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 static int hugetlb_max_hstate;
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;
49 #define for_each_hstate(h) \
50 for ((h) = hstates; (h) < &hstates[hugetlb_max_hstate]; (h)++)
53 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
55 static DEFINE_SPINLOCK(hugetlb_lock);
57 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
59 bool free = (spool->count == 0) && (spool->used_hpages == 0);
61 spin_unlock(&spool->lock);
63 /* If no pages are used, and no other handles to the subpool
64 * remain, free the subpool the subpool remain */
69 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
71 struct hugepage_subpool *spool;
73 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
77 spin_lock_init(&spool->lock);
79 spool->max_hpages = nr_blocks;
80 spool->used_hpages = 0;
85 void hugepage_put_subpool(struct hugepage_subpool *spool)
87 spin_lock(&spool->lock);
88 BUG_ON(!spool->count);
90 unlock_or_release_subpool(spool);
93 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 spin_lock(&spool->lock);
102 if ((spool->used_hpages + delta) <= spool->max_hpages) {
103 spool->used_hpages += delta;
107 spin_unlock(&spool->lock);
112 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
118 spin_lock(&spool->lock);
119 spool->used_hpages -= delta;
120 /* If hugetlbfs_put_super couldn't free spool due to
121 * an outstanding quota reference, free it now. */
122 unlock_or_release_subpool(spool);
125 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
127 return HUGETLBFS_SB(inode->i_sb)->spool;
130 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
132 return subpool_inode(vma->vm_file->f_dentry->d_inode);
136 * Region tracking -- allows tracking of reservations and instantiated pages
137 * across the pages in a mapping.
139 * The region data structures are protected by a combination of the mmap_sem
140 * and the hugetlb_instantion_mutex. To access or modify a region the caller
141 * must either hold the mmap_sem for write, or the mmap_sem for read and
142 * the hugetlb_instantiation mutex:
144 * down_write(&mm->mmap_sem);
146 * down_read(&mm->mmap_sem);
147 * mutex_lock(&hugetlb_instantiation_mutex);
150 struct list_head link;
155 static long region_add(struct list_head *head, long f, long t)
157 struct file_region *rg, *nrg, *trg;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg, head, link)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
171 if (&rg->link == head)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct list_head *head, long f, long t)
193 struct file_region *rg, *nrg;
196 /* Locate the region we are before or in. */
197 list_for_each_entry(rg, head, link)
201 /* If we are below the current region then a new region is required.
202 * Subtle, allocate a new region at the position but make it zero
203 * size such that we can guarantee to record the reservation. */
204 if (&rg->link == head || t < rg->from) {
205 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
210 INIT_LIST_HEAD(&nrg->link);
211 list_add(&nrg->link, rg->link.prev);
216 /* Round our left edge to the current segment if it encloses us. */
221 /* Check for and consume any regions we now overlap with. */
222 list_for_each_entry(rg, rg->link.prev, link) {
223 if (&rg->link == head)
228 /* We overlap with this area, if it extends further than
229 * us then we must extend ourselves. Account for its
230 * existing reservation. */
235 chg -= rg->to - rg->from;
240 static long region_truncate(struct list_head *head, long end)
242 struct file_region *rg, *trg;
245 /* Locate the region we are either in or before. */
246 list_for_each_entry(rg, head, link)
249 if (&rg->link == head)
252 /* If we are in the middle of a region then adjust it. */
253 if (end > rg->from) {
256 rg = list_entry(rg->link.next, typeof(*rg), link);
259 /* Drop any remaining regions. */
260 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
261 if (&rg->link == head)
263 chg += rg->to - rg->from;
270 static long region_count(struct list_head *head, long f, long t)
272 struct file_region *rg;
275 /* Locate each segment we overlap with, and count that overlap. */
276 list_for_each_entry(rg, head, link) {
285 seg_from = max(rg->from, f);
286 seg_to = min(rg->to, t);
288 chg += seg_to - seg_from;
295 * Convert the address within this vma to the page offset within
296 * the mapping, in pagecache page units; huge pages here.
298 static pgoff_t vma_hugecache_offset(struct hstate *h,
299 struct vm_area_struct *vma, unsigned long address)
301 return ((address - vma->vm_start) >> huge_page_shift(h)) +
302 (vma->vm_pgoff >> huge_page_order(h));
305 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
306 unsigned long address)
308 return vma_hugecache_offset(hstate_vma(vma), vma, address);
312 * Return the size of the pages allocated when backing a VMA. In the majority
313 * cases this will be same size as used by the page table entries.
315 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
317 struct hstate *hstate;
319 if (!is_vm_hugetlb_page(vma))
322 hstate = hstate_vma(vma);
324 return 1UL << (hstate->order + PAGE_SHIFT);
326 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
329 * Return the page size being used by the MMU to back a VMA. In the majority
330 * of cases, the page size used by the kernel matches the MMU size. On
331 * architectures where it differs, an architecture-specific version of this
332 * function is required.
334 #ifndef vma_mmu_pagesize
335 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
337 return vma_kernel_pagesize(vma);
342 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
343 * bits of the reservation map pointer, which are always clear due to
346 #define HPAGE_RESV_OWNER (1UL << 0)
347 #define HPAGE_RESV_UNMAPPED (1UL << 1)
348 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
351 * These helpers are used to track how many pages are reserved for
352 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
353 * is guaranteed to have their future faults succeed.
355 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
356 * the reserve counters are updated with the hugetlb_lock held. It is safe
357 * to reset the VMA at fork() time as it is not in use yet and there is no
358 * chance of the global counters getting corrupted as a result of the values.
360 * The private mapping reservation is represented in a subtly different
361 * manner to a shared mapping. A shared mapping has a region map associated
362 * with the underlying file, this region map represents the backing file
363 * pages which have ever had a reservation assigned which this persists even
364 * after the page is instantiated. A private mapping has a region map
365 * associated with the original mmap which is attached to all VMAs which
366 * reference it, this region map represents those offsets which have consumed
367 * reservation ie. where pages have been instantiated.
369 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
371 return (unsigned long)vma->vm_private_data;
374 static void set_vma_private_data(struct vm_area_struct *vma,
377 vma->vm_private_data = (void *)value;
382 struct list_head regions;
385 static struct resv_map *resv_map_alloc(void)
387 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
391 kref_init(&resv_map->refs);
392 INIT_LIST_HEAD(&resv_map->regions);
397 static void resv_map_release(struct kref *ref)
399 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
401 /* Clear out any active regions before we release the map. */
402 region_truncate(&resv_map->regions, 0);
406 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
408 VM_BUG_ON(!is_vm_hugetlb_page(vma));
409 if (!(vma->vm_flags & VM_MAYSHARE))
410 return (struct resv_map *)(get_vma_private_data(vma) &
415 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
417 VM_BUG_ON(!is_vm_hugetlb_page(vma));
418 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
420 set_vma_private_data(vma, (get_vma_private_data(vma) &
421 HPAGE_RESV_MASK) | (unsigned long)map);
424 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
426 VM_BUG_ON(!is_vm_hugetlb_page(vma));
427 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
429 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
432 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
434 VM_BUG_ON(!is_vm_hugetlb_page(vma));
436 return (get_vma_private_data(vma) & flag) != 0;
439 /* Decrement the reserved pages in the hugepage pool by one */
440 static void decrement_hugepage_resv_vma(struct hstate *h,
441 struct vm_area_struct *vma)
443 if (vma->vm_flags & VM_NORESERVE)
446 if (vma->vm_flags & VM_MAYSHARE) {
447 /* Shared mappings always use reserves */
448 h->resv_huge_pages--;
449 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
451 * Only the process that called mmap() has reserves for
454 h->resv_huge_pages--;
458 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
459 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
462 if (!(vma->vm_flags & VM_MAYSHARE))
463 vma->vm_private_data = (void *)0;
466 /* Returns true if the VMA has associated reserve pages */
467 static int vma_has_reserves(struct vm_area_struct *vma)
469 if (vma->vm_flags & VM_MAYSHARE)
471 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
476 static void copy_gigantic_page(struct page *dst, struct page *src)
479 struct hstate *h = page_hstate(src);
480 struct page *dst_base = dst;
481 struct page *src_base = src;
483 for (i = 0; i < pages_per_huge_page(h); ) {
485 copy_highpage(dst, src);
488 dst = mem_map_next(dst, dst_base, i);
489 src = mem_map_next(src, src_base, i);
493 void copy_huge_page(struct page *dst, struct page *src)
496 struct hstate *h = page_hstate(src);
498 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
499 copy_gigantic_page(dst, src);
504 for (i = 0; i < pages_per_huge_page(h); i++) {
506 copy_highpage(dst + i, src + i);
510 static void enqueue_huge_page(struct hstate *h, struct page *page)
512 int nid = page_to_nid(page);
513 list_add(&page->lru, &h->hugepage_freelists[nid]);
514 h->free_huge_pages++;
515 h->free_huge_pages_node[nid]++;
518 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 if (list_empty(&h->hugepage_freelists[nid]))
524 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
525 list_del(&page->lru);
526 set_page_refcounted(page);
527 h->free_huge_pages--;
528 h->free_huge_pages_node[nid]--;
532 static struct page *dequeue_huge_page_vma(struct hstate *h,
533 struct vm_area_struct *vma,
534 unsigned long address, int avoid_reserve)
536 struct page *page = NULL;
537 struct mempolicy *mpol;
538 nodemask_t *nodemask;
539 struct zonelist *zonelist;
542 unsigned int cpuset_mems_cookie;
545 cpuset_mems_cookie = get_mems_allowed();
546 zonelist = huge_zonelist(vma, address,
547 htlb_alloc_mask, &mpol, &nodemask);
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma) &&
554 h->free_huge_pages - h->resv_huge_pages == 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 for_each_zone_zonelist_nodemask(zone, z, zonelist,
562 MAX_NR_ZONES - 1, nodemask) {
563 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
564 page = dequeue_huge_page_node(h, zone_to_nid(zone));
567 decrement_hugepage_resv_vma(h, vma);
574 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
583 static void update_and_free_page(struct hstate *h, struct page *page)
587 VM_BUG_ON(h->order >= MAX_ORDER);
590 h->nr_huge_pages_node[page_to_nid(page)]--;
591 for (i = 0; i < pages_per_huge_page(h); i++) {
592 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
593 1 << PG_referenced | 1 << PG_dirty |
594 1 << PG_active | 1 << PG_reserved |
595 1 << PG_private | 1 << PG_writeback);
597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
599 arch_release_hugepage(page);
600 __free_pages(page, huge_page_order(h));
603 struct hstate *size_to_hstate(unsigned long size)
608 if (huge_page_size(h) == size)
614 static void free_huge_page(struct page *page)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate *h = page_hstate(page);
621 int nid = page_to_nid(page);
622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
625 set_page_private(page, 0);
626 page->mapping = NULL;
627 BUG_ON(page_count(page));
628 BUG_ON(page_mapcount(page));
629 INIT_LIST_HEAD(&page->lru);
631 spin_lock(&hugetlb_lock);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 update_and_free_page(h, page);
634 h->surplus_huge_pages--;
635 h->surplus_huge_pages_node[nid]--;
637 enqueue_huge_page(h, page);
639 spin_unlock(&hugetlb_lock);
640 hugepage_subpool_put_pages(spool, 1);
643 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
645 set_compound_page_dtor(page, free_huge_page);
646 spin_lock(&hugetlb_lock);
648 h->nr_huge_pages_node[nid]++;
649 spin_unlock(&hugetlb_lock);
650 put_page(page); /* free it into the hugepage allocator */
653 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
656 int nr_pages = 1 << order;
657 struct page *p = page + 1;
659 /* we rely on prep_new_huge_page to set the destructor */
660 set_compound_order(page, order);
662 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
664 set_page_count(p, 0);
665 p->first_page = page;
669 int PageHuge(struct page *page)
671 compound_page_dtor *dtor;
673 if (!PageCompound(page))
676 page = compound_head(page);
677 dtor = get_compound_page_dtor(page);
679 return dtor == free_huge_page;
681 EXPORT_SYMBOL_GPL(PageHuge);
683 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
687 if (h->order >= MAX_ORDER)
690 page = alloc_pages_exact_node(nid,
691 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
692 __GFP_REPEAT|__GFP_NOWARN,
695 if (arch_prepare_hugepage(page)) {
696 __free_pages(page, huge_page_order(h));
699 prep_new_huge_page(h, page, nid);
706 * common helper functions for hstate_next_node_to_{alloc|free}.
707 * We may have allocated or freed a huge page based on a different
708 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
709 * be outside of *nodes_allowed. Ensure that we use an allowed
710 * node for alloc or free.
712 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
714 nid = next_node(nid, *nodes_allowed);
715 if (nid == MAX_NUMNODES)
716 nid = first_node(*nodes_allowed);
717 VM_BUG_ON(nid >= MAX_NUMNODES);
722 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
724 if (!node_isset(nid, *nodes_allowed))
725 nid = next_node_allowed(nid, nodes_allowed);
730 * returns the previously saved node ["this node"] from which to
731 * allocate a persistent huge page for the pool and advance the
732 * next node from which to allocate, handling wrap at end of node
735 static int hstate_next_node_to_alloc(struct hstate *h,
736 nodemask_t *nodes_allowed)
740 VM_BUG_ON(!nodes_allowed);
742 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
743 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
748 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
755 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
756 next_nid = start_nid;
759 page = alloc_fresh_huge_page_node(h, next_nid);
764 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
765 } while (next_nid != start_nid);
768 count_vm_event(HTLB_BUDDY_PGALLOC);
770 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
776 * helper for free_pool_huge_page() - return the previously saved
777 * node ["this node"] from which to free a huge page. Advance the
778 * next node id whether or not we find a free huge page to free so
779 * that the next attempt to free addresses the next node.
781 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
785 VM_BUG_ON(!nodes_allowed);
787 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
788 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
794 * Free huge page from pool from next node to free.
795 * Attempt to keep persistent huge pages more or less
796 * balanced over allowed nodes.
797 * Called with hugetlb_lock locked.
799 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
806 start_nid = hstate_next_node_to_free(h, nodes_allowed);
807 next_nid = start_nid;
811 * If we're returning unused surplus pages, only examine
812 * nodes with surplus pages.
814 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
815 !list_empty(&h->hugepage_freelists[next_nid])) {
817 list_entry(h->hugepage_freelists[next_nid].next,
819 list_del(&page->lru);
820 h->free_huge_pages--;
821 h->free_huge_pages_node[next_nid]--;
823 h->surplus_huge_pages--;
824 h->surplus_huge_pages_node[next_nid]--;
826 update_and_free_page(h, page);
830 next_nid = hstate_next_node_to_free(h, nodes_allowed);
831 } while (next_nid != start_nid);
836 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
841 if (h->order >= MAX_ORDER)
845 * Assume we will successfully allocate the surplus page to
846 * prevent racing processes from causing the surplus to exceed
849 * This however introduces a different race, where a process B
850 * tries to grow the static hugepage pool while alloc_pages() is
851 * called by process A. B will only examine the per-node
852 * counters in determining if surplus huge pages can be
853 * converted to normal huge pages in adjust_pool_surplus(). A
854 * won't be able to increment the per-node counter, until the
855 * lock is dropped by B, but B doesn't drop hugetlb_lock until
856 * no more huge pages can be converted from surplus to normal
857 * state (and doesn't try to convert again). Thus, we have a
858 * case where a surplus huge page exists, the pool is grown, and
859 * the surplus huge page still exists after, even though it
860 * should just have been converted to a normal huge page. This
861 * does not leak memory, though, as the hugepage will be freed
862 * once it is out of use. It also does not allow the counters to
863 * go out of whack in adjust_pool_surplus() as we don't modify
864 * the node values until we've gotten the hugepage and only the
865 * per-node value is checked there.
867 spin_lock(&hugetlb_lock);
868 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
869 spin_unlock(&hugetlb_lock);
873 h->surplus_huge_pages++;
875 spin_unlock(&hugetlb_lock);
877 if (nid == NUMA_NO_NODE)
878 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
879 __GFP_REPEAT|__GFP_NOWARN,
882 page = alloc_pages_exact_node(nid,
883 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
884 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
886 if (page && arch_prepare_hugepage(page)) {
887 __free_pages(page, huge_page_order(h));
891 spin_lock(&hugetlb_lock);
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) {
997 list_del(&page->lru);
999 * This page is now managed by the hugetlb allocator and has
1000 * no users -- drop the buddy allocator's reference.
1002 put_page_testzero(page);
1003 VM_BUG_ON(page_count(page));
1004 enqueue_huge_page(h, page);
1007 spin_unlock(&hugetlb_lock);
1009 /* Free unnecessary surplus pages to the buddy allocator */
1010 if (!list_empty(&surplus_list)) {
1011 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1012 list_del(&page->lru);
1016 spin_lock(&hugetlb_lock);
1022 * When releasing a hugetlb pool reservation, any surplus pages that were
1023 * allocated to satisfy the reservation must be explicitly freed if they were
1025 * Called with hugetlb_lock held.
1027 static void return_unused_surplus_pages(struct hstate *h,
1028 unsigned long unused_resv_pages)
1030 unsigned long nr_pages;
1032 /* Uncommit the reservation */
1033 h->resv_huge_pages -= unused_resv_pages;
1035 /* Cannot return gigantic pages currently */
1036 if (h->order >= MAX_ORDER)
1039 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1042 * We want to release as many surplus pages as possible, spread
1043 * evenly across all nodes with memory. Iterate across these nodes
1044 * until we can no longer free unreserved surplus pages. This occurs
1045 * when the nodes with surplus pages have no free pages.
1046 * free_pool_huge_page() will balance the the freed pages across the
1047 * on-line nodes with memory and will handle the hstate accounting.
1049 while (nr_pages--) {
1050 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1056 * Determine if the huge page at addr within the vma has an associated
1057 * reservation. Where it does not we will need to logically increase
1058 * reservation and actually increase subpool usage before an allocation
1059 * can occur. Where any new reservation would be required the
1060 * reservation change is prepared, but not committed. Once the page
1061 * has been allocated from the subpool and instantiated the change should
1062 * be committed via vma_commit_reservation. No action is required on
1065 static long vma_needs_reservation(struct hstate *h,
1066 struct vm_area_struct *vma, unsigned long addr)
1068 struct address_space *mapping = vma->vm_file->f_mapping;
1069 struct inode *inode = mapping->host;
1071 if (vma->vm_flags & VM_MAYSHARE) {
1072 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1073 return region_chg(&inode->i_mapping->private_list,
1076 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1081 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1082 struct resv_map *reservations = vma_resv_map(vma);
1084 err = region_chg(&reservations->regions, idx, idx + 1);
1090 static void vma_commit_reservation(struct hstate *h,
1091 struct vm_area_struct *vma, unsigned long addr)
1093 struct address_space *mapping = vma->vm_file->f_mapping;
1094 struct inode *inode = mapping->host;
1096 if (vma->vm_flags & VM_MAYSHARE) {
1097 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1098 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1100 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1101 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1102 struct resv_map *reservations = vma_resv_map(vma);
1104 /* Mark this page used in the map. */
1105 region_add(&reservations->regions, idx, idx + 1);
1109 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1110 unsigned long addr, int avoid_reserve)
1112 struct hugepage_subpool *spool = subpool_vma(vma);
1113 struct hstate *h = hstate_vma(vma);
1118 * Processes that did not create the mapping will have no
1119 * reserves and will not have accounted against subpool
1120 * limit. Check that the subpool limit can be made before
1121 * satisfying the allocation MAP_NORESERVE mappings may also
1122 * need pages and subpool limit allocated allocated if no reserve
1125 chg = vma_needs_reservation(h, vma, addr);
1127 return ERR_PTR(-ENOMEM);
1129 if (hugepage_subpool_get_pages(spool, chg))
1130 return ERR_PTR(-ENOSPC);
1132 spin_lock(&hugetlb_lock);
1133 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1134 spin_unlock(&hugetlb_lock);
1137 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1139 hugepage_subpool_put_pages(spool, chg);
1140 return ERR_PTR(-ENOSPC);
1144 set_page_private(page, (unsigned long)spool);
1146 vma_commit_reservation(h, vma, addr);
1151 int __weak alloc_bootmem_huge_page(struct hstate *h)
1153 struct huge_bootmem_page *m;
1154 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1159 addr = __alloc_bootmem_node_nopanic(
1160 NODE_DATA(hstate_next_node_to_alloc(h,
1161 &node_states[N_HIGH_MEMORY])),
1162 huge_page_size(h), huge_page_size(h), 0);
1166 * Use the beginning of the huge page to store the
1167 * huge_bootmem_page struct (until gather_bootmem
1168 * puts them into the mem_map).
1178 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1179 /* Put them into a private list first because mem_map is not up yet */
1180 list_add(&m->list, &huge_boot_pages);
1185 static void prep_compound_huge_page(struct page *page, int order)
1187 if (unlikely(order > (MAX_ORDER - 1)))
1188 prep_compound_gigantic_page(page, order);
1190 prep_compound_page(page, order);
1193 /* Put bootmem huge pages into the standard lists after mem_map is up */
1194 static void __init gather_bootmem_prealloc(void)
1196 struct huge_bootmem_page *m;
1198 list_for_each_entry(m, &huge_boot_pages, list) {
1199 struct hstate *h = m->hstate;
1202 #ifdef CONFIG_HIGHMEM
1203 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1204 free_bootmem_late((unsigned long)m,
1205 sizeof(struct huge_bootmem_page));
1207 page = virt_to_page(m);
1209 __ClearPageReserved(page);
1210 WARN_ON(page_count(page) != 1);
1211 prep_compound_huge_page(page, h->order);
1212 prep_new_huge_page(h, page, page_to_nid(page));
1214 * If we had gigantic hugepages allocated at boot time, we need
1215 * to restore the 'stolen' pages to totalram_pages in order to
1216 * fix confusing memory reports from free(1) and another
1217 * side-effects, like CommitLimit going negative.
1219 if (h->order > (MAX_ORDER - 1))
1220 totalram_pages += 1 << h->order;
1224 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1228 for (i = 0; i < h->max_huge_pages; ++i) {
1229 if (h->order >= MAX_ORDER) {
1230 if (!alloc_bootmem_huge_page(h))
1232 } else if (!alloc_fresh_huge_page(h,
1233 &node_states[N_HIGH_MEMORY]))
1236 h->max_huge_pages = i;
1239 static void __init hugetlb_init_hstates(void)
1243 for_each_hstate(h) {
1244 /* oversize hugepages were init'ed in early boot */
1245 if (h->order < MAX_ORDER)
1246 hugetlb_hstate_alloc_pages(h);
1250 static char * __init memfmt(char *buf, unsigned long n)
1252 if (n >= (1UL << 30))
1253 sprintf(buf, "%lu GB", n >> 30);
1254 else if (n >= (1UL << 20))
1255 sprintf(buf, "%lu MB", n >> 20);
1257 sprintf(buf, "%lu KB", n >> 10);
1261 static void __init report_hugepages(void)
1265 for_each_hstate(h) {
1267 printk(KERN_INFO "HugeTLB registered %s page size, "
1268 "pre-allocated %ld pages\n",
1269 memfmt(buf, huge_page_size(h)),
1270 h->free_huge_pages);
1274 #ifdef CONFIG_HIGHMEM
1275 static void try_to_free_low(struct hstate *h, unsigned long count,
1276 nodemask_t *nodes_allowed)
1280 if (h->order >= MAX_ORDER)
1283 for_each_node_mask(i, *nodes_allowed) {
1284 struct page *page, *next;
1285 struct list_head *freel = &h->hugepage_freelists[i];
1286 list_for_each_entry_safe(page, next, freel, lru) {
1287 if (count >= h->nr_huge_pages)
1289 if (PageHighMem(page))
1291 list_del(&page->lru);
1292 update_and_free_page(h, page);
1293 h->free_huge_pages--;
1294 h->free_huge_pages_node[page_to_nid(page)]--;
1299 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1300 nodemask_t *nodes_allowed)
1306 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1307 * balanced by operating on them in a round-robin fashion.
1308 * Returns 1 if an adjustment was made.
1310 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1313 int start_nid, next_nid;
1316 VM_BUG_ON(delta != -1 && delta != 1);
1319 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1321 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1322 next_nid = start_nid;
1328 * To shrink on this node, there must be a surplus page
1330 if (!h->surplus_huge_pages_node[nid]) {
1331 next_nid = hstate_next_node_to_alloc(h,
1338 * Surplus cannot exceed the total number of pages
1340 if (h->surplus_huge_pages_node[nid] >=
1341 h->nr_huge_pages_node[nid]) {
1342 next_nid = hstate_next_node_to_free(h,
1348 h->surplus_huge_pages += delta;
1349 h->surplus_huge_pages_node[nid] += delta;
1352 } while (next_nid != start_nid);
1357 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1358 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1359 nodemask_t *nodes_allowed)
1361 unsigned long min_count, ret;
1363 if (h->order >= MAX_ORDER)
1364 return h->max_huge_pages;
1367 * Increase the pool size
1368 * First take pages out of surplus state. Then make up the
1369 * remaining difference by allocating fresh huge pages.
1371 * We might race with alloc_buddy_huge_page() here and be unable
1372 * to convert a surplus huge page to a normal huge page. That is
1373 * not critical, though, it just means the overall size of the
1374 * pool might be one hugepage larger than it needs to be, but
1375 * within all the constraints specified by the sysctls.
1377 spin_lock(&hugetlb_lock);
1378 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1379 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1383 while (count > persistent_huge_pages(h)) {
1385 * If this allocation races such that we no longer need the
1386 * page, free_huge_page will handle it by freeing the page
1387 * and reducing the surplus.
1389 spin_unlock(&hugetlb_lock);
1390 ret = alloc_fresh_huge_page(h, nodes_allowed);
1391 spin_lock(&hugetlb_lock);
1395 /* Bail for signals. Probably ctrl-c from user */
1396 if (signal_pending(current))
1401 * Decrease the pool size
1402 * First return free pages to the buddy allocator (being careful
1403 * to keep enough around to satisfy reservations). Then place
1404 * pages into surplus state as needed so the pool will shrink
1405 * to the desired size as pages become free.
1407 * By placing pages into the surplus state independent of the
1408 * overcommit value, we are allowing the surplus pool size to
1409 * exceed overcommit. There are few sane options here. Since
1410 * alloc_buddy_huge_page() is checking the global counter,
1411 * though, we'll note that we're not allowed to exceed surplus
1412 * and won't grow the pool anywhere else. Not until one of the
1413 * sysctls are changed, or the surplus pages go out of use.
1415 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1416 min_count = max(count, min_count);
1417 try_to_free_low(h, min_count, nodes_allowed);
1418 while (min_count < persistent_huge_pages(h)) {
1419 if (!free_pool_huge_page(h, nodes_allowed, 0))
1422 while (count < persistent_huge_pages(h)) {
1423 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1427 ret = persistent_huge_pages(h);
1428 spin_unlock(&hugetlb_lock);
1432 #define HSTATE_ATTR_RO(_name) \
1433 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1435 #define HSTATE_ATTR(_name) \
1436 static struct kobj_attribute _name##_attr = \
1437 __ATTR(_name, 0644, _name##_show, _name##_store)
1439 static struct kobject *hugepages_kobj;
1440 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1442 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1444 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1448 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1449 if (hstate_kobjs[i] == kobj) {
1451 *nidp = NUMA_NO_NODE;
1455 return kobj_to_node_hstate(kobj, nidp);
1458 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1459 struct kobj_attribute *attr, char *buf)
1462 unsigned long nr_huge_pages;
1465 h = kobj_to_hstate(kobj, &nid);
1466 if (nid == NUMA_NO_NODE)
1467 nr_huge_pages = h->nr_huge_pages;
1469 nr_huge_pages = h->nr_huge_pages_node[nid];
1471 return sprintf(buf, "%lu\n", nr_huge_pages);
1474 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1475 struct kobject *kobj, struct kobj_attribute *attr,
1476 const char *buf, size_t len)
1480 unsigned long count;
1482 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1484 err = strict_strtoul(buf, 10, &count);
1488 h = kobj_to_hstate(kobj, &nid);
1489 if (h->order >= MAX_ORDER) {
1494 if (nid == NUMA_NO_NODE) {
1496 * global hstate attribute
1498 if (!(obey_mempolicy &&
1499 init_nodemask_of_mempolicy(nodes_allowed))) {
1500 NODEMASK_FREE(nodes_allowed);
1501 nodes_allowed = &node_states[N_HIGH_MEMORY];
1503 } else if (nodes_allowed) {
1505 * per node hstate attribute: adjust count to global,
1506 * but restrict alloc/free to the specified node.
1508 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1509 init_nodemask_of_node(nodes_allowed, nid);
1511 nodes_allowed = &node_states[N_HIGH_MEMORY];
1513 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1515 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1516 NODEMASK_FREE(nodes_allowed);
1520 NODEMASK_FREE(nodes_allowed);
1524 static ssize_t nr_hugepages_show(struct kobject *kobj,
1525 struct kobj_attribute *attr, char *buf)
1527 return nr_hugepages_show_common(kobj, attr, buf);
1530 static ssize_t nr_hugepages_store(struct kobject *kobj,
1531 struct kobj_attribute *attr, const char *buf, size_t len)
1533 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1535 HSTATE_ATTR(nr_hugepages);
1540 * hstate attribute for optionally mempolicy-based constraint on persistent
1541 * huge page alloc/free.
1543 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1544 struct kobj_attribute *attr, char *buf)
1546 return nr_hugepages_show_common(kobj, attr, buf);
1549 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1550 struct kobj_attribute *attr, const char *buf, size_t len)
1552 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1554 HSTATE_ATTR(nr_hugepages_mempolicy);
1558 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1559 struct kobj_attribute *attr, char *buf)
1561 struct hstate *h = kobj_to_hstate(kobj, NULL);
1562 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1565 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1566 struct kobj_attribute *attr, const char *buf, size_t count)
1569 unsigned long input;
1570 struct hstate *h = kobj_to_hstate(kobj, NULL);
1572 if (h->order >= MAX_ORDER)
1575 err = strict_strtoul(buf, 10, &input);
1579 spin_lock(&hugetlb_lock);
1580 h->nr_overcommit_huge_pages = input;
1581 spin_unlock(&hugetlb_lock);
1585 HSTATE_ATTR(nr_overcommit_hugepages);
1587 static ssize_t free_hugepages_show(struct kobject *kobj,
1588 struct kobj_attribute *attr, char *buf)
1591 unsigned long free_huge_pages;
1594 h = kobj_to_hstate(kobj, &nid);
1595 if (nid == NUMA_NO_NODE)
1596 free_huge_pages = h->free_huge_pages;
1598 free_huge_pages = h->free_huge_pages_node[nid];
1600 return sprintf(buf, "%lu\n", free_huge_pages);
1602 HSTATE_ATTR_RO(free_hugepages);
1604 static ssize_t resv_hugepages_show(struct kobject *kobj,
1605 struct kobj_attribute *attr, char *buf)
1607 struct hstate *h = kobj_to_hstate(kobj, NULL);
1608 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1610 HSTATE_ATTR_RO(resv_hugepages);
1612 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1613 struct kobj_attribute *attr, char *buf)
1616 unsigned long surplus_huge_pages;
1619 h = kobj_to_hstate(kobj, &nid);
1620 if (nid == NUMA_NO_NODE)
1621 surplus_huge_pages = h->surplus_huge_pages;
1623 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1625 return sprintf(buf, "%lu\n", surplus_huge_pages);
1627 HSTATE_ATTR_RO(surplus_hugepages);
1629 static struct attribute *hstate_attrs[] = {
1630 &nr_hugepages_attr.attr,
1631 &nr_overcommit_hugepages_attr.attr,
1632 &free_hugepages_attr.attr,
1633 &resv_hugepages_attr.attr,
1634 &surplus_hugepages_attr.attr,
1636 &nr_hugepages_mempolicy_attr.attr,
1641 static struct attribute_group hstate_attr_group = {
1642 .attrs = hstate_attrs,
1645 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1646 struct kobject **hstate_kobjs,
1647 struct attribute_group *hstate_attr_group)
1650 int hi = hstate_index(h);
1652 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1653 if (!hstate_kobjs[hi])
1656 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1658 kobject_put(hstate_kobjs[hi]);
1663 static void __init hugetlb_sysfs_init(void)
1668 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1669 if (!hugepages_kobj)
1672 for_each_hstate(h) {
1673 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1674 hstate_kobjs, &hstate_attr_group);
1676 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1684 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1685 * with node devices in node_devices[] using a parallel array. The array
1686 * index of a node device or _hstate == node id.
1687 * This is here to avoid any static dependency of the node device driver, in
1688 * the base kernel, on the hugetlb module.
1690 struct node_hstate {
1691 struct kobject *hugepages_kobj;
1692 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1694 struct node_hstate node_hstates[MAX_NUMNODES];
1697 * A subset of global hstate attributes for node devices
1699 static struct attribute *per_node_hstate_attrs[] = {
1700 &nr_hugepages_attr.attr,
1701 &free_hugepages_attr.attr,
1702 &surplus_hugepages_attr.attr,
1706 static struct attribute_group per_node_hstate_attr_group = {
1707 .attrs = per_node_hstate_attrs,
1711 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1712 * Returns node id via non-NULL nidp.
1714 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1718 for (nid = 0; nid < nr_node_ids; nid++) {
1719 struct node_hstate *nhs = &node_hstates[nid];
1721 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1722 if (nhs->hstate_kobjs[i] == kobj) {
1734 * Unregister hstate attributes from a single node device.
1735 * No-op if no hstate attributes attached.
1737 void hugetlb_unregister_node(struct node *node)
1740 struct node_hstate *nhs = &node_hstates[node->dev.id];
1742 if (!nhs->hugepages_kobj)
1743 return; /* no hstate attributes */
1745 for_each_hstate(h) {
1746 int idx = hstate_index(h);
1747 if (nhs->hstate_kobjs[idx]) {
1748 kobject_put(nhs->hstate_kobjs[idx]);
1749 nhs->hstate_kobjs[idx] = NULL;
1753 kobject_put(nhs->hugepages_kobj);
1754 nhs->hugepages_kobj = NULL;
1758 * hugetlb module exit: unregister hstate attributes from node devices
1761 static void hugetlb_unregister_all_nodes(void)
1766 * disable node device registrations.
1768 register_hugetlbfs_with_node(NULL, NULL);
1771 * remove hstate attributes from any nodes that have them.
1773 for (nid = 0; nid < nr_node_ids; nid++)
1774 hugetlb_unregister_node(&node_devices[nid]);
1778 * Register hstate attributes for a single node device.
1779 * No-op if attributes already registered.
1781 void hugetlb_register_node(struct node *node)
1784 struct node_hstate *nhs = &node_hstates[node->dev.id];
1787 if (nhs->hugepages_kobj)
1788 return; /* already allocated */
1790 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1792 if (!nhs->hugepages_kobj)
1795 for_each_hstate(h) {
1796 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1798 &per_node_hstate_attr_group);
1800 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1802 h->name, node->dev.id);
1803 hugetlb_unregister_node(node);
1810 * hugetlb init time: register hstate attributes for all registered node
1811 * devices of nodes that have memory. All on-line nodes should have
1812 * registered their associated device by this time.
1814 static void hugetlb_register_all_nodes(void)
1818 for_each_node_state(nid, N_HIGH_MEMORY) {
1819 struct node *node = &node_devices[nid];
1820 if (node->dev.id == nid)
1821 hugetlb_register_node(node);
1825 * Let the node device driver know we're here so it can
1826 * [un]register hstate attributes on node hotplug.
1828 register_hugetlbfs_with_node(hugetlb_register_node,
1829 hugetlb_unregister_node);
1831 #else /* !CONFIG_NUMA */
1833 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1841 static void hugetlb_unregister_all_nodes(void) { }
1843 static void hugetlb_register_all_nodes(void) { }
1847 static void __exit hugetlb_exit(void)
1851 hugetlb_unregister_all_nodes();
1853 for_each_hstate(h) {
1854 kobject_put(hstate_kobjs[hstate_index(h)]);
1857 kobject_put(hugepages_kobj);
1859 module_exit(hugetlb_exit);
1861 static int __init hugetlb_init(void)
1863 /* Some platform decide whether they support huge pages at boot
1864 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1865 * there is no such support
1867 if (HPAGE_SHIFT == 0)
1870 if (!size_to_hstate(default_hstate_size)) {
1871 default_hstate_size = HPAGE_SIZE;
1872 if (!size_to_hstate(default_hstate_size))
1873 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1875 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1876 if (default_hstate_max_huge_pages)
1877 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1879 hugetlb_init_hstates();
1881 gather_bootmem_prealloc();
1885 hugetlb_sysfs_init();
1887 hugetlb_register_all_nodes();
1891 module_init(hugetlb_init);
1893 /* Should be called on processing a hugepagesz=... option */
1894 void __init hugetlb_add_hstate(unsigned order)
1899 if (size_to_hstate(PAGE_SIZE << order)) {
1900 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1903 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1905 h = &hstates[hugetlb_max_hstate++];
1907 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1908 h->nr_huge_pages = 0;
1909 h->free_huge_pages = 0;
1910 for (i = 0; i < MAX_NUMNODES; ++i)
1911 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1912 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1913 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1914 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1915 huge_page_size(h)/1024);
1920 static int __init hugetlb_nrpages_setup(char *s)
1923 static unsigned long *last_mhp;
1926 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1927 * so this hugepages= parameter goes to the "default hstate".
1929 if (!hugetlb_max_hstate)
1930 mhp = &default_hstate_max_huge_pages;
1932 mhp = &parsed_hstate->max_huge_pages;
1934 if (mhp == last_mhp) {
1935 printk(KERN_WARNING "hugepages= specified twice without "
1936 "interleaving hugepagesz=, ignoring\n");
1940 if (sscanf(s, "%lu", mhp) <= 0)
1944 * Global state is always initialized later in hugetlb_init.
1945 * But we need to allocate >= MAX_ORDER hstates here early to still
1946 * use the bootmem allocator.
1948 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1949 hugetlb_hstate_alloc_pages(parsed_hstate);
1955 __setup("hugepages=", hugetlb_nrpages_setup);
1957 static int __init hugetlb_default_setup(char *s)
1959 default_hstate_size = memparse(s, &s);
1962 __setup("default_hugepagesz=", hugetlb_default_setup);
1964 static unsigned int cpuset_mems_nr(unsigned int *array)
1967 unsigned int nr = 0;
1969 for_each_node_mask(node, cpuset_current_mems_allowed)
1975 #ifdef CONFIG_SYSCTL
1976 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1977 struct ctl_table *table, int write,
1978 void __user *buffer, size_t *length, loff_t *ppos)
1980 struct hstate *h = &default_hstate;
1984 tmp = h->max_huge_pages;
1986 if (write && h->order >= MAX_ORDER)
1990 table->maxlen = sizeof(unsigned long);
1991 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1996 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1997 GFP_KERNEL | __GFP_NORETRY);
1998 if (!(obey_mempolicy &&
1999 init_nodemask_of_mempolicy(nodes_allowed))) {
2000 NODEMASK_FREE(nodes_allowed);
2001 nodes_allowed = &node_states[N_HIGH_MEMORY];
2003 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2005 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2006 NODEMASK_FREE(nodes_allowed);
2012 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2013 void __user *buffer, size_t *length, loff_t *ppos)
2016 return hugetlb_sysctl_handler_common(false, table, write,
2017 buffer, length, ppos);
2021 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2022 void __user *buffer, size_t *length, loff_t *ppos)
2024 return hugetlb_sysctl_handler_common(true, table, write,
2025 buffer, length, ppos);
2027 #endif /* CONFIG_NUMA */
2029 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2030 void __user *buffer,
2031 size_t *length, loff_t *ppos)
2033 proc_dointvec(table, write, buffer, length, ppos);
2034 if (hugepages_treat_as_movable)
2035 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2037 htlb_alloc_mask = GFP_HIGHUSER;
2041 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2042 void __user *buffer,
2043 size_t *length, loff_t *ppos)
2045 struct hstate *h = &default_hstate;
2049 tmp = h->nr_overcommit_huge_pages;
2051 if (write && h->order >= MAX_ORDER)
2055 table->maxlen = sizeof(unsigned long);
2056 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2061 spin_lock(&hugetlb_lock);
2062 h->nr_overcommit_huge_pages = tmp;
2063 spin_unlock(&hugetlb_lock);
2069 #endif /* CONFIG_SYSCTL */
2071 void hugetlb_report_meminfo(struct seq_file *m)
2073 struct hstate *h = &default_hstate;
2075 "HugePages_Total: %5lu\n"
2076 "HugePages_Free: %5lu\n"
2077 "HugePages_Rsvd: %5lu\n"
2078 "HugePages_Surp: %5lu\n"
2079 "Hugepagesize: %8lu kB\n",
2083 h->surplus_huge_pages,
2084 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2087 int hugetlb_report_node_meminfo(int nid, char *buf)
2089 struct hstate *h = &default_hstate;
2091 "Node %d HugePages_Total: %5u\n"
2092 "Node %d HugePages_Free: %5u\n"
2093 "Node %d HugePages_Surp: %5u\n",
2094 nid, h->nr_huge_pages_node[nid],
2095 nid, h->free_huge_pages_node[nid],
2096 nid, h->surplus_huge_pages_node[nid]);
2099 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2100 unsigned long hugetlb_total_pages(void)
2102 struct hstate *h = &default_hstate;
2103 return h->nr_huge_pages * pages_per_huge_page(h);
2106 static int hugetlb_acct_memory(struct hstate *h, long delta)
2110 spin_lock(&hugetlb_lock);
2112 * When cpuset is configured, it breaks the strict hugetlb page
2113 * reservation as the accounting is done on a global variable. Such
2114 * reservation is completely rubbish in the presence of cpuset because
2115 * the reservation is not checked against page availability for the
2116 * current cpuset. Application can still potentially OOM'ed by kernel
2117 * with lack of free htlb page in cpuset that the task is in.
2118 * Attempt to enforce strict accounting with cpuset is almost
2119 * impossible (or too ugly) because cpuset is too fluid that
2120 * task or memory node can be dynamically moved between cpusets.
2122 * The change of semantics for shared hugetlb mapping with cpuset is
2123 * undesirable. However, in order to preserve some of the semantics,
2124 * we fall back to check against current free page availability as
2125 * a best attempt and hopefully to minimize the impact of changing
2126 * semantics that cpuset has.
2129 if (gather_surplus_pages(h, delta) < 0)
2132 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2133 return_unused_surplus_pages(h, delta);
2140 return_unused_surplus_pages(h, (unsigned long) -delta);
2143 spin_unlock(&hugetlb_lock);
2147 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2149 struct resv_map *reservations = vma_resv_map(vma);
2152 * This new VMA should share its siblings reservation map if present.
2153 * The VMA will only ever have a valid reservation map pointer where
2154 * it is being copied for another still existing VMA. As that VMA
2155 * has a reference to the reservation map it cannot disappear until
2156 * after this open call completes. It is therefore safe to take a
2157 * new reference here without additional locking.
2160 kref_get(&reservations->refs);
2163 static void resv_map_put(struct vm_area_struct *vma)
2165 struct resv_map *reservations = vma_resv_map(vma);
2169 kref_put(&reservations->refs, resv_map_release);
2172 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2174 struct hstate *h = hstate_vma(vma);
2175 struct resv_map *reservations = vma_resv_map(vma);
2176 struct hugepage_subpool *spool = subpool_vma(vma);
2177 unsigned long reserve;
2178 unsigned long start;
2182 start = vma_hugecache_offset(h, vma, vma->vm_start);
2183 end = vma_hugecache_offset(h, vma, vma->vm_end);
2185 reserve = (end - start) -
2186 region_count(&reservations->regions, start, end);
2191 hugetlb_acct_memory(h, -reserve);
2192 hugepage_subpool_put_pages(spool, reserve);
2198 * We cannot handle pagefaults against hugetlb pages at all. They cause
2199 * handle_mm_fault() to try to instantiate regular-sized pages in the
2200 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2203 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2209 const struct vm_operations_struct hugetlb_vm_ops = {
2210 .fault = hugetlb_vm_op_fault,
2211 .open = hugetlb_vm_op_open,
2212 .close = hugetlb_vm_op_close,
2215 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2222 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2224 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2226 entry = pte_mkyoung(entry);
2227 entry = pte_mkhuge(entry);
2228 entry = arch_make_huge_pte(entry, vma, page, writable);
2233 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2234 unsigned long address, pte_t *ptep)
2238 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2239 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2240 update_mmu_cache(vma, address, ptep);
2244 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2245 struct vm_area_struct *vma)
2247 pte_t *src_pte, *dst_pte, entry;
2248 struct page *ptepage;
2251 struct hstate *h = hstate_vma(vma);
2252 unsigned long sz = huge_page_size(h);
2254 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2256 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2257 src_pte = huge_pte_offset(src, addr);
2260 dst_pte = huge_pte_alloc(dst, addr, sz);
2264 /* If the pagetables are shared don't copy or take references */
2265 if (dst_pte == src_pte)
2268 spin_lock(&dst->page_table_lock);
2269 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2270 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2272 huge_ptep_set_wrprotect(src, addr, src_pte);
2273 entry = huge_ptep_get(src_pte);
2274 ptepage = pte_page(entry);
2276 page_dup_rmap(ptepage);
2277 set_huge_pte_at(dst, addr, dst_pte, entry);
2279 spin_unlock(&src->page_table_lock);
2280 spin_unlock(&dst->page_table_lock);
2288 static int is_hugetlb_entry_migration(pte_t pte)
2292 if (huge_pte_none(pte) || pte_present(pte))
2294 swp = pte_to_swp_entry(pte);
2295 if (non_swap_entry(swp) && is_migration_entry(swp))
2301 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2305 if (huge_pte_none(pte) || pte_present(pte))
2307 swp = pte_to_swp_entry(pte);
2308 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2314 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2315 unsigned long start, unsigned long end,
2316 struct page *ref_page)
2318 int force_flush = 0;
2319 struct mm_struct *mm = vma->vm_mm;
2320 unsigned long address;
2324 struct hstate *h = hstate_vma(vma);
2325 unsigned long sz = huge_page_size(h);
2327 WARN_ON(!is_vm_hugetlb_page(vma));
2328 BUG_ON(start & ~huge_page_mask(h));
2329 BUG_ON(end & ~huge_page_mask(h));
2331 tlb_start_vma(tlb, vma);
2332 mmu_notifier_invalidate_range_start(mm, start, end);
2334 spin_lock(&mm->page_table_lock);
2335 for (address = start; address < end; address += sz) {
2336 ptep = huge_pte_offset(mm, address);
2340 if (huge_pmd_unshare(mm, &address, ptep))
2343 pte = huge_ptep_get(ptep);
2344 if (huge_pte_none(pte))
2348 * HWPoisoned hugepage is already unmapped and dropped reference
2350 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2353 page = pte_page(pte);
2355 * If a reference page is supplied, it is because a specific
2356 * page is being unmapped, not a range. Ensure the page we
2357 * are about to unmap is the actual page of interest.
2360 if (page != ref_page)
2364 * Mark the VMA as having unmapped its page so that
2365 * future faults in this VMA will fail rather than
2366 * looking like data was lost
2368 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2371 pte = huge_ptep_get_and_clear(mm, address, ptep);
2372 tlb_remove_tlb_entry(tlb, ptep, address);
2374 set_page_dirty(page);
2376 page_remove_rmap(page);
2377 force_flush = !__tlb_remove_page(tlb, page);
2380 /* Bail out after unmapping reference page if supplied */
2384 spin_unlock(&mm->page_table_lock);
2386 * mmu_gather ran out of room to batch pages, we break out of
2387 * the PTE lock to avoid doing the potential expensive TLB invalidate
2388 * and page-free while holding it.
2393 if (address < end && !ref_page)
2396 mmu_notifier_invalidate_range_end(mm, start, end);
2397 tlb_end_vma(tlb, vma);
2400 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2401 unsigned long end, struct page *ref_page)
2403 struct mm_struct *mm;
2404 struct mmu_gather tlb;
2408 tlb_gather_mmu(&tlb, mm, 0);
2409 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2410 tlb_finish_mmu(&tlb, start, end);
2414 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2415 * mappping it owns the reserve page for. The intention is to unmap the page
2416 * from other VMAs and let the children be SIGKILLed if they are faulting the
2419 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2420 struct page *page, unsigned long address)
2422 struct hstate *h = hstate_vma(vma);
2423 struct vm_area_struct *iter_vma;
2424 struct address_space *mapping;
2425 struct prio_tree_iter iter;
2429 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2430 * from page cache lookup which is in HPAGE_SIZE units.
2432 address = address & huge_page_mask(h);
2433 pgoff = vma_hugecache_offset(h, vma, address);
2434 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2437 * Take the mapping lock for the duration of the table walk. As
2438 * this mapping should be shared between all the VMAs,
2439 * __unmap_hugepage_range() is called as the lock is already held
2441 mutex_lock(&mapping->i_mmap_mutex);
2442 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2443 /* Do not unmap the current VMA */
2444 if (iter_vma == vma)
2448 * Unmap the page from other VMAs without their own reserves.
2449 * They get marked to be SIGKILLed if they fault in these
2450 * areas. This is because a future no-page fault on this VMA
2451 * could insert a zeroed page instead of the data existing
2452 * from the time of fork. This would look like data corruption
2454 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2455 unmap_hugepage_range(iter_vma, address,
2456 address + huge_page_size(h), page);
2458 mutex_unlock(&mapping->i_mmap_mutex);
2464 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2465 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2466 * cannot race with other handlers or page migration.
2467 * Keep the pte_same checks anyway to make transition from the mutex easier.
2469 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2470 unsigned long address, pte_t *ptep, pte_t pte,
2471 struct page *pagecache_page)
2473 struct hstate *h = hstate_vma(vma);
2474 struct page *old_page, *new_page;
2476 int outside_reserve = 0;
2478 old_page = pte_page(pte);
2481 /* If no-one else is actually using this page, avoid the copy
2482 * and just make the page writable */
2483 avoidcopy = (page_mapcount(old_page) == 1);
2485 if (PageAnon(old_page))
2486 page_move_anon_rmap(old_page, vma, address);
2487 set_huge_ptep_writable(vma, address, ptep);
2492 * If the process that created a MAP_PRIVATE mapping is about to
2493 * perform a COW due to a shared page count, attempt to satisfy
2494 * the allocation without using the existing reserves. The pagecache
2495 * page is used to determine if the reserve at this address was
2496 * consumed or not. If reserves were used, a partial faulted mapping
2497 * at the time of fork() could consume its reserves on COW instead
2498 * of the full address range.
2500 if (!(vma->vm_flags & VM_MAYSHARE) &&
2501 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2502 old_page != pagecache_page)
2503 outside_reserve = 1;
2505 page_cache_get(old_page);
2507 /* Drop page_table_lock as buddy allocator may be called */
2508 spin_unlock(&mm->page_table_lock);
2509 new_page = alloc_huge_page(vma, address, outside_reserve);
2511 if (IS_ERR(new_page)) {
2512 long err = PTR_ERR(new_page);
2513 page_cache_release(old_page);
2516 * If a process owning a MAP_PRIVATE mapping fails to COW,
2517 * it is due to references held by a child and an insufficient
2518 * huge page pool. To guarantee the original mappers
2519 * reliability, unmap the page from child processes. The child
2520 * may get SIGKILLed if it later faults.
2522 if (outside_reserve) {
2523 BUG_ON(huge_pte_none(pte));
2524 if (unmap_ref_private(mm, vma, old_page, address)) {
2525 BUG_ON(huge_pte_none(pte));
2526 spin_lock(&mm->page_table_lock);
2527 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2528 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2529 goto retry_avoidcopy;
2531 * race occurs while re-acquiring page_table_lock, and
2539 /* Caller expects lock to be held */
2540 spin_lock(&mm->page_table_lock);
2542 return VM_FAULT_OOM;
2544 return VM_FAULT_SIGBUS;
2548 * When the original hugepage is shared one, it does not have
2549 * anon_vma prepared.
2551 if (unlikely(anon_vma_prepare(vma))) {
2552 page_cache_release(new_page);
2553 page_cache_release(old_page);
2554 /* Caller expects lock to be held */
2555 spin_lock(&mm->page_table_lock);
2556 return VM_FAULT_OOM;
2559 copy_user_huge_page(new_page, old_page, address, vma,
2560 pages_per_huge_page(h));
2561 __SetPageUptodate(new_page);
2564 * Retake the page_table_lock to check for racing updates
2565 * before the page tables are altered
2567 spin_lock(&mm->page_table_lock);
2568 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2569 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2571 mmu_notifier_invalidate_range_start(mm,
2572 address & huge_page_mask(h),
2573 (address & huge_page_mask(h)) + huge_page_size(h));
2574 huge_ptep_clear_flush(vma, address, ptep);
2575 set_huge_pte_at(mm, address, ptep,
2576 make_huge_pte(vma, new_page, 1));
2577 page_remove_rmap(old_page);
2578 hugepage_add_new_anon_rmap(new_page, vma, address);
2579 /* Make the old page be freed below */
2580 new_page = old_page;
2581 mmu_notifier_invalidate_range_end(mm,
2582 address & huge_page_mask(h),
2583 (address & huge_page_mask(h)) + huge_page_size(h));
2585 page_cache_release(new_page);
2586 page_cache_release(old_page);
2590 /* Return the pagecache page at a given address within a VMA */
2591 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2592 struct vm_area_struct *vma, unsigned long address)
2594 struct address_space *mapping;
2597 mapping = vma->vm_file->f_mapping;
2598 idx = vma_hugecache_offset(h, vma, address);
2600 return find_lock_page(mapping, idx);
2604 * Return whether there is a pagecache page to back given address within VMA.
2605 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2607 static bool hugetlbfs_pagecache_present(struct hstate *h,
2608 struct vm_area_struct *vma, unsigned long address)
2610 struct address_space *mapping;
2614 mapping = vma->vm_file->f_mapping;
2615 idx = vma_hugecache_offset(h, vma, address);
2617 page = find_get_page(mapping, idx);
2620 return page != NULL;
2623 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2624 unsigned long address, pte_t *ptep, unsigned int flags)
2626 struct hstate *h = hstate_vma(vma);
2627 int ret = VM_FAULT_SIGBUS;
2632 struct address_space *mapping;
2636 * Currently, we are forced to kill the process in the event the
2637 * original mapper has unmapped pages from the child due to a failed
2638 * COW. Warn that such a situation has occurred as it may not be obvious
2640 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2642 "PID %d killed due to inadequate hugepage pool\n",
2647 mapping = vma->vm_file->f_mapping;
2648 idx = vma_hugecache_offset(h, vma, address);
2651 * Use page lock to guard against racing truncation
2652 * before we get page_table_lock.
2655 page = find_lock_page(mapping, idx);
2657 size = i_size_read(mapping->host) >> huge_page_shift(h);
2660 page = alloc_huge_page(vma, address, 0);
2662 ret = PTR_ERR(page);
2666 ret = VM_FAULT_SIGBUS;
2669 clear_huge_page(page, address, pages_per_huge_page(h));
2670 __SetPageUptodate(page);
2672 if (vma->vm_flags & VM_MAYSHARE) {
2674 struct inode *inode = mapping->host;
2676 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2684 spin_lock(&inode->i_lock);
2685 inode->i_blocks += blocks_per_huge_page(h);
2686 spin_unlock(&inode->i_lock);
2689 if (unlikely(anon_vma_prepare(vma))) {
2691 goto backout_unlocked;
2697 * If memory error occurs between mmap() and fault, some process
2698 * don't have hwpoisoned swap entry for errored virtual address.
2699 * So we need to block hugepage fault by PG_hwpoison bit check.
2701 if (unlikely(PageHWPoison(page))) {
2702 ret = VM_FAULT_HWPOISON |
2703 VM_FAULT_SET_HINDEX(hstate_index(h));
2704 goto backout_unlocked;
2709 * If we are going to COW a private mapping later, we examine the
2710 * pending reservations for this page now. This will ensure that
2711 * any allocations necessary to record that reservation occur outside
2714 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2715 if (vma_needs_reservation(h, vma, address) < 0) {
2717 goto backout_unlocked;
2720 spin_lock(&mm->page_table_lock);
2721 size = i_size_read(mapping->host) >> huge_page_shift(h);
2726 if (!huge_pte_none(huge_ptep_get(ptep)))
2730 hugepage_add_new_anon_rmap(page, vma, address);
2732 page_dup_rmap(page);
2733 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2734 && (vma->vm_flags & VM_SHARED)));
2735 set_huge_pte_at(mm, address, ptep, new_pte);
2737 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2738 /* Optimization, do the COW without a second fault */
2739 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2742 spin_unlock(&mm->page_table_lock);
2748 spin_unlock(&mm->page_table_lock);
2755 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2756 unsigned long address, unsigned int flags)
2761 struct page *page = NULL;
2762 struct page *pagecache_page = NULL;
2763 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2764 struct hstate *h = hstate_vma(vma);
2766 address &= huge_page_mask(h);
2768 ptep = huge_pte_offset(mm, address);
2770 entry = huge_ptep_get(ptep);
2771 if (unlikely(is_hugetlb_entry_migration(entry))) {
2772 migration_entry_wait(mm, (pmd_t *)ptep, address);
2774 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2775 return VM_FAULT_HWPOISON_LARGE |
2776 VM_FAULT_SET_HINDEX(hstate_index(h));
2779 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2781 return VM_FAULT_OOM;
2784 * Serialize hugepage allocation and instantiation, so that we don't
2785 * get spurious allocation failures if two CPUs race to instantiate
2786 * the same page in the page cache.
2788 mutex_lock(&hugetlb_instantiation_mutex);
2789 entry = huge_ptep_get(ptep);
2790 if (huge_pte_none(entry)) {
2791 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2798 * If we are going to COW the mapping later, we examine the pending
2799 * reservations for this page now. This will ensure that any
2800 * allocations necessary to record that reservation occur outside the
2801 * spinlock. For private mappings, we also lookup the pagecache
2802 * page now as it is used to determine if a reservation has been
2805 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2806 if (vma_needs_reservation(h, vma, address) < 0) {
2811 if (!(vma->vm_flags & VM_MAYSHARE))
2812 pagecache_page = hugetlbfs_pagecache_page(h,
2817 * hugetlb_cow() requires page locks of pte_page(entry) and
2818 * pagecache_page, so here we need take the former one
2819 * when page != pagecache_page or !pagecache_page.
2820 * Note that locking order is always pagecache_page -> page,
2821 * so no worry about deadlock.
2823 page = pte_page(entry);
2825 if (page != pagecache_page)
2828 spin_lock(&mm->page_table_lock);
2829 /* Check for a racing update before calling hugetlb_cow */
2830 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2831 goto out_page_table_lock;
2834 if (flags & FAULT_FLAG_WRITE) {
2835 if (!pte_write(entry)) {
2836 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2838 goto out_page_table_lock;
2840 entry = pte_mkdirty(entry);
2842 entry = pte_mkyoung(entry);
2843 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2844 flags & FAULT_FLAG_WRITE))
2845 update_mmu_cache(vma, address, ptep);
2847 out_page_table_lock:
2848 spin_unlock(&mm->page_table_lock);
2850 if (pagecache_page) {
2851 unlock_page(pagecache_page);
2852 put_page(pagecache_page);
2854 if (page != pagecache_page)
2859 mutex_unlock(&hugetlb_instantiation_mutex);
2864 /* Can be overriden by architectures */
2865 __attribute__((weak)) struct page *
2866 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2867 pud_t *pud, int write)
2873 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2874 struct page **pages, struct vm_area_struct **vmas,
2875 unsigned long *position, int *length, int i,
2878 unsigned long pfn_offset;
2879 unsigned long vaddr = *position;
2880 int remainder = *length;
2881 struct hstate *h = hstate_vma(vma);
2883 spin_lock(&mm->page_table_lock);
2884 while (vaddr < vma->vm_end && remainder) {
2890 * Some archs (sparc64, sh*) have multiple pte_ts to
2891 * each hugepage. We have to make sure we get the
2892 * first, for the page indexing below to work.
2894 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2895 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2898 * When coredumping, it suits get_dump_page if we just return
2899 * an error where there's an empty slot with no huge pagecache
2900 * to back it. This way, we avoid allocating a hugepage, and
2901 * the sparse dumpfile avoids allocating disk blocks, but its
2902 * huge holes still show up with zeroes where they need to be.
2904 if (absent && (flags & FOLL_DUMP) &&
2905 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2911 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2914 spin_unlock(&mm->page_table_lock);
2915 ret = hugetlb_fault(mm, vma, vaddr,
2916 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2917 spin_lock(&mm->page_table_lock);
2918 if (!(ret & VM_FAULT_ERROR))
2925 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2926 page = pte_page(huge_ptep_get(pte));
2929 pages[i] = mem_map_offset(page, pfn_offset);
2940 if (vaddr < vma->vm_end && remainder &&
2941 pfn_offset < pages_per_huge_page(h)) {
2943 * We use pfn_offset to avoid touching the pageframes
2944 * of this compound page.
2949 spin_unlock(&mm->page_table_lock);
2950 *length = remainder;
2953 return i ? i : -EFAULT;
2956 void hugetlb_change_protection(struct vm_area_struct *vma,
2957 unsigned long address, unsigned long end, pgprot_t newprot)
2959 struct mm_struct *mm = vma->vm_mm;
2960 unsigned long start = address;
2963 struct hstate *h = hstate_vma(vma);
2965 BUG_ON(address >= end);
2966 flush_cache_range(vma, address, end);
2968 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2969 spin_lock(&mm->page_table_lock);
2970 for (; address < end; address += huge_page_size(h)) {
2971 ptep = huge_pte_offset(mm, address);
2974 if (huge_pmd_unshare(mm, &address, ptep))
2976 if (!huge_pte_none(huge_ptep_get(ptep))) {
2977 pte = huge_ptep_get_and_clear(mm, address, ptep);
2978 pte = pte_mkhuge(pte_modify(pte, newprot));
2979 set_huge_pte_at(mm, address, ptep, pte);
2982 spin_unlock(&mm->page_table_lock);
2983 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2985 flush_tlb_range(vma, start, end);
2988 int hugetlb_reserve_pages(struct inode *inode,
2990 struct vm_area_struct *vma,
2991 vm_flags_t vm_flags)
2994 struct hstate *h = hstate_inode(inode);
2995 struct hugepage_subpool *spool = subpool_inode(inode);
2998 * Only apply hugepage reservation if asked. At fault time, an
2999 * attempt will be made for VM_NORESERVE to allocate a page
3000 * without using reserves
3002 if (vm_flags & VM_NORESERVE)
3006 * Shared mappings base their reservation on the number of pages that
3007 * are already allocated on behalf of the file. Private mappings need
3008 * to reserve the full area even if read-only as mprotect() may be
3009 * called to make the mapping read-write. Assume !vma is a shm mapping
3011 if (!vma || vma->vm_flags & VM_MAYSHARE)
3012 chg = region_chg(&inode->i_mapping->private_list, from, to);
3014 struct resv_map *resv_map = resv_map_alloc();
3020 set_vma_resv_map(vma, resv_map);
3021 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3029 /* There must be enough pages in the subpool for the mapping */
3030 if (hugepage_subpool_get_pages(spool, chg)) {
3036 * Check enough hugepages are available for the reservation.
3037 * Hand the pages back to the subpool if there are not
3039 ret = hugetlb_acct_memory(h, chg);
3041 hugepage_subpool_put_pages(spool, chg);
3046 * Account for the reservations made. Shared mappings record regions
3047 * that have reservations as they are shared by multiple VMAs.
3048 * When the last VMA disappears, the region map says how much
3049 * the reservation was and the page cache tells how much of
3050 * the reservation was consumed. Private mappings are per-VMA and
3051 * only the consumed reservations are tracked. When the VMA
3052 * disappears, the original reservation is the VMA size and the
3053 * consumed reservations are stored in the map. Hence, nothing
3054 * else has to be done for private mappings here
3056 if (!vma || vma->vm_flags & VM_MAYSHARE)
3057 region_add(&inode->i_mapping->private_list, from, to);
3065 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3067 struct hstate *h = hstate_inode(inode);
3068 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3069 struct hugepage_subpool *spool = subpool_inode(inode);
3071 spin_lock(&inode->i_lock);
3072 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3073 spin_unlock(&inode->i_lock);
3075 hugepage_subpool_put_pages(spool, (chg - freed));
3076 hugetlb_acct_memory(h, -(chg - freed));
3079 #ifdef CONFIG_MEMORY_FAILURE
3081 /* Should be called in hugetlb_lock */
3082 static int is_hugepage_on_freelist(struct page *hpage)
3086 struct hstate *h = page_hstate(hpage);
3087 int nid = page_to_nid(hpage);
3089 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3096 * This function is called from memory failure code.
3097 * Assume the caller holds page lock of the head page.
3099 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3101 struct hstate *h = page_hstate(hpage);
3102 int nid = page_to_nid(hpage);
3105 spin_lock(&hugetlb_lock);
3106 if (is_hugepage_on_freelist(hpage)) {
3107 list_del(&hpage->lru);
3108 set_page_refcounted(hpage);
3109 h->free_huge_pages--;
3110 h->free_huge_pages_node[nid]--;
3113 spin_unlock(&hugetlb_lock);