * We were not able to allocate enough pages to
976 * satisfy the entire reservation so we free what
977 * we've allocated so far.
982 * The surplus_list now contains _at_least_ the number of extra pages
983 * needed to accommodate the reservation. Add the appropriate number
984 * of pages to the hugetlb pool and free the extras back to the buddy
985 * allocator. Commit the entire reservation here to prevent another
986 * process from stealing the pages as they are added to the pool but
987 * before they are reserved.
990 h->resv_huge_pages += delta;
993 /* Free the needed pages to the hugetlb pool */
994 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
998 * This page is now managed by the hugetlb allocator and has
999 * no users -- drop the buddy allocator's reference.
1001 put_page_testzero(page);
1002 VM_BUG_ON(page_count(page));
1003 enqueue_huge_page(h, page);
1006 spin_unlock(&hugetlb_lock);
1008 /* Free unnecessary surplus pages to the buddy allocator */
1009 if (!list_empty(&surplus_list)) {
1010 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1014 spin_lock(&hugetlb_lock);
1020 * When releasing a hugetlb pool reservation, any surplus pages that were
1021 * allocated to satisfy the reservation must be explicitly freed if they were
1023 * Called with hugetlb_lock held.
1025 static void return_unused_surplus_pages(struct hstate *h,
1026 unsigned long unused_resv_pages)
1028 unsigned long nr_pages;
1030 /* Uncommit the reservation */
1031 h->resv_huge_pages -= unused_resv_pages;
1033 /* Cannot return gigantic pages currently */
1034 if (h->order >= MAX_ORDER)
1037 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1040 * We want to release as many surplus pages as possible, spread
1041 * evenly across all nodes with memory. Iterate across these nodes
1042 * until we can no longer free unreserved surplus pages. This occurs
1043 * when the nodes with surplus pages have no free pages.
1044 * free_pool_huge_page() will balance the the freed pages across the
1045 * on-line nodes with memory and will handle the hstate accounting.
1047 while (nr_pages--) {
1048 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1054 * Determine if the huge page at addr within the vma has an associated
1055 * reservation. Where it does not we will need to logically increase
1056 * reservation and actually increase subpool usage before an allocation
1057 * can occur. Where any new reservation would be required the
1058 * reservation change is prepared, but not committed. Once the page
1059 * has been allocated from the subpool and instantiated the change should
1060 * be committed via vma_commit_reservation. No action is required on
1063 static long vma_needs_reservation(struct hstate *h,
1064 struct vm_area_struct *vma, unsigned long addr)
1066 struct address_space *mapping = vma->vm_file->f_mapping;
1067 struct inode *inode = mapping->host;
1069 if (vma->vm_flags & VM_MAYSHARE) {
1070 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1071 return region_chg(&inode->i_mapping->private_list,
1074 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1079 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1080 struct resv_map *reservations = vma_resv_map(vma);
1082 err = region_chg(&reservations->regions, idx, idx + 1);
1088 static void vma_commit_reservation(struct hstate *h,
1089 struct vm_area_struct *vma, unsigned long addr)
1091 struct address_space *mapping = vma->vm_file->f_mapping;
1092 struct inode *inode = mapping->host;
1094 if (vma->vm_flags & VM_MAYSHARE) {
1095 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1096 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1098 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1099 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1100 struct resv_map *reservations = vma_resv_map(vma);
1102 /* Mark this page used in the map. */
1103 region_add(&reservations->regions, idx, idx + 1);
1107 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1108 unsigned long addr, int avoid_reserve)
1110 struct hugepage_subpool *spool = subpool_vma(vma);
1111 struct hstate *h = hstate_vma(vma);
1116 * Processes that did not create the mapping will have no
1117 * reserves and will not have accounted against subpool
1118 * limit. Check that the subpool limit can be made before
1119 * satisfying the allocation MAP_NORESERVE mappings may also
1120 * need pages and subpool limit allocated allocated if no reserve
1123 chg = vma_needs_reservation(h, vma, addr);
1125 return ERR_PTR(-ENOMEM);
1127 if (hugepage_subpool_get_pages(spool, chg))
1128 return ERR_PTR(-ENOSPC);
1130 spin_lock(&hugetlb_lock);
1131 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1132 spin_unlock(&hugetlb_lock);
1135 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1137 hugepage_subpool_put_pages(spool, chg);
1138 return ERR_PTR(-ENOSPC);
1142 set_page_private(page, (unsigned long)spool);
1144 vma_commit_reservation(h, vma, addr);
1149 int __weak alloc_bootmem_huge_page(struct hstate *h)
1151 struct huge_bootmem_page *m;
1152 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1157 addr = __alloc_bootmem_node_nopanic(
1158 NODE_DATA(hstate_next_node_to_alloc(h,
1159 &node_states[N_HIGH_MEMORY])),
1160 huge_page_size(h), huge_page_size(h), 0);
1164 * Use the beginning of the huge page to store the
1165 * huge_bootmem_page struct (until gather_bootmem
1166 * puts them into the mem_map).
1176 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1177 /* Put them into a private list first because mem_map is not up yet */
1178 list_add(&m->list, &huge_boot_pages);
1183 static void prep_compound_huge_page(struct page *page, int order)
1185 if (unlikely(order > (MAX_ORDER - 1)))
1186 prep_compound_gigantic_page(page, order);
1188 prep_compound_page(page, order);
1191 /* Put bootmem huge pages into the standard lists after mem_map is up */
1192 static void __init gather_bootmem_prealloc(void)
1194 struct huge_bootmem_page *m;
1196 list_for_each_entry(m, &huge_boot_pages, list) {
1197 struct hstate *h = m->hstate;
1200 #ifdef CONFIG_HIGHMEM
1201 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1202 free_bootmem_late((unsigned long)m,
1203 sizeof(struct huge_bootmem_page));
1205 page = virt_to_page(m);
1207 __ClearPageReserved(page);
1208 WARN_ON(page_count(page) != 1);
1209 prep_compound_huge_page(page, h->order);
1210 prep_new_huge_page(h, page, page_to_nid(page));
1212 * If we had gigantic hugepages allocated at boot time, we need
1213 * to restore the 'stolen' pages to totalram_pages in order to
1214 * fix confusing memory reports from free(1) and another
1215 * side-effects, like CommitLimit going negative.
1217 if (h->order > (MAX_ORDER - 1))
1218 totalram_pages += 1 << h->order;
1222 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1226 for (i = 0; i < h->max_huge_pages; ++i) {
1227 if (h->order >= MAX_ORDER) {
1228 if (!alloc_bootmem_huge_page(h))
1230 } else if (!alloc_fresh_huge_page(h,
1231 &node_states[N_HIGH_MEMORY]))
1234 h->max_huge_pages = i;
1237 static void __init hugetlb_init_hstates(void)
1241 for_each_hstate(h) {
1242 /* oversize hugepages were init'ed in early boot */
1243 if (h->order < MAX_ORDER)
1244 hugetlb_hstate_alloc_pages(h);
1248 static char * __init memfmt(char *buf, unsigned long n)
1250 if (n >= (1UL << 30))
1251 sprintf(buf, "%lu GB", n >> 30);
1252 else if (n >= (1UL << 20))
1253 sprintf(buf, "%lu MB", n >> 20);
1255 sprintf(buf, "%lu KB", n >> 10);
1259 static void __init report_hugepages(void)
1263 for_each_hstate(h) {
1265 printk(KERN_INFO "HugeTLB registered %s page size, "
1266 "pre-allocated %ld pages\n",
1267 memfmt(buf, huge_page_size(h)),
1268 h->free_huge_pages);
1272 #ifdef CONFIG_HIGHMEM
1273 static void try_to_free_low(struct hstate *h, unsigned long count,
1274 nodemask_t *nodes_allowed)
1278 if (h->order >= MAX_ORDER)
1281 for_each_node_mask(i, *nodes_allowed) {
1282 struct page *page, *next;
1283 struct list_head *freel = &h->hugepage_freelists[i];
1284 list_for_each_entry_safe(page, next, freel, lru) {
1285 if (count >= h->nr_huge_pages)
1287 if (PageHighMem(page))
1289 list_del(&page->lru);
1290 update_and_free_page(h, page);
1291 h->free_huge_pages--;
1292 h->free_huge_pages_node[page_to_nid(page)]--;
1297 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1298 nodemask_t *nodes_allowed)
1304 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1305 * balanced by operating on them in a round-robin fashion.
1306 * Returns 1 if an adjustment was made.
1308 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1311 int start_nid, next_nid;
1314 VM_BUG_ON(delta != -1 && delta != 1);
1317 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1319 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1320 next_nid = start_nid;
1326 * To shrink on this node, there must be a surplus page
1328 if (!h->surplus_huge_pages_node[nid]) {
1329 next_nid = hstate_next_node_to_alloc(h,
1336 * Surplus cannot exceed the total number of pages
1338 if (h->surplus_huge_pages_node[nid] >=
1339 h->nr_huge_pages_node[nid]) {
1340 next_nid = hstate_next_node_to_free(h,
1346 h->surplus_huge_pages += delta;
1347 h->surplus_huge_pages_node[nid] += delta;
1350 } while (next_nid != start_nid);
1355 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1356 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1357 nodemask_t *nodes_allowed)
1359 unsigned long min_count, ret;
1361 if (h->order >= MAX_ORDER)
1362 return h->max_huge_pages;
1365 * Increase the pool size
1366 * First take pages out of surplus state. Then make up the
1367 * remaining difference by allocating fresh huge pages.
1369 * We might race with alloc_buddy_huge_page() here and be unable
1370 * to convert a surplus huge page to a normal huge page. That is
1371 * not critical, though, it just means the overall size of the
1372 * pool might be one hugepage larger than it needs to be, but
1373 * within all the constraints specified by the sysctls.
1375 spin_lock(&hugetlb_lock);
1376 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1377 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1381 while (count > persistent_huge_pages(h)) {
1383 * If this allocation races such that we no longer need the
1384 * page, free_huge_page will handle it by freeing the page
1385 * and reducing the surplus.
1387 spin_unlock(&hugetlb_lock);
1388 ret = alloc_fresh_huge_page(h, nodes_allowed);
1389 spin_lock(&hugetlb_lock);
1393 /* Bail for signals. Probably ctrl-c from user */
1394 if (signal_pending(current))
1399 * Decrease the pool size
1400 * First return free pages to the buddy allocator (being careful
1401 * to keep enough around to satisfy reservations). Then place
1402 * pages into surplus state as needed so the pool will shrink
1403 * to the desired size as pages become free.
1405 * By placing pages into the surplus state independent of the
1406 * overcommit value, we are allowing the surplus pool size to
1407 * exceed overcommit. There are few sane options here. Since
1408 * alloc_buddy_huge_page() is checking the global counter,
1409 * though, we'll note that we're not allowed to exceed surplus
1410 * and won't grow the pool anywhere else. Not until one of the
1411 * sysctls are changed, or the surplus pages go out of use.
1413 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1414 min_count = max(count, min_count);
1415 try_to_free_low(h, min_count, nodes_allowed);
1416 while (min_count < persistent_huge_pages(h)) {
1417 if (!free_pool_huge_page(h, nodes_allowed, 0))
1420 while (count < persistent_huge_pages(h)) {
1421 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1425 ret = persistent_huge_pages(h);
1426 spin_unlock(&hugetlb_lock);
1430 #define HSTATE_ATTR_RO(_name) \
1431 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1433 #define HSTATE_ATTR(_name) \
1434 static struct kobj_attribute _name##_attr = \
1435 __ATTR(_name, 0644, _name##_show, _name##_store)
1437 static struct kobject *hugepages_kobj;
1438 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1440 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1442 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1446 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1447 if (hstate_kobjs[i] == kobj) {
1449 *nidp = NUMA_NO_NODE;
1453 return kobj_to_node_hstate(kobj, nidp);
1456 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1457 struct kobj_attribute *attr, char *buf)
1460 unsigned long nr_huge_pages;
1463 h = kobj_to_hstate(kobj, &nid);
1464 if (nid == NUMA_NO_NODE)
1465 nr_huge_pages = h->nr_huge_pages;
1467 nr_huge_pages = h->nr_huge_pages_node[nid];
1469 return sprintf(buf, "%lu\n", nr_huge_pages);
1472 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1473 struct kobject *kobj, struct kobj_attribute *attr,
1474 const char *buf, size_t len)
1478 unsigned long count;
1480 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1482 err = strict_strtoul(buf, 10, &count);
1486 h = kobj_to_hstate(kobj, &nid);
1487 if (h->order >= MAX_ORDER) {
1492 if (nid == NUMA_NO_NODE) {
1494 * global hstate attribute
1496 if (!(obey_mempolicy &&
1497 init_nodemask_of_mempolicy(nodes_allowed))) {
1498 NODEMASK_FREE(nodes_allowed);
1499 nodes_allowed = &node_states[N_HIGH_MEMORY];
1501 } else if (nodes_allowed) {
1503 * per node hstate attribute: adjust count to global,
1504 * but restrict alloc/free to the specified node.
1506 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1507 init_nodemask_of_node(nodes_allowed, nid);
1509 nodes_allowed = &node_states[N_HIGH_MEMORY];
1511 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1513 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1514 NODEMASK_FREE(nodes_allowed);
1518 NODEMASK_FREE(nodes_allowed);
1522 static ssize_t nr_hugepages_show(struct kobject *kobj,
1523 struct kobj_attribute *attr, char *buf)
1525 return nr_hugepages_show_common(kobj, attr, buf);
1528 static ssize_t nr_hugepages_store(struct kobject *kobj,
1529 struct kobj_attribute *attr, const char *buf, size_t len)
1531 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1533 HSTATE_ATTR(nr_hugepages);
1538 * hstate attribute for optionally mempolicy-based constraint on persistent
1539 * huge page alloc/free.
1541 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1542 struct kobj_attribute *attr, char *buf)
1544 return nr_hugepages_show_common(kobj, attr, buf);
1547 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1548 struct kobj_attribute *attr, const char *buf, size_t len)
1550 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1552 HSTATE_ATTR(nr_hugepages_mempolicy);
1556 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1557 struct kobj_attribute *attr, char *buf)
1559 struct hstate *h = kobj_to_hstate(kobj, NULL);
1560 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1563 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1564 struct kobj_attribute *attr, const char *buf, size_t count)
1567 unsigned long input;
1568 struct hstate *h = kobj_to_hstate(kobj, NULL);
1570 if (h->order >= MAX_ORDER)
1573 err = strict_strtoul(buf, 10, &input);
1577 spin_lock(&hugetlb_lock);
1578 h->nr_overcommit_huge_pages = input;
1579 spin_unlock(&hugetlb_lock);
1583 HSTATE_ATTR(nr_overcommit_hugepages);
1585 static ssize_t free_hugepages_show(struct kobject *kobj,
1586 struct kobj_attribute *attr, char *buf)
1589 unsigned long free_huge_pages;
1592 h = kobj_to_hstate(kobj, &nid);
1593 if (nid == NUMA_NO_NODE)
1594 free_huge_pages = h->free_huge_pages;
1596 free_huge_pages = h->free_huge_pages_node[nid];
1598 return sprintf(buf, "%lu\n", free_huge_pages);
1600 HSTATE_ATTR_RO(free_hugepages);
1602 static ssize_t resv_hugepages_show(struct kobject *kobj,
1603 struct kobj_attribute *attr, char *buf)
1605 struct hstate *h = kobj_to_hstate(kobj, NULL);
1606 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1608 HSTATE_ATTR_RO(resv_hugepages);
1610 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1611 struct kobj_attribute *attr, char *buf)
1614 unsigned long surplus_huge_pages;
1617 h = kobj_to_hstate(kobj, &nid);
1618 if (nid == NUMA_NO_NODE)
1619 surplus_huge_pages = h->surplus_huge_pages;
1621 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1623 return sprintf(buf, "%lu\n", surplus_huge_pages);
1625 HSTATE_ATTR_RO(surplus_hugepages);
1627 static struct attribute *hstate_attrs[] = {
1628 &nr_hugepages_attr.attr,
1629 &nr_overcommit_hugepages_attr.attr,
1630 &free_hugepages_attr.attr,
1631 &resv_hugepages_attr.attr,
1632 &surplus_hugepages_attr.attr,
1634 &nr_hugepages_mempolicy_attr.attr,
1639 static struct attribute_group hstate_attr_group = {
1640 .attrs = hstate_attrs,
1643 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1644 struct kobject **hstate_kobjs,
1645 struct attribute_group *hstate_attr_group)
1648 int hi = hstate_index(h);
1650 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1651 if (!hstate_kobjs[hi])
1654 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1656 kobject_put(hstate_kobjs[hi]);
1661 static void __init hugetlb_sysfs_init(void)
1666 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1667 if (!hugepages_kobj)
1670 for_each_hstate(h) {
1671 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1672 hstate_kobjs, &hstate_attr_group);
1674 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1682 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1683 * with node devices in node_devices[] using a parallel array. The array
1684 * index of a node device or _hstate == node id.
1685 * This is here to avoid any static dependency of the node device driver, in
1686 * the base kernel, on the hugetlb module.
1688 struct node_hstate {
1689 struct kobject *hugepages_kobj;
1690 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1692 struct node_hstate node_hstates[MAX_NUMNODES];
1695 * A subset of global hstate attributes for node devices
1697 static struct attribute *per_node_hstate_attrs[] = {
1698 &nr_hugepages_attr.attr,
1699 &free_hugepages_attr.attr,
1700 &surplus_hugepages_attr.attr,
1704 static struct attribute_group per_node_hstate_attr_group = {
1705 .attrs = per_node_hstate_attrs,
1709 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1710 * Returns node id via non-NULL nidp.
1712 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1716 for (nid = 0; nid < nr_node_ids; nid++) {
1717 struct node_hstate *nhs = &node_hstates[nid];
1719 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1720 if (nhs->hstate_kobjs[i] == kobj) {
1732 * Unregister hstate attributes from a single node device.
1733 * No-op if no hstate attributes attached.
1735 void hugetlb_unregister_node(struct node *node)
1738 struct node_hstate *nhs = &node_hstates[node->dev.id];
1740 if (!nhs->hugepages_kobj)
1741 return; /* no hstate attributes */
1743 for_each_hstate(h) {
1744 int idx = hstate_index(h);
1745 if (nhs->hstate_kobjs[idx]) {
1746 kobject_put(nhs->hstate_kobjs[idx]);
1747 nhs->hstate_kobjs[idx] = NULL;
1751 kobject_put(nhs->hugepages_kobj);
1752 nhs->hugepages_kobj = NULL;
1756 * hugetlb module exit: unregister hstate attributes from node devices
1759 static void hugetlb_unregister_all_nodes(void)
1764 * disable node device registrations.
1766 register_hugetlbfs_with_node(NULL, NULL);
1769 * remove hstate attributes from any nodes that have them.
1771 for (nid = 0; nid < nr_node_ids; nid++)
1772 hugetlb_unregister_node(&node_devices[nid]);
1776 * Register hstate attributes for a single node device.
1777 * No-op if attributes already registered.
1779 void hugetlb_register_node(struct node *node)
1782 struct node_hstate *nhs = &node_hstates[node->dev.id];
1785 if (nhs->hugepages_kobj)
1786 return; /* already allocated */
1788 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1790 if (!nhs->hugepages_kobj)
1793 for_each_hstate(h) {
1794 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1796 &per_node_hstate_attr_group);
1798 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1800 h->name, node->dev.id);
1801 hugetlb_unregister_node(node);
1808 * hugetlb init time: register hstate attributes for all registered node
1809 * devices of nodes that have memory. All on-line nodes should have
1810 * registered their associated device by this time.
1812 static void hugetlb_register_all_nodes(void)
1816 for_each_node_state(nid, N_HIGH_MEMORY) {
1817 struct node *node = &node_devices[nid];
1818 if (node->dev.id == nid)
1819 hugetlb_register_node(node);
1823 * Let the node device driver know we're here so it can
1824 * [un]register hstate attributes on node hotplug.
1826 register_hugetlbfs_with_node(hugetlb_register_node,
1827 hugetlb_unregister_node);
1829 #else /* !CONFIG_NUMA */
1831 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1839 static void hugetlb_unregister_all_nodes(void) { }
1841 static void hugetlb_register_all_nodes(void) { }
1845 static void __exit hugetlb_exit(void)
1849 hugetlb_unregister_all_nodes();
1851 for_each_hstate(h) {
1852 kobject_put(hstate_kobjs[hstate_index(h)]);
1855 kobject_put(hugepages_kobj);
1857 module_exit(hugetlb_exit);
1859 static int __init hugetlb_init(void)
1861 /* Some platform decide whether they support huge pages at boot
1862 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1863 * there is no such support
1865 if (HPAGE_SHIFT == 0)
1868 if (!size_to_hstate(default_hstate_size)) {
1869 default_hstate_size = HPAGE_SIZE;
1870 if (!size_to_hstate(default_hstate_size))
1871 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1873 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1874 if (default_hstate_max_huge_pages)
1875 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1877 hugetlb_init_hstates();
1879 gather_bootmem_prealloc();
1883 hugetlb_sysfs_init();
1885 hugetlb_register_all_nodes();
1889 module_init(hugetlb_init);
1891 /* Should be called on processing a hugepagesz=... option */
1892 void __init hugetlb_add_hstate(unsigned order)
1897 if (size_to_hstate(PAGE_SIZE << order)) {
1898 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1901 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1903 h = &hstates[hugetlb_max_hstate++];
1905 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1906 h->nr_huge_pages = 0;
1907 h->free_huge_pages = 0;
1908 for (i = 0; i < MAX_NUMNODES; ++i)
1909 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1910 INIT_LIST_HEAD(&h->hugepage_activelist);
1911 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1912 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1913 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1914 huge_page_size(h)/1024);
1919 static int __init hugetlb_nrpages_setup(char *s)
1922 static unsigned long *last_mhp;
1925 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1926 * so this hugepages= parameter goes to the "default hstate".
1928 if (!hugetlb_max_hstate)
1929 mhp = &default_hstate_max_huge_pages;
1931 mhp = &parsed_hstate->max_huge_pages;
1933 if (mhp == last_mhp) {
1934 printk(KERN_WARNING "hugepages= specified twice without "
1935 "interleaving hugepagesz=, ignoring\n");
1939 if (sscanf(s, "%lu", mhp) <= 0)
1943 * Global state is always initialized later in hugetlb_init.
1944 * But we need to allocate >= MAX_ORDER hstates here early to still
1945 * use the bootmem allocator.
1947 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1948 hugetlb_hstate_alloc_pages(parsed_hstate);
1954 __setup("hugepages=", hugetlb_nrpages_setup);
1956 static int __init hugetlb_default_setup(char *s)
1958 default_hstate_size = memparse(s, &s);
1961 __setup("default_hugepagesz=", hugetlb_default_setup);
1963 static unsigned int cpuset_mems_nr(unsigned int *array)
1966 unsigned int nr = 0;
1968 for_each_node_mask(node, cpuset_current_mems_allowed)
1974 #ifdef CONFIG_SYSCTL
1975 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1976 struct ctl_table *table, int write,
1977 void __user *buffer, size_t *length, loff_t *ppos)
1979 struct hstate *h = &default_hstate;
1983 tmp = h->max_huge_pages;
1985 if (write && h->order >= MAX_ORDER)
1989 table->maxlen = sizeof(unsigned long);
1990 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1995 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1996 GFP_KERNEL | __GFP_NORETRY);
1997 if (!(obey_mempolicy &&
1998 init_nodemask_of_mempolicy(nodes_allowed))) {
1999 NODEMASK_FREE(nodes_allowed);
2000 nodes_allowed = &node_states[N_HIGH_MEMORY];
2002 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2004 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2005 NODEMASK_FREE(nodes_allowed);
2011 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2012 void __user *buffer, size_t *length, loff_t *ppos)
2015 return hugetlb_sysctl_handler_common(false, table, write,
2016 buffer, length, ppos);
2020 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2021 void __user *buffer, size_t *length, loff_t *ppos)
2023 return hugetlb_sysctl_handler_common(true, table, write,
2024 buffer, length, ppos);
2026 #endif /* CONFIG_NUMA */
2028 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2029 void __user *buffer,
2030 size_t *length, loff_t *ppos)
2032 proc_dointvec(table, write, buffer, length, ppos);
2033 if (hugepages_treat_as_movable)
2034 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2036 htlb_alloc_mask = GFP_HIGHUSER;
2040 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2041 void __user *buffer,
2042 size_t *length, loff_t *ppos)
2044 struct hstate *h = &default_hstate;
2048 tmp = h->nr_overcommit_huge_pages;
2050 if (write && h->order >= MAX_ORDER)
2054 table->maxlen = sizeof(unsigned long);
2055 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2060 spin_lock(&hugetlb_lock);
2061 h->nr_overcommit_huge_pages = tmp;
2062 spin_unlock(&hugetlb_lock);
2068 #endif /* CONFIG_SYSCTL */
2070 void hugetlb_report_meminfo(struct seq_file *m)
2072 struct hstate *h = &default_hstate;
2074 "HugePages_Total: %5lu\n"
2075 "HugePages_Free: %5lu\n"
2076 "HugePages_Rsvd: %5lu\n"
2077 "HugePages_Surp: %5lu\n"
2078 "Hugepagesize: %8lu kB\n",
2082 h->surplus_huge_pages,
2083 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2086 int hugetlb_report_node_meminfo(int nid, char *buf)
2088 struct hstate *h = &default_hstate;
2090 "Node %d HugePages_Total: %5u\n"
2091 "Node %d HugePages_Free: %5u\n"
2092 "Node %d HugePages_Surp: %5u\n",
2093 nid, h->nr_huge_pages_node[nid],
2094 nid, h->free_huge_pages_node[nid],
2095 nid, h->surplus_huge_pages_node[nid]);
2098 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2099 unsigned long hugetlb_total_pages(void)
2101 struct hstate *h = &default_hstate;
2102 return h->nr_huge_pages * pages_per_huge_page(h);
2105 static int hugetlb_acct_memory(struct hstate *h, long delta)
2109 spin_lock(&hugetlb_lock);
2111 * When cpuset is configured, it breaks the strict hugetlb page
2112 * reservation as the accounting is done on a global variable. Such
2113 * reservation is completely rubbish in the presence of cpuset because
2114 * the reservation is not checked against page availability for the
2115 * current cpuset. Application can still potentially OOM'ed by kernel
2116 * with lack of free htlb page in cpuset that the task is in.
2117 * Attempt to enforce strict accounting with cpuset is almost
2118 * impossible (or too ugly) because cpuset is too fluid that
2119 * task or memory node can be dynamically moved between cpusets.
2121 * The change of semantics for shared hugetlb mapping with cpuset is
2122 * undesirable. However, in order to preserve some of the semantics,
2123 * we fall back to check against current free page availability as
2124 * a best attempt and hopefully to minimize the impact of changing
2125 * semantics that cpuset has.
2128 if (gather_surplus_pages(h, delta) < 0)
2131 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2132 return_unused_surplus_pages(h, delta);
2139 return_unused_surplus_pages(h, (unsigned long) -delta);
2142 spin_unlock(&hugetlb_lock);
2146 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2148 struct resv_map *reservations = vma_resv_map(vma);
2151 * This new VMA should share its siblings reservation map if present.
2152 * The VMA will only ever have a valid reservation map pointer where
2153 * it is being copied for another still existing VMA. As that VMA
2154 * has a reference to the reservation map it cannot disappear until
2155 * after this open call completes. It is therefore safe to take a
2156 * new reference here without additional locking.
2159 kref_get(&reservations->refs);
2162 static void resv_map_put(struct vm_area_struct *vma)
2164 struct resv_map *reservations = vma_resv_map(vma);
2168 kref_put(&reservations->refs, resv_map_release);
2171 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2173 struct hstate *h = hstate_vma(vma);
2174 struct resv_map *reservations = vma_resv_map(vma);
2175 struct hugepage_subpool *spool = subpool_vma(vma);
2176 unsigned long reserve;
2177 unsigned long start;
2181 start = vma_hugecache_offset(h, vma, vma->vm_start);
2182 end = vma_hugecache_offset(h, vma, vma->vm_end);
2184 reserve = (end - start) -
2185 region_count(&reservations->regions, start, end);
2190 hugetlb_acct_memory(h, -reserve);
2191 hugepage_subpool_put_pages(spool, reserve);
2197 * We cannot handle pagefaults against hugetlb pages at all. They cause
2198 * handle_mm_fault() to try to instantiate regular-sized pages in the
2199 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2202 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2208 const struct vm_operations_struct hugetlb_vm_ops = {
2209 .fault = hugetlb_vm_op_fault,
2210 .open = hugetlb_vm_op_open,
2211 .close = hugetlb_vm_op_close,
2214 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2221 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2223 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2225 entry = pte_mkyoung(entry);
2226 entry = pte_mkhuge(entry);
2227 entry = arch_make_huge_pte(entry, vma, page, writable);
2232 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2233 unsigned long address, pte_t *ptep)
2237 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2238 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2239 update_mmu_cache(vma, address, ptep);
2243 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2244 struct vm_area_struct *vma)
2246 pte_t *src_pte, *dst_pte, entry;
2247 struct page *ptepage;
2250 struct hstate *h = hstate_vma(vma);
2251 unsigned long sz = huge_page_size(h);
2253 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2255 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2256 src_pte = huge_pte_offset(src, addr);
2259 dst_pte = huge_pte_alloc(dst, addr, sz);
2263 /* If the pagetables are shared don't copy or take references */
2264 if (dst_pte == src_pte)
2267 spin_lock(&dst->page_table_lock);
2268 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2269 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2271 huge_ptep_set_wrprotect(src, addr, src_pte);
2272 entry = huge_ptep_get(src_pte);
2273 ptepage = pte_page(entry);
2275 page_dup_rmap(ptepage);
2276 set_huge_pte_at(dst, addr, dst_pte, entry);
2278 spin_unlock(&src->page_table_lock);
2279 spin_unlock(&dst->page_table_lock);
2287 static int is_hugetlb_entry_migration(pte_t pte)
2291 if (huge_pte_none(pte) || pte_present(pte))
2293 swp = pte_to_swp_entry(pte);
2294 if (non_swap_entry(swp) && is_migration_entry(swp))
2300 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2304 if (huge_pte_none(pte) || pte_present(pte))
2306 swp = pte_to_swp_entry(pte);
2307 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2313 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2314 unsigned long start, unsigned long end,
2315 struct page *ref_page)
2317 int force_flush = 0;
2318 struct mm_struct *mm = vma->vm_mm;
2319 unsigned long address;
2323 struct hstate *h = hstate_vma(vma);
2324 unsigned long sz = huge_page_size(h);
2326 WARN_ON(!is_vm_hugetlb_page(vma));
2327 BUG_ON(start & ~huge_page_mask(h));
2328 BUG_ON(end & ~huge_page_mask(h));
2330 tlb_start_vma(tlb, vma);
2331 mmu_notifier_invalidate_range_start(mm, start, end);
2333 spin_lock(&mm->page_table_lock);
2334 for (address = start; address < end; address += sz) {
2335 ptep = huge_pte_offset(mm, address);
2339 if (huge_pmd_unshare(mm, &address, ptep))
2342 pte = huge_ptep_get(ptep);
2343 if (huge_pte_none(pte))
2347 * HWPoisoned hugepage is already unmapped and dropped reference
2349 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2352 page = pte_page(pte);
2354 * If a reference page is supplied, it is because a specific
2355 * page is being unmapped, not a range. Ensure the page we
2356 * are about to unmap is the actual page of interest.
2359 if (page != ref_page)
2363 * Mark the VMA as having unmapped its page so that
2364 * future faults in this VMA will fail rather than
2365 * looking like data was lost
2367 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2370 pte = huge_ptep_get_and_clear(mm, address, ptep);
2371 tlb_remove_tlb_entry(tlb, ptep, address);
2373 set_page_dirty(page);
2375 page_remove_rmap(page);
2376 force_flush = !__tlb_remove_page(tlb, page);
2379 /* Bail out after unmapping reference page if supplied */
2383 spin_unlock(&mm->page_table_lock);
2385 * mmu_gather ran out of room to batch pages, we break out of
2386 * the PTE lock to avoid doing the potential expensive TLB invalidate
2387 * and page-free while holding it.
2392 if (address < end && !ref_page)
2395 mmu_notifier_invalidate_range_end(mm, start, end);
2396 tlb_end_vma(tlb, vma);
2399 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2400 unsigned long end, struct page *ref_page)
2402 struct mm_struct *mm;
2403 struct mmu_gather tlb;
2407 tlb_gather_mmu(&tlb, mm, 0);
2408 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2409 tlb_finish_mmu(&tlb, start, end);
2413 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2414 * mappping it owns the reserve page for. The intention is to unmap the page
2415 * from other VMAs and let the children be SIGKILLed if they are faulting the
2418 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2419 struct page *page, unsigned long address)
2421 struct hstate *h = hstate_vma(vma);
2422 struct vm_area_struct *iter_vma;
2423 struct address_space *mapping;
2424 struct prio_tree_iter iter;
2428 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2429 * from page cache lookup which is in HPAGE_SIZE units.
2431 address = address & huge_page_mask(h);
2432 pgoff = vma_hugecache_offset(h, vma, address);
2433 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2436 * Take the mapping lock for the duration of the table walk. As
2437 * this mapping should be shared between all the VMAs,
2438 * __unmap_hugepage_range() is called as the lock is already held
2440 mutex_lock(&mapping->i_mmap_mutex);
2441 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2442 /* Do not unmap the current VMA */
2443 if (iter_vma == vma)
2447 * Unmap the page from other VMAs without their own reserves.
2448 * They get marked to be SIGKILLed if they fault in these
2449 * areas. This is because a future no-page fault on this VMA
2450 * could insert a zeroed page instead of the data existing
2451 * from the time of fork. This would look like data corruption
2453 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2454 unmap_hugepage_range(iter_vma, address,
2455 address + huge_page_size(h), page);
2457 mutex_unlock(&mapping->i_mmap_mutex);
2463 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2464 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2465 * cannot race with other handlers or page migration.
2466 * Keep the pte_same checks anyway to make transition from the mutex easier.
2468 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2469 unsigned long address, pte_t *ptep, pte_t pte,
2470 struct page *pagecache_page)
2472 struct hstate *h = hstate_vma(vma);
2473 struct page *old_page, *new_page;
2475 int outside_reserve = 0;
2477 old_page = pte_page(pte);
2480 /* If no-one else is actually using this page, avoid the copy
2481 * and just make the page writable */
2482 avoidcopy = (page_mapcount(old_page) == 1);
2484 if (PageAnon(old_page))
2485 page_move_anon_rmap(old_page, vma, address);
2486 set_huge_ptep_writable(vma, address, ptep);
2491 * If the process that created a MAP_PRIVATE mapping is about to
2492 * perform a COW due to a shared page count, attempt to satisfy
2493 * the allocation without using the existing reserves. The pagecache
2494 * page is used to determine if the reserve at this address was
2495 * consumed or not. If reserves were used, a partial faulted mapping
2496 * at the time of fork() could consume its reserves on COW instead
2497 * of the full address range.
2499 if (!(vma->vm_flags & VM_MAYSHARE) &&
2500 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2501 old_page != pagecache_page)
2502 outside_reserve = 1;
2504 page_cache_get(old_page);
2506 /* Drop page_table_lock as buddy allocator may be called */
2507 spin_unlock(&mm->page_table_lock);
2508 new_page = alloc_huge_page(vma, address, outside_reserve);
2510 if (IS_ERR(new_page)) {
2511 long err = PTR_ERR(new_page);
2512 page_cache_release(old_page);
2515 * If a process owning a MAP_PRIVATE mapping fails to COW,
2516 * it is due to references held by a child and an insufficient
2517 * huge page pool. To guarantee the original mappers
2518 * reliability, unmap the page from child processes. The child
2519 * may get SIGKILLed if it later faults.
2521 if (outside_reserve) {
2522 BUG_ON(huge_pte_none(pte));
2523 if (unmap_ref_private(mm, vma, old_page, address)) {
2524 BUG_ON(huge_pte_none(pte));
2525 spin_lock(&mm->page_table_lock);
2526 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2527 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2528 goto retry_avoidcopy;
2530 * race occurs while re-acquiring page_table_lock, and
2538 /* Caller expects lock to be held */
2539 spin_lock(&mm->page_table_lock);
2541 return VM_FAULT_OOM;
2543 return VM_FAULT_SIGBUS;
2547 * When the original hugepage is shared one, it does not have
2548 * anon_vma prepared.
2550 if (unlikely(anon_vma_prepare(vma))) {
2551 page_cache_release(new_page);
2552 page_cache_release(old_page);
2553 /* Caller expects lock to be held */
2554 spin_lock(&mm->page_table_lock);
2555 return VM_FAULT_OOM;
2558 copy_user_huge_page(new_page, old_page, address, vma,
2559 pages_per_huge_page(h));
2560 __SetPageUptodate(new_page);
2563 * Retake the page_table_lock to check for racing updates
2564 * before the page tables are altered
2566 spin_lock(&mm->page_table_lock);
2567 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2568 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2570 mmu_notifier_invalidate_range_start(mm,
2571 address & huge_page_mask(h),
2572 (address & huge_page_mask(h)) + huge_page_size(h));
2573 huge_ptep_clear_flush(vma, address, ptep);
2574 set_huge_pte_at(mm, address, ptep,
2575 make_huge_pte(vma, new_page, 1));
2576 page_remove_rmap(old_page);
2577 hugepage_add_new_anon_rmap(new_page, vma, address);
2578 /* Make the old page be freed below */
2579 new_page = old_page;
2580 mmu_notifier_invalidate_range_end(mm,
2581 address & huge_page_mask(h),
2582 (address & huge_page_mask(h)) + huge_page_size(h));
2584 page_cache_release(new_page);
2585 page_cache_release(old_page);
2589 /* Return the pagecache page at a given address within a VMA */
2590 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2591 struct vm_area_struct *vma, unsigned long address)
2593 struct address_space *mapping;
2596 mapping = vma->vm_file->f_mapping;
2597 idx = vma_hugecache_offset(h, vma, address);
2599 return find_lock_page(mapping, idx);
2603 * Return whether there is a pagecache page to back given address within VMA.
2604 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2606 static bool hugetlbfs_pagecache_present(struct hstate *h,
2607 struct vm_area_struct *vma, unsigned long address)
2609 struct address_space *mapping;
2613 mapping = vma->vm_file->f_mapping;
2614 idx = vma_hugecache_offset(h, vma, address);
2616 page = find_get_page(mapping, idx);
2619 return page != NULL;
2622 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2623 unsigned long address, pte_t *ptep, unsigned int flags)
2625 struct hstate *h = hstate_vma(vma);
2626 int ret = VM_FAULT_SIGBUS;
2631 struct address_space *mapping;
2635 * Currently, we are forced to kill the process in the event the
2636 * original mapper has unmapped pages from the child due to a failed
2637 * COW. Warn that such a situation has occurred as it may not be obvious
2639 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2641 "PID %d killed due to inadequate hugepage pool\n",
2646 mapping = vma->vm_file->f_mapping;
2647 idx = vma_hugecache_offset(h, vma, address);
2650 * Use page lock to guard against racing truncation
2651 * before we get page_table_lock.
2654 page = find_lock_page(mapping, idx);
2656 size = i_size_read(mapping->host) >> huge_page_shift(h);
2659 page = alloc_huge_page(vma, address, 0);
2661 ret = PTR_ERR(page);
2665 ret = VM_FAULT_SIGBUS;
2668 clear_huge_page(page, address, pages_per_huge_page(h));
2669 __SetPageUptodate(page);
2671 if (vma->vm_flags & VM_MAYSHARE) {
2673 struct inode *inode = mapping->host;
2675 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2683 spin_lock(&inode->i_lock);
2684 inode->i_blocks += blocks_per_huge_page(h);
2685 spin_unlock(&inode->i_lock);
2688 if (unlikely(anon_vma_prepare(vma))) {
2690 goto backout_unlocked;
2696 * If memory error occurs between mmap() and fault, some process
2697 * don't have hwpoisoned swap entry for errored virtual address.
2698 * So we need to block hugepage fault by PG_hwpoison bit check.
2700 if (unlikely(PageHWPoison(page))) {
2701 ret = VM_FAULT_HWPOISON |
2702 VM_FAULT_SET_HINDEX(hstate_index(h));
2703 goto backout_unlocked;
2708 * If we are going to COW a private mapping later, we examine the
2709 * pending reservations for this page now. This will ensure that
2710 * any allocations necessary to record that reservation occur outside
2713 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2714 if (vma_needs_reservation(h, vma, address) < 0) {
2716 goto backout_unlocked;
2719 spin_lock(&mm->page_table_lock);
2720 size = i_size_read(mapping->host) >> huge_page_shift(h);
2725 if (!huge_pte_none(huge_ptep_get(ptep)))
2729 hugepage_add_new_anon_rmap(page, vma, address);
2731 page_dup_rmap(page);
2732 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2733 && (vma->vm_flags & VM_SHARED)));
2734 set_huge_pte_at(mm, address, ptep, new_pte);
2736 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2737 /* Optimization, do the COW without a second fault */
2738 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2741 spin_unlock(&mm->page_table_lock);
2747 spin_unlock(&mm->page_table_lock);
2754 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2755 unsigned long address, unsigned int flags)
2760 struct page *page = NULL;
2761 struct page *pagecache_page = NULL;
2762 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2763 struct hstate *h = hstate_vma(vma);
2765 address &= huge_page_mask(h);
2767 ptep = huge_pte_offset(mm, address);
2769 entry = huge_ptep_get(ptep);
2770 if (unlikely(is_hugetlb_entry_migration(entry))) {
2771 migration_entry_wait(mm, (pmd_t *)ptep, address);
2773 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2774 return VM_FAULT_HWPOISON_LARGE |
2775 VM_FAULT_SET_HINDEX(hstate_index(h));
2778 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2780 return VM_FAULT_OOM;
2783 * Serialize hugepage allocation and instantiation, so that we don't
2784 * get spurious allocation failures if two CPUs race to instantiate
2785 * the same page in the page cache.
2787 mutex_lock(&hugetlb_instantiation_mutex);
2788 entry = huge_ptep_get(ptep);
2789 if (huge_pte_none(entry)) {
2790 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2797 * If we are going to COW the mapping later, we examine the pending
2798 * reservations for this page now. This will ensure that any
2799 * allocations necessary to record that reservation occur outside the
2800 * spinlock. For private mappings, we also lookup the pagecache
2801 * page now as it is used to determine if a reservation has been
2804 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2805 if (vma_needs_reservation(h, vma, address) < 0) {
2810 if (!(vma->vm_flags & VM_MAYSHARE))
2811 pagecache_page = hugetlbfs_pagecache_page(h,
2816 * hugetlb_cow() requires page locks of pte_page(entry) and
2817 * pagecache_page, so here we need take the former one
2818 * when page != pagecache_page or !pagecache_page.
2819 * Note that locking order is always pagecache_page -> page,
2820 * so no worry about deadlock.
2822 page = pte_page(entry);
2824 if (page != pagecache_page)
2827 spin_lock(&mm->page_table_lock);
2828 /* Check for a racing update before calling hugetlb_cow */
2829 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2830 goto out_page_table_lock;
2833 if (flags & FAULT_FLAG_WRITE) {
2834 if (!pte_write(entry)) {
2835 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2837 goto out_page_table_lock;
2839 entry = pte_mkdirty(entry);
2841 entry = pte_mkyoung(entry);
2842 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2843 flags & FAULT_FLAG_WRITE))
2844 update_mmu_cache(vma, address, ptep);
2846 out_page_table_lock:
2847 spin_unlock(&mm->page_table_lock);
2849 if (pagecache_page) {
2850 unlock_page(pagecache_page);
2851 put_page(pagecache_page);
2853 if (page != pagecache_page)
2858 mutex_unlock(&hugetlb_instantiation_mutex);
2863 /* Can be overriden by architectures */
2864 __attribute__((weak)) struct page *
2865 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2866 pud_t *pud, int write)
2872 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2873 struct page **pages, struct vm_area_struct **vmas,
2874 unsigned long *position, int *length, int i,
2877 unsigned long pfn_offset;
2878 unsigned long vaddr = *position;
2879 int remainder = *length;
2880 struct hstate *h = hstate_vma(vma);
2882 spin_lock(&mm->page_table_lock);
2883 while (vaddr < vma->vm_end && remainder) {
2889 * Some archs (sparc64, sh*) have multiple pte_ts to
2890 * each hugepage. We have to make sure we get the
2891 * first, for the page indexing below to work.
2893 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2894 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2897 * When coredumping, it suits get_dump_page if we just return
2898 * an error where there's an empty slot with no huge pagecache
2899 * to back it. This way, we avoid allocating a hugepage, and
2900 * the sparse dumpfile avoids allocating disk blocks, but its
2901 * huge holes still show up with zeroes where they need to be.
2903 if (absent && (flags & FOLL_DUMP) &&
2904 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2910 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2913 spin_unlock(&mm->page_table_lock);
2914 ret = hugetlb_fault(mm, vma, vaddr,
2915 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2916 spin_lock(&mm->page_table_lock);
2917 if (!(ret & VM_FAULT_ERROR))
2924 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2925 page = pte_page(huge_ptep_get(pte));
2928 pages[i] = mem_map_offset(page, pfn_offset);
2939 if (vaddr < vma->vm_end && remainder &&
2940 pfn_offset < pages_per_huge_page(h)) {
2942 * We use pfn_offset to avoid touching the pageframes
2943 * of this compound page.
2948 spin_unlock(&mm->page_table_lock);
2949 *length = remainder;
2952 return i ? i : -EFAULT;
2955 void hugetlb_change_protection(struct vm_area_struct *vma,
2956 unsigned long address, unsigned long end, pgprot_t newprot)
2958 struct mm_struct *mm = vma->vm_mm;
2959 unsigned long start = address;
2962 struct hstate *h = hstate_vma(vma);
2964 BUG_ON(address >= end);
2965 flush_cache_range(vma, address, end);
2967 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2968 spin_lock(&mm->page_table_lock);
2969 for (; address < end; address += huge_page_size(h)) {
2970 ptep = huge_pte_offset(mm, address);
2973 if (huge_pmd_unshare(mm, &address, ptep))
2975 if (!huge_pte_none(huge_ptep_get(ptep))) {
2976 pte = huge_ptep_get_and_clear(mm, address, ptep);
2977 pte = pte_mkhuge(pte_modify(pte, newprot));
2978 set_huge_pte_at(mm, address, ptep, pte);
2981 spin_unlock(&mm->page_table_lock);
2982 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2984 flush_tlb_range(vma, start, end);
2987 int hugetlb_reserve_pages(struct inode *inode,
2989 struct vm_area_struct *vma,
2990 vm_flags_t vm_flags)
2993 struct hstate *h = hstate_inode(inode);
2994 struct hugepage_subpool *spool = subpool_inode(inode);
2997 * Only apply hugepage reservation if asked. At fault time, an
2998 * attempt will be made for VM_NORESERVE to allocate a page
2999 * without using reserves
3001 if (vm_flags & VM_NORESERVE)
3005 * Shared mappings base their reservation on the number of pages that
3006 * are already allocated on behalf of the file. Private mappings need
3007 * to reserve the full area even if read-only as mprotect() may be
3008 * called to make the mapping read-write. Assume !vma is a shm mapping
3010 if (!vma || vma->vm_flags & VM_MAYSHARE)
3011 chg = region_chg(&inode->i_mapping->private_list, from, to);
3013 struct resv_map *resv_map = resv_map_alloc();
3019 set_vma_resv_map(vma, resv_map);
3020 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3028 /* There must be enough pages in the subpool for the mapping */
3029 if (hugepage_subpool_get_pages(spool, chg)) {
3035 * Check enough hugepages are available for the reservation.
3036 * Hand the pages back to the subpool if there are not
3038 ret = hugetlb_acct_memory(h, chg);
3040 hugepage_subpool_put_pages(spool, chg);
3045 * Account for the reservations made. Shared mappings record regions
3046 * that have reservations as they are shared by multiple VMAs.
3047 * When the last VMA disappears, the region map says how much
3048 * the reservation was and the page cache tells how much of
3049 * the reservation was consumed. Private mappings are per-VMA and
3050 * only the consumed reservations are tracked. When the VMA
3051 * disappears, the original reservation is the VMA size and the
3052 * consumed reservations are stored in the map. Hence, nothing
3053 * else has to be done for private mappings here
3055 if (!vma || vma->vm_flags & VM_MAYSHARE)
3056 region_add(&inode->i_mapping->private_list, from, to);
3064 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3066 struct hstate *h = hstate_inode(inode);
3067 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3068 struct hugepage_subpool *spool = subpool_inode(inode);
3070 spin_lock(&inode->i_lock);
3071 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3072 spin_unlock(&inode->i_lock);
3074 hugepage_subpool_put_pages(spool, (chg - freed));
3075 hugetlb_acct_memory(h, -(chg - freed));
3078 #ifdef CONFIG_MEMORY_FAILURE
3080 /* Should be called in hugetlb_lock */
3081 static int is_hugepage_on_freelist(struct page *hpage)
3085 struct hstate *h = page_hstate(hpage);
3086 int nid = page_to_nid(hpage);
3088 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3095 * This function is called from memory failure code.
3096 * Assume the caller holds page lock of the head page.
3098 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3100 struct hstate *h = page_hstate(hpage);
3101 int nid = page_to_nid(hpage);
3104 spin_lock(&hugetlb_lock);
3105 if (is_hugepage_on_freelist(hpage)) {
3106 list_del(&hpage->lru);
3107 set_page_refcounted(hpage);
3108 h->free_huge_pages--;
3109 h->free_huge_pages_node[nid]--;
3112 spin_unlock(&hugetlb_lock);