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>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
41 __initdata LIST_HEAD(huge_boot_pages);
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static 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_add(&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_del(&page->lru);
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 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));
628 INIT_LIST_HEAD(&page->lru);
630 spin_lock(&hugetlb_lock);
631 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
632 update_and_free_page(h, page);
633 h->surplus_huge_pages--;
634 h->surplus_huge_pages_node[nid]--;
636 enqueue_huge_page(h, page);
638 spin_unlock(&hugetlb_lock);
639 hugepage_subpool_put_pages(spool, 1);
642 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
644 set_compound_page_dtor(page, free_huge_page);
645 spin_lock(&hugetlb_lock);
647 h->nr_huge_pages_node[nid]++;
648 spin_unlock(&hugetlb_lock);
649 put_page(page); /* free it into the hugepage allocator */
652 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
655 int nr_pages = 1 << order;
656 struct page *p = page + 1;
658 /* we rely on prep_new_huge_page to set the destructor */
659 set_compound_order(page, order);
661 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
663 set_page_count(p, 0);
664 p->first_page = page;
668 int PageHuge(struct page *page)
670 compound_page_dtor *dtor;
672 if (!PageCompound(page))
675 page = compound_head(page);
676 dtor = get_compound_page_dtor(page);
678 return dtor == free_huge_page;
680 EXPORT_SYMBOL_GPL(PageHuge);
682 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
686 if (h->order >= MAX_ORDER)
689 page = alloc_pages_exact_node(nid,
690 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
691 __GFP_REPEAT|__GFP_NOWARN,
694 if (arch_prepare_hugepage(page)) {
695 __free_pages(page, huge_page_order(h));
698 prep_new_huge_page(h, page, nid);
705 * common helper functions for hstate_next_node_to_{alloc|free}.
706 * We may have allocated or freed a huge page based on a different
707 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
708 * be outside of *nodes_allowed. Ensure that we use an allowed
709 * node for alloc or free.
711 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
713 nid = next_node(nid, *nodes_allowed);
714 if (nid == MAX_NUMNODES)
715 nid = first_node(*nodes_allowed);
716 VM_BUG_ON(nid >= MAX_NUMNODES);
721 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
723 if (!node_isset(nid, *nodes_allowed))
724 nid = next_node_allowed(nid, nodes_allowed);
729 * returns the previously saved node ["this node"] from which to
730 * allocate a persistent huge page for the pool and advance the
731 * next node from which to allocate, handling wrap at end of node
734 static int hstate_next_node_to_alloc(struct hstate *h,
735 nodemask_t *nodes_allowed)
739 VM_BUG_ON(!nodes_allowed);
741 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
742 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
747 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
754 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
755 next_nid = start_nid;
758 page = alloc_fresh_huge_page_node(h, next_nid);
763 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
764 } while (next_nid != start_nid);
767 count_vm_event(HTLB_BUDDY_PGALLOC);
769 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
775 * helper for free_pool_huge_page() - return the previously saved
776 * node ["this node"] from which to free a huge page. Advance the
777 * next node id whether or not we find a free huge page to free so
778 * that the next attempt to free addresses the next node.
780 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
784 VM_BUG_ON(!nodes_allowed);
786 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
787 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
793 * Free huge page from pool from next node to free.
794 * Attempt to keep persistent huge pages more or less
795 * balanced over allowed nodes.
796 * Called with hugetlb_lock locked.
798 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
805 start_nid = hstate_next_node_to_free(h, nodes_allowed);
806 next_nid = start_nid;
810 * If we're returning unused surplus pages, only examine
811 * nodes with surplus pages.
813 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
814 !list_empty(&h->hugepage_freelists[next_nid])) {
816 list_entry(h->hugepage_freelists[next_nid].next,
818 list_del(&page->lru);
819 h->free_huge_pages--;
820 h->free_huge_pages_node[next_nid]--;
822 h->surplus_huge_pages--;
823 h->surplus_huge_pages_node[next_nid]--;
825 update_and_free_page(h, page);
829 next_nid = hstate_next_node_to_free(h, nodes_allowed);
830 } while (next_nid != start_nid);
835 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
840 if (h->order >= MAX_ORDER)
844 * Assume we will successfully allocate the surplus page to
845 * prevent racing processes from causing the surplus to exceed
848 * This however introduces a different race, where a process B
849 * tries to grow the static hugepage pool while alloc_pages() is
850 * called by process A. B will only examine the per-node
851 * counters in determining if surplus huge pages can be
852 * converted to normal huge pages in adjust_pool_surplus(). A
853 * won't be able to increment the per-node counter, until the
854 * lock is dropped by B, but B doesn't drop hugetlb_lock until
855 * no more huge pages can be converted from surplus to normal
856 * state (and doesn't try to convert again). Thus, we have a
857 * case where a surplus huge page exists, the pool is grown, and
858 * the surplus huge page still exists after, even though it
859 * should just have been converted to a normal huge page. This
860 * does not leak memory, though, as the hugepage will be freed
861 * once it is out of use. It also does not allow the counters to
862 * go out of whack in adjust_pool_surplus() as we don't modify
863 * the node values until we've gotten the hugepage and only the
864 * per-node value is checked there.
866 spin_lock(&hugetlb_lock);
867 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
868 spin_unlock(&hugetlb_lock);
872 h->surplus_huge_pages++;
874 spin_unlock(&hugetlb_lock);
876 if (nid == NUMA_NO_NODE)
877 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
878 __GFP_REPEAT|__GFP_NOWARN,
881 page = alloc_pages_exact_node(nid,
882 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
883 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
885 if (page && arch_prepare_hugepage(page)) {
886 __free_pages(page, huge_page_order(h));
890 spin_lock(&hugetlb_lock);
892 r_nid = page_to_nid(page);
893 set_compound_page_dtor(page, free_huge_page);
895 * We incremented the global counters already
897 h->nr_huge_pages_node[r_nid]++;
898 h->surplus_huge_pages_node[r_nid]++;
899 __count_vm_event(HTLB_BUDDY_PGALLOC);
902 h->surplus_huge_pages--;
903 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
905 spin_unlock(&hugetlb_lock);
911 * This allocation function is useful in the context where vma is irrelevant.
912 * E.g. soft-offlining uses this function because it only cares physical
913 * address of error page.
915 struct page *alloc_huge_page_node(struct hstate *h, int nid)
919 spin_lock(&hugetlb_lock);
920 page = dequeue_huge_page_node(h, nid);
921 spin_unlock(&hugetlb_lock);
924 page = alloc_buddy_huge_page(h, nid);
930 * Increase the hugetlb pool such that it can accommodate a reservation
933 static int gather_surplus_pages(struct hstate *h, int delta)
935 struct list_head surplus_list;
936 struct page *page, *tmp;
938 int needed, allocated;
940 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
942 h->resv_huge_pages += delta;
947 INIT_LIST_HEAD(&surplus_list);
951 spin_unlock(&hugetlb_lock);
952 for (i = 0; i < needed; i++) {
953 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
956 * We were not able to allocate enough pages to
957 * satisfy the entire reservation so we free what
958 * we've allocated so far.
962 list_add(&page->lru, &surplus_list);
967 * After retaking hugetlb_lock, we need to recalculate 'needed'
968 * because either resv_huge_pages or free_huge_pages may have changed.
970 spin_lock(&hugetlb_lock);
971 needed = (h->resv_huge_pages + delta) -
972 (h->free_huge_pages + allocated);
977 * The surplus_list now contains _at_least_ the number of extra pages
978 * needed to accommodate the reservation. Add the appropriate number
979 * of pages to the hugetlb pool and free the extras back to the buddy
980 * allocator. Commit the entire reservation here to prevent another
981 * process from stealing the pages as they are added to the pool but
982 * before they are reserved.
985 h->resv_huge_pages += delta;
988 /* Free the needed pages to the hugetlb pool */
989 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
992 list_del(&page->lru);
994 * This page is now managed by the hugetlb allocator and has
995 * no users -- drop the buddy allocator's reference.
997 put_page_testzero(page);
998 VM_BUG_ON(page_count(page));
999 enqueue_huge_page(h, page);
1001 spin_unlock(&hugetlb_lock);
1003 /* Free unnecessary surplus pages to the buddy allocator */
1005 if (!list_empty(&surplus_list)) {
1006 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1007 list_del(&page->lru);
1011 spin_lock(&hugetlb_lock);
1017 * When releasing a hugetlb pool reservation, any surplus pages that were
1018 * allocated to satisfy the reservation must be explicitly freed if they were
1020 * Called with hugetlb_lock held.
1022 static void return_unused_surplus_pages(struct hstate *h,
1023 unsigned long unused_resv_pages)
1025 unsigned long nr_pages;
1027 /* Uncommit the reservation */
1028 h->resv_huge_pages -= unused_resv_pages;
1030 /* Cannot return gigantic pages currently */
1031 if (h->order >= MAX_ORDER)
1034 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1037 * We want to release as many surplus pages as possible, spread
1038 * evenly across all nodes with memory. Iterate across these nodes
1039 * until we can no longer free unreserved surplus pages. This occurs
1040 * when the nodes with surplus pages have no free pages.
1041 * free_pool_huge_page() will balance the the freed pages across the
1042 * on-line nodes with memory and will handle the hstate accounting.
1044 while (nr_pages--) {
1045 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1051 * Determine if the huge page at addr within the vma has an associated
1052 * reservation. Where it does not we will need to logically increase
1053 * reservation and actually increase subpool usage before an allocation
1054 * can occur. Where any new reservation would be required the
1055 * reservation change is prepared, but not committed. Once the page
1056 * has been allocated from the subpool and instantiated the change should
1057 * be committed via vma_commit_reservation. No action is required on
1060 static long vma_needs_reservation(struct hstate *h,
1061 struct vm_area_struct *vma, unsigned long addr)
1063 struct address_space *mapping = vma->vm_file->f_mapping;
1064 struct inode *inode = mapping->host;
1066 if (vma->vm_flags & VM_MAYSHARE) {
1067 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1068 return region_chg(&inode->i_mapping->private_list,
1071 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1076 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1077 struct resv_map *reservations = vma_resv_map(vma);
1079 err = region_chg(&reservations->regions, idx, idx + 1);
1085 static void vma_commit_reservation(struct hstate *h,
1086 struct vm_area_struct *vma, unsigned long addr)
1088 struct address_space *mapping = vma->vm_file->f_mapping;
1089 struct inode *inode = mapping->host;
1091 if (vma->vm_flags & VM_MAYSHARE) {
1092 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1093 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1095 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1096 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1097 struct resv_map *reservations = vma_resv_map(vma);
1099 /* Mark this page used in the map. */
1100 region_add(&reservations->regions, idx, idx + 1);
1104 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1105 unsigned long addr, int avoid_reserve)
1107 struct hugepage_subpool *spool = subpool_vma(vma);
1108 struct hstate *h = hstate_vma(vma);
1113 * Processes that did not create the mapping will have no
1114 * reserves and will not have accounted against subpool
1115 * limit. Check that the subpool limit can be made before
1116 * satisfying the allocation MAP_NORESERVE mappings may also
1117 * need pages and subpool limit allocated allocated if no reserve
1120 chg = vma_needs_reservation(h, vma, addr);
1122 return ERR_PTR(-VM_FAULT_OOM);
1124 if (hugepage_subpool_get_pages(spool, chg))
1125 return ERR_PTR(-VM_FAULT_SIGBUS);
1127 spin_lock(&hugetlb_lock);
1128 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1129 spin_unlock(&hugetlb_lock);
1132 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1134 hugepage_subpool_put_pages(spool, chg);
1135 return ERR_PTR(-VM_FAULT_SIGBUS);
1139 set_page_private(page, (unsigned long)spool);
1141 vma_commit_reservation(h, vma, addr);
1146 int __weak alloc_bootmem_huge_page(struct hstate *h)
1148 struct huge_bootmem_page *m;
1149 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1154 addr = __alloc_bootmem_node_nopanic(
1155 NODE_DATA(hstate_next_node_to_alloc(h,
1156 &node_states[N_HIGH_MEMORY])),
1157 huge_page_size(h), huge_page_size(h), 0);
1161 * Use the beginning of the huge page to store the
1162 * huge_bootmem_page struct (until gather_bootmem
1163 * puts them into the mem_map).
1173 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1174 /* Put them into a private list first because mem_map is not up yet */
1175 list_add(&m->list, &huge_boot_pages);
1180 static void prep_compound_huge_page(struct page *page, int order)
1182 if (unlikely(order > (MAX_ORDER - 1)))
1183 prep_compound_gigantic_page(page, order);
1185 prep_compound_page(page, order);
1188 /* Put bootmem huge pages into the standard lists after mem_map is up */
1189 static void __init gather_bootmem_prealloc(void)
1191 struct huge_bootmem_page *m;
1193 list_for_each_entry(m, &huge_boot_pages, list) {
1194 struct hstate *h = m->hstate;
1197 #ifdef CONFIG_HIGHMEM
1198 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1199 free_bootmem_late((unsigned long)m,
1200 sizeof(struct huge_bootmem_page));
1202 page = virt_to_page(m);
1204 __ClearPageReserved(page);
1205 WARN_ON(page_count(page) != 1);
1206 prep_compound_huge_page(page, h->order);
1207 prep_new_huge_page(h, page, page_to_nid(page));
1209 * If we had gigantic hugepages allocated at boot time, we need
1210 * to restore the 'stolen' pages to totalram_pages in order to
1211 * fix confusing memory reports from free(1) and another
1212 * side-effects, like CommitLimit going negative.
1214 if (h->order > (MAX_ORDER - 1))
1215 totalram_pages += 1 << h->order;
1219 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1223 for (i = 0; i < h->max_huge_pages; ++i) {
1224 if (h->order >= MAX_ORDER) {
1225 if (!alloc_bootmem_huge_page(h))
1227 } else if (!alloc_fresh_huge_page(h,
1228 &node_states[N_HIGH_MEMORY]))
1231 h->max_huge_pages = i;
1234 static void __init hugetlb_init_hstates(void)
1238 for_each_hstate(h) {
1239 /* oversize hugepages were init'ed in early boot */
1240 if (h->order < MAX_ORDER)
1241 hugetlb_hstate_alloc_pages(h);
1245 static char * __init memfmt(char *buf, unsigned long n)
1247 if (n >= (1UL << 30))
1248 sprintf(buf, "%lu GB", n >> 30);
1249 else if (n >= (1UL << 20))
1250 sprintf(buf, "%lu MB", n >> 20);
1252 sprintf(buf, "%lu KB", n >> 10);
1256 static void __init report_hugepages(void)
1260 for_each_hstate(h) {
1262 printk(KERN_INFO "HugeTLB registered %s page size, "
1263 "pre-allocated %ld pages\n",
1264 memfmt(buf, huge_page_size(h)),
1265 h->free_huge_pages);
1269 #ifdef CONFIG_HIGHMEM
1270 static void try_to_free_low(struct hstate *h, unsigned long count,
1271 nodemask_t *nodes_allowed)
1275 if (h->order >= MAX_ORDER)
1278 for_each_node_mask(i, *nodes_allowed) {
1279 struct page *page, *next;
1280 struct list_head *freel = &h->hugepage_freelists[i];
1281 list_for_each_entry_safe(page, next, freel, lru) {
1282 if (count >= h->nr_huge_pages)
1284 if (PageHighMem(page))
1286 list_del(&page->lru);
1287 update_and_free_page(h, page);
1288 h->free_huge_pages--;
1289 h->free_huge_pages_node[page_to_nid(page)]--;
1294 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1295 nodemask_t *nodes_allowed)
1301 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1302 * balanced by operating on them in a round-robin fashion.
1303 * Returns 1 if an adjustment was made.
1305 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1308 int start_nid, next_nid;
1311 VM_BUG_ON(delta != -1 && delta != 1);
1314 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1316 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1317 next_nid = start_nid;
1323 * To shrink on this node, there must be a surplus page
1325 if (!h->surplus_huge_pages_node[nid]) {
1326 next_nid = hstate_next_node_to_alloc(h,
1333 * Surplus cannot exceed the total number of pages
1335 if (h->surplus_huge_pages_node[nid] >=
1336 h->nr_huge_pages_node[nid]) {
1337 next_nid = hstate_next_node_to_free(h,
1343 h->surplus_huge_pages += delta;
1344 h->surplus_huge_pages_node[nid] += delta;
1347 } while (next_nid != start_nid);
1352 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1353 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1354 nodemask_t *nodes_allowed)
1356 unsigned long min_count, ret;
1358 if (h->order >= MAX_ORDER)
1359 return h->max_huge_pages;
1362 * Increase the pool size
1363 * First take pages out of surplus state. Then make up the
1364 * remaining difference by allocating fresh huge pages.
1366 * We might race with alloc_buddy_huge_page() here and be unable
1367 * to convert a surplus huge page to a normal huge page. That is
1368 * not critical, though, it just means the overall size of the
1369 * pool might be one hugepage larger than it needs to be, but
1370 * within all the constraints specified by the sysctls.
1372 spin_lock(&hugetlb_lock);
1373 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1374 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1378 while (count > persistent_huge_pages(h)) {
1380 * If this allocation races such that we no longer need the
1381 * page, free_huge_page will handle it by freeing the page
1382 * and reducing the surplus.
1384 spin_unlock(&hugetlb_lock);
1385 ret = alloc_fresh_huge_page(h, nodes_allowed);
1386 spin_lock(&hugetlb_lock);
1390 /* Bail for signals. Probably ctrl-c from user */
1391 if (signal_pending(current))
1396 * Decrease the pool size
1397 * First return free pages to the buddy allocator (being careful
1398 * to keep enough around to satisfy reservations). Then place
1399 * pages into surplus state as needed so the pool will shrink
1400 * to the desired size as pages become free.
1402 * By placing pages into the surplus state independent of the
1403 * overcommit value, we are allowing the surplus pool size to
1404 * exceed overcommit. There are few sane options here. Since
1405 * alloc_buddy_huge_page() is checking the global counter,
1406 * though, we'll note that we're not allowed to exceed surplus
1407 * and won't grow the pool anywhere else. Not until one of the
1408 * sysctls are changed, or the surplus pages go out of use.
1410 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1411 min_count = max(count, min_count);
1412 try_to_free_low(h, min_count, nodes_allowed);
1413 while (min_count < persistent_huge_pages(h)) {
1414 if (!free_pool_huge_page(h, nodes_allowed, 0))
1417 while (count < persistent_huge_pages(h)) {
1418 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1422 ret = persistent_huge_pages(h);
1423 spin_unlock(&hugetlb_lock);
1427 #define HSTATE_ATTR_RO(_name) \
1428 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1430 #define HSTATE_ATTR(_name) \
1431 static struct kobj_attribute _name##_attr = \
1432 __ATTR(_name, 0644, _name##_show, _name##_store)
1434 static struct kobject *hugepages_kobj;
1435 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1437 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1439 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1443 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1444 if (hstate_kobjs[i] == kobj) {
1446 *nidp = NUMA_NO_NODE;
1450 return kobj_to_node_hstate(kobj, nidp);
1453 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1454 struct kobj_attribute *attr, char *buf)
1457 unsigned long nr_huge_pages;
1460 h = kobj_to_hstate(kobj, &nid);
1461 if (nid == NUMA_NO_NODE)
1462 nr_huge_pages = h->nr_huge_pages;
1464 nr_huge_pages = h->nr_huge_pages_node[nid];
1466 return sprintf(buf, "%lu\n", nr_huge_pages);
1469 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1470 struct kobject *kobj, struct kobj_attribute *attr,
1471 const char *buf, size_t len)
1475 unsigned long count;
1477 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1479 err = strict_strtoul(buf, 10, &count);
1483 h = kobj_to_hstate(kobj, &nid);
1484 if (h->order >= MAX_ORDER) {
1489 if (nid == NUMA_NO_NODE) {
1491 * global hstate attribute
1493 if (!(obey_mempolicy &&
1494 init_nodemask_of_mempolicy(nodes_allowed))) {
1495 NODEMASK_FREE(nodes_allowed);
1496 nodes_allowed = &node_states[N_HIGH_MEMORY];
1498 } else if (nodes_allowed) {
1500 * per node hstate attribute: adjust count to global,
1501 * but restrict alloc/free to the specified node.
1503 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1504 init_nodemask_of_node(nodes_allowed, nid);
1506 nodes_allowed = &node_states[N_HIGH_MEMORY];
1508 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1510 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1511 NODEMASK_FREE(nodes_allowed);
1515 NODEMASK_FREE(nodes_allowed);
1519 static ssize_t nr_hugepages_show(struct kobject *kobj,
1520 struct kobj_attribute *attr, char *buf)
1522 return nr_hugepages_show_common(kobj, attr, buf);
1525 static ssize_t nr_hugepages_store(struct kobject *kobj,
1526 struct kobj_attribute *attr, const char *buf, size_t len)
1528 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1530 HSTATE_ATTR(nr_hugepages);
1535 * hstate attribute for optionally mempolicy-based constraint on persistent
1536 * huge page alloc/free.
1538 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1539 struct kobj_attribute *attr, char *buf)
1541 return nr_hugepages_show_common(kobj, attr, buf);
1544 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1545 struct kobj_attribute *attr, const char *buf, size_t len)
1547 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1549 HSTATE_ATTR(nr_hugepages_mempolicy);
1553 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1554 struct kobj_attribute *attr, char *buf)
1556 struct hstate *h = kobj_to_hstate(kobj, NULL);
1557 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1560 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1561 struct kobj_attribute *attr, const char *buf, size_t count)
1564 unsigned long input;
1565 struct hstate *h = kobj_to_hstate(kobj, NULL);
1567 if (h->order >= MAX_ORDER)
1570 err = strict_strtoul(buf, 10, &input);
1574 spin_lock(&hugetlb_lock);
1575 h->nr_overcommit_huge_pages = input;
1576 spin_unlock(&hugetlb_lock);
1580 HSTATE_ATTR(nr_overcommit_hugepages);
1582 static ssize_t free_hugepages_show(struct kobject *kobj,
1583 struct kobj_attribute *attr, char *buf)
1586 unsigned long free_huge_pages;
1589 h = kobj_to_hstate(kobj, &nid);
1590 if (nid == NUMA_NO_NODE)
1591 free_huge_pages = h->free_huge_pages;
1593 free_huge_pages = h->free_huge_pages_node[nid];
1595 return sprintf(buf, "%lu\n", free_huge_pages);
1597 HSTATE_ATTR_RO(free_hugepages);
1599 static ssize_t resv_hugepages_show(struct kobject *kobj,
1600 struct kobj_attribute *attr, char *buf)
1602 struct hstate *h = kobj_to_hstate(kobj, NULL);
1603 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1605 HSTATE_ATTR_RO(resv_hugepages);
1607 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1608 struct kobj_attribute *attr, char *buf)
1611 unsigned long surplus_huge_pages;
1614 h = kobj_to_hstate(kobj, &nid);
1615 if (nid == NUMA_NO_NODE)
1616 surplus_huge_pages = h->surplus_huge_pages;
1618 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1620 return sprintf(buf, "%lu\n", surplus_huge_pages);
1622 HSTATE_ATTR_RO(surplus_hugepages);
1624 static struct attribute *hstate_attrs[] = {
1625 &nr_hugepages_attr.attr,
1626 &nr_overcommit_hugepages_attr.attr,
1627 &free_hugepages_attr.attr,
1628 &resv_hugepages_attr.attr,
1629 &surplus_hugepages_attr.attr,
1631 &nr_hugepages_mempolicy_attr.attr,
1636 static struct attribute_group hstate_attr_group = {
1637 .attrs = hstate_attrs,
1640 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1641 struct kobject **hstate_kobjs,
1642 struct attribute_group *hstate_attr_group)
1645 int hi = h - hstates;
1647 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1648 if (!hstate_kobjs[hi])
1651 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1653 kobject_put(hstate_kobjs[hi]);
1658 static void __init hugetlb_sysfs_init(void)
1663 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1664 if (!hugepages_kobj)
1667 for_each_hstate(h) {
1668 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1669 hstate_kobjs, &hstate_attr_group);
1671 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1679 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1680 * with node sysdevs in node_devices[] using a parallel array. The array
1681 * index of a node sysdev or _hstate == node id.
1682 * This is here to avoid any static dependency of the node sysdev driver, in
1683 * the base kernel, on the hugetlb module.
1685 struct node_hstate {
1686 struct kobject *hugepages_kobj;
1687 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1689 struct node_hstate node_hstates[MAX_NUMNODES];
1692 * A subset of global hstate attributes for node sysdevs
1694 static struct attribute *per_node_hstate_attrs[] = {
1695 &nr_hugepages_attr.attr,
1696 &free_hugepages_attr.attr,
1697 &surplus_hugepages_attr.attr,
1701 static struct attribute_group per_node_hstate_attr_group = {
1702 .attrs = per_node_hstate_attrs,
1706 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1707 * Returns node id via non-NULL nidp.
1709 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1713 for (nid = 0; nid < nr_node_ids; nid++) {
1714 struct node_hstate *nhs = &node_hstates[nid];
1716 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1717 if (nhs->hstate_kobjs[i] == kobj) {
1729 * Unregister hstate attributes from a single node sysdev.
1730 * No-op if no hstate attributes attached.
1732 void hugetlb_unregister_node(struct node *node)
1735 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1737 if (!nhs->hugepages_kobj)
1738 return; /* no hstate attributes */
1741 if (nhs->hstate_kobjs[h - hstates]) {
1742 kobject_put(nhs->hstate_kobjs[h - hstates]);
1743 nhs->hstate_kobjs[h - hstates] = NULL;
1746 kobject_put(nhs->hugepages_kobj);
1747 nhs->hugepages_kobj = NULL;
1751 * hugetlb module exit: unregister hstate attributes from node sysdevs
1754 static void hugetlb_unregister_all_nodes(void)
1759 * disable node sysdev registrations.
1761 register_hugetlbfs_with_node(NULL, NULL);
1764 * remove hstate attributes from any nodes that have them.
1766 for (nid = 0; nid < nr_node_ids; nid++)
1767 hugetlb_unregister_node(&node_devices[nid]);
1771 * Register hstate attributes for a single node sysdev.
1772 * No-op if attributes already registered.
1774 void hugetlb_register_node(struct node *node)
1777 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1780 if (nhs->hugepages_kobj)
1781 return; /* already allocated */
1783 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1784 &node->sysdev.kobj);
1785 if (!nhs->hugepages_kobj)
1788 for_each_hstate(h) {
1789 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1791 &per_node_hstate_attr_group);
1793 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1795 h->name, node->sysdev.id);
1796 hugetlb_unregister_node(node);
1803 * hugetlb init time: register hstate attributes for all registered node
1804 * sysdevs of nodes that have memory. All on-line nodes should have
1805 * registered their associated sysdev by this time.
1807 static void hugetlb_register_all_nodes(void)
1811 for_each_node_state(nid, N_HIGH_MEMORY) {
1812 struct node *node = &node_devices[nid];
1813 if (node->sysdev.id == nid)
1814 hugetlb_register_node(node);
1818 * Let the node sysdev driver know we're here so it can
1819 * [un]register hstate attributes on node hotplug.
1821 register_hugetlbfs_with_node(hugetlb_register_node,
1822 hugetlb_unregister_node);
1824 #else /* !CONFIG_NUMA */
1826 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1834 static void hugetlb_unregister_all_nodes(void) { }
1836 static void hugetlb_register_all_nodes(void) { }
1840 static void __exit hugetlb_exit(void)
1844 hugetlb_unregister_all_nodes();
1846 for_each_hstate(h) {
1847 kobject_put(hstate_kobjs[h - hstates]);
1850 kobject_put(hugepages_kobj);
1852 module_exit(hugetlb_exit);
1854 static int __init hugetlb_init(void)
1856 /* Some platform decide whether they support huge pages at boot
1857 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1858 * there is no such support
1860 if (HPAGE_SHIFT == 0)
1863 if (!size_to_hstate(default_hstate_size)) {
1864 default_hstate_size = HPAGE_SIZE;
1865 if (!size_to_hstate(default_hstate_size))
1866 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1868 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1869 if (default_hstate_max_huge_pages)
1870 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1872 hugetlb_init_hstates();
1874 gather_bootmem_prealloc();
1878 hugetlb_sysfs_init();
1880 hugetlb_register_all_nodes();
1884 module_init(hugetlb_init);
1886 /* Should be called on processing a hugepagesz=... option */
1887 void __init hugetlb_add_hstate(unsigned order)
1892 if (size_to_hstate(PAGE_SIZE << order)) {
1893 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1896 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1898 h = &hstates[max_hstate++];
1900 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1901 h->nr_huge_pages = 0;
1902 h->free_huge_pages = 0;
1903 for (i = 0; i < MAX_NUMNODES; ++i)
1904 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1905 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1906 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1907 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1908 huge_page_size(h)/1024);
1913 static int __init hugetlb_nrpages_setup(char *s)
1916 static unsigned long *last_mhp;
1919 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1920 * so this hugepages= parameter goes to the "default hstate".
1923 mhp = &default_hstate_max_huge_pages;
1925 mhp = &parsed_hstate->max_huge_pages;
1927 if (mhp == last_mhp) {
1928 printk(KERN_WARNING "hugepages= specified twice without "
1929 "interleaving hugepagesz=, ignoring\n");
1933 if (sscanf(s, "%lu", mhp) <= 0)
1937 * Global state is always initialized later in hugetlb_init.
1938 * But we need to allocate >= MAX_ORDER hstates here early to still
1939 * use the bootmem allocator.
1941 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1942 hugetlb_hstate_alloc_pages(parsed_hstate);
1948 __setup("hugepages=", hugetlb_nrpages_setup);
1950 static int __init hugetlb_default_setup(char *s)
1952 default_hstate_size = memparse(s, &s);
1955 __setup("default_hugepagesz=", hugetlb_default_setup);
1957 static unsigned int cpuset_mems_nr(unsigned int *array)
1960 unsigned int nr = 0;
1962 for_each_node_mask(node, cpuset_current_mems_allowed)
1968 #ifdef CONFIG_SYSCTL
1969 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1970 struct ctl_table *table, int write,
1971 void __user *buffer, size_t *length, loff_t *ppos)
1973 struct hstate *h = &default_hstate;
1977 tmp = h->max_huge_pages;
1979 if (write && h->order >= MAX_ORDER)
1983 table->maxlen = sizeof(unsigned long);
1984 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1989 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1990 GFP_KERNEL | __GFP_NORETRY);
1991 if (!(obey_mempolicy &&
1992 init_nodemask_of_mempolicy(nodes_allowed))) {
1993 NODEMASK_FREE(nodes_allowed);
1994 nodes_allowed = &node_states[N_HIGH_MEMORY];
1996 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1998 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1999 NODEMASK_FREE(nodes_allowed);
2005 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2006 void __user *buffer, size_t *length, loff_t *ppos)
2009 return hugetlb_sysctl_handler_common(false, table, write,
2010 buffer, length, ppos);
2014 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2015 void __user *buffer, size_t *length, loff_t *ppos)
2017 return hugetlb_sysctl_handler_common(true, table, write,
2018 buffer, length, ppos);
2020 #endif /* CONFIG_NUMA */
2022 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2023 void __user *buffer,
2024 size_t *length, loff_t *ppos)
2026 proc_dointvec(table, write, buffer, length, ppos);
2027 if (hugepages_treat_as_movable)
2028 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2030 htlb_alloc_mask = GFP_HIGHUSER;
2034 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2035 void __user *buffer,
2036 size_t *length, loff_t *ppos)
2038 struct hstate *h = &default_hstate;
2042 tmp = h->nr_overcommit_huge_pages;
2044 if (write && h->order >= MAX_ORDER)
2048 table->maxlen = sizeof(unsigned long);
2049 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2054 spin_lock(&hugetlb_lock);
2055 h->nr_overcommit_huge_pages = tmp;
2056 spin_unlock(&hugetlb_lock);
2062 #endif /* CONFIG_SYSCTL */
2064 void hugetlb_report_meminfo(struct seq_file *m)
2066 struct hstate *h = &default_hstate;
2068 "HugePages_Total: %5lu\n"
2069 "HugePages_Free: %5lu\n"
2070 "HugePages_Rsvd: %5lu\n"
2071 "HugePages_Surp: %5lu\n"
2072 "Hugepagesize: %8lu kB\n",
2076 h->surplus_huge_pages,
2077 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2080 int hugetlb_report_node_meminfo(int nid, char *buf)
2082 struct hstate *h = &default_hstate;
2084 "Node %d HugePages_Total: %5u\n"
2085 "Node %d HugePages_Free: %5u\n"
2086 "Node %d HugePages_Surp: %5u\n",
2087 nid, h->nr_huge_pages_node[nid],
2088 nid, h->free_huge_pages_node[nid],
2089 nid, h->surplus_huge_pages_node[nid]);
2092 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2093 unsigned long hugetlb_total_pages(void)
2095 struct hstate *h = &default_hstate;
2096 return h->nr_huge_pages * pages_per_huge_page(h);
2099 static int hugetlb_acct_memory(struct hstate *h, long delta)
2103 spin_lock(&hugetlb_lock);
2105 * When cpuset is configured, it breaks the strict hugetlb page
2106 * reservation as the accounting is done on a global variable. Such
2107 * reservation is completely rubbish in the presence of cpuset because
2108 * the reservation is not checked against page availability for the
2109 * current cpuset. Application can still potentially OOM'ed by kernel
2110 * with lack of free htlb page in cpuset that the task is in.
2111 * Attempt to enforce strict accounting with cpuset is almost
2112 * impossible (or too ugly) because cpuset is too fluid that
2113 * task or memory node can be dynamically moved between cpusets.
2115 * The change of semantics for shared hugetlb mapping with cpuset is
2116 * undesirable. However, in order to preserve some of the semantics,
2117 * we fall back to check against current free page availability as
2118 * a best attempt and hopefully to minimize the impact of changing
2119 * semantics that cpuset has.
2122 if (gather_surplus_pages(h, delta) < 0)
2125 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2126 return_unused_surplus_pages(h, delta);
2133 return_unused_surplus_pages(h, (unsigned long) -delta);
2136 spin_unlock(&hugetlb_lock);
2140 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2142 struct resv_map *reservations = vma_resv_map(vma);
2145 * This new VMA should share its siblings reservation map if present.
2146 * The VMA will only ever have a valid reservation map pointer where
2147 * it is being copied for another still existing VMA. As that VMA
2148 * has a reference to the reservation map it cannot disappear until
2149 * after this open call completes. It is therefore safe to take a
2150 * new reference here without additional locking.
2153 kref_get(&reservations->refs);
2156 static void resv_map_put(struct vm_area_struct *vma)
2158 struct resv_map *reservations = vma_resv_map(vma);
2162 kref_put(&reservations->refs, resv_map_release);
2165 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2167 struct hstate *h = hstate_vma(vma);
2168 struct resv_map *reservations = vma_resv_map(vma);
2169 struct hugepage_subpool *spool = subpool_vma(vma);
2170 unsigned long reserve;
2171 unsigned long start;
2175 start = vma_hugecache_offset(h, vma, vma->vm_start);
2176 end = vma_hugecache_offset(h, vma, vma->vm_end);
2178 reserve = (end - start) -
2179 region_count(&reservations->regions, start, end);
2184 hugetlb_acct_memory(h, -reserve);
2185 hugepage_subpool_put_pages(spool, reserve);
2191 * We cannot handle pagefaults against hugetlb pages at all. They cause
2192 * handle_mm_fault() to try to instantiate regular-sized pages in the
2193 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2196 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2202 const struct vm_operations_struct hugetlb_vm_ops = {
2203 .fault = hugetlb_vm_op_fault,
2204 .open = hugetlb_vm_op_open,
2205 .close = hugetlb_vm_op_close,
2208 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2215 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2217 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2219 entry = pte_mkyoung(entry);
2220 entry = pte_mkhuge(entry);
2225 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2226 unsigned long address, pte_t *ptep)
2230 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2231 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2232 update_mmu_cache(vma, address, ptep);
2236 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2237 struct vm_area_struct *vma)
2239 pte_t *src_pte, *dst_pte, entry;
2240 struct page *ptepage;
2243 struct hstate *h = hstate_vma(vma);
2244 unsigned long sz = huge_page_size(h);
2246 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2248 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2249 src_pte = huge_pte_offset(src, addr);
2252 dst_pte = huge_pte_alloc(dst, addr, sz);
2256 /* If the pagetables are shared don't copy or take references */
2257 if (dst_pte == src_pte)
2260 spin_lock(&dst->page_table_lock);
2261 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2262 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2264 huge_ptep_set_wrprotect(src, addr, src_pte);
2265 entry = huge_ptep_get(src_pte);
2266 ptepage = pte_page(entry);
2268 page_dup_rmap(ptepage);
2269 set_huge_pte_at(dst, addr, dst_pte, entry);
2271 spin_unlock(&src->page_table_lock);
2272 spin_unlock(&dst->page_table_lock);
2280 static int is_hugetlb_entry_migration(pte_t pte)
2284 if (huge_pte_none(pte) || pte_present(pte))
2286 swp = pte_to_swp_entry(pte);
2287 if (non_swap_entry(swp) && is_migration_entry(swp))
2293 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2297 if (huge_pte_none(pte) || pte_present(pte))
2299 swp = pte_to_swp_entry(pte);
2300 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2306 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2307 unsigned long end, struct page *ref_page)
2309 struct mm_struct *mm = vma->vm_mm;
2310 unsigned long address;
2315 struct hstate *h = hstate_vma(vma);
2316 unsigned long sz = huge_page_size(h);
2319 * A page gathering list, protected by per file i_mmap_mutex. The
2320 * lock is used to avoid list corruption from multiple unmapping
2321 * of the same page since we are using page->lru.
2323 LIST_HEAD(page_list);
2325 WARN_ON(!is_vm_hugetlb_page(vma));
2326 BUG_ON(start & ~huge_page_mask(h));
2327 BUG_ON(end & ~huge_page_mask(h));
2329 mmu_notifier_invalidate_range_start(mm, start, end);
2330 spin_lock(&mm->page_table_lock);
2331 for (address = start; address < end; address += sz) {
2332 ptep = huge_pte_offset(mm, address);
2336 if (huge_pmd_unshare(mm, &address, ptep))
2340 * If a reference page is supplied, it is because a specific
2341 * page is being unmapped, not a range. Ensure the page we
2342 * are about to unmap is the actual page of interest.
2345 pte = huge_ptep_get(ptep);
2346 if (huge_pte_none(pte))
2348 page = pte_page(pte);
2349 if (page != ref_page)
2353 * Mark the VMA as having unmapped its page so that
2354 * future faults in this VMA will fail rather than
2355 * looking like data was lost
2357 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2360 pte = huge_ptep_get_and_clear(mm, address, 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 set_page_dirty(page);
2373 list_add(&page->lru, &page_list);
2375 spin_unlock(&mm->page_table_lock);
2376 flush_tlb_range(vma, start, end);
2377 mmu_notifier_invalidate_range_end(mm, start, end);
2378 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2379 page_remove_rmap(page);
2380 list_del(&page->lru);
2385 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2386 unsigned long start, unsigned long end,
2387 struct page *ref_page)
2389 __unmap_hugepage_range(vma, start, end, ref_page);
2392 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2393 * test will fail on a vma being torn down, and not grab a page table
2394 * on its way out. We're lucky that the flag has such an appropriate
2395 * name, and can in fact be safely cleared here. We could clear it
2396 * before the __unmap_hugepage_range above, but all that's necessary
2397 * is to clear it before releasing the i_mmap_mutex. This works
2398 * because in the context this is called, the VMA is about to be
2399 * destroyed and the i_mmap_mutex is held.
2401 vma->vm_flags &= ~VM_MAYSHARE;
2404 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2405 unsigned long end, struct page *ref_page)
2407 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2408 __unmap_hugepage_range(vma, start, end, ref_page);
2409 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2413 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2414 * mappping it owns the reserve page for. The intention is to unmap the page
2415 * from other VMAs and let the children be SIGKILLed if they are faulting the
2418 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2419 struct page *page, unsigned long address)
2421 struct hstate *h = hstate_vma(vma);
2422 struct vm_area_struct *iter_vma;
2423 struct address_space *mapping;
2424 struct prio_tree_iter iter;
2428 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2429 * from page cache lookup which is in HPAGE_SIZE units.
2431 address = address & huge_page_mask(h);
2432 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2434 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2437 * Take the mapping lock for the duration of the table walk. As
2438 * this mapping should be shared between all the VMAs,
2439 * __unmap_hugepage_range() is called as the lock is already held
2441 mutex_lock(&mapping->i_mmap_mutex);
2442 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2443 /* Do not unmap the current VMA */
2444 if (iter_vma == vma)
2448 * Unmap the page from other VMAs without their own reserves.
2449 * They get marked to be SIGKILLed if they fault in these
2450 * areas. This is because a future no-page fault on this VMA
2451 * could insert a zeroed page instead of the data existing
2452 * from the time of fork. This would look like data corruption
2454 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2455 __unmap_hugepage_range(iter_vma,
2456 address, address + huge_page_size(h),
2459 mutex_unlock(&mapping->i_mmap_mutex);
2465 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2467 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2468 unsigned long address, pte_t *ptep, pte_t pte,
2469 struct page *pagecache_page)
2471 struct hstate *h = hstate_vma(vma);
2472 struct page *old_page, *new_page;
2474 int outside_reserve = 0;
2476 old_page = pte_page(pte);
2479 /* If no-one else is actually using this page, avoid the copy
2480 * and just make the page writable */
2481 avoidcopy = (page_mapcount(old_page) == 1);
2483 if (PageAnon(old_page))
2484 page_move_anon_rmap(old_page, vma, address);
2485 set_huge_ptep_writable(vma, address, ptep);
2490 * If the process that created a MAP_PRIVATE mapping is about to
2491 * perform a COW due to a shared page count, attempt to satisfy
2492 * the allocation without using the existing reserves. The pagecache
2493 * page is used to determine if the reserve at this address was
2494 * consumed or not. If reserves were used, a partial faulted mapping
2495 * at the time of fork() could consume its reserves on COW instead
2496 * of the full address range.
2498 if (!(vma->vm_flags & VM_MAYSHARE) &&
2499 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2500 old_page != pagecache_page)
2501 outside_reserve = 1;
2503 page_cache_get(old_page);
2505 /* Drop page_table_lock as buddy allocator may be called */
2506 spin_unlock(&mm->page_table_lock);
2507 new_page = alloc_huge_page(vma, address, outside_reserve);
2509 if (IS_ERR(new_page)) {
2510 page_cache_release(old_page);
2513 * If a process owning a MAP_PRIVATE mapping fails to COW,
2514 * it is due to references held by a child and an insufficient
2515 * huge page pool. To guarantee the original mappers
2516 * reliability, unmap the page from child processes. The child
2517 * may get SIGKILLed if it later faults.
2519 if (outside_reserve) {
2520 BUG_ON(huge_pte_none(pte));
2521 if (unmap_ref_private(mm, vma, old_page, address)) {
2522 BUG_ON(huge_pte_none(pte));
2523 spin_lock(&mm->page_table_lock);
2524 goto retry_avoidcopy;
2529 /* Caller expects lock to be held */
2530 spin_lock(&mm->page_table_lock);
2531 return -PTR_ERR(new_page);
2535 * When the original hugepage is shared one, it does not have
2536 * anon_vma prepared.
2538 if (unlikely(anon_vma_prepare(vma))) {
2539 page_cache_release(new_page);
2540 page_cache_release(old_page);
2541 /* Caller expects lock to be held */
2542 spin_lock(&mm->page_table_lock);
2543 return VM_FAULT_OOM;
2546 copy_user_huge_page(new_page, old_page, address, vma,
2547 pages_per_huge_page(h));
2548 __SetPageUptodate(new_page);
2551 * Retake the page_table_lock to check for racing updates
2552 * before the page tables are altered
2554 spin_lock(&mm->page_table_lock);
2555 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2556 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2558 mmu_notifier_invalidate_range_start(mm,
2559 address & huge_page_mask(h),
2560 (address & huge_page_mask(h)) + huge_page_size(h));
2561 huge_ptep_clear_flush(vma, address, ptep);
2562 set_huge_pte_at(mm, address, ptep,
2563 make_huge_pte(vma, new_page, 1));
2564 page_remove_rmap(old_page);
2565 hugepage_add_new_anon_rmap(new_page, vma, address);
2566 /* Make the old page be freed below */
2567 new_page = old_page;
2568 mmu_notifier_invalidate_range_end(mm,
2569 address & huge_page_mask(h),
2570 (address & huge_page_mask(h)) + huge_page_size(h));
2572 page_cache_release(new_page);
2573 page_cache_release(old_page);
2577 /* Return the pagecache page at a given address within a VMA */
2578 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2579 struct vm_area_struct *vma, unsigned long address)
2581 struct address_space *mapping;
2584 mapping = vma->vm_file->f_mapping;
2585 idx = vma_hugecache_offset(h, vma, address);
2587 return find_lock_page(mapping, idx);
2591 * Return whether there is a pagecache page to back given address within VMA.
2592 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2594 static bool hugetlbfs_pagecache_present(struct hstate *h,
2595 struct vm_area_struct *vma, unsigned long address)
2597 struct address_space *mapping;
2601 mapping = vma->vm_file->f_mapping;
2602 idx = vma_hugecache_offset(h, vma, address);
2604 page = find_get_page(mapping, idx);
2607 return page != NULL;
2610 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2611 unsigned long address, pte_t *ptep, unsigned int flags)
2613 struct hstate *h = hstate_vma(vma);
2614 int ret = VM_FAULT_SIGBUS;
2618 struct address_space *mapping;
2622 * Currently, we are forced to kill the process in the event the
2623 * original mapper has unmapped pages from the child due to a failed
2624 * COW. Warn that such a situation has occurred as it may not be obvious
2626 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2628 "PID %d killed due to inadequate hugepage pool\n",
2633 mapping = vma->vm_file->f_mapping;
2634 idx = vma_hugecache_offset(h, vma, address);
2637 * Use page lock to guard against racing truncation
2638 * before we get page_table_lock.
2641 page = find_lock_page(mapping, idx);
2643 size = i_size_read(mapping->host) >> huge_page_shift(h);
2646 page = alloc_huge_page(vma, address, 0);
2648 ret = -PTR_ERR(page);
2651 clear_huge_page(page, address, pages_per_huge_page(h));
2652 __SetPageUptodate(page);
2654 if (vma->vm_flags & VM_MAYSHARE) {
2656 struct inode *inode = mapping->host;
2658 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2666 spin_lock(&inode->i_lock);
2667 inode->i_blocks += blocks_per_huge_page(h);
2668 spin_unlock(&inode->i_lock);
2669 page_dup_rmap(page);
2672 if (unlikely(anon_vma_prepare(vma))) {
2674 goto backout_unlocked;
2676 hugepage_add_new_anon_rmap(page, vma, address);
2680 * If memory error occurs between mmap() and fault, some process
2681 * don't have hwpoisoned swap entry for errored virtual address.
2682 * So we need to block hugepage fault by PG_hwpoison bit check.
2684 if (unlikely(PageHWPoison(page))) {
2685 ret = VM_FAULT_HWPOISON |
2686 VM_FAULT_SET_HINDEX(h - hstates);
2687 goto backout_unlocked;
2689 page_dup_rmap(page);
2693 * If we are going to COW a private mapping later, we examine the
2694 * pending reservations for this page now. This will ensure that
2695 * any allocations necessary to record that reservation occur outside
2698 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2699 if (vma_needs_reservation(h, vma, address) < 0) {
2701 goto backout_unlocked;
2704 spin_lock(&mm->page_table_lock);
2705 size = i_size_read(mapping->host) >> huge_page_shift(h);
2710 if (!huge_pte_none(huge_ptep_get(ptep)))
2713 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2714 && (vma->vm_flags & VM_SHARED)));
2715 set_huge_pte_at(mm, address, ptep, new_pte);
2717 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2718 /* Optimization, do the COW without a second fault */
2719 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2722 spin_unlock(&mm->page_table_lock);
2728 spin_unlock(&mm->page_table_lock);
2735 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2736 unsigned long address, unsigned int flags)
2741 struct page *page = NULL;
2742 struct page *pagecache_page = NULL;
2743 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2744 struct hstate *h = hstate_vma(vma);
2746 ptep = huge_pte_offset(mm, address);
2748 entry = huge_ptep_get(ptep);
2749 if (unlikely(is_hugetlb_entry_migration(entry))) {
2750 migration_entry_wait(mm, (pmd_t *)ptep, address);
2752 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2753 return VM_FAULT_HWPOISON_LARGE |
2754 VM_FAULT_SET_HINDEX(h - hstates);
2757 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2759 return VM_FAULT_OOM;
2762 * Serialize hugepage allocation and instantiation, so that we don't
2763 * get spurious allocation failures if two CPUs race to instantiate
2764 * the same page in the page cache.
2766 mutex_lock(&hugetlb_instantiation_mutex);
2767 entry = huge_ptep_get(ptep);
2768 if (huge_pte_none(entry)) {
2769 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2776 * If we are going to COW the mapping later, we examine the pending
2777 * reservations for this page now. This will ensure that any
2778 * allocations necessary to record that reservation occur outside the
2779 * spinlock. For private mappings, we also lookup the pagecache
2780 * page now as it is used to determine if a reservation has been
2783 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2784 if (vma_needs_reservation(h, vma, address) < 0) {
2789 if (!(vma->vm_flags & VM_MAYSHARE))
2790 pagecache_page = hugetlbfs_pagecache_page(h,
2795 * hugetlb_cow() requires page locks of pte_page(entry) and
2796 * pagecache_page, so here we need take the former one
2797 * when page != pagecache_page or !pagecache_page.
2798 * Note that locking order is always pagecache_page -> page,
2799 * so no worry about deadlock.
2801 page = pte_page(entry);
2803 if (page != pagecache_page)
2806 spin_lock(&mm->page_table_lock);
2807 /* Check for a racing update before calling hugetlb_cow */
2808 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2809 goto out_page_table_lock;
2812 if (flags & FAULT_FLAG_WRITE) {
2813 if (!pte_write(entry)) {
2814 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2816 goto out_page_table_lock;
2818 entry = pte_mkdirty(entry);
2820 entry = pte_mkyoung(entry);
2821 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2822 flags & FAULT_FLAG_WRITE))
2823 update_mmu_cache(vma, address, ptep);
2825 out_page_table_lock:
2826 spin_unlock(&mm->page_table_lock);
2828 if (pagecache_page) {
2829 unlock_page(pagecache_page);
2830 put_page(pagecache_page);
2832 if (page != pagecache_page)
2837 mutex_unlock(&hugetlb_instantiation_mutex);
2842 /* Can be overriden by architectures */
2843 __attribute__((weak)) struct page *
2844 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2845 pud_t *pud, int write)
2851 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2852 struct page **pages, struct vm_area_struct **vmas,
2853 unsigned long *position, int *length, int i,
2856 unsigned long pfn_offset;
2857 unsigned long vaddr = *position;
2858 int remainder = *length;
2859 struct hstate *h = hstate_vma(vma);
2861 spin_lock(&mm->page_table_lock);
2862 while (vaddr < vma->vm_end && remainder) {
2868 * Some archs (sparc64, sh*) have multiple pte_ts to
2869 * each hugepage. We have to make sure we get the
2870 * first, for the page indexing below to work.
2872 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2873 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2876 * When coredumping, it suits get_dump_page if we just return
2877 * an error where there's an empty slot with no huge pagecache
2878 * to back it. This way, we avoid allocating a hugepage, and
2879 * the sparse dumpfile avoids allocating disk blocks, but its
2880 * huge holes still show up with zeroes where they need to be.
2882 if (absent && (flags & FOLL_DUMP) &&
2883 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2889 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2892 spin_unlock(&mm->page_table_lock);
2893 ret = hugetlb_fault(mm, vma, vaddr,
2894 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2895 spin_lock(&mm->page_table_lock);
2896 if (!(ret & VM_FAULT_ERROR))
2903 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2904 page = pte_page(huge_ptep_get(pte));
2907 pages[i] = mem_map_offset(page, pfn_offset);
2918 if (vaddr < vma->vm_end && remainder &&
2919 pfn_offset < pages_per_huge_page(h)) {
2921 * We use pfn_offset to avoid touching the pageframes
2922 * of this compound page.
2927 spin_unlock(&mm->page_table_lock);
2928 *length = remainder;
2931 return i ? i : -EFAULT;
2934 void hugetlb_change_protection(struct vm_area_struct *vma,
2935 unsigned long address, unsigned long end, pgprot_t newprot)
2937 struct mm_struct *mm = vma->vm_mm;
2938 unsigned long start = address;
2941 struct hstate *h = hstate_vma(vma);
2943 BUG_ON(address >= end);
2944 flush_cache_range(vma, address, end);
2946 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2947 spin_lock(&mm->page_table_lock);
2948 for (; address < end; address += huge_page_size(h)) {
2949 ptep = huge_pte_offset(mm, address);
2952 if (huge_pmd_unshare(mm, &address, ptep))
2954 if (!huge_pte_none(huge_ptep_get(ptep))) {
2955 pte = huge_ptep_get_and_clear(mm, address, ptep);
2956 pte = pte_mkhuge(pte_modify(pte, newprot));
2957 set_huge_pte_at(mm, address, ptep, pte);
2960 spin_unlock(&mm->page_table_lock);
2962 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2963 * may have cleared our pud entry and done put_page on the page table:
2964 * once we release i_mmap_mutex, another task can do the final put_page
2965 * and that page table be reused and filled with junk.
2967 flush_tlb_range(vma, start, end);
2968 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2971 int hugetlb_reserve_pages(struct inode *inode,
2973 struct vm_area_struct *vma,
2974 vm_flags_t vm_flags)
2977 struct hstate *h = hstate_inode(inode);
2978 struct hugepage_subpool *spool = subpool_inode(inode);
2981 * Only apply hugepage reservation if asked. At fault time, an
2982 * attempt will be made for VM_NORESERVE to allocate a page
2983 * without using reserves
2985 if (vm_flags & VM_NORESERVE)
2989 * Shared mappings base their reservation on the number of pages that
2990 * are already allocated on behalf of the file. Private mappings need
2991 * to reserve the full area even if read-only as mprotect() may be
2992 * called to make the mapping read-write. Assume !vma is a shm mapping
2994 if (!vma || vma->vm_flags & VM_MAYSHARE)
2995 chg = region_chg(&inode->i_mapping->private_list, from, to);
2997 struct resv_map *resv_map = resv_map_alloc();
3003 set_vma_resv_map(vma, resv_map);
3004 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3012 /* There must be enough pages in the subpool for the mapping */
3013 if (hugepage_subpool_get_pages(spool, chg)) {
3019 * Check enough hugepages are available for the reservation.
3020 * Hand the pages back to the subpool if there are not
3022 ret = hugetlb_acct_memory(h, chg);
3024 hugepage_subpool_put_pages(spool, chg);
3029 * Account for the reservations made. Shared mappings record regions
3030 * that have reservations as they are shared by multiple VMAs.
3031 * When the last VMA disappears, the region map says how much
3032 * the reservation was and the page cache tells how much of
3033 * the reservation was consumed. Private mappings are per-VMA and
3034 * only the consumed reservations are tracked. When the VMA
3035 * disappears, the original reservation is the VMA size and the
3036 * consumed reservations are stored in the map. Hence, nothing
3037 * else has to be done for private mappings here
3039 if (!vma || vma->vm_flags & VM_MAYSHARE)
3040 region_add(&inode->i_mapping->private_list, from, to);
3048 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3050 struct hstate *h = hstate_inode(inode);
3051 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3052 struct hugepage_subpool *spool = subpool_inode(inode);
3054 spin_lock(&inode->i_lock);
3055 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3056 spin_unlock(&inode->i_lock);
3058 hugepage_subpool_put_pages(spool, (chg - freed));
3059 hugetlb_acct_memory(h, -(chg - freed));
3062 #ifdef CONFIG_MEMORY_FAILURE
3064 /* Should be called in hugetlb_lock */
3065 static int is_hugepage_on_freelist(struct page *hpage)
3069 struct hstate *h = page_hstate(hpage);
3070 int nid = page_to_nid(hpage);
3072 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3079 * This function is called from memory failure code.
3080 * Assume the caller holds page lock of the head page.
3082 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3084 struct hstate *h = page_hstate(hpage);
3085 int nid = page_to_nid(hpage);
3088 spin_lock(&hugetlb_lock);
3089 if (is_hugepage_on_freelist(hpage)) {
3090 list_del(&hpage->lru);
3091 set_page_refcounted(hpage);
3092 h->free_huge_pages--;
3093 h->free_huge_pages_node[nid]--;
3096 spin_unlock(&hugetlb_lock);