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 hugetlb_cgroup_uncharge_page(hstate_index(h),
631 pages_per_huge_page(h), page);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 /* remove the page from active list */
634 list_del(&page->lru);
635 update_and_free_page(h, page);
636 h->surplus_huge_pages--;
637 h->surplus_huge_pages_node[nid]--;
639 enqueue_huge_page(h, page);
641 spin_unlock(&hugetlb_lock);
642 hugepage_subpool_put_pages(spool, 1);
645 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
647 INIT_LIST_HEAD(&page->lru);
648 set_compound_page_dtor(page, free_huge_page);
649 spin_lock(&hugetlb_lock);
650 set_hugetlb_cgroup(page, NULL);
652 h->nr_huge_pages_node[nid]++;
653 spin_unlock(&hugetlb_lock);
654 put_page(page); /* free it into the hugepage allocator */
657 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
660 int nr_pages = 1 << order;
661 struct page *p = page + 1;
663 /* we rely on prep_new_huge_page to set the destructor */
664 set_compound_order(page, order);
666 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
668 set_page_count(p, 0);
669 p->first_page = page;
673 int PageHuge(struct page *page)
675 compound_page_dtor *dtor;
677 if (!PageCompound(page))
680 page = compound_head(page);
681 dtor = get_compound_page_dtor(page);
683 return dtor == free_huge_page;
685 EXPORT_SYMBOL_GPL(PageHuge);
687 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
691 if (h->order >= MAX_ORDER)
694 page = alloc_pages_exact_node(nid,
695 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
696 __GFP_REPEAT|__GFP_NOWARN,
699 if (arch_prepare_hugepage(page)) {
700 __free_pages(page, huge_page_order(h));
703 prep_new_huge_page(h, page, nid);
710 * common helper functions for hstate_next_node_to_{alloc|free}.
711 * We may have allocated or freed a huge page based on a different
712 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
713 * be outside of *nodes_allowed. Ensure that we use an allowed
714 * node for alloc or free.
716 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
718 nid = next_node(nid, *nodes_allowed);
719 if (nid == MAX_NUMNODES)
720 nid = first_node(*nodes_allowed);
721 VM_BUG_ON(nid >= MAX_NUMNODES);
726 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
728 if (!node_isset(nid, *nodes_allowed))
729 nid = next_node_allowed(nid, nodes_allowed);
734 * returns the previously saved node ["this node"] from which to
735 * allocate a persistent huge page for the pool and advance the
736 * next node from which to allocate, handling wrap at end of node
739 static int hstate_next_node_to_alloc(struct hstate *h,
740 nodemask_t *nodes_allowed)
744 VM_BUG_ON(!nodes_allowed);
746 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
747 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
752 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
759 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
760 next_nid = start_nid;
763 page = alloc_fresh_huge_page_node(h, next_nid);
768 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
769 } while (next_nid != start_nid);
772 count_vm_event(HTLB_BUDDY_PGALLOC);
774 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
780 * helper for free_pool_huge_page() - return the previously saved
781 * node ["this node"] from which to free a huge page. Advance the
782 * next node id whether or not we find a free huge page to free so
783 * that the next attempt to free addresses the next node.
785 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
789 VM_BUG_ON(!nodes_allowed);
791 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
792 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
798 * Free huge page from pool from next node to free.
799 * Attempt to keep persistent huge pages more or less
800 * balanced over allowed nodes.
801 * Called with hugetlb_lock locked.
803 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
810 start_nid = hstate_next_node_to_free(h, nodes_allowed);
811 next_nid = start_nid;
815 * If we're returning unused surplus pages, only examine
816 * nodes with surplus pages.
818 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
819 !list_empty(&h->hugepage_freelists[next_nid])) {
821 list_entry(h->hugepage_freelists[next_nid].next,
823 list_del(&page->lru);
824 h->free_huge_pages--;
825 h->free_huge_pages_node[next_nid]--;
827 h->surplus_huge_pages--;
828 h->surplus_huge_pages_node[next_nid]--;
830 update_and_free_page(h, page);
834 next_nid = hstate_next_node_to_free(h, nodes_allowed);
835 } while (next_nid != start_nid);
840 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
845 if (h->order >= MAX_ORDER)
849 * Assume we will successfully allocate the surplus page to
850 * prevent racing processes from causing the surplus to exceed
853 * This however introduces a different race, where a process B
854 * tries to grow the static hugepage pool while alloc_pages() is
855 * called by process A. B will only examine the per-node
856 * counters in determining if surplus huge pages can be
857 * converted to normal huge pages in adjust_pool_surplus(). A
858 * won't be able to increment the per-node counter, until the
859 * lock is dropped by B, but B doesn't drop hugetlb_lock until
860 * no more huge pages can be converted from surplus to normal
861 * state (and doesn't try to convert again). Thus, we have a
862 * case where a surplus huge page exists, the pool is grown, and
863 * the surplus huge page still exists after, even though it
864 * should just have been converted to a normal huge page. This
865 * does not leak memory, though, as the hugepage will be freed
866 * once it is out of use. It also does not allow the counters to
867 * go out of whack in adjust_pool_surplus() as we don't modify
868 * the node values until we've gotten the hugepage and only the
869 * per-node value is checked there.
871 spin_lock(&hugetlb_lock);
872 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
873 spin_unlock(&hugetlb_lock);
877 h->surplus_huge_pages++;
879 spin_unlock(&hugetlb_lock);
881 if (nid == NUMA_NO_NODE)
882 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
883 __GFP_REPEAT|__GFP_NOWARN,
886 page = alloc_pages_exact_node(nid,
887 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
888 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
890 if (page && arch_prepare_hugepage(page)) {
891 __free_pages(page, huge_page_order(h));
895 spin_lock(&hugetlb_lock);
897 INIT_LIST_HEAD(&page->lru);
898 r_nid = page_to_nid(page);
899 set_compound_page_dtor(page, free_huge_page);
900 set_hugetlb_cgroup(page, NULL);
902 * We incremented the global counters already
904 h->nr_huge_pages_node[r_nid]++;
905 h->surplus_huge_pages_node[r_nid]++;
906 __count_vm_event(HTLB_BUDDY_PGALLOC);
909 h->surplus_huge_pages--;
910 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
912 spin_unlock(&hugetlb_lock);
918 * This allocation function is useful in the context where vma is irrelevant.
919 * E.g. soft-offlining uses this function because it only cares physical
920 * address of error page.
922 struct page *alloc_huge_page_node(struct hstate *h, int nid)
926 spin_lock(&hugetlb_lock);
927 page = dequeue_huge_page_node(h, nid);
928 spin_unlock(&hugetlb_lock);
931 page = alloc_buddy_huge_page(h, nid);
937 * Increase the hugetlb pool such that it can accommodate a reservation
940 static int gather_surplus_pages(struct hstate *h, int delta)
942 struct list_head surplus_list;
943 struct page *page, *tmp;
945 int needed, allocated;
946 bool alloc_ok = true;
948 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
950 h->resv_huge_pages += delta;
955 INIT_LIST_HEAD(&surplus_list);
959 spin_unlock(&hugetlb_lock);
960 for (i = 0; i < needed; i++) {
961 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
966 list_add(&page->lru, &surplus_list);
971 * After retaking hugetlb_lock, we need to recalculate 'needed'
972 * because either resv_huge_pages or free_huge_pages may have changed.
974 spin_lock(&hugetlb_lock);
975 needed = (h->resv_huge_pages + delta) -
976 (h->free_huge_pages + allocated);
981 * We were not able to allocate enough pages to
982 * satisfy the entire reservation so we free what
983 * we've allocated so far.
988 * The surplus_list now contains _at_least_ the number of extra pages
989 * needed to accommodate the reservation. Add the appropriate number
990 * of pages to the hugetlb pool and free the extras back to the buddy
991 * allocator. Commit the entire reservation here to prevent another
992 * process from stealing the pages as they are added to the pool but
993 * before they are reserved.
996 h->resv_huge_pages += delta;
999 /* Free the needed pages to the hugetlb pool */
1000 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1004 * This page is now managed by the hugetlb allocator and has
1005 * no users -- drop the buddy allocator's reference.
1007 put_page_testzero(page);
1008 VM_BUG_ON(page_count(page));
1009 enqueue_huge_page(h, page);
1012 spin_unlock(&hugetlb_lock);
1014 /* Free unnecessary surplus pages to the buddy allocator */
1015 if (!list_empty(&surplus_list)) {
1016 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1020 spin_lock(&hugetlb_lock);
1026 * When releasing a hugetlb pool reservation, any surplus pages that were
1027 * allocated to satisfy the reservation must be explicitly freed if they were
1029 * Called with hugetlb_lock held.
1031 static void return_unused_surplus_pages(struct hstate *h,
1032 unsigned long unused_resv_pages)
1034 unsigned long nr_pages;
1036 /* Uncommit the reservation */
1037 h->resv_huge_pages -= unused_resv_pages;
1039 /* Cannot return gigantic pages currently */
1040 if (h->order >= MAX_ORDER)
1043 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1046 * We want to release as many surplus pages as possible, spread
1047 * evenly across all nodes with memory. Iterate across these nodes
1048 * until we can no longer free unreserved surplus pages. This occurs
1049 * when the nodes with surplus pages have no free pages.
1050 * free_pool_huge_page() will balance the the freed pages across the
1051 * on-line nodes with memory and will handle the hstate accounting.
1053 while (nr_pages--) {
1054 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1060 * Determine if the huge page at addr within the vma has an associated
1061 * reservation. Where it does not we will need to logically increase
1062 * reservation and actually increase subpool usage before an allocation
1063 * can occur. Where any new reservation would be required the
1064 * reservation change is prepared, but not committed. Once the page
1065 * has been allocated from the subpool and instantiated the change should
1066 * be committed via vma_commit_reservation. No action is required on
1069 static long vma_needs_reservation(struct hstate *h,
1070 struct vm_area_struct *vma, unsigned long addr)
1072 struct address_space *mapping = vma->vm_file->f_mapping;
1073 struct inode *inode = mapping->host;
1075 if (vma->vm_flags & VM_MAYSHARE) {
1076 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1077 return region_chg(&inode->i_mapping->private_list,
1080 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1085 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1086 struct resv_map *reservations = vma_resv_map(vma);
1088 err = region_chg(&reservations->regions, idx, idx + 1);
1094 static void vma_commit_reservation(struct hstate *h,
1095 struct vm_area_struct *vma, unsigned long addr)
1097 struct address_space *mapping = vma->vm_file->f_mapping;
1098 struct inode *inode = mapping->host;
1100 if (vma->vm_flags & VM_MAYSHARE) {
1101 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1102 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1104 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1105 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1106 struct resv_map *reservations = vma_resv_map(vma);
1108 /* Mark this page used in the map. */
1109 region_add(&reservations->regions, idx, idx + 1);
1113 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1114 unsigned long addr, int avoid_reserve)
1116 struct hugepage_subpool *spool = subpool_vma(vma);
1117 struct hstate *h = hstate_vma(vma);
1121 struct hugetlb_cgroup *h_cg;
1123 idx = hstate_index(h);
1125 * Processes that did not create the mapping will have no
1126 * reserves and will not have accounted against subpool
1127 * limit. Check that the subpool limit can be made before
1128 * satisfying the allocation MAP_NORESERVE mappings may also
1129 * need pages and subpool limit allocated allocated if no reserve
1132 chg = vma_needs_reservation(h, vma, addr);
1134 return ERR_PTR(-ENOMEM);
1136 if (hugepage_subpool_get_pages(spool, chg))
1137 return ERR_PTR(-ENOSPC);
1139 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1141 hugepage_subpool_put_pages(spool, chg);
1142 return ERR_PTR(-ENOSPC);
1144 spin_lock(&hugetlb_lock);
1145 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1146 spin_unlock(&hugetlb_lock);
1149 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1151 hugetlb_cgroup_uncharge_cgroup(idx,
1152 pages_per_huge_page(h),
1154 hugepage_subpool_put_pages(spool, chg);
1155 return ERR_PTR(-ENOSPC);
1159 set_page_private(page, (unsigned long)spool);
1161 vma_commit_reservation(h, vma, addr);
1162 /* update page cgroup details */
1163 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1167 int __weak alloc_bootmem_huge_page(struct hstate *h)
1169 struct huge_bootmem_page *m;
1170 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1175 addr = __alloc_bootmem_node_nopanic(
1176 NODE_DATA(hstate_next_node_to_alloc(h,
1177 &node_states[N_HIGH_MEMORY])),
1178 huge_page_size(h), huge_page_size(h), 0);
1182 * Use the beginning of the huge page to store the
1183 * huge_bootmem_page struct (until gather_bootmem
1184 * puts them into the mem_map).
1194 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1195 /* Put them into a private list first because mem_map is not up yet */
1196 list_add(&m->list, &huge_boot_pages);
1201 static void prep_compound_huge_page(struct page *page, int order)
1203 if (unlikely(order > (MAX_ORDER - 1)))
1204 prep_compound_gigantic_page(page, order);
1206 prep_compound_page(page, order);
1209 /* Put bootmem huge pages into the standard lists after mem_map is up */
1210 static void __init gather_bootmem_prealloc(void)
1212 struct huge_bootmem_page *m;
1214 list_for_each_entry(m, &huge_boot_pages, list) {
1215 struct hstate *h = m->hstate;
1218 #ifdef CONFIG_HIGHMEM
1219 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1220 free_bootmem_late((unsigned long)m,
1221 sizeof(struct huge_bootmem_page));
1223 page = virt_to_page(m);
1225 __ClearPageReserved(page);
1226 WARN_ON(page_count(page) != 1);
1227 prep_compound_huge_page(page, h->order);
1228 prep_new_huge_page(h, page, page_to_nid(page));
1230 * If we had gigantic hugepages allocated at boot time, we need
1231 * to restore the 'stolen' pages to totalram_pages in order to
1232 * fix confusing memory reports from free(1) and another
1233 * side-effects, like CommitLimit going negative.
1235 if (h->order > (MAX_ORDER - 1))
1236 totalram_pages += 1 << h->order;
1240 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1244 for (i = 0; i < h->max_huge_pages; ++i) {
1245 if (h->order >= MAX_ORDER) {
1246 if (!alloc_bootmem_huge_page(h))
1248 } else if (!alloc_fresh_huge_page(h,
1249 &node_states[N_HIGH_MEMORY]))
1252 h->max_huge_pages = i;
1255 static void __init hugetlb_init_hstates(void)
1259 for_each_hstate(h) {
1260 /* oversize hugepages were init'ed in early boot */
1261 if (h->order < MAX_ORDER)
1262 hugetlb_hstate_alloc_pages(h);
1266 static char * __init memfmt(char *buf, unsigned long n)
1268 if (n >= (1UL << 30))
1269 sprintf(buf, "%lu GB", n >> 30);
1270 else if (n >= (1UL << 20))
1271 sprintf(buf, "%lu MB", n >> 20);
1273 sprintf(buf, "%lu KB", n >> 10);
1277 static void __init report_hugepages(void)
1281 for_each_hstate(h) {
1283 printk(KERN_INFO "HugeTLB registered %s page size, "
1284 "pre-allocated %ld pages\n",
1285 memfmt(buf, huge_page_size(h)),
1286 h->free_huge_pages);
1290 #ifdef CONFIG_HIGHMEM
1291 static void try_to_free_low(struct hstate *h, unsigned long count,
1292 nodemask_t *nodes_allowed)
1296 if (h->order >= MAX_ORDER)
1299 for_each_node_mask(i, *nodes_allowed) {
1300 struct page *page, *next;
1301 struct list_head *freel = &h->hugepage_freelists[i];
1302 list_for_each_entry_safe(page, next, freel, lru) {
1303 if (count >= h->nr_huge_pages)
1305 if (PageHighMem(page))
1307 list_del(&page->lru);
1308 update_and_free_page(h, page);
1309 h->free_huge_pages--;
1310 h->free_huge_pages_node[page_to_nid(page)]--;
1315 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1316 nodemask_t *nodes_allowed)
1322 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1323 * balanced by operating on them in a round-robin fashion.
1324 * Returns 1 if an adjustment was made.
1326 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1329 int start_nid, next_nid;
1332 VM_BUG_ON(delta != -1 && delta != 1);
1335 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1337 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1338 next_nid = start_nid;
1344 * To shrink on this node, there must be a surplus page
1346 if (!h->surplus_huge_pages_node[nid]) {
1347 next_nid = hstate_next_node_to_alloc(h,
1354 * Surplus cannot exceed the total number of pages
1356 if (h->surplus_huge_pages_node[nid] >=
1357 h->nr_huge_pages_node[nid]) {
1358 next_nid = hstate_next_node_to_free(h,
1364 h->surplus_huge_pages += delta;
1365 h->surplus_huge_pages_node[nid] += delta;
1368 } while (next_nid != start_nid);
1373 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1374 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1375 nodemask_t *nodes_allowed)
1377 unsigned long min_count, ret;
1379 if (h->order >= MAX_ORDER)
1380 return h->max_huge_pages;
1383 * Increase the pool size
1384 * First take pages out of surplus state. Then make up the
1385 * remaining difference by allocating fresh huge pages.
1387 * We might race with alloc_buddy_huge_page() here and be unable
1388 * to convert a surplus huge page to a normal huge page. That is
1389 * not critical, though, it just means the overall size of the
1390 * pool might be one hugepage larger than it needs to be, but
1391 * within all the constraints specified by the sysctls.
1393 spin_lock(&hugetlb_lock);
1394 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1395 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1399 while (count > persistent_huge_pages(h)) {
1401 * If this allocation races such that we no longer need the
1402 * page, free_huge_page will handle it by freeing the page
1403 * and reducing the surplus.
1405 spin_unlock(&hugetlb_lock);
1406 ret = alloc_fresh_huge_page(h, nodes_allowed);
1407 spin_lock(&hugetlb_lock);
1411 /* Bail for signals. Probably ctrl-c from user */
1412 if (signal_pending(current))
1417 * Decrease the pool size
1418 * First return free pages to the buddy allocator (being careful
1419 * to keep enough around to satisfy reservations). Then place
1420 * pages into surplus state as needed so the pool will shrink
1421 * to the desired size as pages become free.
1423 * By placing pages into the surplus state independent of the
1424 * overcommit value, we are allowing the surplus pool size to
1425 * exceed overcommit. There are few sane options here. Since
1426 * alloc_buddy_huge_page() is checking the global counter,
1427 * though, we'll note that we're not allowed to exceed surplus
1428 * and won't grow the pool anywhere else. Not until one of the
1429 * sysctls are changed, or the surplus pages go out of use.
1431 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1432 min_count = max(count, min_count);
1433 try_to_free_low(h, min_count, nodes_allowed);
1434 while (min_count < persistent_huge_pages(h)) {
1435 if (!free_pool_huge_page(h, nodes_allowed, 0))
1438 while (count < persistent_huge_pages(h)) {
1439 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1443 ret = persistent_huge_pages(h);
1444 spin_unlock(&hugetlb_lock);
1448 #define HSTATE_ATTR_RO(_name) \
1449 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1451 #define HSTATE_ATTR(_name) \
1452 static struct kobj_attribute _name##_attr = \
1453 __ATTR(_name, 0644, _name##_show, _name##_store)
1455 static struct kobject *hugepages_kobj;
1456 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1458 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1460 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1464 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1465 if (hstate_kobjs[i] == kobj) {
1467 *nidp = NUMA_NO_NODE;
1471 return kobj_to_node_hstate(kobj, nidp);
1474 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1475 struct kobj_attribute *attr, char *buf)
1478 unsigned long nr_huge_pages;
1481 h = kobj_to_hstate(kobj, &nid);
1482 if (nid == NUMA_NO_NODE)
1483 nr_huge_pages = h->nr_huge_pages;
1485 nr_huge_pages = h->nr_huge_pages_node[nid];
1487 return sprintf(buf, "%lu\n", nr_huge_pages);
1490 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1491 struct kobject *kobj, struct kobj_attribute *attr,
1492 const char *buf, size_t len)
1496 unsigned long count;
1498 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1500 err = strict_strtoul(buf, 10, &count);
1504 h = kobj_to_hstate(kobj, &nid);
1505 if (h->order >= MAX_ORDER) {
1510 if (nid == NUMA_NO_NODE) {
1512 * global hstate attribute
1514 if (!(obey_mempolicy &&
1515 init_nodemask_of_mempolicy(nodes_allowed))) {
1516 NODEMASK_FREE(nodes_allowed);
1517 nodes_allowed = &node_states[N_HIGH_MEMORY];
1519 } else if (nodes_allowed) {
1521 * per node hstate attribute: adjust count to global,
1522 * but restrict alloc/free to the specified node.
1524 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1525 init_nodemask_of_node(nodes_allowed, nid);
1527 nodes_allowed = &node_states[N_HIGH_MEMORY];
1529 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1531 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1532 NODEMASK_FREE(nodes_allowed);
1536 NODEMASK_FREE(nodes_allowed);
1540 static ssize_t nr_hugepages_show(struct kobject *kobj,
1541 struct kobj_attribute *attr, char *buf)
1543 return nr_hugepages_show_common(kobj, attr, buf);
1546 static ssize_t nr_hugepages_store(struct kobject *kobj,
1547 struct kobj_attribute *attr, const char *buf, size_t len)
1549 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1551 HSTATE_ATTR(nr_hugepages);
1556 * hstate attribute for optionally mempolicy-based constraint on persistent
1557 * huge page alloc/free.
1559 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1560 struct kobj_attribute *attr, char *buf)
1562 return nr_hugepages_show_common(kobj, attr, buf);
1565 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1566 struct kobj_attribute *attr, const char *buf, size_t len)
1568 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1570 HSTATE_ATTR(nr_hugepages_mempolicy);
1574 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1575 struct kobj_attribute *attr, char *buf)
1577 struct hstate *h = kobj_to_hstate(kobj, NULL);
1578 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1581 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1582 struct kobj_attribute *attr, const char *buf, size_t count)
1585 unsigned long input;
1586 struct hstate *h = kobj_to_hstate(kobj, NULL);
1588 if (h->order >= MAX_ORDER)
1591 err = strict_strtoul(buf, 10, &input);
1595 spin_lock(&hugetlb_lock);
1596 h->nr_overcommit_huge_pages = input;
1597 spin_unlock(&hugetlb_lock);
1601 HSTATE_ATTR(nr_overcommit_hugepages);
1603 static ssize_t free_hugepages_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1607 unsigned long free_huge_pages;
1610 h = kobj_to_hstate(kobj, &nid);
1611 if (nid == NUMA_NO_NODE)
1612 free_huge_pages = h->free_huge_pages;
1614 free_huge_pages = h->free_huge_pages_node[nid];
1616 return sprintf(buf, "%lu\n", free_huge_pages);
1618 HSTATE_ATTR_RO(free_hugepages);
1620 static ssize_t resv_hugepages_show(struct kobject *kobj,
1621 struct kobj_attribute *attr, char *buf)
1623 struct hstate *h = kobj_to_hstate(kobj, NULL);
1624 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1626 HSTATE_ATTR_RO(resv_hugepages);
1628 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1629 struct kobj_attribute *attr, char *buf)
1632 unsigned long surplus_huge_pages;
1635 h = kobj_to_hstate(kobj, &nid);
1636 if (nid == NUMA_NO_NODE)
1637 surplus_huge_pages = h->surplus_huge_pages;
1639 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1641 return sprintf(buf, "%lu\n", surplus_huge_pages);
1643 HSTATE_ATTR_RO(surplus_hugepages);
1645 static struct attribute *hstate_attrs[] = {
1646 &nr_hugepages_attr.attr,
1647 &nr_overcommit_hugepages_attr.attr,
1648 &free_hugepages_attr.attr,
1649 &resv_hugepages_attr.attr,
1650 &surplus_hugepages_attr.attr,
1652 &nr_hugepages_mempolicy_attr.attr,
1657 static struct attribute_group hstate_attr_group = {
1658 .attrs = hstate_attrs,
1661 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1662 struct kobject **hstate_kobjs,
1663 struct attribute_group *hstate_attr_group)
1666 int hi = hstate_index(h);
1668 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1669 if (!hstate_kobjs[hi])
1672 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1674 kobject_put(hstate_kobjs[hi]);
1679 static void __init hugetlb_sysfs_init(void)
1684 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1685 if (!hugepages_kobj)
1688 for_each_hstate(h) {
1689 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1690 hstate_kobjs, &hstate_attr_group);
1692 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1700 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1701 * with node devices in node_devices[] using a parallel array. The array
1702 * index of a node device or _hstate == node id.
1703 * This is here to avoid any static dependency of the node device driver, in
1704 * the base kernel, on the hugetlb module.
1706 struct node_hstate {
1707 struct kobject *hugepages_kobj;
1708 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1710 struct node_hstate node_hstates[MAX_NUMNODES];
1713 * A subset of global hstate attributes for node devices
1715 static struct attribute *per_node_hstate_attrs[] = {
1716 &nr_hugepages_attr.attr,
1717 &free_hugepages_attr.attr,
1718 &surplus_hugepages_attr.attr,
1722 static struct attribute_group per_node_hstate_attr_group = {
1723 .attrs = per_node_hstate_attrs,
1727 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1728 * Returns node id via non-NULL nidp.
1730 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1734 for (nid = 0; nid < nr_node_ids; nid++) {
1735 struct node_hstate *nhs = &node_hstates[nid];
1737 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1738 if (nhs->hstate_kobjs[i] == kobj) {
1750 * Unregister hstate attributes from a single node device.
1751 * No-op if no hstate attributes attached.
1753 void hugetlb_unregister_node(struct node *node)
1756 struct node_hstate *nhs = &node_hstates[node->dev.id];
1758 if (!nhs->hugepages_kobj)
1759 return; /* no hstate attributes */
1761 for_each_hstate(h) {
1762 int idx = hstate_index(h);
1763 if (nhs->hstate_kobjs[idx]) {
1764 kobject_put(nhs->hstate_kobjs[idx]);
1765 nhs->hstate_kobjs[idx] = NULL;
1769 kobject_put(nhs->hugepages_kobj);
1770 nhs->hugepages_kobj = NULL;
1774 * hugetlb module exit: unregister hstate attributes from node devices
1777 static void hugetlb_unregister_all_nodes(void)
1782 * disable node device registrations.
1784 register_hugetlbfs_with_node(NULL, NULL);
1787 * remove hstate attributes from any nodes that have them.
1789 for (nid = 0; nid < nr_node_ids; nid++)
1790 hugetlb_unregister_node(&node_devices[nid]);
1794 * Register hstate attributes for a single node device.
1795 * No-op if attributes already registered.
1797 void hugetlb_register_node(struct node *node)
1800 struct node_hstate *nhs = &node_hstates[node->dev.id];
1803 if (nhs->hugepages_kobj)
1804 return; /* already allocated */
1806 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1808 if (!nhs->hugepages_kobj)
1811 for_each_hstate(h) {
1812 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1814 &per_node_hstate_attr_group);
1816 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1818 h->name, node->dev.id);
1819 hugetlb_unregister_node(node);
1826 * hugetlb init time: register hstate attributes for all registered node
1827 * devices of nodes that have memory. All on-line nodes should have
1828 * registered their associated device by this time.
1830 static void hugetlb_register_all_nodes(void)
1834 for_each_node_state(nid, N_HIGH_MEMORY) {
1835 struct node *node = &node_devices[nid];
1836 if (node->dev.id == nid)
1837 hugetlb_register_node(node);
1841 * Let the node device driver know we're here so it can
1842 * [un]register hstate attributes on node hotplug.
1844 register_hugetlbfs_with_node(hugetlb_register_node,
1845 hugetlb_unregister_node);
1847 #else /* !CONFIG_NUMA */
1849 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1857 static void hugetlb_unregister_all_nodes(void) { }
1859 static void hugetlb_register_all_nodes(void) { }
1863 static void __exit hugetlb_exit(void)
1867 hugetlb_unregister_all_nodes();
1869 for_each_hstate(h) {
1870 kobject_put(hstate_kobjs[hstate_index(h)]);
1873 kobject_put(hugepages_kobj);
1875 module_exit(hugetlb_exit);
1877 static int __init hugetlb_init(void)
1879 /* Some platform decide whether they support huge pages at boot
1880 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1881 * there is no such support
1883 if (HPAGE_SHIFT == 0)
1886 if (!size_to_hstate(default_hstate_size)) {
1887 default_hstate_size = HPAGE_SIZE;
1888 if (!size_to_hstate(default_hstate_size))
1889 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1891 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1892 if (default_hstate_max_huge_pages)
1893 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1895 hugetlb_init_hstates();
1897 gather_bootmem_prealloc();
1901 hugetlb_sysfs_init();
1903 hugetlb_register_all_nodes();
1907 module_init(hugetlb_init);
1909 /* Should be called on processing a hugepagesz=... option */
1910 void __init hugetlb_add_hstate(unsigned order)
1915 if (size_to_hstate(PAGE_SIZE << order)) {
1916 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1919 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1921 h = &hstates[hugetlb_max_hstate++];
1923 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1924 h->nr_huge_pages = 0;
1925 h->free_huge_pages = 0;
1926 for (i = 0; i < MAX_NUMNODES; ++i)
1927 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1928 INIT_LIST_HEAD(&h->hugepage_activelist);
1929 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1930 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1931 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1932 huge_page_size(h)/1024);
1937 static int __init hugetlb_nrpages_setup(char *s)
1940 static unsigned long *last_mhp;
1943 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1944 * so this hugepages= parameter goes to the "default hstate".
1946 if (!hugetlb_max_hstate)
1947 mhp = &default_hstate_max_huge_pages;
1949 mhp = &parsed_hstate->max_huge_pages;
1951 if (mhp == last_mhp) {
1952 printk(KERN_WARNING "hugepages= specified twice without "
1953 "interleaving hugepagesz=, ignoring\n");
1957 if (sscanf(s, "%lu", mhp) <= 0)
1961 * Global state is always initialized later in hugetlb_init.
1962 * But we need to allocate >= MAX_ORDER hstates here early to still
1963 * use the bootmem allocator.
1965 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1966 hugetlb_hstate_alloc_pages(parsed_hstate);
1972 __setup("hugepages=", hugetlb_nrpages_setup);
1974 static int __init hugetlb_default_setup(char *s)
1976 default_hstate_size = memparse(s, &s);
1979 __setup("default_hugepagesz=", hugetlb_default_setup);
1981 static unsigned int cpuset_mems_nr(unsigned int *array)
1984 unsigned int nr = 0;
1986 for_each_node_mask(node, cpuset_current_mems_allowed)
1992 #ifdef CONFIG_SYSCTL
1993 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1994 struct ctl_table *table, int write,
1995 void __user *buffer, size_t *length, loff_t *ppos)
1997 struct hstate *h = &default_hstate;
2001 tmp = h->max_huge_pages;
2003 if (write && h->order >= MAX_ORDER)
2007 table->maxlen = sizeof(unsigned long);
2008 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2013 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2014 GFP_KERNEL | __GFP_NORETRY);
2015 if (!(obey_mempolicy &&
2016 init_nodemask_of_mempolicy(nodes_allowed))) {
2017 NODEMASK_FREE(nodes_allowed);
2018 nodes_allowed = &node_states[N_HIGH_MEMORY];
2020 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2022 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2023 NODEMASK_FREE(nodes_allowed);
2029 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2030 void __user *buffer, size_t *length, loff_t *ppos)
2033 return hugetlb_sysctl_handler_common(false, table, write,
2034 buffer, length, ppos);
2038 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2039 void __user *buffer, size_t *length, loff_t *ppos)
2041 return hugetlb_sysctl_handler_common(true, table, write,
2042 buffer, length, ppos);
2044 #endif /* CONFIG_NUMA */
2046 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2047 void __user *buffer,
2048 size_t *length, loff_t *ppos)
2050 proc_dointvec(table, write, buffer, length, ppos);
2051 if (hugepages_treat_as_movable)
2052 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2054 htlb_alloc_mask = GFP_HIGHUSER;
2058 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2059 void __user *buffer,
2060 size_t *length, loff_t *ppos)
2062 struct hstate *h = &default_hstate;
2066 tmp = h->nr_overcommit_huge_pages;
2068 if (write && h->order >= MAX_ORDER)
2072 table->maxlen = sizeof(unsigned long);
2073 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2078 spin_lock(&hugetlb_lock);
2079 h->nr_overcommit_huge_pages = tmp;
2080 spin_unlock(&hugetlb_lock);
2086 #endif /* CONFIG_SYSCTL */
2088 void hugetlb_report_meminfo(struct seq_file *m)
2090 struct hstate *h = &default_hstate;
2092 "HugePages_Total: %5lu\n"
2093 "HugePages_Free: %5lu\n"
2094 "HugePages_Rsvd: %5lu\n"
2095 "HugePages_Surp: %5lu\n"
2096 "Hugepagesize: %8lu kB\n",
2100 h->surplus_huge_pages,
2101 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2104 int hugetlb_report_node_meminfo(int nid, char *buf)
2106 struct hstate *h = &default_hstate;
2108 "Node %d HugePages_Total: %5u\n"
2109 "Node %d HugePages_Free: %5u\n"
2110 "Node %d HugePages_Surp: %5u\n",
2111 nid, h->nr_huge_pages_node[nid],
2112 nid, h->free_huge_pages_node[nid],
2113 nid, h->surplus_huge_pages_node[nid]);
2116 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2117 unsigned long hugetlb_total_pages(void)
2119 struct hstate *h = &default_hstate;
2120 return h->nr_huge_pages * pages_per_huge_page(h);
2123 static int hugetlb_acct_memory(struct hstate *h, long delta)
2127 spin_lock(&hugetlb_lock);
2129 * When cpuset is configured, it breaks the strict hugetlb page
2130 * reservation as the accounting is done on a global variable. Such
2131 * reservation is completely rubbish in the presence of cpuset because
2132 * the reservation is not checked against page availability for the
2133 * current cpuset. Application can still potentially OOM'ed by kernel
2134 * with lack of free htlb page in cpuset that the task is in.
2135 * Attempt to enforce strict accounting with cpuset is almost
2136 * impossible (or too ugly) because cpuset is too fluid that
2137 * task or memory node can be dynamically moved between cpusets.
2139 * The change of semantics for shared hugetlb mapping with cpuset is
2140 * undesirable. However, in order to preserve some of the semantics,
2141 * we fall back to check against current free page availability as
2142 * a best attempt and hopefully to minimize the impact of changing
2143 * semantics that cpuset has.
2146 if (gather_surplus_pages(h, delta) < 0)
2149 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2150 return_unused_surplus_pages(h, delta);
2157 return_unused_surplus_pages(h, (unsigned long) -delta);
2160 spin_unlock(&hugetlb_lock);
2164 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2166 struct resv_map *reservations = vma_resv_map(vma);
2169 * This new VMA should share its siblings reservation map if present.
2170 * The VMA will only ever have a valid reservation map pointer where
2171 * it is being copied for another still existing VMA. As that VMA
2172 * has a reference to the reservation map it cannot disappear until
2173 * after this open call completes. It is therefore safe to take a
2174 * new reference here without additional locking.
2177 kref_get(&reservations->refs);
2180 static void resv_map_put(struct vm_area_struct *vma)
2182 struct resv_map *reservations = vma_resv_map(vma);
2186 kref_put(&reservations->refs, resv_map_release);
2189 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2191 struct hstate *h = hstate_vma(vma);
2192 struct resv_map *reservations = vma_resv_map(vma);
2193 struct hugepage_subpool *spool = subpool_vma(vma);
2194 unsigned long reserve;
2195 unsigned long start;
2199 start = vma_hugecache_offset(h, vma, vma->vm_start);
2200 end = vma_hugecache_offset(h, vma, vma->vm_end);
2202 reserve = (end - start) -
2203 region_count(&reservations->regions, start, end);
2208 hugetlb_acct_memory(h, -reserve);
2209 hugepage_subpool_put_pages(spool, reserve);
2215 * We cannot handle pagefaults against hugetlb pages at all. They cause
2216 * handle_mm_fault() to try to instantiate regular-sized pages in the
2217 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2220 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2226 const struct vm_operations_struct hugetlb_vm_ops = {
2227 .fault = hugetlb_vm_op_fault,
2228 .open = hugetlb_vm_op_open,
2229 .close = hugetlb_vm_op_close,
2232 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2239 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2241 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2243 entry = pte_mkyoung(entry);
2244 entry = pte_mkhuge(entry);
2245 entry = arch_make_huge_pte(entry, vma, page, writable);
2250 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2251 unsigned long address, pte_t *ptep)
2255 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2256 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2257 update_mmu_cache(vma, address, ptep);
2261 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2262 struct vm_area_struct *vma)
2264 pte_t *src_pte, *dst_pte, entry;
2265 struct page *ptepage;
2268 struct hstate *h = hstate_vma(vma);
2269 unsigned long sz = huge_page_size(h);
2271 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2273 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2274 src_pte = huge_pte_offset(src, addr);
2277 dst_pte = huge_pte_alloc(dst, addr, sz);
2281 /* If the pagetables are shared don't copy or take references */
2282 if (dst_pte == src_pte)
2285 spin_lock(&dst->page_table_lock);
2286 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2287 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2289 huge_ptep_set_wrprotect(src, addr, src_pte);
2290 entry = huge_ptep_get(src_pte);
2291 ptepage = pte_page(entry);
2293 page_dup_rmap(ptepage);
2294 set_huge_pte_at(dst, addr, dst_pte, entry);
2296 spin_unlock(&src->page_table_lock);
2297 spin_unlock(&dst->page_table_lock);
2305 static int is_hugetlb_entry_migration(pte_t pte)
2309 if (huge_pte_none(pte) || pte_present(pte))
2311 swp = pte_to_swp_entry(pte);
2312 if (non_swap_entry(swp) && is_migration_entry(swp))
2318 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2322 if (huge_pte_none(pte) || pte_present(pte))
2324 swp = pte_to_swp_entry(pte);
2325 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2331 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2332 unsigned long start, unsigned long end,
2333 struct page *ref_page)
2335 int force_flush = 0;
2336 struct mm_struct *mm = vma->vm_mm;
2337 unsigned long address;
2341 struct hstate *h = hstate_vma(vma);
2342 unsigned long sz = huge_page_size(h);
2344 WARN_ON(!is_vm_hugetlb_page(vma));
2345 BUG_ON(start & ~huge_page_mask(h));
2346 BUG_ON(end & ~huge_page_mask(h));
2348 tlb_start_vma(tlb, vma);
2349 mmu_notifier_invalidate_range_start(mm, start, end);
2351 spin_lock(&mm->page_table_lock);
2352 for (address = start; address < end; address += sz) {
2353 ptep = huge_pte_offset(mm, address);
2357 if (huge_pmd_unshare(mm, &address, ptep))
2360 pte = huge_ptep_get(ptep);
2361 if (huge_pte_none(pte))
2365 * HWPoisoned hugepage is already unmapped and dropped reference
2367 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2370 page = pte_page(pte);
2372 * If a reference page is supplied, it is because a specific
2373 * page is being unmapped, not a range. Ensure the page we
2374 * are about to unmap is the actual page of interest.
2377 if (page != ref_page)
2381 * Mark the VMA as having unmapped its page so that
2382 * future faults in this VMA will fail rather than
2383 * looking like data was lost
2385 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2388 pte = huge_ptep_get_and_clear(mm, address, ptep);
2389 tlb_remove_tlb_entry(tlb, ptep, address);
2391 set_page_dirty(page);
2393 page_remove_rmap(page);
2394 force_flush = !__tlb_remove_page(tlb, page);
2397 /* Bail out after unmapping reference page if supplied */
2401 spin_unlock(&mm->page_table_lock);
2403 * mmu_gather ran out of room to batch pages, we break out of
2404 * the PTE lock to avoid doing the potential expensive TLB invalidate
2405 * and page-free while holding it.
2410 if (address < end && !ref_page)
2413 mmu_notifier_invalidate_range_end(mm, start, end);
2414 tlb_end_vma(tlb, vma);
2417 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2418 unsigned long end, struct page *ref_page)
2420 struct mm_struct *mm;
2421 struct mmu_gather tlb;
2425 tlb_gather_mmu(&tlb, mm, 0);
2426 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2427 tlb_finish_mmu(&tlb, start, end);
2431 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2432 * mappping it owns the reserve page for. The intention is to unmap the page
2433 * from other VMAs and let the children be SIGKILLed if they are faulting the
2436 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2437 struct page *page, unsigned long address)
2439 struct hstate *h = hstate_vma(vma);
2440 struct vm_area_struct *iter_vma;
2441 struct address_space *mapping;
2442 struct prio_tree_iter iter;
2446 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2447 * from page cache lookup which is in HPAGE_SIZE units.
2449 address = address & huge_page_mask(h);
2450 pgoff = vma_hugecache_offset(h, vma, address);
2451 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2454 * Take the mapping lock for the duration of the table walk. As
2455 * this mapping should be shared between all the VMAs,
2456 * __unmap_hugepage_range() is called as the lock is already held
2458 mutex_lock(&mapping->i_mmap_mutex);
2459 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2460 /* Do not unmap the current VMA */
2461 if (iter_vma == vma)
2465 * Unmap the page from other VMAs without their own reserves.
2466 * They get marked to be SIGKILLed if they fault in these
2467 * areas. This is because a future no-page fault on this VMA
2468 * could insert a zeroed page instead of the data existing
2469 * from the time of fork. This would look like data corruption
2471 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2472 unmap_hugepage_range(iter_vma, address,
2473 address + huge_page_size(h), page);
2475 mutex_unlock(&mapping->i_mmap_mutex);
2481 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2482 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2483 * cannot race with other handlers or page migration.
2484 * Keep the pte_same checks anyway to make transition from the mutex easier.
2486 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2487 unsigned long address, pte_t *ptep, pte_t pte,
2488 struct page *pagecache_page)
2490 struct hstate *h = hstate_vma(vma);
2491 struct page *old_page, *new_page;
2493 int outside_reserve = 0;
2495 old_page = pte_page(pte);
2498 /* If no-one else is actually using this page, avoid the copy
2499 * and just make the page writable */
2500 avoidcopy = (page_mapcount(old_page) == 1);
2502 if (PageAnon(old_page))
2503 page_move_anon_rmap(old_page, vma, address);
2504 set_huge_ptep_writable(vma, address, ptep);
2509 * If the process that created a MAP_PRIVATE mapping is about to
2510 * perform a COW due to a shared page count, attempt to satisfy
2511 * the allocation without using the existing reserves. The pagecache
2512 * page is used to determine if the reserve at this address was
2513 * consumed or not. If reserves were used, a partial faulted mapping
2514 * at the time of fork() could consume its reserves on COW instead
2515 * of the full address range.
2517 if (!(vma->vm_flags & VM_MAYSHARE) &&
2518 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2519 old_page != pagecache_page)
2520 outside_reserve = 1;
2522 page_cache_get(old_page);
2524 /* Drop page_table_lock as buddy allocator may be called */
2525 spin_unlock(&mm->page_table_lock);
2526 new_page = alloc_huge_page(vma, address, outside_reserve);
2528 if (IS_ERR(new_page)) {
2529 long err = PTR_ERR(new_page);
2530 page_cache_release(old_page);
2533 * If a process owning a MAP_PRIVATE mapping fails to COW,
2534 * it is due to references held by a child and an insufficient
2535 * huge page pool. To guarantee the original mappers
2536 * reliability, unmap the page from child processes. The child
2537 * may get SIGKILLed if it later faults.
2539 if (outside_reserve) {
2540 BUG_ON(huge_pte_none(pte));
2541 if (unmap_ref_private(mm, vma, old_page, address)) {
2542 BUG_ON(huge_pte_none(pte));
2543 spin_lock(&mm->page_table_lock);
2544 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2545 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2546 goto retry_avoidcopy;
2548 * race occurs while re-acquiring page_table_lock, and
2556 /* Caller expects lock to be held */
2557 spin_lock(&mm->page_table_lock);
2559 return VM_FAULT_OOM;
2561 return VM_FAULT_SIGBUS;
2565 * When the original hugepage is shared one, it does not have
2566 * anon_vma prepared.
2568 if (unlikely(anon_vma_prepare(vma))) {
2569 page_cache_release(new_page);
2570 page_cache_release(old_page);
2571 /* Caller expects lock to be held */
2572 spin_lock(&mm->page_table_lock);
2573 return VM_FAULT_OOM;
2576 copy_user_huge_page(new_page, old_page, address, vma,
2577 pages_per_huge_page(h));
2578 __SetPageUptodate(new_page);
2581 * Retake the page_table_lock to check for racing updates
2582 * before the page tables are altered
2584 spin_lock(&mm->page_table_lock);
2585 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2586 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2588 mmu_notifier_invalidate_range_start(mm,
2589 address & huge_page_mask(h),
2590 (address & huge_page_mask(h)) + huge_page_size(h));
2591 huge_ptep_clear_flush(vma, address, ptep);
2592 set_huge_pte_at(mm, address, ptep,
2593 make_huge_pte(vma, new_page, 1));
2594 page_remove_rmap(old_page);
2595 hugepage_add_new_anon_rmap(new_page, vma, address);
2596 /* Make the old page be freed below */
2597 new_page = old_page;
2598 mmu_notifier_invalidate_range_end(mm,
2599 address & huge_page_mask(h),
2600 (address & huge_page_mask(h)) + huge_page_size(h));
2602 page_cache_release(new_page);
2603 page_cache_release(old_page);
2607 /* Return the pagecache page at a given address within a VMA */
2608 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2609 struct vm_area_struct *vma, unsigned long address)
2611 struct address_space *mapping;
2614 mapping = vma->vm_file->f_mapping;
2615 idx = vma_hugecache_offset(h, vma, address);
2617 return find_lock_page(mapping, idx);
2621 * Return whether there is a pagecache page to back given address within VMA.
2622 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2624 static bool hugetlbfs_pagecache_present(struct hstate *h,
2625 struct vm_area_struct *vma, unsigned long address)
2627 struct address_space *mapping;
2631 mapping = vma->vm_file->f_mapping;
2632 idx = vma_hugecache_offset(h, vma, address);
2634 page = find_get_page(mapping, idx);
2637 return page != NULL;
2640 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2641 unsigned long address, pte_t *ptep, unsigned int flags)
2643 struct hstate *h = hstate_vma(vma);
2644 int ret = VM_FAULT_SIGBUS;
2649 struct address_space *mapping;
2653 * Currently, we are forced to kill the process in the event the
2654 * original mapper has unmapped pages from the child due to a failed
2655 * COW. Warn that such a situation has occurred as it may not be obvious
2657 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2659 "PID %d killed due to inadequate hugepage pool\n",
2664 mapping = vma->vm_file->f_mapping;
2665 idx = vma_hugecache_offset(h, vma, address);
2668 * Use page lock to guard against racing truncation
2669 * before we get page_table_lock.
2672 page = find_lock_page(mapping, idx);
2674 size = i_size_read(mapping->host) >> huge_page_shift(h);
2677 page = alloc_huge_page(vma, address, 0);
2679 ret = PTR_ERR(page);
2683 ret = VM_FAULT_SIGBUS;
2686 clear_huge_page(page, address, pages_per_huge_page(h));
2687 __SetPageUptodate(page);
2689 if (vma->vm_flags & VM_MAYSHARE) {
2691 struct inode *inode = mapping->host;
2693 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2701 spin_lock(&inode->i_lock);
2702 inode->i_blocks += blocks_per_huge_page(h);
2703 spin_unlock(&inode->i_lock);
2706 if (unlikely(anon_vma_prepare(vma))) {
2708 goto backout_unlocked;
2714 * If memory error occurs between mmap() and fault, some process
2715 * don't have hwpoisoned swap entry for errored virtual address.
2716 * So we need to block hugepage fault by PG_hwpoison bit check.
2718 if (unlikely(PageHWPoison(page))) {
2719 ret = VM_FAULT_HWPOISON |
2720 VM_FAULT_SET_HINDEX(hstate_index(h));
2721 goto backout_unlocked;
2726 * If we are going to COW a private mapping later, we examine the
2727 * pending reservations for this page now. This will ensure that
2728 * any allocations necessary to record that reservation occur outside
2731 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2732 if (vma_needs_reservation(h, vma, address) < 0) {
2734 goto backout_unlocked;
2737 spin_lock(&mm->page_table_lock);
2738 size = i_size_read(mapping->host) >> huge_page_shift(h);
2743 if (!huge_pte_none(huge_ptep_get(ptep)))
2747 hugepage_add_new_anon_rmap(page, vma, address);
2749 page_dup_rmap(page);
2750 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2751 && (vma->vm_flags & VM_SHARED)));
2752 set_huge_pte_at(mm, address, ptep, new_pte);
2754 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2755 /* Optimization, do the COW without a second fault */
2756 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2759 spin_unlock(&mm->page_table_lock);
2765 spin_unlock(&mm->page_table_lock);
2772 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2773 unsigned long address, unsigned int flags)
2778 struct page *page = NULL;
2779 struct page *pagecache_page = NULL;
2780 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2781 struct hstate *h = hstate_vma(vma);
2783 address &= huge_page_mask(h);
2785 ptep = huge_pte_offset(mm, address);
2787 entry = huge_ptep_get(ptep);
2788 if (unlikely(is_hugetlb_entry_migration(entry))) {
2789 migration_entry_wait(mm, (pmd_t *)ptep, address);
2791 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2792 return VM_FAULT_HWPOISON_LARGE |
2793 VM_FAULT_SET_HINDEX(hstate_index(h));
2796 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2798 return VM_FAULT_OOM;
2801 * Serialize hugepage allocation and instantiation, so that we don't
2802 * get spurious allocation failures if two CPUs race to instantiate
2803 * the same page in the page cache.
2805 mutex_lock(&hugetlb_instantiation_mutex);
2806 entry = huge_ptep_get(ptep);
2807 if (huge_pte_none(entry)) {
2808 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2815 * If we are going to COW the mapping later, we examine the pending
2816 * reservations for this page now. This will ensure that any
2817 * allocations necessary to record that reservation occur outside the
2818 * spinlock. For private mappings, we also lookup the pagecache
2819 * page now as it is used to determine if a reservation has been
2822 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2823 if (vma_needs_reservation(h, vma, address) < 0) {
2828 if (!(vma->vm_flags & VM_MAYSHARE))
2829 pagecache_page = hugetlbfs_pagecache_page(h,
2834 * hugetlb_cow() requires page locks of pte_page(entry) and
2835 * pagecache_page, so here we need take the former one
2836 * when page != pagecache_page or !pagecache_page.
2837 * Note that locking order is always pagecache_page -> page,
2838 * so no worry about deadlock.
2840 page = pte_page(entry);
2842 if (page != pagecache_page)
2845 spin_lock(&mm->page_table_lock);
2846 /* Check for a racing update before calling hugetlb_cow */
2847 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2848 goto out_page_table_lock;
2851 if (flags & FAULT_FLAG_WRITE) {
2852 if (!pte_write(entry)) {
2853 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2855 goto out_page_table_lock;
2857 entry = pte_mkdirty(entry);
2859 entry = pte_mkyoung(entry);
2860 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2861 flags & FAULT_FLAG_WRITE))
2862 update_mmu_cache(vma, address, ptep);
2864 out_page_table_lock:
2865 spin_unlock(&mm->page_table_lock);
2867 if (pagecache_page) {
2868 unlock_page(pagecache_page);
2869 put_page(pagecache_page);
2871 if (page != pagecache_page)
2876 mutex_unlock(&hugetlb_instantiation_mutex);
2881 /* Can be overriden by architectures */
2882 __attribute__((weak)) struct page *
2883 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2884 pud_t *pud, int write)
2890 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2891 struct page **pages, struct vm_area_struct **vmas,
2892 unsigned long *position, int *length, int i,
2895 unsigned long pfn_offset;
2896 unsigned long vaddr = *position;
2897 int remainder = *length;
2898 struct hstate *h = hstate_vma(vma);
2900 spin_lock(&mm->page_table_lock);
2901 while (vaddr < vma->vm_end && remainder) {
2907 * Some archs (sparc64, sh*) have multiple pte_ts to
2908 * each hugepage. We have to make sure we get the
2909 * first, for the page indexing below to work.
2911 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2912 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2915 * When coredumping, it suits get_dump_page if we just return
2916 * an error where there's an empty slot with no huge pagecache
2917 * to back it. This way, we avoid allocating a hugepage, and
2918 * the sparse dumpfile avoids allocating disk blocks, but its
2919 * huge holes still show up with zeroes where they need to be.
2921 if (absent && (flags & FOLL_DUMP) &&
2922 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2928 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2931 spin_unlock(&mm->page_table_lock);
2932 ret = hugetlb_fault(mm, vma, vaddr,
2933 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2934 spin_lock(&mm->page_table_lock);
2935 if (!(ret & VM_FAULT_ERROR))
2942 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2943 page = pte_page(huge_ptep_get(pte));
2946 pages[i] = mem_map_offset(page, pfn_offset);
2957 if (vaddr < vma->vm_end && remainder &&
2958 pfn_offset < pages_per_huge_page(h)) {
2960 * We use pfn_offset to avoid touching the pageframes
2961 * of this compound page.
2966 spin_unlock(&mm->page_table_lock);
2967 *length = remainder;
2970 return i ? i : -EFAULT;
2973 void hugetlb_change_protection(struct vm_area_struct *vma,
2974 unsigned long address, unsigned long end, pgprot_t newprot)
2976 struct mm_struct *mm = vma->vm_mm;
2977 unsigned long start = address;
2980 struct hstate *h = hstate_vma(vma);
2982 BUG_ON(address >= end);
2983 flush_cache_range(vma, address, end);
2985 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2986 spin_lock(&mm->page_table_lock);
2987 for (; address < end; address += huge_page_size(h)) {
2988 ptep = huge_pte_offset(mm, address);
2991 if (huge_pmd_unshare(mm, &address, ptep))
2993 if (!huge_pte_none(huge_ptep_get(ptep))) {
2994 pte = huge_ptep_get_and_clear(mm, address, ptep);
2995 pte = pte_mkhuge(pte_modify(pte, newprot));
2996 set_huge_pte_at(mm, address, ptep, pte);
2999 spin_unlock(&mm->page_table_lock);
3000 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3002 flush_tlb_range(vma, start, end);
3005 int hugetlb_reserve_pages(struct inode *inode,
3007 struct vm_area_struct *vma,
3008 vm_flags_t vm_flags)
3011 struct hstate *h = hstate_inode(inode);
3012 struct hugepage_subpool *spool = subpool_inode(inode);
3015 * Only apply hugepage reservation if asked. At fault time, an
3016 * attempt will be made for VM_NORESERVE to allocate a page
3017 * without using reserves
3019 if (vm_flags & VM_NORESERVE)
3023 * Shared mappings base their reservation on the number of pages that
3024 * are already allocated on behalf of the file. Private mappings need
3025 * to reserve the full area even if read-only as mprotect() may be
3026 * called to make the mapping read-write. Assume !vma is a shm mapping
3028 if (!vma || vma->vm_flags & VM_MAYSHARE)
3029 chg = region_chg(&inode->i_mapping->private_list, from, to);
3031 struct resv_map *resv_map = resv_map_alloc();
3037 set_vma_resv_map(vma, resv_map);
3038 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3046 /* There must be enough pages in the subpool for the mapping */
3047 if (hugepage_subpool_get_pages(spool, chg)) {
3053 * Check enough hugepages are available for the reservation.
3054 * Hand the pages back to the subpool if there are not
3056 ret = hugetlb_acct_memory(h, chg);
3058 hugepage_subpool_put_pages(spool, chg);
3063 * Account for the reservations made. Shared mappings record regions
3064 * that have reservations as they are shared by multiple VMAs.
3065 * When the last VMA disappears, the region map says how much
3066 * the reservation was and the page cache tells how much of
3067 * the reservation was consumed. Private mappings are per-VMA and
3068 * only the consumed reservations are tracked. When the VMA
3069 * disappears, the original reservation is the VMA size and the
3070 * consumed reservations are stored in the map. Hence, nothing
3071 * else has to be done for private mappings here
3073 if (!vma || vma->vm_flags & VM_MAYSHARE)
3074 region_add(&inode->i_mapping->private_list, from, to);
3082 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3084 struct hstate *h = hstate_inode(inode);
3085 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3086 struct hugepage_subpool *spool = subpool_inode(inode);
3088 spin_lock(&inode->i_lock);
3089 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3090 spin_unlock(&inode->i_lock);
3092 hugepage_subpool_put_pages(spool, (chg - freed));
3093 hugetlb_acct_memory(h, -(chg - freed));
3096 #ifdef CONFIG_MEMORY_FAILURE
3098 /* Should be called in hugetlb_lock */
3099 static int is_hugepage_on_freelist(struct page *hpage)
3103 struct hstate *h = page_hstate(hpage);
3104 int nid = page_to_nid(hpage);
3106 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3113 * This function is called from memory failure code.
3114 * Assume the caller holds page lock of the head page.
3116 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3118 struct hstate *h = page_hstate(hpage);
3119 int nid = page_to_nid(hpage);
3122 spin_lock(&hugetlb_lock);
3123 if (is_hugepage_on_freelist(hpage)) {
3124 list_del(&hpage->lru);
3125 set_page_refcounted(hpage);
3126 h->free_huge_pages--;
3127 h->free_huge_pages_node[nid]--;
3130 spin_unlock(&hugetlb_lock);