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/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
36 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37 unsigned long hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 __initdata LIST_HEAD(huge_boot_pages);
45 /* for command line parsing */
46 static struct hstate * __initdata parsed_hstate;
47 static unsigned long __initdata default_hstate_max_huge_pages;
48 static unsigned long __initdata default_hstate_size;
51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 DEFINE_SPINLOCK(hugetlb_lock);
55 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59 spin_unlock(&spool->lock);
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
67 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 struct hugepage_subpool *spool;
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
75 spin_lock_init(&spool->lock);
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
83 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
88 unlock_or_release_subpool(spool);
91 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
105 spin_unlock(&spool->lock);
110 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
123 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 return HUGETLBFS_SB(inode->i_sb)->spool;
128 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 return subpool_inode(vma->vm_file->f_dentry->d_inode);
134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantion_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation mutex:
142 * down_write(&mm->mmap_sem);
144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
148 struct list_head link;
153 static long region_add(struct list_head *head, long f, long t)
155 struct file_region *rg, *nrg, *trg;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg, head, link)
162 /* Round our left edge to the current segment if it encloses us. */
166 /* Check for and consume any regions we now overlap with. */
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
189 static long region_chg(struct list_head *head, long f, long t)
191 struct file_region *rg, *nrg;
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg, head, link)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg->link == head || t < rg->from) {
203 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
208 INIT_LIST_HEAD(&nrg->link);
209 list_add(&nrg->link, rg->link.prev);
214 /* Round our left edge to the current segment if it encloses us. */
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg, rg->link.prev, link) {
221 if (&rg->link == head)
226 /* We overlap with this area, if it extends further than
227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
233 chg -= rg->to - rg->from;
238 static long region_truncate(struct list_head *head, long end)
240 struct file_region *rg, *trg;
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg, head, link)
247 if (&rg->link == head)
250 /* If we are in the middle of a region then adjust it. */
251 if (end > rg->from) {
254 rg = list_entry(rg->link.next, typeof(*rg), link);
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
259 if (&rg->link == head)
261 chg += rg->to - rg->from;
268 static long region_count(struct list_head *head, long f, long t)
270 struct file_region *rg;
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg, head, link) {
283 seg_from = max(rg->from, f);
284 seg_to = min(rg->to, t);
286 chg += seg_to - seg_from;
293 * Convert the address within this vma to the page offset within
294 * the mapping, in pagecache page units; huge pages here.
296 static pgoff_t vma_hugecache_offset(struct hstate *h,
297 struct vm_area_struct *vma, unsigned long address)
299 return ((address - vma->vm_start) >> huge_page_shift(h)) +
300 (vma->vm_pgoff >> huge_page_order(h));
303 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
304 unsigned long address)
306 return vma_hugecache_offset(hstate_vma(vma), vma, address);
310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
313 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 struct hstate *hstate;
317 if (!is_vm_hugetlb_page(vma))
320 hstate = hstate_vma(vma);
322 return 1UL << (hstate->order + PAGE_SHIFT);
324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
332 #ifndef vma_mmu_pagesize
333 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 return vma_kernel_pagesize(vma);
340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
344 #define HPAGE_RESV_OWNER (1UL << 0)
345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
367 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 return (unsigned long)vma->vm_private_data;
372 static void set_vma_private_data(struct vm_area_struct *vma,
375 vma->vm_private_data = (void *)value;
380 struct list_head regions;
383 static struct resv_map *resv_map_alloc(void)
385 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
389 kref_init(&resv_map->refs);
390 INIT_LIST_HEAD(&resv_map->regions);
395 static void resv_map_release(struct kref *ref)
397 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map->regions, 0);
404 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma));
407 if (!(vma->vm_flags & VM_MAYSHARE))
408 return (struct resv_map *)(get_vma_private_data(vma) &
413 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma));
416 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418 set_vma_private_data(vma, (get_vma_private_data(vma) &
419 HPAGE_RESV_MASK) | (unsigned long)map);
422 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma));
425 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
430 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma));
434 return (get_vma_private_data(vma) & flag) != 0;
437 /* Decrement the reserved pages in the hugepage pool by one */
438 static void decrement_hugepage_resv_vma(struct hstate *h,
439 struct vm_area_struct *vma)
441 if (vma->vm_flags & VM_NORESERVE)
444 if (vma->vm_flags & VM_MAYSHARE) {
445 /* Shared mappings always use reserves */
446 h->resv_huge_pages--;
447 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
449 * Only the process that called mmap() has reserves for
452 h->resv_huge_pages--;
456 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
457 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459 VM_BUG_ON(!is_vm_hugetlb_page(vma));
460 if (!(vma->vm_flags & VM_MAYSHARE))
461 vma->vm_private_data = (void *)0;
464 /* Returns true if the VMA has associated reserve pages */
465 static int vma_has_reserves(struct vm_area_struct *vma)
467 if (vma->vm_flags & VM_MAYSHARE)
469 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474 static void copy_gigantic_page(struct page *dst, struct page *src)
477 struct hstate *h = page_hstate(src);
478 struct page *dst_base = dst;
479 struct page *src_base = src;
481 for (i = 0; i < pages_per_huge_page(h); ) {
483 copy_highpage(dst, src);
486 dst = mem_map_next(dst, dst_base, i);
487 src = mem_map_next(src, src_base, i);
491 void copy_huge_page(struct page *dst, struct page *src)
494 struct hstate *h = page_hstate(src);
496 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
497 copy_gigantic_page(dst, src);
502 for (i = 0; i < pages_per_huge_page(h); i++) {
504 copy_highpage(dst + i, src + i);
508 static void enqueue_huge_page(struct hstate *h, struct page *page)
510 int nid = page_to_nid(page);
511 list_move(&page->lru, &h->hugepage_freelists[nid]);
512 h->free_huge_pages++;
513 h->free_huge_pages_node[nid]++;
516 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
520 if (list_empty(&h->hugepage_freelists[nid]))
522 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
523 list_move(&page->lru, &h->hugepage_activelist);
524 set_page_refcounted(page);
525 h->free_huge_pages--;
526 h->free_huge_pages_node[nid]--;
530 static struct page *dequeue_huge_page_vma(struct hstate *h,
531 struct vm_area_struct *vma,
532 unsigned long address, int avoid_reserve)
534 struct page *page = NULL;
535 struct mempolicy *mpol;
536 nodemask_t *nodemask;
537 struct zonelist *zonelist;
540 unsigned int cpuset_mems_cookie;
543 cpuset_mems_cookie = get_mems_allowed();
544 zonelist = huge_zonelist(vma, address,
545 htlb_alloc_mask, &mpol, &nodemask);
547 * A child process with MAP_PRIVATE mappings created by their parent
548 * have no page reserves. This check ensures that reservations are
549 * not "stolen". The child may still get SIGKILLed
551 if (!vma_has_reserves(vma) &&
552 h->free_huge_pages - h->resv_huge_pages == 0)
555 /* If reserves cannot be used, ensure enough pages are in the pool */
556 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
559 for_each_zone_zonelist_nodemask(zone, z, zonelist,
560 MAX_NR_ZONES - 1, nodemask) {
561 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
562 page = dequeue_huge_page_node(h, zone_to_nid(zone));
565 decrement_hugepage_resv_vma(h, vma);
572 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
581 static void update_and_free_page(struct hstate *h, struct page *page)
585 VM_BUG_ON(h->order >= MAX_ORDER);
588 h->nr_huge_pages_node[page_to_nid(page)]--;
589 for (i = 0; i < pages_per_huge_page(h); i++) {
590 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
591 1 << PG_referenced | 1 << PG_dirty |
592 1 << PG_active | 1 << PG_reserved |
593 1 << PG_private | 1 << PG_writeback);
595 VM_BUG_ON(hugetlb_cgroup_from_page(page));
596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
598 arch_release_hugepage(page);
599 __free_pages(page, huge_page_order(h));
602 struct hstate *size_to_hstate(unsigned long size)
607 if (huge_page_size(h) == size)
613 static void free_huge_page(struct page *page)
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
619 struct hstate *h = page_hstate(page);
620 int nid = page_to_nid(page);
621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
624 set_page_private(page, 0);
625 page->mapping = NULL;
626 BUG_ON(page_count(page));
627 BUG_ON(page_mapcount(page));
629 spin_lock(&hugetlb_lock);
630 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
631 /* remove the page from active list */
632 list_del(&page->lru);
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 INIT_LIST_HEAD(&page->lru);
646 set_compound_page_dtor(page, free_huge_page);
647 spin_lock(&hugetlb_lock);
648 set_hugetlb_cgroup(page, NULL);
650 h->nr_huge_pages_node[nid]++;
651 spin_unlock(&hugetlb_lock);
652 put_page(page); /* free it into the hugepage allocator */
655 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
658 int nr_pages = 1 << order;
659 struct page *p = page + 1;
661 /* we rely on prep_new_huge_page to set the destructor */
662 set_compound_order(page, order);
664 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
666 set_page_count(p, 0);
667 p->first_page = page;
671 int PageHuge(struct page *page)
673 compound_page_dtor *dtor;
675 if (!PageCompound(page))
678 page = compound_head(page);
679 dtor = get_compound_page_dtor(page);
681 return dtor == free_huge_page;
683 EXPORT_SYMBOL_GPL(PageHuge);
685 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
689 if (h->order >= MAX_ORDER)
692 page = alloc_pages_exact_node(nid,
693 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
694 __GFP_REPEAT|__GFP_NOWARN,
697 if (arch_prepare_hugepage(page)) {
698 __free_pages(page, huge_page_order(h));
701 prep_new_huge_page(h, page, nid);
708 * common helper functions for hstate_next_node_to_{alloc|free}.
709 * We may have allocated or freed a huge page based on a different
710 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
711 * be outside of *nodes_allowed. Ensure that we use an allowed
712 * node for alloc or free.
714 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
716 nid = next_node(nid, *nodes_allowed);
717 if (nid == MAX_NUMNODES)
718 nid = first_node(*nodes_allowed);
719 VM_BUG_ON(nid >= MAX_NUMNODES);
724 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
726 if (!node_isset(nid, *nodes_allowed))
727 nid = next_node_allowed(nid, nodes_allowed);
732 * returns the previously saved node ["this node"] from which to
733 * allocate a persistent huge page for the pool and advance the
734 * next node from which to allocate, handling wrap at end of node
737 static int hstate_next_node_to_alloc(struct hstate *h,
738 nodemask_t *nodes_allowed)
742 VM_BUG_ON(!nodes_allowed);
744 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
745 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
750 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
757 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
758 next_nid = start_nid;
761 page = alloc_fresh_huge_page_node(h, next_nid);
766 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
767 } while (next_nid != start_nid);
770 count_vm_event(HTLB_BUDDY_PGALLOC);
772 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
778 * helper for free_pool_huge_page() - return the previously saved
779 * node ["this node"] from which to free a huge page. Advance the
780 * next node id whether or not we find a free huge page to free so
781 * that the next attempt to free addresses the next node.
783 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
787 VM_BUG_ON(!nodes_allowed);
789 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
790 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
796 * Free huge page from pool from next node to free.
797 * Attempt to keep persistent huge pages more or less
798 * balanced over allowed nodes.
799 * Called with hugetlb_lock locked.
801 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
808 start_nid = hstate_next_node_to_free(h, nodes_allowed);
809 next_nid = start_nid;
813 * If we're returning unused surplus pages, only examine
814 * nodes with surplus pages.
816 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
817 !list_empty(&h->hugepage_freelists[next_nid])) {
819 list_entry(h->hugepage_freelists[next_nid].next,
821 list_del(&page->lru);
822 h->free_huge_pages--;
823 h->free_huge_pages_node[next_nid]--;
825 h->surplus_huge_pages--;
826 h->surplus_huge_pages_node[next_nid]--;
828 update_and_free_page(h, page);
832 next_nid = hstate_next_node_to_free(h, nodes_allowed);
833 } while (next_nid != start_nid);
838 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
843 if (h->order >= MAX_ORDER)
847 * Assume we will successfully allocate the surplus page to
848 * prevent racing processes from causing the surplus to exceed
851 * This however introduces a different race, where a process B
852 * tries to grow the static hugepage pool while alloc_pages() is
853 * called by process A. B will only examine the per-node
854 * counters in determining if surplus huge pages can be
855 * converted to normal huge pages in adjust_pool_surplus(). A
856 * won't be able to increment the per-node counter, until the
857 * lock is dropped by B, but B doesn't drop hugetlb_lock until
858 * no more huge pages can be converted from surplus to normal
859 * state (and doesn't try to convert again). Thus, we have a
860 * case where a surplus huge page exists, the pool is grown, and
861 * the surplus huge page still exists after, even though it
862 * should just have been converted to a normal huge page. This
863 * does not leak memory, though, as the hugepage will be freed
864 * once it is out of use. It also does not allow the counters to
865 * go out of whack in adjust_pool_surplus() as we don't modify
866 * the node values until we've gotten the hugepage and only the
867 * per-node value is checked there.
869 spin_lock(&hugetlb_lock);
870 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
871 spin_unlock(&hugetlb_lock);
875 h->surplus_huge_pages++;
877 spin_unlock(&hugetlb_lock);
879 if (nid == NUMA_NO_NODE)
880 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
881 __GFP_REPEAT|__GFP_NOWARN,
884 page = alloc_pages_exact_node(nid,
885 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
886 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
888 if (page && arch_prepare_hugepage(page)) {
889 __free_pages(page, huge_page_order(h));
893 spin_lock(&hugetlb_lock);
895 INIT_LIST_HEAD(&page->lru);
896 r_nid = page_to_nid(page);
897 set_compound_page_dtor(page, free_huge_page);
898 set_hugetlb_cgroup(page, NULL);
900 * We incremented the global counters already
902 h->nr_huge_pages_node[r_nid]++;
903 h->surplus_huge_pages_node[r_nid]++;
904 __count_vm_event(HTLB_BUDDY_PGALLOC);
907 h->surplus_huge_pages--;
908 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
910 spin_unlock(&hugetlb_lock);
916 * This allocation function is useful in the context where vma is irrelevant.
917 * E.g. soft-offlining uses this function because it only cares physical
918 * address of error page.
920 struct page *alloc_huge_page_node(struct hstate *h, int nid)
924 spin_lock(&hugetlb_lock);
925 page = dequeue_huge_page_node(h, nid);
926 spin_unlock(&hugetlb_lock);
929 page = alloc_buddy_huge_page(h, nid);
935 * Increase the hugetlb pool such that it can accommodate a reservation
938 static int gather_surplus_pages(struct hstate *h, int delta)
940 struct list_head surplus_list;
941 struct page *page, *tmp;
943 int needed, allocated;
944 bool alloc_ok = true;
946 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
948 h->resv_huge_pages += delta;
953 INIT_LIST_HEAD(&surplus_list);
957 spin_unlock(&hugetlb_lock);
958 for (i = 0; i < needed; i++) {
959 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
964 list_add(&page->lru, &surplus_list);
969 * After retaking hugetlb_lock, we need to recalculate 'needed'
970 * because either resv_huge_pages or free_huge_pages may have changed.
972 spin_lock(&hugetlb_lock);
973 needed = (h->resv_huge_pages + delta) -
974 (h->free_huge_pages + allocated);
979 * We were not able to allocate enough pages to
980 * satisfy the entire reservation so we free what
981 * we've allocated so far.
986 * The surplus_list now contains _at_least_ the number of extra pages
987 * needed to accommodate the reservation. Add the appropriate number
988 * of pages to the hugetlb pool and free the extras back to the buddy
989 * allocator. Commit the entire reservation here to prevent another
990 * process from stealing the pages as they are added to the pool but
991 * before they are reserved.
994 h->resv_huge_pages += delta;
997 /* Free the needed pages to the hugetlb pool */
998 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1002 * This page is now managed by the hugetlb allocator and has
1003 * no users -- drop the buddy allocator's reference.
1005 put_page_testzero(page);
1006 VM_BUG_ON(page_count(page));
1007 enqueue_huge_page(h, page);
1010 spin_unlock(&hugetlb_lock);
1012 /* Free unnecessary surplus pages to the buddy allocator */
1013 if (!list_empty(&surplus_list)) {
1014 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1018 spin_lock(&hugetlb_lock);
1024 * When releasing a hugetlb pool reservation, any surplus pages that were
1025 * allocated to satisfy the reservation must be explicitly freed if they were
1027 * Called with hugetlb_lock held.
1029 static void return_unused_surplus_pages(struct hstate *h,
1030 unsigned long unused_resv_pages)
1032 unsigned long nr_pages;
1034 /* Uncommit the reservation */
1035 h->resv_huge_pages -= unused_resv_pages;
1037 /* Cannot return gigantic pages currently */
1038 if (h->order >= MAX_ORDER)
1041 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1044 * We want to release as many surplus pages as possible, spread
1045 * evenly across all nodes with memory. Iterate across these nodes
1046 * until we can no longer free unreserved surplus pages. This occurs
1047 * when the nodes with surplus pages have no free pages.
1048 * free_pool_huge_page() will balance the the freed pages across the
1049 * on-line nodes with memory and will handle the hstate accounting.
1051 while (nr_pages--) {
1052 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1058 * Determine if the huge page at addr within the vma has an associated
1059 * reservation. Where it does not we will need to logically increase
1060 * reservation and actually increase subpool usage before an allocation
1061 * can occur. Where any new reservation would be required the
1062 * reservation change is prepared, but not committed. Once the page
1063 * has been allocated from the subpool and instantiated the change should
1064 * be committed via vma_commit_reservation. No action is required on
1067 static long vma_needs_reservation(struct hstate *h,
1068 struct vm_area_struct *vma, unsigned long addr)
1070 struct address_space *mapping = vma->vm_file->f_mapping;
1071 struct inode *inode = mapping->host;
1073 if (vma->vm_flags & VM_MAYSHARE) {
1074 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1075 return region_chg(&inode->i_mapping->private_list,
1078 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1083 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1084 struct resv_map *reservations = vma_resv_map(vma);
1086 err = region_chg(&reservations->regions, idx, idx + 1);
1092 static void vma_commit_reservation(struct hstate *h,
1093 struct vm_area_struct *vma, unsigned long addr)
1095 struct address_space *mapping = vma->vm_file->f_mapping;
1096 struct inode *inode = mapping->host;
1098 if (vma->vm_flags & VM_MAYSHARE) {
1099 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1100 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1102 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1103 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1104 struct resv_map *reservations = vma_resv_map(vma);
1106 /* Mark this page used in the map. */
1107 region_add(&reservations->regions, idx, idx + 1);
1111 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1112 unsigned long addr, int avoid_reserve)
1114 struct hugepage_subpool *spool = subpool_vma(vma);
1115 struct hstate *h = hstate_vma(vma);
1120 * Processes that did not create the mapping will have no
1121 * reserves and will not have accounted against subpool
1122 * limit. Check that the subpool limit can be made before
1123 * satisfying the allocation MAP_NORESERVE mappings may also
1124 * need pages and subpool limit allocated allocated if no reserve
1127 chg = vma_needs_reservation(h, vma, addr);
1129 return ERR_PTR(-ENOMEM);
1131 if (hugepage_subpool_get_pages(spool, chg))
1132 return ERR_PTR(-ENOSPC);
1134 spin_lock(&hugetlb_lock);
1135 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1136 spin_unlock(&hugetlb_lock);
1139 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1141 hugepage_subpool_put_pages(spool, chg);
1142 return ERR_PTR(-ENOSPC);
1146 set_page_private(page, (unsigned long)spool);
1148 vma_commit_reservation(h, vma, addr);
1153 int __weak alloc_bootmem_huge_page(struct hstate *h)
1155 struct huge_bootmem_page *m;
1156 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1161 addr = __alloc_bootmem_node_nopanic(
1162 NODE_DATA(hstate_next_node_to_alloc(h,
1163 &node_states[N_HIGH_MEMORY])),
1164 huge_page_size(h), huge_page_size(h), 0);
1168 * Use the beginning of the huge page to store the
1169 * huge_bootmem_page struct (until gather_bootmem
1170 * puts them into the mem_map).
1180 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1181 /* Put them into a private list first because mem_map is not up yet */
1182 list_add(&m->list, &huge_boot_pages);
1187 static void prep_compound_huge_page(struct page *page, int order)
1189 if (unlikely(order > (MAX_ORDER - 1)))
1190 prep_compound_gigantic_page(page, order);
1192 prep_compound_page(page, order);
1195 /* Put bootmem huge pages into the standard lists after mem_map is up */
1196 static void __init gather_bootmem_prealloc(void)
1198 struct huge_bootmem_page *m;
1200 list_for_each_entry(m, &huge_boot_pages, list) {
1201 struct hstate *h = m->hstate;
1204 #ifdef CONFIG_HIGHMEM
1205 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1206 free_bootmem_late((unsigned long)m,
1207 sizeof(struct huge_bootmem_page));
1209 page = virt_to_page(m);
1211 __ClearPageReserved(page);
1212 WARN_ON(page_count(page) != 1);
1213 prep_compound_huge_page(page, h->order);
1214 prep_new_huge_page(h, page, page_to_nid(page));
1216 * If we had gigantic hugepages allocated at boot time, we need
1217 * to restore the 'stolen' pages to totalram_pages in order to
1218 * fix confusing memory reports from free(1) and another
1219 * side-effects, like CommitLimit going negative.
1221 if (h->order > (MAX_ORDER - 1))
1222 totalram_pages += 1 << h->order;
1226 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1230 for (i = 0; i < h->max_huge_pages; ++i) {
1231 if (h->order >= MAX_ORDER) {
1232 if (!alloc_bootmem_huge_page(h))
1234 } else if (!alloc_fresh_huge_page(h,
1235 &node_states[N_HIGH_MEMORY]))
1238 h->max_huge_pages = i;
1241 static void __init hugetlb_init_hstates(void)
1245 for_each_hstate(h) {
1246 /* oversize hugepages were init'ed in early boot */
1247 if (h->order < MAX_ORDER)
1248 hugetlb_hstate_alloc_pages(h);
1252 static char * __init memfmt(char *buf, unsigned long n)
1254 if (n >= (1UL << 30))
1255 sprintf(buf, "%lu GB", n >> 30);
1256 else if (n >= (1UL << 20))
1257 sprintf(buf, "%lu MB", n >> 20);
1259 sprintf(buf, "%lu KB", n >> 10);
1263 static void __init report_hugepages(void)
1267 for_each_hstate(h) {
1269 printk(KERN_INFO "HugeTLB registered %s page size, "
1270 "pre-allocated %ld pages\n",
1271 memfmt(buf, huge_page_size(h)),
1272 h->free_huge_pages);
1276 #ifdef CONFIG_HIGHMEM
1277 static void try_to_free_low(struct hstate *h, unsigned long count,
1278 nodemask_t *nodes_allowed)
1282 if (h->order >= MAX_ORDER)
1285 for_each_node_mask(i, *nodes_allowed) {
1286 struct page *page, *next;
1287 struct list_head *freel = &h->hugepage_freelists[i];
1288 list_for_each_entry_safe(page, next, freel, lru) {
1289 if (count >= h->nr_huge_pages)
1291 if (PageHighMem(page))
1293 list_del(&page->lru);
1294 update_and_free_page(h, page);
1295 h->free_huge_pages--;
1296 h->free_huge_pages_node[page_to_nid(page)]--;
1301 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1302 nodemask_t *nodes_allowed)
1308 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1309 * balanced by operating on them in a round-robin fashion.
1310 * Returns 1 if an adjustment was made.
1312 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1315 int start_nid, next_nid;
1318 VM_BUG_ON(delta != -1 && delta != 1);
1321 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1323 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1324 next_nid = start_nid;
1330 * To shrink on this node, there must be a surplus page
1332 if (!h->surplus_huge_pages_node[nid]) {
1333 next_nid = hstate_next_node_to_alloc(h,
1340 * Surplus cannot exceed the total number of pages
1342 if (h->surplus_huge_pages_node[nid] >=
1343 h->nr_huge_pages_node[nid]) {
1344 next_nid = hstate_next_node_to_free(h,
1350 h->surplus_huge_pages += delta;
1351 h->surplus_huge_pages_node[nid] += delta;
1354 } while (next_nid != start_nid);
1359 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1360 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1361 nodemask_t *nodes_allowed)
1363 unsigned long min_count, ret;
1365 if (h->order >= MAX_ORDER)
1366 return h->max_huge_pages;
1369 * Increase the pool size
1370 * First take pages out of surplus state. Then make up the
1371 * remaining difference by allocating fresh huge pages.
1373 * We might race with alloc_buddy_huge_page() here and be unable
1374 * to convert a surplus huge page to a normal huge page. That is
1375 * not critical, though, it just means the overall size of the
1376 * pool might be one hugepage larger than it needs to be, but
1377 * within all the constraints specified by the sysctls.
1379 spin_lock(&hugetlb_lock);
1380 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1381 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1385 while (count > persistent_huge_pages(h)) {
1387 * If this allocation races such that we no longer need the
1388 * page, free_huge_page will handle it by freeing the page
1389 * and reducing the surplus.
1391 spin_unlock(&hugetlb_lock);
1392 ret = alloc_fresh_huge_page(h, nodes_allowed);
1393 spin_lock(&hugetlb_lock);
1397 /* Bail for signals. Probably ctrl-c from user */
1398 if (signal_pending(current))
1403 * Decrease the pool size
1404 * First return free pages to the buddy allocator (being careful
1405 * to keep enough around to satisfy reservations). Then place
1406 * pages into surplus state as needed so the pool will shrink
1407 * to the desired size as pages become free.
1409 * By placing pages into the surplus state independent of the
1410 * overcommit value, we are allowing the surplus pool size to
1411 * exceed overcommit. There are few sane options here. Since
1412 * alloc_buddy_huge_page() is checking the global counter,
1413 * though, we'll note that we're not allowed to exceed surplus
1414 * and won't grow the pool anywhere else. Not until one of the
1415 * sysctls are changed, or the surplus pages go out of use.
1417 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1418 min_count = max(count, min_count);
1419 try_to_free_low(h, min_count, nodes_allowed);
1420 while (min_count < persistent_huge_pages(h)) {
1421 if (!free_pool_huge_page(h, nodes_allowed, 0))
1424 while (count < persistent_huge_pages(h)) {
1425 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1429 ret = persistent_huge_pages(h);
1430 spin_unlock(&hugetlb_lock);
1434 #define HSTATE_ATTR_RO(_name) \
1435 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1437 #define HSTATE_ATTR(_name) \
1438 static struct kobj_attribute _name##_attr = \
1439 __ATTR(_name, 0644, _name##_show, _name##_store)
1441 static struct kobject *hugepages_kobj;
1442 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1444 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1446 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1450 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1451 if (hstate_kobjs[i] == kobj) {
1453 *nidp = NUMA_NO_NODE;
1457 return kobj_to_node_hstate(kobj, nidp);
1460 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1461 struct kobj_attribute *attr, char *buf)
1464 unsigned long nr_huge_pages;
1467 h = kobj_to_hstate(kobj, &nid);
1468 if (nid == NUMA_NO_NODE)
1469 nr_huge_pages = h->nr_huge_pages;
1471 nr_huge_pages = h->nr_huge_pages_node[nid];
1473 return sprintf(buf, "%lu\n", nr_huge_pages);
1476 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1477 struct kobject *kobj, struct kobj_attribute *attr,
1478 const char *buf, size_t len)
1482 unsigned long count;
1484 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1486 err = strict_strtoul(buf, 10, &count);
1490 h = kobj_to_hstate(kobj, &nid);
1491 if (h->order >= MAX_ORDER) {
1496 if (nid == NUMA_NO_NODE) {
1498 * global hstate attribute
1500 if (!(obey_mempolicy &&
1501 init_nodemask_of_mempolicy(nodes_allowed))) {
1502 NODEMASK_FREE(nodes_allowed);
1503 nodes_allowed = &node_states[N_HIGH_MEMORY];
1505 } else if (nodes_allowed) {
1507 * per node hstate attribute: adjust count to global,
1508 * but restrict alloc/free to the specified node.
1510 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1511 init_nodemask_of_node(nodes_allowed, nid);
1513 nodes_allowed = &node_states[N_HIGH_MEMORY];
1515 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1517 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1518 NODEMASK_FREE(nodes_allowed);
1522 NODEMASK_FREE(nodes_allowed);
1526 static ssize_t nr_hugepages_show(struct kobject *kobj,
1527 struct kobj_attribute *attr, char *buf)
1529 return nr_hugepages_show_common(kobj, attr, buf);
1532 static ssize_t nr_hugepages_store(struct kobject *kobj,
1533 struct kobj_attribute *attr, const char *buf, size_t len)
1535 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1537 HSTATE_ATTR(nr_hugepages);
1542 * hstate attribute for optionally mempolicy-based constraint on persistent
1543 * huge page alloc/free.
1545 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1546 struct kobj_attribute *attr, char *buf)
1548 return nr_hugepages_show_common(kobj, attr, buf);
1551 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1552 struct kobj_attribute *attr, const char *buf, size_t len)
1554 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1556 HSTATE_ATTR(nr_hugepages_mempolicy);
1560 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1561 struct kobj_attribute *attr, char *buf)
1563 struct hstate *h = kobj_to_hstate(kobj, NULL);
1564 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1567 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1568 struct kobj_attribute *attr, const char *buf, size_t count)
1571 unsigned long input;
1572 struct hstate *h = kobj_to_hstate(kobj, NULL);
1574 if (h->order >= MAX_ORDER)
1577 err = strict_strtoul(buf, 10, &input);
1581 spin_lock(&hugetlb_lock);
1582 h->nr_overcommit_huge_pages = input;
1583 spin_unlock(&hugetlb_lock);
1587 HSTATE_ATTR(nr_overcommit_hugepages);
1589 static ssize_t free_hugepages_show(struct kobject *kobj,
1590 struct kobj_attribute *attr, char *buf)
1593 unsigned long free_huge_pages;
1596 h = kobj_to_hstate(kobj, &nid);
1597 if (nid == NUMA_NO_NODE)
1598 free_huge_pages = h->free_huge_pages;
1600 free_huge_pages = h->free_huge_pages_node[nid];
1602 return sprintf(buf, "%lu\n", free_huge_pages);
1604 HSTATE_ATTR_RO(free_hugepages);
1606 static ssize_t resv_hugepages_show(struct kobject *kobj,
1607 struct kobj_attribute *attr, char *buf)
1609 struct hstate *h = kobj_to_hstate(kobj, NULL);
1610 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1612 HSTATE_ATTR_RO(resv_hugepages);
1614 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1615 struct kobj_attribute *attr, char *buf)
1618 unsigned long surplus_huge_pages;
1621 h = kobj_to_hstate(kobj, &nid);
1622 if (nid == NUMA_NO_NODE)
1623 surplus_huge_pages = h->surplus_huge_pages;
1625 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1627 return sprintf(buf, "%lu\n", surplus_huge_pages);
1629 HSTATE_ATTR_RO(surplus_hugepages);
1631 static struct attribute *hstate_attrs[] = {
1632 &nr_hugepages_attr.attr,
1633 &nr_overcommit_hugepages_attr.attr,
1634 &free_hugepages_attr.attr,
1635 &resv_hugepages_attr.attr,
1636 &surplus_hugepages_attr.attr,
1638 &nr_hugepages_mempolicy_attr.attr,
1643 static struct attribute_group hstate_attr_group = {
1644 .attrs = hstate_attrs,
1647 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1648 struct kobject **hstate_kobjs,
1649 struct attribute_group *hstate_attr_group)
1652 int hi = hstate_index(h);
1654 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1655 if (!hstate_kobjs[hi])
1658 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1660 kobject_put(hstate_kobjs[hi]);
1665 static void __init hugetlb_sysfs_init(void)
1670 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1671 if (!hugepages_kobj)
1674 for_each_hstate(h) {
1675 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1676 hstate_kobjs, &hstate_attr_group);
1678 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1686 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1687 * with node devices in node_devices[] using a parallel array. The array
1688 * index of a node device or _hstate == node id.
1689 * This is here to avoid any static dependency of the node device driver, in
1690 * the base kernel, on the hugetlb module.
1692 struct node_hstate {
1693 struct kobject *hugepages_kobj;
1694 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1696 struct node_hstate node_hstates[MAX_NUMNODES];
1699 * A subset of global hstate attributes for node devices
1701 static struct attribute *per_node_hstate_attrs[] = {
1702 &nr_hugepages_attr.attr,
1703 &free_hugepages_attr.attr,
1704 &surplus_hugepages_attr.attr,
1708 static struct attribute_group per_node_hstate_attr_group = {
1709 .attrs = per_node_hstate_attrs,
1713 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1714 * Returns node id via non-NULL nidp.
1716 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1720 for (nid = 0; nid < nr_node_ids; nid++) {
1721 struct node_hstate *nhs = &node_hstates[nid];
1723 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1724 if (nhs->hstate_kobjs[i] == kobj) {
1736 * Unregister hstate attributes from a single node device.
1737 * No-op if no hstate attributes attached.
1739 void hugetlb_unregister_node(struct node *node)
1742 struct node_hstate *nhs = &node_hstates[node->dev.id];
1744 if (!nhs->hugepages_kobj)
1745 return; /* no hstate attributes */
1747 for_each_hstate(h) {
1748 int idx = hstate_index(h);
1749 if (nhs->hstate_kobjs[idx]) {
1750 kobject_put(nhs->hstate_kobjs[idx]);
1751 nhs->hstate_kobjs[idx] = NULL;
1755 kobject_put(nhs->hugepages_kobj);
1756 nhs->hugepages_kobj = NULL;
1760 * hugetlb module exit: unregister hstate attributes from node devices
1763 static void hugetlb_unregister_all_nodes(void)
1768 * disable node device registrations.
1770 register_hugetlbfs_with_node(NULL, NULL);
1773 * remove hstate attributes from any nodes that have them.
1775 for (nid = 0; nid < nr_node_ids; nid++)
1776 hugetlb_unregister_node(&node_devices[nid]);
1780 * Register hstate attributes for a single node device.
1781 * No-op if attributes already registered.
1783 void hugetlb_register_node(struct node *node)
1786 struct node_hstate *nhs = &node_hstates[node->dev.id];
1789 if (nhs->hugepages_kobj)
1790 return; /* already allocated */
1792 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1794 if (!nhs->hugepages_kobj)
1797 for_each_hstate(h) {
1798 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1800 &per_node_hstate_attr_group);
1802 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1804 h->name, node->dev.id);
1805 hugetlb_unregister_node(node);
1812 * hugetlb init time: register hstate attributes for all registered node
1813 * devices of nodes that have memory. All on-line nodes should have
1814 * registered their associated device by this time.
1816 static void hugetlb_register_all_nodes(void)
1820 for_each_node_state(nid, N_HIGH_MEMORY) {
1821 struct node *node = &node_devices[nid];
1822 if (node->dev.id == nid)
1823 hugetlb_register_node(node);
1827 * Let the node device driver know we're here so it can
1828 * [un]register hstate attributes on node hotplug.
1830 register_hugetlbfs_with_node(hugetlb_register_node,
1831 hugetlb_unregister_node);
1833 #else /* !CONFIG_NUMA */
1835 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1843 static void hugetlb_unregister_all_nodes(void) { }
1845 static void hugetlb_register_all_nodes(void) { }
1849 static void __exit hugetlb_exit(void)
1853 hugetlb_unregister_all_nodes();
1855 for_each_hstate(h) {
1856 kobject_put(hstate_kobjs[hstate_index(h)]);
1859 kobject_put(hugepages_kobj);
1861 module_exit(hugetlb_exit);
1863 static int __init hugetlb_init(void)
1865 /* Some platform decide whether they support huge pages at boot
1866 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1867 * there is no such support
1869 if (HPAGE_SHIFT == 0)
1872 if (!size_to_hstate(default_hstate_size)) {
1873 default_hstate_size = HPAGE_SIZE;
1874 if (!size_to_hstate(default_hstate_size))
1875 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1877 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1878 if (default_hstate_max_huge_pages)
1879 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1881 hugetlb_init_hstates();
1883 gather_bootmem_prealloc();
1887 hugetlb_sysfs_init();
1889 hugetlb_register_all_nodes();
1893 module_init(hugetlb_init);
1895 /* Should be called on processing a hugepagesz=... option */
1896 void __init hugetlb_add_hstate(unsigned order)
1901 if (size_to_hstate(PAGE_SIZE << order)) {
1902 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1905 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1907 h = &hstates[hugetlb_max_hstate++];
1909 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1910 h->nr_huge_pages = 0;
1911 h->free_huge_pages = 0;
1912 for (i = 0; i < MAX_NUMNODES; ++i)
1913 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1914 INIT_LIST_HEAD(&h->hugepage_activelist);
1915 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1916 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1917 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1918 huge_page_size(h)/1024);
1923 static int __init hugetlb_nrpages_setup(char *s)
1926 static unsigned long *last_mhp;
1929 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1930 * so this hugepages= parameter goes to the "default hstate".
1932 if (!hugetlb_max_hstate)
1933 mhp = &default_hstate_max_huge_pages;
1935 mhp = &parsed_hstate->max_huge_pages;
1937 if (mhp == last_mhp) {
1938 printk(KERN_WARNING "hugepages= specified twice without "
1939 "interleaving hugepagesz=, ignoring\n");
1943 if (sscanf(s, "%lu", mhp) <= 0)
1947 * Global state is always initialized later in hugetlb_init.
1948 * But we need to allocate >= MAX_ORDER hstates here early to still
1949 * use the bootmem allocator.
1951 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1952 hugetlb_hstate_alloc_pages(parsed_hstate);
1958 __setup("hugepages=", hugetlb_nrpages_setup);
1960 static int __init hugetlb_default_setup(char *s)
1962 default_hstate_size = memparse(s, &s);
1965 __setup("default_hugepagesz=", hugetlb_default_setup);
1967 static unsigned int cpuset_mems_nr(unsigned int *array)
1970 unsigned int nr = 0;
1972 for_each_node_mask(node, cpuset_current_mems_allowed)
1978 #ifdef CONFIG_SYSCTL
1979 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1980 struct ctl_table *table, int write,
1981 void __user *buffer, size_t *length, loff_t *ppos)
1983 struct hstate *h = &default_hstate;
1987 tmp = h->max_huge_pages;
1989 if (write && h->order >= MAX_ORDER)
1993 table->maxlen = sizeof(unsigned long);
1994 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1999 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2000 GFP_KERNEL | __GFP_NORETRY);
2001 if (!(obey_mempolicy &&
2002 init_nodemask_of_mempolicy(nodes_allowed))) {
2003 NODEMASK_FREE(nodes_allowed);
2004 nodes_allowed = &node_states[N_HIGH_MEMORY];
2006 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2008 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2009 NODEMASK_FREE(nodes_allowed);
2015 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2016 void __user *buffer, size_t *length, loff_t *ppos)
2019 return hugetlb_sysctl_handler_common(false, table, write,
2020 buffer, length, ppos);
2024 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2025 void __user *buffer, size_t *length, loff_t *ppos)
2027 return hugetlb_sysctl_handler_common(true, table, write,
2028 buffer, length, ppos);
2030 #endif /* CONFIG_NUMA */
2032 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2033 void __user *buffer,
2034 size_t *length, loff_t *ppos)
2036 proc_dointvec(table, write, buffer, length, ppos);
2037 if (hugepages_treat_as_movable)
2038 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2040 htlb_alloc_mask = GFP_HIGHUSER;
2044 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2045 void __user *buffer,
2046 size_t *length, loff_t *ppos)
2048 struct hstate *h = &default_hstate;
2052 tmp = h->nr_overcommit_huge_pages;
2054 if (write && h->order >= MAX_ORDER)
2058 table->maxlen = sizeof(unsigned long);
2059 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2064 spin_lock(&hugetlb_lock);
2065 h->nr_overcommit_huge_pages = tmp;
2066 spin_unlock(&hugetlb_lock);
2072 #endif /* CONFIG_SYSCTL */
2074 void hugetlb_report_meminfo(struct seq_file *m)
2076 struct hstate *h = &default_hstate;
2078 "HugePages_Total: %5lu\n"
2079 "HugePages_Free: %5lu\n"
2080 "HugePages_Rsvd: %5lu\n"
2081 "HugePages_Surp: %5lu\n"
2082 "Hugepagesize: %8lu kB\n",
2086 h->surplus_huge_pages,
2087 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2090 int hugetlb_report_node_meminfo(int nid, char *buf)
2092 struct hstate *h = &default_hstate;
2094 "Node %d HugePages_Total: %5u\n"
2095 "Node %d HugePages_Free: %5u\n"
2096 "Node %d HugePages_Surp: %5u\n",
2097 nid, h->nr_huge_pages_node[nid],
2098 nid, h->free_huge_pages_node[nid],
2099 nid, h->surplus_huge_pages_node[nid]);
2102 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2103 unsigned long hugetlb_total_pages(void)
2105 struct hstate *h = &default_hstate;
2106 return h->nr_huge_pages * pages_per_huge_page(h);
2109 static int hugetlb_acct_memory(struct hstate *h, long delta)
2113 spin_lock(&hugetlb_lock);
2115 * When cpuset is configured, it breaks the strict hugetlb page
2116 * reservation as the accounting is done on a global variable. Such
2117 * reservation is completely rubbish in the presence of cpuset because
2118 * the reservation is not checked against page availability for the
2119 * current cpuset. Application can still potentially OOM'ed by kernel
2120 * with lack of free htlb page in cpuset that the task is in.
2121 * Attempt to enforce strict accounting with cpuset is almost
2122 * impossible (or too ugly) because cpuset is too fluid that
2123 * task or memory node can be dynamically moved between cpusets.
2125 * The change of semantics for shared hugetlb mapping with cpuset is
2126 * undesirable. However, in order to preserve some of the semantics,
2127 * we fall back to check against current free page availability as
2128 * a best attempt and hopefully to minimize the impact of changing
2129 * semantics that cpuset has.
2132 if (gather_surplus_pages(h, delta) < 0)
2135 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2136 return_unused_surplus_pages(h, delta);
2143 return_unused_surplus_pages(h, (unsigned long) -delta);
2146 spin_unlock(&hugetlb_lock);
2150 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2152 struct resv_map *reservations = vma_resv_map(vma);
2155 * This new VMA should share its siblings reservation map if present.
2156 * The VMA will only ever have a valid reservation map pointer where
2157 * it is being copied for another still existing VMA. As that VMA
2158 * has a reference to the reservation map it cannot disappear until
2159 * after this open call completes. It is therefore safe to take a
2160 * new reference here without additional locking.
2163 kref_get(&reservations->refs);
2166 static void resv_map_put(struct vm_area_struct *vma)
2168 struct resv_map *reservations = vma_resv_map(vma);
2172 kref_put(&reservations->refs, resv_map_release);
2175 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2177 struct hstate *h = hstate_vma(vma);
2178 struct resv_map *reservations = vma_resv_map(vma);
2179 struct hugepage_subpool *spool = subpool_vma(vma);
2180 unsigned long reserve;
2181 unsigned long start;
2185 start = vma_hugecache_offset(h, vma, vma->vm_start);
2186 end = vma_hugecache_offset(h, vma, vma->vm_end);
2188 reserve = (end - start) -
2189 region_count(&reservations->regions, start, end);
2194 hugetlb_acct_memory(h, -reserve);
2195 hugepage_subpool_put_pages(spool, reserve);
2201 * We cannot handle pagefaults against hugetlb pages at all. They cause
2202 * handle_mm_fault() to try to instantiate regular-sized pages in the
2203 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2206 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2212 const struct vm_operations_struct hugetlb_vm_ops = {
2213 .fault = hugetlb_vm_op_fault,
2214 .open = hugetlb_vm_op_open,
2215 .close = hugetlb_vm_op_close,
2218 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2225 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2227 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2229 entry = pte_mkyoung(entry);
2230 entry = pte_mkhuge(entry);
2231 entry = arch_make_huge_pte(entry, vma, page, writable);
2236 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2237 unsigned long address, pte_t *ptep)
2241 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2242 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2243 update_mmu_cache(vma, address, ptep);
2247 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2248 struct vm_area_struct *vma)
2250 pte_t *src_pte, *dst_pte, entry;
2251 struct page *ptepage;
2254 struct hstate *h = hstate_vma(vma);
2255 unsigned long sz = huge_page_size(h);
2257 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2259 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2260 src_pte = huge_pte_offset(src, addr);
2263 dst_pte = huge_pte_alloc(dst, addr, sz);
2267 /* If the pagetables are shared don't copy or take references */
2268 if (dst_pte == src_pte)
2271 spin_lock(&dst->page_table_lock);
2272 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2273 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2275 huge_ptep_set_wrprotect(src, addr, src_pte);
2276 entry = huge_ptep_get(src_pte);
2277 ptepage = pte_page(entry);
2279 page_dup_rmap(ptepage);
2280 set_huge_pte_at(dst, addr, dst_pte, entry);
2282 spin_unlock(&src->page_table_lock);
2283 spin_unlock(&dst->page_table_lock);
2291 static int is_hugetlb_entry_migration(pte_t pte)
2295 if (huge_pte_none(pte) || pte_present(pte))
2297 swp = pte_to_swp_entry(pte);
2298 if (non_swap_entry(swp) && is_migration_entry(swp))
2304 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2308 if (huge_pte_none(pte) || pte_present(pte))
2310 swp = pte_to_swp_entry(pte);
2311 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2317 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2318 unsigned long start, unsigned long end,
2319 struct page *ref_page)
2321 int force_flush = 0;
2322 struct mm_struct *mm = vma->vm_mm;
2323 unsigned long address;
2327 struct hstate *h = hstate_vma(vma);
2328 unsigned long sz = huge_page_size(h);
2330 WARN_ON(!is_vm_hugetlb_page(vma));
2331 BUG_ON(start & ~huge_page_mask(h));
2332 BUG_ON(end & ~huge_page_mask(h));
2334 tlb_start_vma(tlb, vma);
2335 mmu_notifier_invalidate_range_start(mm, start, end);
2337 spin_lock(&mm->page_table_lock);
2338 for (address = start; address < end; address += sz) {
2339 ptep = huge_pte_offset(mm, address);
2343 if (huge_pmd_unshare(mm, &address, ptep))
2346 pte = huge_ptep_get(ptep);
2347 if (huge_pte_none(pte))
2351 * HWPoisoned hugepage is already unmapped and dropped reference
2353 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2356 page = pte_page(pte);
2358 * If a reference page is supplied, it is because a specific
2359 * page is being unmapped, not a range. Ensure the page we
2360 * are about to unmap is the actual page of interest.
2363 if (page != ref_page)
2367 * Mark the VMA as having unmapped its page so that
2368 * future faults in this VMA will fail rather than
2369 * looking like data was lost
2371 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2374 pte = huge_ptep_get_and_clear(mm, address, ptep);
2375 tlb_remove_tlb_entry(tlb, ptep, address);
2377 set_page_dirty(page);
2379 page_remove_rmap(page);
2380 force_flush = !__tlb_remove_page(tlb, page);
2383 /* Bail out after unmapping reference page if supplied */
2387 spin_unlock(&mm->page_table_lock);
2389 * mmu_gather ran out of room to batch pages, we break out of
2390 * the PTE lock to avoid doing the potential expensive TLB invalidate
2391 * and page-free while holding it.
2396 if (address < end && !ref_page)
2399 mmu_notifier_invalidate_range_end(mm, start, end);
2400 tlb_end_vma(tlb, vma);
2403 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2404 unsigned long end, struct page *ref_page)
2406 struct mm_struct *mm;
2407 struct mmu_gather tlb;
2411 tlb_gather_mmu(&tlb, mm, 0);
2412 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2413 tlb_finish_mmu(&tlb, start, end);
2417 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2418 * mappping it owns the reserve page for. The intention is to unmap the page
2419 * from other VMAs and let the children be SIGKILLed if they are faulting the
2422 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2423 struct page *page, unsigned long address)
2425 struct hstate *h = hstate_vma(vma);
2426 struct vm_area_struct *iter_vma;
2427 struct address_space *mapping;
2428 struct prio_tree_iter iter;
2432 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2433 * from page cache lookup which is in HPAGE_SIZE units.
2435 address = address & huge_page_mask(h);
2436 pgoff = vma_hugecache_offset(h, vma, address);
2437 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2440 * Take the mapping lock for the duration of the table walk. As
2441 * this mapping should be shared between all the VMAs,
2442 * __unmap_hugepage_range() is called as the lock is already held
2444 mutex_lock(&mapping->i_mmap_mutex);
2445 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2446 /* Do not unmap the current VMA */
2447 if (iter_vma == vma)
2451 * Unmap the page from other VMAs without their own reserves.
2452 * They get marked to be SIGKILLed if they fault in these
2453 * areas. This is because a future no-page fault on this VMA
2454 * could insert a zeroed page instead of the data existing
2455 * from the time of fork. This would look like data corruption
2457 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2458 unmap_hugepage_range(iter_vma, address,
2459 address + huge_page_size(h), page);
2461 mutex_unlock(&mapping->i_mmap_mutex);
2467 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2468 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2469 * cannot race with other handlers or page migration.
2470 * Keep the pte_same checks anyway to make transition from the mutex easier.
2472 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2473 unsigned long address, pte_t *ptep, pte_t pte,
2474 struct page *pagecache_page)
2476 struct hstate *h = hstate_vma(vma);
2477 struct page *old_page, *new_page;
2479 int outside_reserve = 0;
2481 old_page = pte_page(pte);
2484 /* If no-one else is actually using this page, avoid the copy
2485 * and just make the page writable */
2486 avoidcopy = (page_mapcount(old_page) == 1);
2488 if (PageAnon(old_page))
2489 page_move_anon_rmap(old_page, vma, address);
2490 set_huge_ptep_writable(vma, address, ptep);
2495 * If the process that created a MAP_PRIVATE mapping is about to
2496 * perform a COW due to a shared page count, attempt to satisfy
2497 * the allocation without using the existing reserves. The pagecache
2498 * page is used to determine if the reserve at this address was
2499 * consumed or not. If reserves were used, a partial faulted mapping
2500 * at the time of fork() could consume its reserves on COW instead
2501 * of the full address range.
2503 if (!(vma->vm_flags & VM_MAYSHARE) &&
2504 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2505 old_page != pagecache_page)
2506 outside_reserve = 1;
2508 page_cache_get(old_page);
2510 /* Drop page_table_lock as buddy allocator may be called */
2511 spin_unlock(&mm->page_table_lock);
2512 new_page = alloc_huge_page(vma, address, outside_reserve);
2514 if (IS_ERR(new_page)) {
2515 long err = PTR_ERR(new_page);
2516 page_cache_release(old_page);
2519 * If a process owning a MAP_PRIVATE mapping fails to COW,
2520 * it is due to references held by a child and an insufficient
2521 * huge page pool. To guarantee the original mappers
2522 * reliability, unmap the page from child processes. The child
2523 * may get SIGKILLed if it later faults.
2525 if (outside_reserve) {
2526 BUG_ON(huge_pte_none(pte));
2527 if (unmap_ref_private(mm, vma, old_page, address)) {
2528 BUG_ON(huge_pte_none(pte));
2529 spin_lock(&mm->page_table_lock);
2530 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2531 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2532 goto retry_avoidcopy;
2534 * race occurs while re-acquiring page_table_lock, and
2542 /* Caller expects lock to be held */
2543 spin_lock(&mm->page_table_lock);
2545 return VM_FAULT_OOM;
2547 return VM_FAULT_SIGBUS;
2551 * When the original hugepage is shared one, it does not have
2552 * anon_vma prepared.
2554 if (unlikely(anon_vma_prepare(vma))) {
2555 page_cache_release(new_page);
2556 page_cache_release(old_page);
2557 /* Caller expects lock to be held */
2558 spin_lock(&mm->page_table_lock);
2559 return VM_FAULT_OOM;
2562 copy_user_huge_page(new_page, old_page, address, vma,
2563 pages_per_huge_page(h));
2564 __SetPageUptodate(new_page);
2567 * Retake the page_table_lock to check for racing updates
2568 * before the page tables are altered
2570 spin_lock(&mm->page_table_lock);
2571 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2572 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2574 mmu_notifier_invalidate_range_start(mm,
2575 address & huge_page_mask(h),
2576 (address & huge_page_mask(h)) + huge_page_size(h));
2577 huge_ptep_clear_flush(vma, address, ptep);
2578 set_huge_pte_at(mm, address, ptep,
2579 make_huge_pte(vma, new_page, 1));
2580 page_remove_rmap(old_page);
2581 hugepage_add_new_anon_rmap(new_page, vma, address);
2582 /* Make the old page be freed below */
2583 new_page = old_page;
2584 mmu_notifier_invalidate_range_end(mm,
2585 address & huge_page_mask(h),
2586 (address & huge_page_mask(h)) + huge_page_size(h));
2588 page_cache_release(new_page);
2589 page_cache_release(old_page);
2593 /* Return the pagecache page at a given address within a VMA */
2594 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2595 struct vm_area_struct *vma, unsigned long address)
2597 struct address_space *mapping;
2600 mapping = vma->vm_file->f_mapping;
2601 idx = vma_hugecache_offset(h, vma, address);
2603 return find_lock_page(mapping, idx);
2607 * Return whether there is a pagecache page to back given address within VMA.
2608 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2610 static bool hugetlbfs_pagecache_present(struct hstate *h,
2611 struct vm_area_struct *vma, unsigned long address)
2613 struct address_space *mapping;
2617 mapping = vma->vm_file->f_mapping;
2618 idx = vma_hugecache_offset(h, vma, address);
2620 page = find_get_page(mapping, idx);
2623 return page != NULL;
2626 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2627 unsigned long address, pte_t *ptep, unsigned int flags)
2629 struct hstate *h = hstate_vma(vma);
2630 int ret = VM_FAULT_SIGBUS;
2635 struct address_space *mapping;
2639 * Currently, we are forced to kill the process in the event the
2640 * original mapper has unmapped pages from the child due to a failed
2641 * COW. Warn that such a situation has occurred as it may not be obvious
2643 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2645 "PID %d killed due to inadequate hugepage pool\n",
2650 mapping = vma->vm_file->f_mapping;
2651 idx = vma_hugecache_offset(h, vma, address);
2654 * Use page lock to guard against racing truncation
2655 * before we get page_table_lock.
2658 page = find_lock_page(mapping, idx);
2660 size = i_size_read(mapping->host) >> huge_page_shift(h);
2663 page = alloc_huge_page(vma, address, 0);
2665 ret = PTR_ERR(page);
2669 ret = VM_FAULT_SIGBUS;
2672 clear_huge_page(page, address, pages_per_huge_page(h));
2673 __SetPageUptodate(page);
2675 if (vma->vm_flags & VM_MAYSHARE) {
2677 struct inode *inode = mapping->host;
2679 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2687 spin_lock(&inode->i_lock);
2688 inode->i_blocks += blocks_per_huge_page(h);
2689 spin_unlock(&inode->i_lock);
2692 if (unlikely(anon_vma_prepare(vma))) {
2694 goto backout_unlocked;
2700 * If memory error occurs between mmap() and fault, some process
2701 * don't have hwpoisoned swap entry for errored virtual address.
2702 * So we need to block hugepage fault by PG_hwpoison bit check.
2704 if (unlikely(PageHWPoison(page))) {
2705 ret = VM_FAULT_HWPOISON |
2706 VM_FAULT_SET_HINDEX(hstate_index(h));
2707 goto backout_unlocked;
2712 * If we are going to COW a private mapping later, we examine the
2713 * pending reservations for this page now. This will ensure that
2714 * any allocations necessary to record that reservation occur outside
2717 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2718 if (vma_needs_reservation(h, vma, address) < 0) {
2720 goto backout_unlocked;
2723 spin_lock(&mm->page_table_lock);
2724 size = i_size_read(mapping->host) >> huge_page_shift(h);
2729 if (!huge_pte_none(huge_ptep_get(ptep)))
2733 hugepage_add_new_anon_rmap(page, vma, address);
2735 page_dup_rmap(page);
2736 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2737 && (vma->vm_flags & VM_SHARED)));
2738 set_huge_pte_at(mm, address, ptep, new_pte);
2740 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2741 /* Optimization, do the COW without a second fault */
2742 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2745 spin_unlock(&mm->page_table_lock);
2751 spin_unlock(&mm->page_table_lock);
2758 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2759 unsigned long address, unsigned int flags)
2764 struct page *page = NULL;
2765 struct page *pagecache_page = NULL;
2766 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2767 struct hstate *h = hstate_vma(vma);
2769 address &= huge_page_mask(h);
2771 ptep = huge_pte_offset(mm, address);
2773 entry = huge_ptep_get(ptep);
2774 if (unlikely(is_hugetlb_entry_migration(entry))) {
2775 migration_entry_wait(mm, (pmd_t *)ptep, address);
2777 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2778 return VM_FAULT_HWPOISON_LARGE |
2779 VM_FAULT_SET_HINDEX(hstate_index(h));
2782 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2784 return VM_FAULT_OOM;
2787 * Serialize hugepage allocation and instantiation, so that we don't
2788 * get spurious allocation failures if two CPUs race to instantiate
2789 * the same page in the page cache.
2791 mutex_lock(&hugetlb_instantiation_mutex);
2792 entry = huge_ptep_get(ptep);
2793 if (huge_pte_none(entry)) {
2794 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2801 * If we are going to COW the mapping later, we examine the pending
2802 * reservations for this page now. This will ensure that any
2803 * allocations necessary to record that reservation occur outside the
2804 * spinlock. For private mappings, we also lookup the pagecache
2805 * page now as it is used to determine if a reservation has been
2808 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2809 if (vma_needs_reservation(h, vma, address) < 0) {
2814 if (!(vma->vm_flags & VM_MAYSHARE))
2815 pagecache_page = hugetlbfs_pagecache_page(h,
2820 * hugetlb_cow() requires page locks of pte_page(entry) and
2821 * pagecache_page, so here we need take the former one
2822 * when page != pagecache_page or !pagecache_page.
2823 * Note that locking order is always pagecache_page -> page,
2824 * so no worry about deadlock.
2826 page = pte_page(entry);
2828 if (page != pagecache_page)
2831 spin_lock(&mm->page_table_lock);
2832 /* Check for a racing update before calling hugetlb_cow */
2833 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2834 goto out_page_table_lock;
2837 if (flags & FAULT_FLAG_WRITE) {
2838 if (!pte_write(entry)) {
2839 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2841 goto out_page_table_lock;
2843 entry = pte_mkdirty(entry);
2845 entry = pte_mkyoung(entry);
2846 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2847 flags & FAULT_FLAG_WRITE))
2848 update_mmu_cache(vma, address, ptep);
2850 out_page_table_lock:
2851 spin_unlock(&mm->page_table_lock);
2853 if (pagecache_page) {
2854 unlock_page(pagecache_page);
2855 put_page(pagecache_page);
2857 if (page != pagecache_page)
2862 mutex_unlock(&hugetlb_instantiation_mutex);
2867 /* Can be overriden by architectures */
2868 __attribute__((weak)) struct page *
2869 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2870 pud_t *pud, int write)
2876 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2877 struct page **pages, struct vm_area_struct **vmas,
2878 unsigned long *position, int *length, int i,
2881 unsigned long pfn_offset;
2882 unsigned long vaddr = *position;
2883 int remainder = *length;
2884 struct hstate *h = hstate_vma(vma);
2886 spin_lock(&mm->page_table_lock);
2887 while (vaddr < vma->vm_end && remainder) {
2893 * Some archs (sparc64, sh*) have multiple pte_ts to
2894 * each hugepage. We have to make sure we get the
2895 * first, for the page indexing below to work.
2897 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2898 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2901 * When coredumping, it suits get_dump_page if we just return
2902 * an error where there's an empty slot with no huge pagecache
2903 * to back it. This way, we avoid allocating a hugepage, and
2904 * the sparse dumpfile avoids allocating disk blocks, but its
2905 * huge holes still show up with zeroes where they need to be.
2907 if (absent && (flags & FOLL_DUMP) &&
2908 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2914 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2917 spin_unlock(&mm->page_table_lock);
2918 ret = hugetlb_fault(mm, vma, vaddr,
2919 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2920 spin_lock(&mm->page_table_lock);
2921 if (!(ret & VM_FAULT_ERROR))
2928 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2929 page = pte_page(huge_ptep_get(pte));
2932 pages[i] = mem_map_offset(page, pfn_offset);
2943 if (vaddr < vma->vm_end && remainder &&
2944 pfn_offset < pages_per_huge_page(h)) {
2946 * We use pfn_offset to avoid touching the pageframes
2947 * of this compound page.
2952 spin_unlock(&mm->page_table_lock);
2953 *length = remainder;
2956 return i ? i : -EFAULT;
2959 void hugetlb_change_protection(struct vm_area_struct *vma,
2960 unsigned long address, unsigned long end, pgprot_t newprot)
2962 struct mm_struct *mm = vma->vm_mm;
2963 unsigned long start = address;
2966 struct hstate *h = hstate_vma(vma);
2968 BUG_ON(address >= end);
2969 flush_cache_range(vma, address, end);
2971 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2972 spin_lock(&mm->page_table_lock);
2973 for (; address < end; address += huge_page_size(h)) {
2974 ptep = huge_pte_offset(mm, address);
2977 if (huge_pmd_unshare(mm, &address, ptep))
2979 if (!huge_pte_none(huge_ptep_get(ptep))) {
2980 pte = huge_ptep_get_and_clear(mm, address, ptep);
2981 pte = pte_mkhuge(pte_modify(pte, newprot));
2982 set_huge_pte_at(mm, address, ptep, pte);
2985 spin_unlock(&mm->page_table_lock);
2986 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2988 flush_tlb_range(vma, start, end);
2991 int hugetlb_reserve_pages(struct inode *inode,
2993 struct vm_area_struct *vma,
2994 vm_flags_t vm_flags)
2997 struct hstate *h = hstate_inode(inode);
2998 struct hugepage_subpool *spool = subpool_inode(inode);
3001 * Only apply hugepage reservation if asked. At fault time, an
3002 * attempt will be made for VM_NORESERVE to allocate a page
3003 * without using reserves
3005 if (vm_flags & VM_NORESERVE)
3009 * Shared mappings base their reservation on the number of pages that
3010 * are already allocated on behalf of the file. Private mappings need
3011 * to reserve the full area even if read-only as mprotect() may be
3012 * called to make the mapping read-write. Assume !vma is a shm mapping
3014 if (!vma || vma->vm_flags & VM_MAYSHARE)
3015 chg = region_chg(&inode->i_mapping->private_list, from, to);
3017 struct resv_map *resv_map = resv_map_alloc();
3023 set_vma_resv_map(vma, resv_map);
3024 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3032 /* There must be enough pages in the subpool for the mapping */
3033 if (hugepage_subpool_get_pages(spool, chg)) {
3039 * Check enough hugepages are available for the reservation.
3040 * Hand the pages back to the subpool if there are not
3042 ret = hugetlb_acct_memory(h, chg);
3044 hugepage_subpool_put_pages(spool, chg);
3049 * Account for the reservations made. Shared mappings record regions
3050 * that have reservations as they are shared by multiple VMAs.
3051 * When the last VMA disappears, the region map says how much
3052 * the reservation was and the page cache tells how much of
3053 * the reservation was consumed. Private mappings are per-VMA and
3054 * only the consumed reservations are tracked. When the VMA
3055 * disappears, the original reservation is the VMA size and the
3056 * consumed reservations are stored in the map. Hence, nothing
3057 * else has to be done for private mappings here
3059 if (!vma || vma->vm_flags & VM_MAYSHARE)
3060 region_add(&inode->i_mapping->private_list, from, to);
3068 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3070 struct hstate *h = hstate_inode(inode);
3071 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3072 struct hugepage_subpool *spool = subpool_inode(inode);
3074 spin_lock(&inode->i_lock);
3075 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3076 spin_unlock(&inode->i_lock);
3078 hugepage_subpool_put_pages(spool, (chg - freed));
3079 hugetlb_acct_memory(h, -(chg - freed));
3082 #ifdef CONFIG_MEMORY_FAILURE
3084 /* Should be called in hugetlb_lock */
3085 static int is_hugepage_on_freelist(struct page *hpage)
3089 struct hstate *h = page_hstate(hpage);
3090 int nid = page_to_nid(hpage);
3092 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3099 * This function is called from memory failure code.
3100 * Assume the caller holds page lock of the head page.
3102 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3104 struct hstate *h = page_hstate(hpage);
3105 int nid = page_to_nid(hpage);
3108 spin_lock(&hugetlb_lock);
3109 if (is_hugepage_on_freelist(hpage)) {
3110 list_del(&hpage->lru);
3111 set_page_refcounted(hpage);
3112 h->free_huge_pages--;
3113 h->free_huge_pages_node[nid]--;
3116 spin_unlock(&hugetlb_lock);