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>
33 #include <linux/hugetlb_cgroup.h>
36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
37 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
38 unsigned long hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 __initdata LIST_HEAD(huge_boot_pages);
46 /* for command line parsing */
47 static struct hstate * __initdata parsed_hstate;
48 static unsigned long __initdata default_hstate_max_huge_pages;
49 static unsigned long __initdata default_hstate_size;
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 DEFINE_SPINLOCK(hugetlb_lock);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
60 spin_unlock(&spool->lock);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
70 struct hugepage_subpool *spool;
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
76 spin_lock_init(&spool->lock);
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
89 unlock_or_release_subpool(spool);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
106 spin_unlock(&spool->lock);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
126 return HUGETLBFS_SB(inode->i_sb)->spool;
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
131 return subpool_inode(vma->vm_file->f_dentry->d_inode);
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link;
154 static long region_add(struct list_head *head, long f, long t)
156 struct file_region *rg, *nrg, *trg;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head *head, long f, long t)
192 struct file_region *rg, *nrg;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg -= rg->to - rg->from;
239 static long region_truncate(struct list_head *head, long end)
241 struct file_region *rg, *trg;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
248 if (&rg->link == head)
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
255 rg = list_entry(rg->link.next, typeof(*rg), link);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
262 chg += rg->to - rg->from;
269 static long region_count(struct list_head *head, long f, long t)
271 struct file_region *rg;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
287 chg += seg_to - seg_from;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
316 struct hstate *hstate;
318 if (!is_vm_hugetlb_page(vma))
321 hstate = hstate_vma(vma);
323 return 1UL << (hstate->order + PAGE_SHIFT);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
336 return vma_kernel_pagesize(vma);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
370 return (unsigned long)vma->vm_private_data;
373 static void set_vma_private_data(struct vm_area_struct *vma,
376 vma->vm_private_data = (void *)value;
381 struct list_head regions;
384 static struct resv_map *resv_map_alloc(void)
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
396 static void resv_map_release(struct kref *ref)
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
435 return (get_vma_private_data(vma) & flag) != 0;
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
442 if (vma->vm_flags & VM_NORESERVE)
445 if (vma->vm_flags & VM_MAYSHARE) {
446 /* Shared mappings always use reserves */
447 h->resv_huge_pages--;
448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
450 * Only the process that called mmap() has reserves for
453 h->resv_huge_pages--;
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
461 if (!(vma->vm_flags & VM_MAYSHARE))
462 vma->vm_private_data = (void *)0;
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct *vma)
468 if (vma->vm_flags & VM_MAYSHARE)
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
475 static void copy_gigantic_page(struct page *dst, struct page *src)
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
482 for (i = 0; i < pages_per_huge_page(h); ) {
484 copy_highpage(dst, src);
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
492 void copy_huge_page(struct page *dst, struct page *src)
495 struct hstate *h = page_hstate(src);
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
503 for (i = 0; i < pages_per_huge_page(h); i++) {
505 copy_highpage(dst + i, src + i);
509 static void enqueue_huge_page(struct hstate *h, struct page *page)
511 int nid = page_to_nid(page);
512 list_move(&page->lru, &h->hugepage_freelists[nid]);
513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
517 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
521 if (list_empty(&h->hugepage_freelists[nid]))
523 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
524 list_move(&page->lru, &h->hugepage_activelist);
525 set_page_refcounted(page);
526 h->free_huge_pages--;
527 h->free_huge_pages_node[nid]--;
531 static struct page *dequeue_huge_page_vma(struct hstate *h,
532 struct vm_area_struct *vma,
533 unsigned long address, int avoid_reserve)
535 struct page *page = NULL;
536 struct mempolicy *mpol;
537 nodemask_t *nodemask;
538 struct zonelist *zonelist;
541 unsigned int cpuset_mems_cookie;
544 cpuset_mems_cookie = get_mems_allowed();
545 zonelist = huge_zonelist(vma, address,
546 htlb_alloc_mask, &mpol, &nodemask);
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma) &&
553 h->free_huge_pages - h->resv_huge_pages == 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
566 decrement_hugepage_resv_vma(h, vma);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
582 static void update_and_free_page(struct hstate *h, struct page *page)
586 VM_BUG_ON(h->order >= MAX_ORDER);
589 h->nr_huge_pages_node[page_to_nid(page)]--;
590 for (i = 0; i < pages_per_huge_page(h); i++) {
591 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
592 1 << PG_referenced | 1 << PG_dirty |
593 1 << PG_active | 1 << PG_reserved |
594 1 << PG_private | 1 << PG_writeback);
596 VM_BUG_ON(hugetlb_cgroup_from_page(page));
597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
599 arch_release_hugepage(page);
600 __free_pages(page, huge_page_order(h));
603 struct hstate *size_to_hstate(unsigned long size)
608 if (huge_page_size(h) == size)
614 static void free_huge_page(struct page *page)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate *h = page_hstate(page);
621 int nid = page_to_nid(page);
622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
625 set_page_private(page, 0);
626 page->mapping = NULL;
627 BUG_ON(page_count(page));
628 BUG_ON(page_mapcount(page));
630 spin_lock(&hugetlb_lock);
631 hugetlb_cgroup_uncharge_page(hstate_index(h),
632 pages_per_huge_page(h), page);
633 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
634 /* remove the page from active list */
635 list_del(&page->lru);
636 update_and_free_page(h, page);
637 h->surplus_huge_pages--;
638 h->surplus_huge_pages_node[nid]--;
640 enqueue_huge_page(h, page);
642 spin_unlock(&hugetlb_lock);
643 hugepage_subpool_put_pages(spool, 1);
646 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
648 INIT_LIST_HEAD(&page->lru);
649 set_compound_page_dtor(page, free_huge_page);
650 spin_lock(&hugetlb_lock);
651 set_hugetlb_cgroup(page, NULL);
653 h->nr_huge_pages_node[nid]++;
654 spin_unlock(&hugetlb_lock);
655 put_page(page); /* free it into the hugepage allocator */
658 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
661 int nr_pages = 1 << order;
662 struct page *p = page + 1;
664 /* we rely on prep_new_huge_page to set the destructor */
665 set_compound_order(page, order);
667 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
669 set_page_count(p, 0);
670 p->first_page = page;
674 int PageHuge(struct page *page)
676 compound_page_dtor *dtor;
678 if (!PageCompound(page))
681 page = compound_head(page);
682 dtor = get_compound_page_dtor(page);
684 return dtor == free_huge_page;
686 EXPORT_SYMBOL_GPL(PageHuge);
688 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
692 if (h->order >= MAX_ORDER)
695 page = alloc_pages_exact_node(nid,
696 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
697 __GFP_REPEAT|__GFP_NOWARN,
700 if (arch_prepare_hugepage(page)) {
701 __free_pages(page, huge_page_order(h));
704 prep_new_huge_page(h, page, nid);
711 * common helper functions for hstate_next_node_to_{alloc|free}.
712 * We may have allocated or freed a huge page based on a different
713 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
714 * be outside of *nodes_allowed. Ensure that we use an allowed
715 * node for alloc or free.
717 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
719 nid = next_node(nid, *nodes_allowed);
720 if (nid == MAX_NUMNODES)
721 nid = first_node(*nodes_allowed);
722 VM_BUG_ON(nid >= MAX_NUMNODES);
727 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
729 if (!node_isset(nid, *nodes_allowed))
730 nid = next_node_allowed(nid, nodes_allowed);
735 * returns the previously saved node ["this node"] from which to
736 * allocate a persistent huge page for the pool and advance the
737 * next node from which to allocate, handling wrap at end of node
740 static int hstate_next_node_to_alloc(struct hstate *h,
741 nodemask_t *nodes_allowed)
745 VM_BUG_ON(!nodes_allowed);
747 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
748 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
753 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
760 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
761 next_nid = start_nid;
764 page = alloc_fresh_huge_page_node(h, next_nid);
769 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
770 } while (next_nid != start_nid);
773 count_vm_event(HTLB_BUDDY_PGALLOC);
775 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
781 * helper for free_pool_huge_page() - return the previously saved
782 * node ["this node"] from which to free a huge page. Advance the
783 * next node id whether or not we find a free huge page to free so
784 * that the next attempt to free addresses the next node.
786 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
790 VM_BUG_ON(!nodes_allowed);
792 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
793 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
799 * Free huge page from pool from next node to free.
800 * Attempt to keep persistent huge pages more or less
801 * balanced over allowed nodes.
802 * Called with hugetlb_lock locked.
804 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
811 start_nid = hstate_next_node_to_free(h, nodes_allowed);
812 next_nid = start_nid;
816 * If we're returning unused surplus pages, only examine
817 * nodes with surplus pages.
819 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
820 !list_empty(&h->hugepage_freelists[next_nid])) {
822 list_entry(h->hugepage_freelists[next_nid].next,
824 list_del(&page->lru);
825 h->free_huge_pages--;
826 h->free_huge_pages_node[next_nid]--;
828 h->surplus_huge_pages--;
829 h->surplus_huge_pages_node[next_nid]--;
831 update_and_free_page(h, page);
835 next_nid = hstate_next_node_to_free(h, nodes_allowed);
836 } while (next_nid != start_nid);
841 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
846 if (h->order >= MAX_ORDER)
850 * Assume we will successfully allocate the surplus page to
851 * prevent racing processes from causing the surplus to exceed
854 * This however introduces a different race, where a process B
855 * tries to grow the static hugepage pool while alloc_pages() is
856 * called by process A. B will only examine the per-node
857 * counters in determining if surplus huge pages can be
858 * converted to normal huge pages in adjust_pool_surplus(). A
859 * won't be able to increment the per-node counter, until the
860 * lock is dropped by B, but B doesn't drop hugetlb_lock until
861 * no more huge pages can be converted from surplus to normal
862 * state (and doesn't try to convert again). Thus, we have a
863 * case where a surplus huge page exists, the pool is grown, and
864 * the surplus huge page still exists after, even though it
865 * should just have been converted to a normal huge page. This
866 * does not leak memory, though, as the hugepage will be freed
867 * once it is out of use. It also does not allow the counters to
868 * go out of whack in adjust_pool_surplus() as we don't modify
869 * the node values until we've gotten the hugepage and only the
870 * per-node value is checked there.
872 spin_lock(&hugetlb_lock);
873 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
874 spin_unlock(&hugetlb_lock);
878 h->surplus_huge_pages++;
880 spin_unlock(&hugetlb_lock);
882 if (nid == NUMA_NO_NODE)
883 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
884 __GFP_REPEAT|__GFP_NOWARN,
887 page = alloc_pages_exact_node(nid,
888 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
889 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
891 if (page && arch_prepare_hugepage(page)) {
892 __free_pages(page, huge_page_order(h));
896 spin_lock(&hugetlb_lock);
898 INIT_LIST_HEAD(&page->lru);
899 r_nid = page_to_nid(page);
900 set_compound_page_dtor(page, free_huge_page);
901 set_hugetlb_cgroup(page, NULL);
903 * We incremented the global counters already
905 h->nr_huge_pages_node[r_nid]++;
906 h->surplus_huge_pages_node[r_nid]++;
907 __count_vm_event(HTLB_BUDDY_PGALLOC);
910 h->surplus_huge_pages--;
911 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
913 spin_unlock(&hugetlb_lock);
919 * This allocation function is useful in the context where vma is irrelevant.
920 * E.g. soft-offlining uses this function because it only cares physical
921 * address of error page.
923 struct page *alloc_huge_page_node(struct hstate *h, int nid)
927 spin_lock(&hugetlb_lock);
928 page = dequeue_huge_page_node(h, nid);
929 spin_unlock(&hugetlb_lock);
932 page = alloc_buddy_huge_page(h, nid);
934 spin_lock(&hugetlb_lock);
935 list_move(&page->lru, &h->hugepage_activelist);
936 spin_unlock(&hugetlb_lock);
944 * Increase the hugetlb pool such that it can accommodate a reservation
947 static int gather_surplus_pages(struct hstate *h, int delta)
949 struct list_head surplus_list;
950 struct page *page, *tmp;
952 int needed, allocated;
953 bool alloc_ok = true;
955 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
957 h->resv_huge_pages += delta;
962 INIT_LIST_HEAD(&surplus_list);
966 spin_unlock(&hugetlb_lock);
967 for (i = 0; i < needed; i++) {
968 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
973 list_add(&page->lru, &surplus_list);
978 * After retaking hugetlb_lock, we need to recalculate 'needed'
979 * because either resv_huge_pages or free_huge_pages may have changed.
981 spin_lock(&hugetlb_lock);
982 needed = (h->resv_huge_pages + delta) -
983 (h->free_huge_pages + allocated);
988 * We were not able to allocate enough pages to
989 * satisfy the entire reservation so we free what
990 * we've allocated so far.
995 * The surplus_list now contains _at_least_ the number of extra pages
996 * needed to accommodate the reservation. Add the appropriate number
997 * of pages to the hugetlb pool and free the extras back to the buddy
998 * allocator. Commit the entire reservation here to prevent another
999 * process from stealing the pages as they are added to the pool but
1000 * before they are reserved.
1002 needed += allocated;
1003 h->resv_huge_pages += delta;
1006 /* Free the needed pages to the hugetlb pool */
1007 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1011 * This page is now managed by the hugetlb allocator and has
1012 * no users -- drop the buddy allocator's reference.
1014 put_page_testzero(page);
1015 VM_BUG_ON(page_count(page));
1016 enqueue_huge_page(h, page);
1019 spin_unlock(&hugetlb_lock);
1021 /* Free unnecessary surplus pages to the buddy allocator */
1022 if (!list_empty(&surplus_list)) {
1023 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1027 spin_lock(&hugetlb_lock);
1033 * When releasing a hugetlb pool reservation, any surplus pages that were
1034 * allocated to satisfy the reservation must be explicitly freed if they were
1036 * Called with hugetlb_lock held.
1038 static void return_unused_surplus_pages(struct hstate *h,
1039 unsigned long unused_resv_pages)
1041 unsigned long nr_pages;
1043 /* Uncommit the reservation */
1044 h->resv_huge_pages -= unused_resv_pages;
1046 /* Cannot return gigantic pages currently */
1047 if (h->order >= MAX_ORDER)
1050 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1053 * We want to release as many surplus pages as possible, spread
1054 * evenly across all nodes with memory. Iterate across these nodes
1055 * until we can no longer free unreserved surplus pages. This occurs
1056 * when the nodes with surplus pages have no free pages.
1057 * free_pool_huge_page() will balance the the freed pages across the
1058 * on-line nodes with memory and will handle the hstate accounting.
1060 while (nr_pages--) {
1061 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1067 * Determine if the huge page at addr within the vma has an associated
1068 * reservation. Where it does not we will need to logically increase
1069 * reservation and actually increase subpool usage before an allocation
1070 * can occur. Where any new reservation would be required the
1071 * reservation change is prepared, but not committed. Once the page
1072 * has been allocated from the subpool and instantiated the change should
1073 * be committed via vma_commit_reservation. No action is required on
1076 static long vma_needs_reservation(struct hstate *h,
1077 struct vm_area_struct *vma, unsigned long addr)
1079 struct address_space *mapping = vma->vm_file->f_mapping;
1080 struct inode *inode = mapping->host;
1082 if (vma->vm_flags & VM_MAYSHARE) {
1083 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1084 return region_chg(&inode->i_mapping->private_list,
1087 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1092 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1093 struct resv_map *reservations = vma_resv_map(vma);
1095 err = region_chg(&reservations->regions, idx, idx + 1);
1101 static void vma_commit_reservation(struct hstate *h,
1102 struct vm_area_struct *vma, unsigned long addr)
1104 struct address_space *mapping = vma->vm_file->f_mapping;
1105 struct inode *inode = mapping->host;
1107 if (vma->vm_flags & VM_MAYSHARE) {
1108 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1109 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1111 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1112 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1113 struct resv_map *reservations = vma_resv_map(vma);
1115 /* Mark this page used in the map. */
1116 region_add(&reservations->regions, idx, idx + 1);
1120 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1121 unsigned long addr, int avoid_reserve)
1123 struct hugepage_subpool *spool = subpool_vma(vma);
1124 struct hstate *h = hstate_vma(vma);
1128 struct hugetlb_cgroup *h_cg;
1130 idx = hstate_index(h);
1132 * Processes that did not create the mapping will have no
1133 * reserves and will not have accounted against subpool
1134 * limit. Check that the subpool limit can be made before
1135 * satisfying the allocation MAP_NORESERVE mappings may also
1136 * need pages and subpool limit allocated allocated if no reserve
1139 chg = vma_needs_reservation(h, vma, addr);
1141 return ERR_PTR(-ENOMEM);
1143 if (hugepage_subpool_get_pages(spool, chg))
1144 return ERR_PTR(-ENOSPC);
1146 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1148 hugepage_subpool_put_pages(spool, chg);
1149 return ERR_PTR(-ENOSPC);
1151 spin_lock(&hugetlb_lock);
1152 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1153 spin_unlock(&hugetlb_lock);
1156 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1158 hugetlb_cgroup_uncharge_cgroup(idx,
1159 pages_per_huge_page(h),
1161 hugepage_subpool_put_pages(spool, chg);
1162 return ERR_PTR(-ENOSPC);
1164 spin_lock(&hugetlb_lock);
1165 list_move(&page->lru, &h->hugepage_activelist);
1166 spin_unlock(&hugetlb_lock);
1169 set_page_private(page, (unsigned long)spool);
1171 vma_commit_reservation(h, vma, addr);
1172 /* update page cgroup details */
1173 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1177 int __weak alloc_bootmem_huge_page(struct hstate *h)
1179 struct huge_bootmem_page *m;
1180 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1185 addr = __alloc_bootmem_node_nopanic(
1186 NODE_DATA(hstate_next_node_to_alloc(h,
1187 &node_states[N_HIGH_MEMORY])),
1188 huge_page_size(h), huge_page_size(h), 0);
1192 * Use the beginning of the huge page to store the
1193 * huge_bootmem_page struct (until gather_bootmem
1194 * puts them into the mem_map).
1204 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1205 /* Put them into a private list first because mem_map is not up yet */
1206 list_add(&m->list, &huge_boot_pages);
1211 static void prep_compound_huge_page(struct page *page, int order)
1213 if (unlikely(order > (MAX_ORDER - 1)))
1214 prep_compound_gigantic_page(page, order);
1216 prep_compound_page(page, order);
1219 /* Put bootmem huge pages into the standard lists after mem_map is up */
1220 static void __init gather_bootmem_prealloc(void)
1222 struct huge_bootmem_page *m;
1224 list_for_each_entry(m, &huge_boot_pages, list) {
1225 struct hstate *h = m->hstate;
1228 #ifdef CONFIG_HIGHMEM
1229 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1230 free_bootmem_late((unsigned long)m,
1231 sizeof(struct huge_bootmem_page));
1233 page = virt_to_page(m);
1235 __ClearPageReserved(page);
1236 WARN_ON(page_count(page) != 1);
1237 prep_compound_huge_page(page, h->order);
1238 prep_new_huge_page(h, page, page_to_nid(page));
1240 * If we had gigantic hugepages allocated at boot time, we need
1241 * to restore the 'stolen' pages to totalram_pages in order to
1242 * fix confusing memory reports from free(1) and another
1243 * side-effects, like CommitLimit going negative.
1245 if (h->order > (MAX_ORDER - 1))
1246 totalram_pages += 1 << h->order;
1250 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1254 for (i = 0; i < h->max_huge_pages; ++i) {
1255 if (h->order >= MAX_ORDER) {
1256 if (!alloc_bootmem_huge_page(h))
1258 } else if (!alloc_fresh_huge_page(h,
1259 &node_states[N_HIGH_MEMORY]))
1262 h->max_huge_pages = i;
1265 static void __init hugetlb_init_hstates(void)
1269 for_each_hstate(h) {
1270 /* oversize hugepages were init'ed in early boot */
1271 if (h->order < MAX_ORDER)
1272 hugetlb_hstate_alloc_pages(h);
1276 static char * __init memfmt(char *buf, unsigned long n)
1278 if (n >= (1UL << 30))
1279 sprintf(buf, "%lu GB", n >> 30);
1280 else if (n >= (1UL << 20))
1281 sprintf(buf, "%lu MB", n >> 20);
1283 sprintf(buf, "%lu KB", n >> 10);
1287 static void __init report_hugepages(void)
1291 for_each_hstate(h) {
1293 printk(KERN_INFO "HugeTLB registered %s page size, "
1294 "pre-allocated %ld pages\n",
1295 memfmt(buf, huge_page_size(h)),
1296 h->free_huge_pages);
1300 #ifdef CONFIG_HIGHMEM
1301 static void try_to_free_low(struct hstate *h, unsigned long count,
1302 nodemask_t *nodes_allowed)
1306 if (h->order >= MAX_ORDER)
1309 for_each_node_mask(i, *nodes_allowed) {
1310 struct page *page, *next;
1311 struct list_head *freel = &h->hugepage_freelists[i];
1312 list_for_each_entry_safe(page, next, freel, lru) {
1313 if (count >= h->nr_huge_pages)
1315 if (PageHighMem(page))
1317 list_del(&page->lru);
1318 update_and_free_page(h, page);
1319 h->free_huge_pages--;
1320 h->free_huge_pages_node[page_to_nid(page)]--;
1325 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1326 nodemask_t *nodes_allowed)
1332 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1333 * balanced by operating on them in a round-robin fashion.
1334 * Returns 1 if an adjustment was made.
1336 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1339 int start_nid, next_nid;
1342 VM_BUG_ON(delta != -1 && delta != 1);
1345 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1347 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1348 next_nid = start_nid;
1354 * To shrink on this node, there must be a surplus page
1356 if (!h->surplus_huge_pages_node[nid]) {
1357 next_nid = hstate_next_node_to_alloc(h,
1364 * Surplus cannot exceed the total number of pages
1366 if (h->surplus_huge_pages_node[nid] >=
1367 h->nr_huge_pages_node[nid]) {
1368 next_nid = hstate_next_node_to_free(h,
1374 h->surplus_huge_pages += delta;
1375 h->surplus_huge_pages_node[nid] += delta;
1378 } while (next_nid != start_nid);
1383 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1384 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1385 nodemask_t *nodes_allowed)
1387 unsigned long min_count, ret;
1389 if (h->order >= MAX_ORDER)
1390 return h->max_huge_pages;
1393 * Increase the pool size
1394 * First take pages out of surplus state. Then make up the
1395 * remaining difference by allocating fresh huge pages.
1397 * We might race with alloc_buddy_huge_page() here and be unable
1398 * to convert a surplus huge page to a normal huge page. That is
1399 * not critical, though, it just means the overall size of the
1400 * pool might be one hugepage larger than it needs to be, but
1401 * within all the constraints specified by the sysctls.
1403 spin_lock(&hugetlb_lock);
1404 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1405 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1409 while (count > persistent_huge_pages(h)) {
1411 * If this allocation races such that we no longer need the
1412 * page, free_huge_page will handle it by freeing the page
1413 * and reducing the surplus.
1415 spin_unlock(&hugetlb_lock);
1416 ret = alloc_fresh_huge_page(h, nodes_allowed);
1417 spin_lock(&hugetlb_lock);
1421 /* Bail for signals. Probably ctrl-c from user */
1422 if (signal_pending(current))
1427 * Decrease the pool size
1428 * First return free pages to the buddy allocator (being careful
1429 * to keep enough around to satisfy reservations). Then place
1430 * pages into surplus state as needed so the pool will shrink
1431 * to the desired size as pages become free.
1433 * By placing pages into the surplus state independent of the
1434 * overcommit value, we are allowing the surplus pool size to
1435 * exceed overcommit. There are few sane options here. Since
1436 * alloc_buddy_huge_page() is checking the global counter,
1437 * though, we'll note that we're not allowed to exceed surplus
1438 * and won't grow the pool anywhere else. Not until one of the
1439 * sysctls are changed, or the surplus pages go out of use.
1441 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1442 min_count = max(count, min_count);
1443 try_to_free_low(h, min_count, nodes_allowed);
1444 while (min_count < persistent_huge_pages(h)) {
1445 if (!free_pool_huge_page(h, nodes_allowed, 0))
1448 while (count < persistent_huge_pages(h)) {
1449 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1453 ret = persistent_huge_pages(h);
1454 spin_unlock(&hugetlb_lock);
1458 #define HSTATE_ATTR_RO(_name) \
1459 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1461 #define HSTATE_ATTR(_name) \
1462 static struct kobj_attribute _name##_attr = \
1463 __ATTR(_name, 0644, _name##_show, _name##_store)
1465 static struct kobject *hugepages_kobj;
1466 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1468 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1470 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1474 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1475 if (hstate_kobjs[i] == kobj) {
1477 *nidp = NUMA_NO_NODE;
1481 return kobj_to_node_hstate(kobj, nidp);
1484 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1485 struct kobj_attribute *attr, char *buf)
1488 unsigned long nr_huge_pages;
1491 h = kobj_to_hstate(kobj, &nid);
1492 if (nid == NUMA_NO_NODE)
1493 nr_huge_pages = h->nr_huge_pages;
1495 nr_huge_pages = h->nr_huge_pages_node[nid];
1497 return sprintf(buf, "%lu\n", nr_huge_pages);
1500 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1501 struct kobject *kobj, struct kobj_attribute *attr,
1502 const char *buf, size_t len)
1506 unsigned long count;
1508 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1510 err = strict_strtoul(buf, 10, &count);
1514 h = kobj_to_hstate(kobj, &nid);
1515 if (h->order >= MAX_ORDER) {
1520 if (nid == NUMA_NO_NODE) {
1522 * global hstate attribute
1524 if (!(obey_mempolicy &&
1525 init_nodemask_of_mempolicy(nodes_allowed))) {
1526 NODEMASK_FREE(nodes_allowed);
1527 nodes_allowed = &node_states[N_HIGH_MEMORY];
1529 } else if (nodes_allowed) {
1531 * per node hstate attribute: adjust count to global,
1532 * but restrict alloc/free to the specified node.
1534 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1535 init_nodemask_of_node(nodes_allowed, nid);
1537 nodes_allowed = &node_states[N_HIGH_MEMORY];
1539 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1541 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1542 NODEMASK_FREE(nodes_allowed);
1546 NODEMASK_FREE(nodes_allowed);
1550 static ssize_t nr_hugepages_show(struct kobject *kobj,
1551 struct kobj_attribute *attr, char *buf)
1553 return nr_hugepages_show_common(kobj, attr, buf);
1556 static ssize_t nr_hugepages_store(struct kobject *kobj,
1557 struct kobj_attribute *attr, const char *buf, size_t len)
1559 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1561 HSTATE_ATTR(nr_hugepages);
1566 * hstate attribute for optionally mempolicy-based constraint on persistent
1567 * huge page alloc/free.
1569 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1570 struct kobj_attribute *attr, char *buf)
1572 return nr_hugepages_show_common(kobj, attr, buf);
1575 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1576 struct kobj_attribute *attr, const char *buf, size_t len)
1578 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1580 HSTATE_ATTR(nr_hugepages_mempolicy);
1584 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1585 struct kobj_attribute *attr, char *buf)
1587 struct hstate *h = kobj_to_hstate(kobj, NULL);
1588 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1591 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1592 struct kobj_attribute *attr, const char *buf, size_t count)
1595 unsigned long input;
1596 struct hstate *h = kobj_to_hstate(kobj, NULL);
1598 if (h->order >= MAX_ORDER)
1601 err = strict_strtoul(buf, 10, &input);
1605 spin_lock(&hugetlb_lock);
1606 h->nr_overcommit_huge_pages = input;
1607 spin_unlock(&hugetlb_lock);
1611 HSTATE_ATTR(nr_overcommit_hugepages);
1613 static ssize_t free_hugepages_show(struct kobject *kobj,
1614 struct kobj_attribute *attr, char *buf)
1617 unsigned long free_huge_pages;
1620 h = kobj_to_hstate(kobj, &nid);
1621 if (nid == NUMA_NO_NODE)
1622 free_huge_pages = h->free_huge_pages;
1624 free_huge_pages = h->free_huge_pages_node[nid];
1626 return sprintf(buf, "%lu\n", free_huge_pages);
1628 HSTATE_ATTR_RO(free_hugepages);
1630 static ssize_t resv_hugepages_show(struct kobject *kobj,
1631 struct kobj_attribute *attr, char *buf)
1633 struct hstate *h = kobj_to_hstate(kobj, NULL);
1634 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1636 HSTATE_ATTR_RO(resv_hugepages);
1638 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1639 struct kobj_attribute *attr, char *buf)
1642 unsigned long surplus_huge_pages;
1645 h = kobj_to_hstate(kobj, &nid);
1646 if (nid == NUMA_NO_NODE)
1647 surplus_huge_pages = h->surplus_huge_pages;
1649 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1651 return sprintf(buf, "%lu\n", surplus_huge_pages);
1653 HSTATE_ATTR_RO(surplus_hugepages);
1655 static struct attribute *hstate_attrs[] = {
1656 &nr_hugepages_attr.attr,
1657 &nr_overcommit_hugepages_attr.attr,
1658 &free_hugepages_attr.attr,
1659 &resv_hugepages_attr.attr,
1660 &surplus_hugepages_attr.attr,
1662 &nr_hugepages_mempolicy_attr.attr,
1667 static struct attribute_group hstate_attr_group = {
1668 .attrs = hstate_attrs,
1671 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1672 struct kobject **hstate_kobjs,
1673 struct attribute_group *hstate_attr_group)
1676 int hi = hstate_index(h);
1678 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1679 if (!hstate_kobjs[hi])
1682 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1684 kobject_put(hstate_kobjs[hi]);
1689 static void __init hugetlb_sysfs_init(void)
1694 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1695 if (!hugepages_kobj)
1698 for_each_hstate(h) {
1699 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1700 hstate_kobjs, &hstate_attr_group);
1702 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1710 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1711 * with node devices in node_devices[] using a parallel array. The array
1712 * index of a node device or _hstate == node id.
1713 * This is here to avoid any static dependency of the node device driver, in
1714 * the base kernel, on the hugetlb module.
1716 struct node_hstate {
1717 struct kobject *hugepages_kobj;
1718 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1720 struct node_hstate node_hstates[MAX_NUMNODES];
1723 * A subset of global hstate attributes for node devices
1725 static struct attribute *per_node_hstate_attrs[] = {
1726 &nr_hugepages_attr.attr,
1727 &free_hugepages_attr.attr,
1728 &surplus_hugepages_attr.attr,
1732 static struct attribute_group per_node_hstate_attr_group = {
1733 .attrs = per_node_hstate_attrs,
1737 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1738 * Returns node id via non-NULL nidp.
1740 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1744 for (nid = 0; nid < nr_node_ids; nid++) {
1745 struct node_hstate *nhs = &node_hstates[nid];
1747 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1748 if (nhs->hstate_kobjs[i] == kobj) {
1760 * Unregister hstate attributes from a single node device.
1761 * No-op if no hstate attributes attached.
1763 void hugetlb_unregister_node(struct node *node)
1766 struct node_hstate *nhs = &node_hstates[node->dev.id];
1768 if (!nhs->hugepages_kobj)
1769 return; /* no hstate attributes */
1771 for_each_hstate(h) {
1772 int idx = hstate_index(h);
1773 if (nhs->hstate_kobjs[idx]) {
1774 kobject_put(nhs->hstate_kobjs[idx]);
1775 nhs->hstate_kobjs[idx] = NULL;
1779 kobject_put(nhs->hugepages_kobj);
1780 nhs->hugepages_kobj = NULL;
1784 * hugetlb module exit: unregister hstate attributes from node devices
1787 static void hugetlb_unregister_all_nodes(void)
1792 * disable node device registrations.
1794 register_hugetlbfs_with_node(NULL, NULL);
1797 * remove hstate attributes from any nodes that have them.
1799 for (nid = 0; nid < nr_node_ids; nid++)
1800 hugetlb_unregister_node(&node_devices[nid]);
1804 * Register hstate attributes for a single node device.
1805 * No-op if attributes already registered.
1807 void hugetlb_register_node(struct node *node)
1810 struct node_hstate *nhs = &node_hstates[node->dev.id];
1813 if (nhs->hugepages_kobj)
1814 return; /* already allocated */
1816 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1818 if (!nhs->hugepages_kobj)
1821 for_each_hstate(h) {
1822 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1824 &per_node_hstate_attr_group);
1826 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1828 h->name, node->dev.id);
1829 hugetlb_unregister_node(node);
1836 * hugetlb init time: register hstate attributes for all registered node
1837 * devices of nodes that have memory. All on-line nodes should have
1838 * registered their associated device by this time.
1840 static void hugetlb_register_all_nodes(void)
1844 for_each_node_state(nid, N_HIGH_MEMORY) {
1845 struct node *node = &node_devices[nid];
1846 if (node->dev.id == nid)
1847 hugetlb_register_node(node);
1851 * Let the node device driver know we're here so it can
1852 * [un]register hstate attributes on node hotplug.
1854 register_hugetlbfs_with_node(hugetlb_register_node,
1855 hugetlb_unregister_node);
1857 #else /* !CONFIG_NUMA */
1859 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1867 static void hugetlb_unregister_all_nodes(void) { }
1869 static void hugetlb_register_all_nodes(void) { }
1873 static void __exit hugetlb_exit(void)
1877 hugetlb_unregister_all_nodes();
1879 for_each_hstate(h) {
1880 kobject_put(hstate_kobjs[hstate_index(h)]);
1883 kobject_put(hugepages_kobj);
1885 module_exit(hugetlb_exit);
1887 static int __init hugetlb_init(void)
1889 /* Some platform decide whether they support huge pages at boot
1890 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1891 * there is no such support
1893 if (HPAGE_SHIFT == 0)
1896 if (!size_to_hstate(default_hstate_size)) {
1897 default_hstate_size = HPAGE_SIZE;
1898 if (!size_to_hstate(default_hstate_size))
1899 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1901 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1902 if (default_hstate_max_huge_pages)
1903 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1905 hugetlb_init_hstates();
1907 gather_bootmem_prealloc();
1911 hugetlb_sysfs_init();
1913 hugetlb_register_all_nodes();
1917 module_init(hugetlb_init);
1919 /* Should be called on processing a hugepagesz=... option */
1920 void __init hugetlb_add_hstate(unsigned order)
1925 if (size_to_hstate(PAGE_SIZE << order)) {
1926 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1929 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1931 h = &hstates[hugetlb_max_hstate++];
1933 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1934 h->nr_huge_pages = 0;
1935 h->free_huge_pages = 0;
1936 for (i = 0; i < MAX_NUMNODES; ++i)
1937 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1938 INIT_LIST_HEAD(&h->hugepage_activelist);
1939 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1940 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1941 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1942 huge_page_size(h)/1024);
1944 * Add cgroup control files only if the huge page consists
1945 * of more than two normal pages. This is because we use
1946 * page[2].lru.next for storing cgoup details.
1948 if (order >= HUGETLB_CGROUP_MIN_ORDER)
1949 hugetlb_cgroup_file_init(hugetlb_max_hstate - 1);
1954 static int __init hugetlb_nrpages_setup(char *s)
1957 static unsigned long *last_mhp;
1960 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1961 * so this hugepages= parameter goes to the "default hstate".
1963 if (!hugetlb_max_hstate)
1964 mhp = &default_hstate_max_huge_pages;
1966 mhp = &parsed_hstate->max_huge_pages;
1968 if (mhp == last_mhp) {
1969 printk(KERN_WARNING "hugepages= specified twice without "
1970 "interleaving hugepagesz=, ignoring\n");
1974 if (sscanf(s, "%lu", mhp) <= 0)
1978 * Global state is always initialized later in hugetlb_init.
1979 * But we need to allocate >= MAX_ORDER hstates here early to still
1980 * use the bootmem allocator.
1982 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1983 hugetlb_hstate_alloc_pages(parsed_hstate);
1989 __setup("hugepages=", hugetlb_nrpages_setup);
1991 static int __init hugetlb_default_setup(char *s)
1993 default_hstate_size = memparse(s, &s);
1996 __setup("default_hugepagesz=", hugetlb_default_setup);
1998 static unsigned int cpuset_mems_nr(unsigned int *array)
2001 unsigned int nr = 0;
2003 for_each_node_mask(node, cpuset_current_mems_allowed)
2009 #ifdef CONFIG_SYSCTL
2010 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2011 struct ctl_table *table, int write,
2012 void __user *buffer, size_t *length, loff_t *ppos)
2014 struct hstate *h = &default_hstate;
2018 tmp = h->max_huge_pages;
2020 if (write && h->order >= MAX_ORDER)
2024 table->maxlen = sizeof(unsigned long);
2025 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2030 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2031 GFP_KERNEL | __GFP_NORETRY);
2032 if (!(obey_mempolicy &&
2033 init_nodemask_of_mempolicy(nodes_allowed))) {
2034 NODEMASK_FREE(nodes_allowed);
2035 nodes_allowed = &node_states[N_HIGH_MEMORY];
2037 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2039 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2040 NODEMASK_FREE(nodes_allowed);
2046 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2047 void __user *buffer, size_t *length, loff_t *ppos)
2050 return hugetlb_sysctl_handler_common(false, table, write,
2051 buffer, length, ppos);
2055 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2056 void __user *buffer, size_t *length, loff_t *ppos)
2058 return hugetlb_sysctl_handler_common(true, table, write,
2059 buffer, length, ppos);
2061 #endif /* CONFIG_NUMA */
2063 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2064 void __user *buffer,
2065 size_t *length, loff_t *ppos)
2067 proc_dointvec(table, write, buffer, length, ppos);
2068 if (hugepages_treat_as_movable)
2069 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2071 htlb_alloc_mask = GFP_HIGHUSER;
2075 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2076 void __user *buffer,
2077 size_t *length, loff_t *ppos)
2079 struct hstate *h = &default_hstate;
2083 tmp = h->nr_overcommit_huge_pages;
2085 if (write && h->order >= MAX_ORDER)
2089 table->maxlen = sizeof(unsigned long);
2090 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2095 spin_lock(&hugetlb_lock);
2096 h->nr_overcommit_huge_pages = tmp;
2097 spin_unlock(&hugetlb_lock);
2103 #endif /* CONFIG_SYSCTL */
2105 void hugetlb_report_meminfo(struct seq_file *m)
2107 struct hstate *h = &default_hstate;
2109 "HugePages_Total: %5lu\n"
2110 "HugePages_Free: %5lu\n"
2111 "HugePages_Rsvd: %5lu\n"
2112 "HugePages_Surp: %5lu\n"
2113 "Hugepagesize: %8lu kB\n",
2117 h->surplus_huge_pages,
2118 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2121 int hugetlb_report_node_meminfo(int nid, char *buf)
2123 struct hstate *h = &default_hstate;
2125 "Node %d HugePages_Total: %5u\n"
2126 "Node %d HugePages_Free: %5u\n"
2127 "Node %d HugePages_Surp: %5u\n",
2128 nid, h->nr_huge_pages_node[nid],
2129 nid, h->free_huge_pages_node[nid],
2130 nid, h->surplus_huge_pages_node[nid]);
2133 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2134 unsigned long hugetlb_total_pages(void)
2136 struct hstate *h = &default_hstate;
2137 return h->nr_huge_pages * pages_per_huge_page(h);
2140 static int hugetlb_acct_memory(struct hstate *h, long delta)
2144 spin_lock(&hugetlb_lock);
2146 * When cpuset is configured, it breaks the strict hugetlb page
2147 * reservation as the accounting is done on a global variable. Such
2148 * reservation is completely rubbish in the presence of cpuset because
2149 * the reservation is not checked against page availability for the
2150 * current cpuset. Application can still potentially OOM'ed by kernel
2151 * with lack of free htlb page in cpuset that the task is in.
2152 * Attempt to enforce strict accounting with cpuset is almost
2153 * impossible (or too ugly) because cpuset is too fluid that
2154 * task or memory node can be dynamically moved between cpusets.
2156 * The change of semantics for shared hugetlb mapping with cpuset is
2157 * undesirable. However, in order to preserve some of the semantics,
2158 * we fall back to check against current free page availability as
2159 * a best attempt and hopefully to minimize the impact of changing
2160 * semantics that cpuset has.
2163 if (gather_surplus_pages(h, delta) < 0)
2166 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2167 return_unused_surplus_pages(h, delta);
2174 return_unused_surplus_pages(h, (unsigned long) -delta);
2177 spin_unlock(&hugetlb_lock);
2181 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2183 struct resv_map *reservations = vma_resv_map(vma);
2186 * This new VMA should share its siblings reservation map if present.
2187 * The VMA will only ever have a valid reservation map pointer where
2188 * it is being copied for another still existing VMA. As that VMA
2189 * has a reference to the reservation map it cannot disappear until
2190 * after this open call completes. It is therefore safe to take a
2191 * new reference here without additional locking.
2194 kref_get(&reservations->refs);
2197 static void resv_map_put(struct vm_area_struct *vma)
2199 struct resv_map *reservations = vma_resv_map(vma);
2203 kref_put(&reservations->refs, resv_map_release);
2206 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2208 struct hstate *h = hstate_vma(vma);
2209 struct resv_map *reservations = vma_resv_map(vma);
2210 struct hugepage_subpool *spool = subpool_vma(vma);
2211 unsigned long reserve;
2212 unsigned long start;
2216 start = vma_hugecache_offset(h, vma, vma->vm_start);
2217 end = vma_hugecache_offset(h, vma, vma->vm_end);
2219 reserve = (end - start) -
2220 region_count(&reservations->regions, start, end);
2225 hugetlb_acct_memory(h, -reserve);
2226 hugepage_subpool_put_pages(spool, reserve);
2232 * We cannot handle pagefaults against hugetlb pages at all. They cause
2233 * handle_mm_fault() to try to instantiate regular-sized pages in the
2234 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2237 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2243 const struct vm_operations_struct hugetlb_vm_ops = {
2244 .fault = hugetlb_vm_op_fault,
2245 .open = hugetlb_vm_op_open,
2246 .close = hugetlb_vm_op_close,
2249 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2256 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2258 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2260 entry = pte_mkyoung(entry);
2261 entry = pte_mkhuge(entry);
2262 entry = arch_make_huge_pte(entry, vma, page, writable);
2267 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2268 unsigned long address, pte_t *ptep)
2272 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2273 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2274 update_mmu_cache(vma, address, ptep);
2278 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2279 struct vm_area_struct *vma)
2281 pte_t *src_pte, *dst_pte, entry;
2282 struct page *ptepage;
2285 struct hstate *h = hstate_vma(vma);
2286 unsigned long sz = huge_page_size(h);
2288 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2290 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2291 src_pte = huge_pte_offset(src, addr);
2294 dst_pte = huge_pte_alloc(dst, addr, sz);
2298 /* If the pagetables are shared don't copy or take references */
2299 if (dst_pte == src_pte)
2302 spin_lock(&dst->page_table_lock);
2303 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2304 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2306 huge_ptep_set_wrprotect(src, addr, src_pte);
2307 entry = huge_ptep_get(src_pte);
2308 ptepage = pte_page(entry);
2310 page_dup_rmap(ptepage);
2311 set_huge_pte_at(dst, addr, dst_pte, entry);
2313 spin_unlock(&src->page_table_lock);
2314 spin_unlock(&dst->page_table_lock);
2322 static int is_hugetlb_entry_migration(pte_t pte)
2326 if (huge_pte_none(pte) || pte_present(pte))
2328 swp = pte_to_swp_entry(pte);
2329 if (non_swap_entry(swp) && is_migration_entry(swp))
2335 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2339 if (huge_pte_none(pte) || pte_present(pte))
2341 swp = pte_to_swp_entry(pte);
2342 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2348 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2349 unsigned long start, unsigned long end,
2350 struct page *ref_page)
2352 int force_flush = 0;
2353 struct mm_struct *mm = vma->vm_mm;
2354 unsigned long address;
2358 struct hstate *h = hstate_vma(vma);
2359 unsigned long sz = huge_page_size(h);
2361 WARN_ON(!is_vm_hugetlb_page(vma));
2362 BUG_ON(start & ~huge_page_mask(h));
2363 BUG_ON(end & ~huge_page_mask(h));
2365 tlb_start_vma(tlb, vma);
2366 mmu_notifier_invalidate_range_start(mm, start, end);
2368 spin_lock(&mm->page_table_lock);
2369 for (address = start; address < end; address += sz) {
2370 ptep = huge_pte_offset(mm, address);
2374 if (huge_pmd_unshare(mm, &address, ptep))
2377 pte = huge_ptep_get(ptep);
2378 if (huge_pte_none(pte))
2382 * HWPoisoned hugepage is already unmapped and dropped reference
2384 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2387 page = pte_page(pte);
2389 * If a reference page is supplied, it is because a specific
2390 * page is being unmapped, not a range. Ensure the page we
2391 * are about to unmap is the actual page of interest.
2394 if (page != ref_page)
2398 * Mark the VMA as having unmapped its page so that
2399 * future faults in this VMA will fail rather than
2400 * looking like data was lost
2402 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2405 pte = huge_ptep_get_and_clear(mm, address, ptep);
2406 tlb_remove_tlb_entry(tlb, ptep, address);
2408 set_page_dirty(page);
2410 page_remove_rmap(page);
2411 force_flush = !__tlb_remove_page(tlb, page);
2414 /* Bail out after unmapping reference page if supplied */
2418 spin_unlock(&mm->page_table_lock);
2420 * mmu_gather ran out of room to batch pages, we break out of
2421 * the PTE lock to avoid doing the potential expensive TLB invalidate
2422 * and page-free while holding it.
2427 if (address < end && !ref_page)
2430 mmu_notifier_invalidate_range_end(mm, start, end);
2431 tlb_end_vma(tlb, vma);
2434 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2435 unsigned long end, struct page *ref_page)
2437 struct mm_struct *mm;
2438 struct mmu_gather tlb;
2442 tlb_gather_mmu(&tlb, mm, 0);
2443 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2444 tlb_finish_mmu(&tlb, start, end);
2448 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2449 * mappping it owns the reserve page for. The intention is to unmap the page
2450 * from other VMAs and let the children be SIGKILLed if they are faulting the
2453 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2454 struct page *page, unsigned long address)
2456 struct hstate *h = hstate_vma(vma);
2457 struct vm_area_struct *iter_vma;
2458 struct address_space *mapping;
2459 struct prio_tree_iter iter;
2463 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2464 * from page cache lookup which is in HPAGE_SIZE units.
2466 address = address & huge_page_mask(h);
2467 pgoff = vma_hugecache_offset(h, vma, address);
2468 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2471 * Take the mapping lock for the duration of the table walk. As
2472 * this mapping should be shared between all the VMAs,
2473 * __unmap_hugepage_range() is called as the lock is already held
2475 mutex_lock(&mapping->i_mmap_mutex);
2476 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2477 /* Do not unmap the current VMA */
2478 if (iter_vma == vma)
2482 * Unmap the page from other VMAs without their own reserves.
2483 * They get marked to be SIGKILLed if they fault in these
2484 * areas. This is because a future no-page fault on this VMA
2485 * could insert a zeroed page instead of the data existing
2486 * from the time of fork. This would look like data corruption
2488 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2489 unmap_hugepage_range(iter_vma, address,
2490 address + huge_page_size(h), page);
2492 mutex_unlock(&mapping->i_mmap_mutex);
2498 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2499 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2500 * cannot race with other handlers or page migration.
2501 * Keep the pte_same checks anyway to make transition from the mutex easier.
2503 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2504 unsigned long address, pte_t *ptep, pte_t pte,
2505 struct page *pagecache_page)
2507 struct hstate *h = hstate_vma(vma);
2508 struct page *old_page, *new_page;
2510 int outside_reserve = 0;
2512 old_page = pte_page(pte);
2515 /* If no-one else is actually using this page, avoid the copy
2516 * and just make the page writable */
2517 avoidcopy = (page_mapcount(old_page) == 1);
2519 if (PageAnon(old_page))
2520 page_move_anon_rmap(old_page, vma, address);
2521 set_huge_ptep_writable(vma, address, ptep);
2526 * If the process that created a MAP_PRIVATE mapping is about to
2527 * perform a COW due to a shared page count, attempt to satisfy
2528 * the allocation without using the existing reserves. The pagecache
2529 * page is used to determine if the reserve at this address was
2530 * consumed or not. If reserves were used, a partial faulted mapping
2531 * at the time of fork() could consume its reserves on COW instead
2532 * of the full address range.
2534 if (!(vma->vm_flags & VM_MAYSHARE) &&
2535 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2536 old_page != pagecache_page)
2537 outside_reserve = 1;
2539 page_cache_get(old_page);
2541 /* Drop page_table_lock as buddy allocator may be called */
2542 spin_unlock(&mm->page_table_lock);
2543 new_page = alloc_huge_page(vma, address, outside_reserve);
2545 if (IS_ERR(new_page)) {
2546 long err = PTR_ERR(new_page);
2547 page_cache_release(old_page);
2550 * If a process owning a MAP_PRIVATE mapping fails to COW,
2551 * it is due to references held by a child and an insufficient
2552 * huge page pool. To guarantee the original mappers
2553 * reliability, unmap the page from child processes. The child
2554 * may get SIGKILLed if it later faults.
2556 if (outside_reserve) {
2557 BUG_ON(huge_pte_none(pte));
2558 if (unmap_ref_private(mm, vma, old_page, address)) {
2559 BUG_ON(huge_pte_none(pte));
2560 spin_lock(&mm->page_table_lock);
2561 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2562 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2563 goto retry_avoidcopy;
2565 * race occurs while re-acquiring page_table_lock, and
2573 /* Caller expects lock to be held */
2574 spin_lock(&mm->page_table_lock);
2576 return VM_FAULT_OOM;
2578 return VM_FAULT_SIGBUS;
2582 * When the original hugepage is shared one, it does not have
2583 * anon_vma prepared.
2585 if (unlikely(anon_vma_prepare(vma))) {
2586 page_cache_release(new_page);
2587 page_cache_release(old_page);
2588 /* Caller expects lock to be held */
2589 spin_lock(&mm->page_table_lock);
2590 return VM_FAULT_OOM;
2593 copy_user_huge_page(new_page, old_page, address, vma,
2594 pages_per_huge_page(h));
2595 __SetPageUptodate(new_page);
2598 * Retake the page_table_lock to check for racing updates
2599 * before the page tables are altered
2601 spin_lock(&mm->page_table_lock);
2602 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2603 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2605 mmu_notifier_invalidate_range_start(mm,
2606 address & huge_page_mask(h),
2607 (address & huge_page_mask(h)) + huge_page_size(h));
2608 huge_ptep_clear_flush(vma, address, ptep);
2609 set_huge_pte_at(mm, address, ptep,
2610 make_huge_pte(vma, new_page, 1));
2611 page_remove_rmap(old_page);
2612 hugepage_add_new_anon_rmap(new_page, vma, address);
2613 /* Make the old page be freed below */
2614 new_page = old_page;
2615 mmu_notifier_invalidate_range_end(mm,
2616 address & huge_page_mask(h),
2617 (address & huge_page_mask(h)) + huge_page_size(h));
2619 page_cache_release(new_page);
2620 page_cache_release(old_page);
2624 /* Return the pagecache page at a given address within a VMA */
2625 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2626 struct vm_area_struct *vma, unsigned long address)
2628 struct address_space *mapping;
2631 mapping = vma->vm_file->f_mapping;
2632 idx = vma_hugecache_offset(h, vma, address);
2634 return find_lock_page(mapping, idx);
2638 * Return whether there is a pagecache page to back given address within VMA.
2639 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2641 static bool hugetlbfs_pagecache_present(struct hstate *h,
2642 struct vm_area_struct *vma, unsigned long address)
2644 struct address_space *mapping;
2648 mapping = vma->vm_file->f_mapping;
2649 idx = vma_hugecache_offset(h, vma, address);
2651 page = find_get_page(mapping, idx);
2654 return page != NULL;
2657 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2658 unsigned long address, pte_t *ptep, unsigned int flags)
2660 struct hstate *h = hstate_vma(vma);
2661 int ret = VM_FAULT_SIGBUS;
2666 struct address_space *mapping;
2670 * Currently, we are forced to kill the process in the event the
2671 * original mapper has unmapped pages from the child due to a failed
2672 * COW. Warn that such a situation has occurred as it may not be obvious
2674 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2676 "PID %d killed due to inadequate hugepage pool\n",
2681 mapping = vma->vm_file->f_mapping;
2682 idx = vma_hugecache_offset(h, vma, address);
2685 * Use page lock to guard against racing truncation
2686 * before we get page_table_lock.
2689 page = find_lock_page(mapping, idx);
2691 size = i_size_read(mapping->host) >> huge_page_shift(h);
2694 page = alloc_huge_page(vma, address, 0);
2696 ret = PTR_ERR(page);
2700 ret = VM_FAULT_SIGBUS;
2703 clear_huge_page(page, address, pages_per_huge_page(h));
2704 __SetPageUptodate(page);
2706 if (vma->vm_flags & VM_MAYSHARE) {
2708 struct inode *inode = mapping->host;
2710 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2718 spin_lock(&inode->i_lock);
2719 inode->i_blocks += blocks_per_huge_page(h);
2720 spin_unlock(&inode->i_lock);
2723 if (unlikely(anon_vma_prepare(vma))) {
2725 goto backout_unlocked;
2731 * If memory error occurs between mmap() and fault, some process
2732 * don't have hwpoisoned swap entry for errored virtual address.
2733 * So we need to block hugepage fault by PG_hwpoison bit check.
2735 if (unlikely(PageHWPoison(page))) {
2736 ret = VM_FAULT_HWPOISON |
2737 VM_FAULT_SET_HINDEX(hstate_index(h));
2738 goto backout_unlocked;
2743 * If we are going to COW a private mapping later, we examine the
2744 * pending reservations for this page now. This will ensure that
2745 * any allocations necessary to record that reservation occur outside
2748 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2749 if (vma_needs_reservation(h, vma, address) < 0) {
2751 goto backout_unlocked;
2754 spin_lock(&mm->page_table_lock);
2755 size = i_size_read(mapping->host) >> huge_page_shift(h);
2760 if (!huge_pte_none(huge_ptep_get(ptep)))
2764 hugepage_add_new_anon_rmap(page, vma, address);
2766 page_dup_rmap(page);
2767 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2768 && (vma->vm_flags & VM_SHARED)));
2769 set_huge_pte_at(mm, address, ptep, new_pte);
2771 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2772 /* Optimization, do the COW without a second fault */
2773 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2776 spin_unlock(&mm->page_table_lock);
2782 spin_unlock(&mm->page_table_lock);
2789 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2790 unsigned long address, unsigned int flags)
2795 struct page *page = NULL;
2796 struct page *pagecache_page = NULL;
2797 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2798 struct hstate *h = hstate_vma(vma);
2800 address &= huge_page_mask(h);
2802 ptep = huge_pte_offset(mm, address);
2804 entry = huge_ptep_get(ptep);
2805 if (unlikely(is_hugetlb_entry_migration(entry))) {
2806 migration_entry_wait(mm, (pmd_t *)ptep, address);
2808 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2809 return VM_FAULT_HWPOISON_LARGE |
2810 VM_FAULT_SET_HINDEX(hstate_index(h));
2813 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2815 return VM_FAULT_OOM;
2818 * Serialize hugepage allocation and instantiation, so that we don't
2819 * get spurious allocation failures if two CPUs race to instantiate
2820 * the same page in the page cache.
2822 mutex_lock(&hugetlb_instantiation_mutex);
2823 entry = huge_ptep_get(ptep);
2824 if (huge_pte_none(entry)) {
2825 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2832 * If we are going to COW the mapping later, we examine the pending
2833 * reservations for this page now. This will ensure that any
2834 * allocations necessary to record that reservation occur outside the
2835 * spinlock. For private mappings, we also lookup the pagecache
2836 * page now as it is used to determine if a reservation has been
2839 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2840 if (vma_needs_reservation(h, vma, address) < 0) {
2845 if (!(vma->vm_flags & VM_MAYSHARE))
2846 pagecache_page = hugetlbfs_pagecache_page(h,
2851 * hugetlb_cow() requires page locks of pte_page(entry) and
2852 * pagecache_page, so here we need take the former one
2853 * when page != pagecache_page or !pagecache_page.
2854 * Note that locking order is always pagecache_page -> page,
2855 * so no worry about deadlock.
2857 page = pte_page(entry);
2859 if (page != pagecache_page)
2862 spin_lock(&mm->page_table_lock);
2863 /* Check for a racing update before calling hugetlb_cow */
2864 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2865 goto out_page_table_lock;
2868 if (flags & FAULT_FLAG_WRITE) {
2869 if (!pte_write(entry)) {
2870 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2872 goto out_page_table_lock;
2874 entry = pte_mkdirty(entry);
2876 entry = pte_mkyoung(entry);
2877 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2878 flags & FAULT_FLAG_WRITE))
2879 update_mmu_cache(vma, address, ptep);
2881 out_page_table_lock:
2882 spin_unlock(&mm->page_table_lock);
2884 if (pagecache_page) {
2885 unlock_page(pagecache_page);
2886 put_page(pagecache_page);
2888 if (page != pagecache_page)
2893 mutex_unlock(&hugetlb_instantiation_mutex);
2898 /* Can be overriden by architectures */
2899 __attribute__((weak)) struct page *
2900 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2901 pud_t *pud, int write)
2907 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2908 struct page **pages, struct vm_area_struct **vmas,
2909 unsigned long *position, int *length, int i,
2912 unsigned long pfn_offset;
2913 unsigned long vaddr = *position;
2914 int remainder = *length;
2915 struct hstate *h = hstate_vma(vma);
2917 spin_lock(&mm->page_table_lock);
2918 while (vaddr < vma->vm_end && remainder) {
2924 * Some archs (sparc64, sh*) have multiple pte_ts to
2925 * each hugepage. We have to make sure we get the
2926 * first, for the page indexing below to work.
2928 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2929 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2932 * When coredumping, it suits get_dump_page if we just return
2933 * an error where there's an empty slot with no huge pagecache
2934 * to back it. This way, we avoid allocating a hugepage, and
2935 * the sparse dumpfile avoids allocating disk blocks, but its
2936 * huge holes still show up with zeroes where they need to be.
2938 if (absent && (flags & FOLL_DUMP) &&
2939 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2945 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2948 spin_unlock(&mm->page_table_lock);
2949 ret = hugetlb_fault(mm, vma, vaddr,
2950 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2951 spin_lock(&mm->page_table_lock);
2952 if (!(ret & VM_FAULT_ERROR))
2959 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2960 page = pte_page(huge_ptep_get(pte));
2963 pages[i] = mem_map_offset(page, pfn_offset);
2974 if (vaddr < vma->vm_end && remainder &&
2975 pfn_offset < pages_per_huge_page(h)) {
2977 * We use pfn_offset to avoid touching the pageframes
2978 * of this compound page.
2983 spin_unlock(&mm->page_table_lock);
2984 *length = remainder;
2987 return i ? i : -EFAULT;
2990 void hugetlb_change_protection(struct vm_area_struct *vma,
2991 unsigned long address, unsigned long end, pgprot_t newprot)
2993 struct mm_struct *mm = vma->vm_mm;
2994 unsigned long start = address;
2997 struct hstate *h = hstate_vma(vma);
2999 BUG_ON(address >= end);
3000 flush_cache_range(vma, address, end);
3002 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3003 spin_lock(&mm->page_table_lock);
3004 for (; address < end; address += huge_page_size(h)) {
3005 ptep = huge_pte_offset(mm, address);
3008 if (huge_pmd_unshare(mm, &address, ptep))
3010 if (!huge_pte_none(huge_ptep_get(ptep))) {
3011 pte = huge_ptep_get_and_clear(mm, address, ptep);
3012 pte = pte_mkhuge(pte_modify(pte, newprot));
3013 set_huge_pte_at(mm, address, ptep, pte);
3016 spin_unlock(&mm->page_table_lock);
3017 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3019 flush_tlb_range(vma, start, end);
3022 int hugetlb_reserve_pages(struct inode *inode,
3024 struct vm_area_struct *vma,
3025 vm_flags_t vm_flags)
3028 struct hstate *h = hstate_inode(inode);
3029 struct hugepage_subpool *spool = subpool_inode(inode);
3032 * Only apply hugepage reservation if asked. At fault time, an
3033 * attempt will be made for VM_NORESERVE to allocate a page
3034 * without using reserves
3036 if (vm_flags & VM_NORESERVE)
3040 * Shared mappings base their reservation on the number of pages that
3041 * are already allocated on behalf of the file. Private mappings need
3042 * to reserve the full area even if read-only as mprotect() may be
3043 * called to make the mapping read-write. Assume !vma is a shm mapping
3045 if (!vma || vma->vm_flags & VM_MAYSHARE)
3046 chg = region_chg(&inode->i_mapping->private_list, from, to);
3048 struct resv_map *resv_map = resv_map_alloc();
3054 set_vma_resv_map(vma, resv_map);
3055 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3063 /* There must be enough pages in the subpool for the mapping */
3064 if (hugepage_subpool_get_pages(spool, chg)) {
3070 * Check enough hugepages are available for the reservation.
3071 * Hand the pages back to the subpool if there are not
3073 ret = hugetlb_acct_memory(h, chg);
3075 hugepage_subpool_put_pages(spool, chg);
3080 * Account for the reservations made. Shared mappings record regions
3081 * that have reservations as they are shared by multiple VMAs.
3082 * When the last VMA disappears, the region map says how much
3083 * the reservation was and the page cache tells how much of
3084 * the reservation was consumed. Private mappings are per-VMA and
3085 * only the consumed reservations are tracked. When the VMA
3086 * disappears, the original reservation is the VMA size and the
3087 * consumed reservations are stored in the map. Hence, nothing
3088 * else has to be done for private mappings here
3090 if (!vma || vma->vm_flags & VM_MAYSHARE)
3091 region_add(&inode->i_mapping->private_list, from, to);
3099 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3101 struct hstate *h = hstate_inode(inode);
3102 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3103 struct hugepage_subpool *spool = subpool_inode(inode);
3105 spin_lock(&inode->i_lock);
3106 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3107 spin_unlock(&inode->i_lock);
3109 hugepage_subpool_put_pages(spool, (chg - freed));
3110 hugetlb_acct_memory(h, -(chg - freed));
3113 #ifdef CONFIG_MEMORY_FAILURE
3115 /* Should be called in hugetlb_lock */
3116 static int is_hugepage_on_freelist(struct page *hpage)
3120 struct hstate *h = page_hstate(hpage);
3121 int nid = page_to_nid(hpage);
3123 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3130 * This function is called from memory failure code.
3131 * Assume the caller holds page lock of the head page.
3133 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3135 struct hstate *h = page_hstate(hpage);
3136 int nid = page_to_nid(hpage);
3139 spin_lock(&hugetlb_lock);
3140 if (is_hugepage_on_freelist(hpage)) {
3141 list_del(&hpage->lru);
3142 set_page_refcounted(hpage);
3143 h->free_huge_pages--;
3144 h->free_huge_pages_node[nid]--;
3147 spin_unlock(&hugetlb_lock);