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 arch_clear_hugepage_flags(page);
637 enqueue_huge_page(h, page);
639 spin_unlock(&hugetlb_lock);
640 hugepage_subpool_put_pages(spool, 1);
643 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
645 set_compound_page_dtor(page, free_huge_page);
646 spin_lock(&hugetlb_lock);
648 h->nr_huge_pages_node[nid]++;
649 spin_unlock(&hugetlb_lock);
650 put_page(page); /* free it into the hugepage allocator */
653 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
656 int nr_pages = 1 << order;
657 struct page *p = page + 1;
659 /* we rely on prep_new_huge_page to set the destructor */
660 set_compound_order(page, order);
662 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
664 set_page_count(p, 0);
665 p->first_page = page;
669 int PageHuge(struct page *page)
671 compound_page_dtor *dtor;
673 if (!PageCompound(page))
676 page = compound_head(page);
677 dtor = get_compound_page_dtor(page);
679 return dtor == free_huge_page;
681 EXPORT_SYMBOL_GPL(PageHuge);
683 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
687 if (h->order >= MAX_ORDER)
690 page = alloc_pages_exact_node(nid,
691 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
692 __GFP_REPEAT|__GFP_NOWARN,
695 if (arch_prepare_hugepage(page)) {
696 __free_pages(page, huge_page_order(h));
699 prep_new_huge_page(h, page, nid);
706 * common helper functions for hstate_next_node_to_{alloc|free}.
707 * We may have allocated or freed a huge page based on a different
708 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
709 * be outside of *nodes_allowed. Ensure that we use an allowed
710 * node for alloc or free.
712 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
714 nid = next_node(nid, *nodes_allowed);
715 if (nid == MAX_NUMNODES)
716 nid = first_node(*nodes_allowed);
717 VM_BUG_ON(nid >= MAX_NUMNODES);
722 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
724 if (!node_isset(nid, *nodes_allowed))
725 nid = next_node_allowed(nid, nodes_allowed);
730 * returns the previously saved node ["this node"] from which to
731 * allocate a persistent huge page for the pool and advance the
732 * next node from which to allocate, handling wrap at end of node
735 static int hstate_next_node_to_alloc(struct hstate *h,
736 nodemask_t *nodes_allowed)
740 VM_BUG_ON(!nodes_allowed);
742 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
743 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
748 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
755 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
756 next_nid = start_nid;
759 page = alloc_fresh_huge_page_node(h, next_nid);
764 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
765 } while (next_nid != start_nid);
768 count_vm_event(HTLB_BUDDY_PGALLOC);
770 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
776 * helper for free_pool_huge_page() - return the previously saved
777 * node ["this node"] from which to free a huge page. Advance the
778 * next node id whether or not we find a free huge page to free so
779 * that the next attempt to free addresses the next node.
781 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
785 VM_BUG_ON(!nodes_allowed);
787 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
788 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
794 * Free huge page from pool from next node to free.
795 * Attempt to keep persistent huge pages more or less
796 * balanced over allowed nodes.
797 * Called with hugetlb_lock locked.
799 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
806 start_nid = hstate_next_node_to_free(h, nodes_allowed);
807 next_nid = start_nid;
811 * If we're returning unused surplus pages, only examine
812 * nodes with surplus pages.
814 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
815 !list_empty(&h->hugepage_freelists[next_nid])) {
817 list_entry(h->hugepage_freelists[next_nid].next,
819 list_del(&page->lru);
820 h->free_huge_pages--;
821 h->free_huge_pages_node[next_nid]--;
823 h->surplus_huge_pages--;
824 h->surplus_huge_pages_node[next_nid]--;
826 update_and_free_page(h, page);
830 next_nid = hstate_next_node_to_free(h, nodes_allowed);
831 } while (next_nid != start_nid);
836 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
841 if (h->order >= MAX_ORDER)
845 * Assume we will successfully allocate the surplus page to
846 * prevent racing processes from causing the surplus to exceed
849 * This however introduces a different race, where a process B
850 * tries to grow the static hugepage pool while alloc_pages() is
851 * called by process A. B will only examine the per-node
852 * counters in determining if surplus huge pages can be
853 * converted to normal huge pages in adjust_pool_surplus(). A
854 * won't be able to increment the per-node counter, until the
855 * lock is dropped by B, but B doesn't drop hugetlb_lock until
856 * no more huge pages can be converted from surplus to normal
857 * state (and doesn't try to convert again). Thus, we have a
858 * case where a surplus huge page exists, the pool is grown, and
859 * the surplus huge page still exists after, even though it
860 * should just have been converted to a normal huge page. This
861 * does not leak memory, though, as the hugepage will be freed
862 * once it is out of use. It also does not allow the counters to
863 * go out of whack in adjust_pool_surplus() as we don't modify
864 * the node values until we've gotten the hugepage and only the
865 * per-node value is checked there.
867 spin_lock(&hugetlb_lock);
868 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
869 spin_unlock(&hugetlb_lock);
873 h->surplus_huge_pages++;
875 spin_unlock(&hugetlb_lock);
877 if (nid == NUMA_NO_NODE)
878 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
879 __GFP_REPEAT|__GFP_NOWARN,
882 page = alloc_pages_exact_node(nid,
883 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
884 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
886 if (page && arch_prepare_hugepage(page)) {
887 __free_pages(page, huge_page_order(h));
891 spin_lock(&hugetlb_lock);
893 r_nid = page_to_nid(page);
894 set_compound_page_dtor(page, free_huge_page);
896 * We incremented the global counters already
898 h->nr_huge_pages_node[r_nid]++;
899 h->surplus_huge_pages_node[r_nid]++;
900 __count_vm_event(HTLB_BUDDY_PGALLOC);
903 h->surplus_huge_pages--;
904 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
906 spin_unlock(&hugetlb_lock);
912 * This allocation function is useful in the context where vma is irrelevant.
913 * E.g. soft-offlining uses this function because it only cares physical
914 * address of error page.
916 struct page *alloc_huge_page_node(struct hstate *h, int nid)
920 spin_lock(&hugetlb_lock);
921 page = dequeue_huge_page_node(h, nid);
922 spin_unlock(&hugetlb_lock);
925 page = alloc_buddy_huge_page(h, nid);
931 * Increase the hugetlb pool such that it can accommodate a reservation
934 static int gather_surplus_pages(struct hstate *h, int delta)
936 struct list_head surplus_list;
937 struct page *page, *tmp;
939 int needed, allocated;
941 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
943 h->resv_huge_pages += delta;
948 INIT_LIST_HEAD(&surplus_list);
952 spin_unlock(&hugetlb_lock);
953 for (i = 0; i < needed; i++) {
954 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
957 * We were not able to allocate enough pages to
958 * satisfy the entire reservation so we free what
959 * we've allocated so far.
963 list_add(&page->lru, &surplus_list);
968 * After retaking hugetlb_lock, we need to recalculate 'needed'
969 * because either resv_huge_pages or free_huge_pages may have changed.
971 spin_lock(&hugetlb_lock);
972 needed = (h->resv_huge_pages + delta) -
973 (h->free_huge_pages + allocated);
978 * The surplus_list now contains _at_least_ the number of extra pages
979 * needed to accommodate the reservation. Add the appropriate number
980 * of pages to the hugetlb pool and free the extras back to the buddy
981 * allocator. Commit the entire reservation here to prevent another
982 * process from stealing the pages as they are added to the pool but
983 * before they are reserved.
986 h->resv_huge_pages += delta;
989 /* Free the needed pages to the hugetlb pool */
990 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
993 list_del(&page->lru);
995 * This page is now managed by the hugetlb allocator and has
996 * no users -- drop the buddy allocator's reference.
998 put_page_testzero(page);
999 VM_BUG_ON(page_count(page));
1000 enqueue_huge_page(h, page);
1002 spin_unlock(&hugetlb_lock);
1004 /* Free unnecessary surplus pages to the buddy allocator */
1006 if (!list_empty(&surplus_list)) {
1007 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1008 list_del(&page->lru);
1012 spin_lock(&hugetlb_lock);
1018 * When releasing a hugetlb pool reservation, any surplus pages that were
1019 * allocated to satisfy the reservation must be explicitly freed if they were
1021 * Called with hugetlb_lock held.
1023 static void return_unused_surplus_pages(struct hstate *h,
1024 unsigned long unused_resv_pages)
1026 unsigned long nr_pages;
1028 /* Uncommit the reservation */
1029 h->resv_huge_pages -= unused_resv_pages;
1031 /* Cannot return gigantic pages currently */
1032 if (h->order >= MAX_ORDER)
1035 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1038 * We want to release as many surplus pages as possible, spread
1039 * evenly across all nodes with memory. Iterate across these nodes
1040 * until we can no longer free unreserved surplus pages. This occurs
1041 * when the nodes with surplus pages have no free pages.
1042 * free_pool_huge_page() will balance the the freed pages across the
1043 * on-line nodes with memory and will handle the hstate accounting.
1045 while (nr_pages--) {
1046 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1052 * Determine if the huge page at addr within the vma has an associated
1053 * reservation. Where it does not we will need to logically increase
1054 * reservation and actually increase subpool usage before an allocation
1055 * can occur. Where any new reservation would be required the
1056 * reservation change is prepared, but not committed. Once the page
1057 * has been allocated from the subpool and instantiated the change should
1058 * be committed via vma_commit_reservation. No action is required on
1061 static long vma_needs_reservation(struct hstate *h,
1062 struct vm_area_struct *vma, unsigned long addr)
1064 struct address_space *mapping = vma->vm_file->f_mapping;
1065 struct inode *inode = mapping->host;
1067 if (vma->vm_flags & VM_MAYSHARE) {
1068 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1069 return region_chg(&inode->i_mapping->private_list,
1072 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1077 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1078 struct resv_map *reservations = vma_resv_map(vma);
1080 err = region_chg(&reservations->regions, idx, idx + 1);
1086 static void vma_commit_reservation(struct hstate *h,
1087 struct vm_area_struct *vma, unsigned long addr)
1089 struct address_space *mapping = vma->vm_file->f_mapping;
1090 struct inode *inode = mapping->host;
1092 if (vma->vm_flags & VM_MAYSHARE) {
1093 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1094 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1096 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1097 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1098 struct resv_map *reservations = vma_resv_map(vma);
1100 /* Mark this page used in the map. */
1101 region_add(&reservations->regions, idx, idx + 1);
1105 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1106 unsigned long addr, int avoid_reserve)
1108 struct hugepage_subpool *spool = subpool_vma(vma);
1109 struct hstate *h = hstate_vma(vma);
1114 * Processes that did not create the mapping will have no
1115 * reserves and will not have accounted against subpool
1116 * limit. Check that the subpool limit can be made before
1117 * satisfying the allocation MAP_NORESERVE mappings may also
1118 * need pages and subpool limit allocated allocated if no reserve
1121 chg = vma_needs_reservation(h, vma, addr);
1123 return ERR_PTR(-VM_FAULT_OOM);
1125 if (hugepage_subpool_get_pages(spool, chg))
1126 return ERR_PTR(-VM_FAULT_SIGBUS);
1128 spin_lock(&hugetlb_lock);
1129 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1130 spin_unlock(&hugetlb_lock);
1133 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1135 hugepage_subpool_put_pages(spool, chg);
1136 return ERR_PTR(-VM_FAULT_SIGBUS);
1140 set_page_private(page, (unsigned long)spool);
1142 vma_commit_reservation(h, vma, addr);
1147 int __weak alloc_bootmem_huge_page(struct hstate *h)
1149 struct huge_bootmem_page *m;
1150 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1155 addr = __alloc_bootmem_node_nopanic(
1156 NODE_DATA(hstate_next_node_to_alloc(h,
1157 &node_states[N_HIGH_MEMORY])),
1158 huge_page_size(h), huge_page_size(h), 0);
1162 * Use the beginning of the huge page to store the
1163 * huge_bootmem_page struct (until gather_bootmem
1164 * puts them into the mem_map).
1174 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1175 /* Put them into a private list first because mem_map is not up yet */
1176 list_add(&m->list, &huge_boot_pages);
1181 static void prep_compound_huge_page(struct page *page, int order)
1183 if (unlikely(order > (MAX_ORDER - 1)))
1184 prep_compound_gigantic_page(page, order);
1186 prep_compound_page(page, order);
1189 /* Put bootmem huge pages into the standard lists after mem_map is up */
1190 static void __init gather_bootmem_prealloc(void)
1192 struct huge_bootmem_page *m;
1194 list_for_each_entry(m, &huge_boot_pages, list) {
1195 struct hstate *h = m->hstate;
1198 #ifdef CONFIG_HIGHMEM
1199 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1200 free_bootmem_late((unsigned long)m,
1201 sizeof(struct huge_bootmem_page));
1203 page = virt_to_page(m);
1205 __ClearPageReserved(page);
1206 WARN_ON(page_count(page) != 1);
1207 prep_compound_huge_page(page, h->order);
1208 prep_new_huge_page(h, page, page_to_nid(page));
1210 * If we had gigantic hugepages allocated at boot time, we need
1211 * to restore the 'stolen' pages to totalram_pages in order to
1212 * fix confusing memory reports from free(1) and another
1213 * side-effects, like CommitLimit going negative.
1215 if (h->order > (MAX_ORDER - 1))
1216 totalram_pages += 1 << h->order;
1220 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1224 for (i = 0; i < h->max_huge_pages; ++i) {
1225 if (h->order >= MAX_ORDER) {
1226 if (!alloc_bootmem_huge_page(h))
1228 } else if (!alloc_fresh_huge_page(h,
1229 &node_states[N_HIGH_MEMORY]))
1232 h->max_huge_pages = i;
1235 static void __init hugetlb_init_hstates(void)
1239 for_each_hstate(h) {
1240 /* oversize hugepages were init'ed in early boot */
1241 if (h->order < MAX_ORDER)
1242 hugetlb_hstate_alloc_pages(h);
1246 static char * __init memfmt(char *buf, unsigned long n)
1248 if (n >= (1UL << 30))
1249 sprintf(buf, "%lu GB", n >> 30);
1250 else if (n >= (1UL << 20))
1251 sprintf(buf, "%lu MB", n >> 20);
1253 sprintf(buf, "%lu KB", n >> 10);
1257 static void __init report_hugepages(void)
1261 for_each_hstate(h) {
1263 printk(KERN_INFO "HugeTLB registered %s page size, "
1264 "pre-allocated %ld pages\n",
1265 memfmt(buf, huge_page_size(h)),
1266 h->free_huge_pages);
1270 #ifdef CONFIG_HIGHMEM
1271 static void try_to_free_low(struct hstate *h, unsigned long count,
1272 nodemask_t *nodes_allowed)
1276 if (h->order >= MAX_ORDER)
1279 for_each_node_mask(i, *nodes_allowed) {
1280 struct page *page, *next;
1281 struct list_head *freel = &h->hugepage_freelists[i];
1282 list_for_each_entry_safe(page, next, freel, lru) {
1283 if (count >= h->nr_huge_pages)
1285 if (PageHighMem(page))
1287 list_del(&page->lru);
1288 update_and_free_page(h, page);
1289 h->free_huge_pages--;
1290 h->free_huge_pages_node[page_to_nid(page)]--;
1295 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1296 nodemask_t *nodes_allowed)
1302 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1303 * balanced by operating on them in a round-robin fashion.
1304 * Returns 1 if an adjustment was made.
1306 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1309 int start_nid, next_nid;
1312 VM_BUG_ON(delta != -1 && delta != 1);
1315 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1317 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1318 next_nid = start_nid;
1324 * To shrink on this node, there must be a surplus page
1326 if (!h->surplus_huge_pages_node[nid]) {
1327 next_nid = hstate_next_node_to_alloc(h,
1334 * Surplus cannot exceed the total number of pages
1336 if (h->surplus_huge_pages_node[nid] >=
1337 h->nr_huge_pages_node[nid]) {
1338 next_nid = hstate_next_node_to_free(h,
1344 h->surplus_huge_pages += delta;
1345 h->surplus_huge_pages_node[nid] += delta;
1348 } while (next_nid != start_nid);
1353 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1354 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1355 nodemask_t *nodes_allowed)
1357 unsigned long min_count, ret;
1359 if (h->order >= MAX_ORDER)
1360 return h->max_huge_pages;
1363 * Increase the pool size
1364 * First take pages out of surplus state. Then make up the
1365 * remaining difference by allocating fresh huge pages.
1367 * We might race with alloc_buddy_huge_page() here and be unable
1368 * to convert a surplus huge page to a normal huge page. That is
1369 * not critical, though, it just means the overall size of the
1370 * pool might be one hugepage larger than it needs to be, but
1371 * within all the constraints specified by the sysctls.
1373 spin_lock(&hugetlb_lock);
1374 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1375 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1379 while (count > persistent_huge_pages(h)) {
1381 * If this allocation races such that we no longer need the
1382 * page, free_huge_page will handle it by freeing the page
1383 * and reducing the surplus.
1385 spin_unlock(&hugetlb_lock);
1386 ret = alloc_fresh_huge_page(h, nodes_allowed);
1387 spin_lock(&hugetlb_lock);
1391 /* Bail for signals. Probably ctrl-c from user */
1392 if (signal_pending(current))
1397 * Decrease the pool size
1398 * First return free pages to the buddy allocator (being careful
1399 * to keep enough around to satisfy reservations). Then place
1400 * pages into surplus state as needed so the pool will shrink
1401 * to the desired size as pages become free.
1403 * By placing pages into the surplus state independent of the
1404 * overcommit value, we are allowing the surplus pool size to
1405 * exceed overcommit. There are few sane options here. Since
1406 * alloc_buddy_huge_page() is checking the global counter,
1407 * though, we'll note that we're not allowed to exceed surplus
1408 * and won't grow the pool anywhere else. Not until one of the
1409 * sysctls are changed, or the surplus pages go out of use.
1411 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1412 min_count = max(count, min_count);
1413 try_to_free_low(h, min_count, nodes_allowed);
1414 while (min_count < persistent_huge_pages(h)) {
1415 if (!free_pool_huge_page(h, nodes_allowed, 0))
1418 while (count < persistent_huge_pages(h)) {
1419 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1423 ret = persistent_huge_pages(h);
1424 spin_unlock(&hugetlb_lock);
1428 #define HSTATE_ATTR_RO(_name) \
1429 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1431 #define HSTATE_ATTR(_name) \
1432 static struct kobj_attribute _name##_attr = \
1433 __ATTR(_name, 0644, _name##_show, _name##_store)
1435 static struct kobject *hugepages_kobj;
1436 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1438 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1440 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1444 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1445 if (hstate_kobjs[i] == kobj) {
1447 *nidp = NUMA_NO_NODE;
1451 return kobj_to_node_hstate(kobj, nidp);
1454 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1455 struct kobj_attribute *attr, char *buf)
1458 unsigned long nr_huge_pages;
1461 h = kobj_to_hstate(kobj, &nid);
1462 if (nid == NUMA_NO_NODE)
1463 nr_huge_pages = h->nr_huge_pages;
1465 nr_huge_pages = h->nr_huge_pages_node[nid];
1467 return sprintf(buf, "%lu\n", nr_huge_pages);
1470 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1471 struct kobject *kobj, struct kobj_attribute *attr,
1472 const char *buf, size_t len)
1476 unsigned long count;
1478 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1480 err = strict_strtoul(buf, 10, &count);
1484 h = kobj_to_hstate(kobj, &nid);
1485 if (h->order >= MAX_ORDER) {
1490 if (nid == NUMA_NO_NODE) {
1492 * global hstate attribute
1494 if (!(obey_mempolicy &&
1495 init_nodemask_of_mempolicy(nodes_allowed))) {
1496 NODEMASK_FREE(nodes_allowed);
1497 nodes_allowed = &node_states[N_HIGH_MEMORY];
1499 } else if (nodes_allowed) {
1501 * per node hstate attribute: adjust count to global,
1502 * but restrict alloc/free to the specified node.
1504 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1505 init_nodemask_of_node(nodes_allowed, nid);
1507 nodes_allowed = &node_states[N_HIGH_MEMORY];
1509 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1511 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1512 NODEMASK_FREE(nodes_allowed);
1516 NODEMASK_FREE(nodes_allowed);
1520 static ssize_t nr_hugepages_show(struct kobject *kobj,
1521 struct kobj_attribute *attr, char *buf)
1523 return nr_hugepages_show_common(kobj, attr, buf);
1526 static ssize_t nr_hugepages_store(struct kobject *kobj,
1527 struct kobj_attribute *attr, const char *buf, size_t len)
1529 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1531 HSTATE_ATTR(nr_hugepages);
1536 * hstate attribute for optionally mempolicy-based constraint on persistent
1537 * huge page alloc/free.
1539 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1540 struct kobj_attribute *attr, char *buf)
1542 return nr_hugepages_show_common(kobj, attr, buf);
1545 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1546 struct kobj_attribute *attr, const char *buf, size_t len)
1548 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1550 HSTATE_ATTR(nr_hugepages_mempolicy);
1554 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1555 struct kobj_attribute *attr, char *buf)
1557 struct hstate *h = kobj_to_hstate(kobj, NULL);
1558 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1561 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1562 struct kobj_attribute *attr, const char *buf, size_t count)
1565 unsigned long input;
1566 struct hstate *h = kobj_to_hstate(kobj, NULL);
1568 if (h->order >= MAX_ORDER)
1571 err = strict_strtoul(buf, 10, &input);
1575 spin_lock(&hugetlb_lock);
1576 h->nr_overcommit_huge_pages = input;
1577 spin_unlock(&hugetlb_lock);
1581 HSTATE_ATTR(nr_overcommit_hugepages);
1583 static ssize_t free_hugepages_show(struct kobject *kobj,
1584 struct kobj_attribute *attr, char *buf)
1587 unsigned long free_huge_pages;
1590 h = kobj_to_hstate(kobj, &nid);
1591 if (nid == NUMA_NO_NODE)
1592 free_huge_pages = h->free_huge_pages;
1594 free_huge_pages = h->free_huge_pages_node[nid];
1596 return sprintf(buf, "%lu\n", free_huge_pages);
1598 HSTATE_ATTR_RO(free_hugepages);
1600 static ssize_t resv_hugepages_show(struct kobject *kobj,
1601 struct kobj_attribute *attr, char *buf)
1603 struct hstate *h = kobj_to_hstate(kobj, NULL);
1604 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1606 HSTATE_ATTR_RO(resv_hugepages);
1608 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1609 struct kobj_attribute *attr, char *buf)
1612 unsigned long surplus_huge_pages;
1615 h = kobj_to_hstate(kobj, &nid);
1616 if (nid == NUMA_NO_NODE)
1617 surplus_huge_pages = h->surplus_huge_pages;
1619 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1621 return sprintf(buf, "%lu\n", surplus_huge_pages);
1623 HSTATE_ATTR_RO(surplus_hugepages);
1625 static struct attribute *hstate_attrs[] = {
1626 &nr_hugepages_attr.attr,
1627 &nr_overcommit_hugepages_attr.attr,
1628 &free_hugepages_attr.attr,
1629 &resv_hugepages_attr.attr,
1630 &surplus_hugepages_attr.attr,
1632 &nr_hugepages_mempolicy_attr.attr,
1637 static struct attribute_group hstate_attr_group = {
1638 .attrs = hstate_attrs,
1641 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1642 struct kobject **hstate_kobjs,
1643 struct attribute_group *hstate_attr_group)
1646 int hi = h - hstates;
1648 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1649 if (!hstate_kobjs[hi])
1652 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1654 kobject_put(hstate_kobjs[hi]);
1659 static void __init hugetlb_sysfs_init(void)
1664 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1665 if (!hugepages_kobj)
1668 for_each_hstate(h) {
1669 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1670 hstate_kobjs, &hstate_attr_group);
1672 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1680 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1681 * with node sysdevs in node_devices[] using a parallel array. The array
1682 * index of a node sysdev or _hstate == node id.
1683 * This is here to avoid any static dependency of the node sysdev driver, in
1684 * the base kernel, on the hugetlb module.
1686 struct node_hstate {
1687 struct kobject *hugepages_kobj;
1688 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1690 struct node_hstate node_hstates[MAX_NUMNODES];
1693 * A subset of global hstate attributes for node sysdevs
1695 static struct attribute *per_node_hstate_attrs[] = {
1696 &nr_hugepages_attr.attr,
1697 &free_hugepages_attr.attr,
1698 &surplus_hugepages_attr.attr,
1702 static struct attribute_group per_node_hstate_attr_group = {
1703 .attrs = per_node_hstate_attrs,
1707 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1708 * Returns node id via non-NULL nidp.
1710 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1714 for (nid = 0; nid < nr_node_ids; nid++) {
1715 struct node_hstate *nhs = &node_hstates[nid];
1717 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1718 if (nhs->hstate_kobjs[i] == kobj) {
1730 * Unregister hstate attributes from a single node sysdev.
1731 * No-op if no hstate attributes attached.
1733 void hugetlb_unregister_node(struct node *node)
1736 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1738 if (!nhs->hugepages_kobj)
1739 return; /* no hstate attributes */
1742 if (nhs->hstate_kobjs[h - hstates]) {
1743 kobject_put(nhs->hstate_kobjs[h - hstates]);
1744 nhs->hstate_kobjs[h - hstates] = NULL;
1747 kobject_put(nhs->hugepages_kobj);
1748 nhs->hugepages_kobj = NULL;
1752 * hugetlb module exit: unregister hstate attributes from node sysdevs
1755 static void hugetlb_unregister_all_nodes(void)
1760 * disable node sysdev registrations.
1762 register_hugetlbfs_with_node(NULL, NULL);
1765 * remove hstate attributes from any nodes that have them.
1767 for (nid = 0; nid < nr_node_ids; nid++)
1768 hugetlb_unregister_node(&node_devices[nid]);
1772 * Register hstate attributes for a single node sysdev.
1773 * No-op if attributes already registered.
1775 void hugetlb_register_node(struct node *node)
1778 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1781 if (nhs->hugepages_kobj)
1782 return; /* already allocated */
1784 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1785 &node->sysdev.kobj);
1786 if (!nhs->hugepages_kobj)
1789 for_each_hstate(h) {
1790 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1792 &per_node_hstate_attr_group);
1794 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1796 h->name, node->sysdev.id);
1797 hugetlb_unregister_node(node);
1804 * hugetlb init time: register hstate attributes for all registered node
1805 * sysdevs of nodes that have memory. All on-line nodes should have
1806 * registered their associated sysdev by this time.
1808 static void hugetlb_register_all_nodes(void)
1812 for_each_node_state(nid, N_HIGH_MEMORY) {
1813 struct node *node = &node_devices[nid];
1814 if (node->sysdev.id == nid)
1815 hugetlb_register_node(node);
1819 * Let the node sysdev driver know we're here so it can
1820 * [un]register hstate attributes on node hotplug.
1822 register_hugetlbfs_with_node(hugetlb_register_node,
1823 hugetlb_unregister_node);
1825 #else /* !CONFIG_NUMA */
1827 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1835 static void hugetlb_unregister_all_nodes(void) { }
1837 static void hugetlb_register_all_nodes(void) { }
1841 static void __exit hugetlb_exit(void)
1845 hugetlb_unregister_all_nodes();
1847 for_each_hstate(h) {
1848 kobject_put(hstate_kobjs[h - hstates]);
1851 kobject_put(hugepages_kobj);
1853 module_exit(hugetlb_exit);
1855 static int __init hugetlb_init(void)
1857 /* Some platform decide whether they support huge pages at boot
1858 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1859 * there is no such support
1861 if (HPAGE_SHIFT == 0)
1864 if (!size_to_hstate(default_hstate_size)) {
1865 default_hstate_size = HPAGE_SIZE;
1866 if (!size_to_hstate(default_hstate_size))
1867 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1869 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1870 if (default_hstate_max_huge_pages)
1871 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1873 hugetlb_init_hstates();
1875 gather_bootmem_prealloc();
1879 hugetlb_sysfs_init();
1881 hugetlb_register_all_nodes();
1885 module_init(hugetlb_init);
1887 /* Should be called on processing a hugepagesz=... option */
1888 void __init hugetlb_add_hstate(unsigned order)
1893 if (size_to_hstate(PAGE_SIZE << order)) {
1894 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1897 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1899 h = &hstates[max_hstate++];
1901 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1902 h->nr_huge_pages = 0;
1903 h->free_huge_pages = 0;
1904 for (i = 0; i < MAX_NUMNODES; ++i)
1905 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1906 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1907 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1908 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1909 huge_page_size(h)/1024);
1914 static int __init hugetlb_nrpages_setup(char *s)
1917 static unsigned long *last_mhp;
1920 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1921 * so this hugepages= parameter goes to the "default hstate".
1924 mhp = &default_hstate_max_huge_pages;
1926 mhp = &parsed_hstate->max_huge_pages;
1928 if (mhp == last_mhp) {
1929 printk(KERN_WARNING "hugepages= specified twice without "
1930 "interleaving hugepagesz=, ignoring\n");
1934 if (sscanf(s, "%lu", mhp) <= 0)
1938 * Global state is always initialized later in hugetlb_init.
1939 * But we need to allocate >= MAX_ORDER hstates here early to still
1940 * use the bootmem allocator.
1942 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1943 hugetlb_hstate_alloc_pages(parsed_hstate);
1949 __setup("hugepages=", hugetlb_nrpages_setup);
1951 static int __init hugetlb_default_setup(char *s)
1953 default_hstate_size = memparse(s, &s);
1956 __setup("default_hugepagesz=", hugetlb_default_setup);
1958 static unsigned int cpuset_mems_nr(unsigned int *array)
1961 unsigned int nr = 0;
1963 for_each_node_mask(node, cpuset_current_mems_allowed)
1969 #ifdef CONFIG_SYSCTL
1970 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1971 struct ctl_table *table, int write,
1972 void __user *buffer, size_t *length, loff_t *ppos)
1974 struct hstate *h = &default_hstate;
1978 tmp = h->max_huge_pages;
1980 if (write && h->order >= MAX_ORDER)
1984 table->maxlen = sizeof(unsigned long);
1985 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1990 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1991 GFP_KERNEL | __GFP_NORETRY);
1992 if (!(obey_mempolicy &&
1993 init_nodemask_of_mempolicy(nodes_allowed))) {
1994 NODEMASK_FREE(nodes_allowed);
1995 nodes_allowed = &node_states[N_HIGH_MEMORY];
1997 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1999 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2000 NODEMASK_FREE(nodes_allowed);
2006 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2007 void __user *buffer, size_t *length, loff_t *ppos)
2010 return hugetlb_sysctl_handler_common(false, table, write,
2011 buffer, length, ppos);
2015 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2016 void __user *buffer, size_t *length, loff_t *ppos)
2018 return hugetlb_sysctl_handler_common(true, table, write,
2019 buffer, length, ppos);
2021 #endif /* CONFIG_NUMA */
2023 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2024 void __user *buffer,
2025 size_t *length, loff_t *ppos)
2027 proc_dointvec(table, write, buffer, length, ppos);
2028 if (hugepages_treat_as_movable)
2029 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2031 htlb_alloc_mask = GFP_HIGHUSER;
2035 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2036 void __user *buffer,
2037 size_t *length, loff_t *ppos)
2039 struct hstate *h = &default_hstate;
2043 tmp = h->nr_overcommit_huge_pages;
2045 if (write && h->order >= MAX_ORDER)
2049 table->maxlen = sizeof(unsigned long);
2050 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2055 spin_lock(&hugetlb_lock);
2056 h->nr_overcommit_huge_pages = tmp;
2057 spin_unlock(&hugetlb_lock);
2063 #endif /* CONFIG_SYSCTL */
2065 void hugetlb_report_meminfo(struct seq_file *m)
2067 struct hstate *h = &default_hstate;
2069 "HugePages_Total: %5lu\n"
2070 "HugePages_Free: %5lu\n"
2071 "HugePages_Rsvd: %5lu\n"
2072 "HugePages_Surp: %5lu\n"
2073 "Hugepagesize: %8lu kB\n",
2077 h->surplus_huge_pages,
2078 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2081 int hugetlb_report_node_meminfo(int nid, char *buf)
2083 struct hstate *h = &default_hstate;
2085 "Node %d HugePages_Total: %5u\n"
2086 "Node %d HugePages_Free: %5u\n"
2087 "Node %d HugePages_Surp: %5u\n",
2088 nid, h->nr_huge_pages_node[nid],
2089 nid, h->free_huge_pages_node[nid],
2090 nid, h->surplus_huge_pages_node[nid]);
2093 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2094 unsigned long hugetlb_total_pages(void)
2096 struct hstate *h = &default_hstate;
2097 return h->nr_huge_pages * pages_per_huge_page(h);
2100 static int hugetlb_acct_memory(struct hstate *h, long delta)
2104 spin_lock(&hugetlb_lock);
2106 * When cpuset is configured, it breaks the strict hugetlb page
2107 * reservation as the accounting is done on a global variable. Such
2108 * reservation is completely rubbish in the presence of cpuset because
2109 * the reservation is not checked against page availability for the
2110 * current cpuset. Application can still potentially OOM'ed by kernel
2111 * with lack of free htlb page in cpuset that the task is in.
2112 * Attempt to enforce strict accounting with cpuset is almost
2113 * impossible (or too ugly) because cpuset is too fluid that
2114 * task or memory node can be dynamically moved between cpusets.
2116 * The change of semantics for shared hugetlb mapping with cpuset is
2117 * undesirable. However, in order to preserve some of the semantics,
2118 * we fall back to check against current free page availability as
2119 * a best attempt and hopefully to minimize the impact of changing
2120 * semantics that cpuset has.
2123 if (gather_surplus_pages(h, delta) < 0)
2126 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2127 return_unused_surplus_pages(h, delta);
2134 return_unused_surplus_pages(h, (unsigned long) -delta);
2137 spin_unlock(&hugetlb_lock);
2141 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2143 struct resv_map *reservations = vma_resv_map(vma);
2146 * This new VMA should share its siblings reservation map if present.
2147 * The VMA will only ever have a valid reservation map pointer where
2148 * it is being copied for another still existing VMA. As that VMA
2149 * has a reference to the reservation map it cannot disappear until
2150 * after this open call completes. It is therefore safe to take a
2151 * new reference here without additional locking.
2154 kref_get(&reservations->refs);
2157 static void resv_map_put(struct vm_area_struct *vma)
2159 struct resv_map *reservations = vma_resv_map(vma);
2163 kref_put(&reservations->refs, resv_map_release);
2166 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2168 struct hstate *h = hstate_vma(vma);
2169 struct resv_map *reservations = vma_resv_map(vma);
2170 struct hugepage_subpool *spool = subpool_vma(vma);
2171 unsigned long reserve;
2172 unsigned long start;
2176 start = vma_hugecache_offset(h, vma, vma->vm_start);
2177 end = vma_hugecache_offset(h, vma, vma->vm_end);
2179 reserve = (end - start) -
2180 region_count(&reservations->regions, start, end);
2185 hugetlb_acct_memory(h, -reserve);
2186 hugepage_subpool_put_pages(spool, reserve);
2192 * We cannot handle pagefaults against hugetlb pages at all. They cause
2193 * handle_mm_fault() to try to instantiate regular-sized pages in the
2194 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2197 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2203 const struct vm_operations_struct hugetlb_vm_ops = {
2204 .fault = hugetlb_vm_op_fault,
2205 .open = hugetlb_vm_op_open,
2206 .close = hugetlb_vm_op_close,
2209 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2216 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2218 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2220 entry = pte_mkyoung(entry);
2221 entry = pte_mkhuge(entry);
2226 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2227 unsigned long address, pte_t *ptep)
2231 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2232 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2233 update_mmu_cache(vma, address, ptep);
2237 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2238 struct vm_area_struct *vma)
2240 pte_t *src_pte, *dst_pte, entry;
2241 struct page *ptepage;
2244 struct hstate *h = hstate_vma(vma);
2245 unsigned long sz = huge_page_size(h);
2247 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2249 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2250 src_pte = huge_pte_offset(src, addr);
2253 dst_pte = huge_pte_alloc(dst, addr, sz);
2257 /* If the pagetables are shared don't copy or take references */
2258 if (dst_pte == src_pte)
2261 spin_lock(&dst->page_table_lock);
2262 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2263 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2265 huge_ptep_set_wrprotect(src, addr, src_pte);
2266 entry = huge_ptep_get(src_pte);
2267 ptepage = pte_page(entry);
2269 page_dup_rmap(ptepage);
2270 set_huge_pte_at(dst, addr, dst_pte, entry);
2272 spin_unlock(&src->page_table_lock);
2273 spin_unlock(&dst->page_table_lock);
2281 static int is_hugetlb_entry_migration(pte_t pte)
2285 if (huge_pte_none(pte) || pte_present(pte))
2287 swp = pte_to_swp_entry(pte);
2288 if (non_swap_entry(swp) && is_migration_entry(swp))
2294 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2298 if (huge_pte_none(pte) || pte_present(pte))
2300 swp = pte_to_swp_entry(pte);
2301 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2307 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2308 unsigned long end, struct page *ref_page)
2310 struct mm_struct *mm = vma->vm_mm;
2311 unsigned long address;
2316 struct hstate *h = hstate_vma(vma);
2317 unsigned long sz = huge_page_size(h);
2320 * A page gathering list, protected by per file i_mmap_mutex. The
2321 * lock is used to avoid list corruption from multiple unmapping
2322 * of the same page since we are using page->lru.
2324 LIST_HEAD(page_list);
2326 WARN_ON(!is_vm_hugetlb_page(vma));
2327 BUG_ON(start & ~huge_page_mask(h));
2328 BUG_ON(end & ~huge_page_mask(h));
2330 mmu_notifier_invalidate_range_start(mm, start, end);
2331 spin_lock(&mm->page_table_lock);
2332 for (address = start; address < end; address += sz) {
2333 ptep = huge_pte_offset(mm, address);
2337 if (huge_pmd_unshare(mm, &address, ptep))
2341 * If a reference page is supplied, it is because a specific
2342 * page is being unmapped, not a range. Ensure the page we
2343 * are about to unmap is the actual page of interest.
2346 pte = huge_ptep_get(ptep);
2347 if (huge_pte_none(pte))
2349 page = pte_page(pte);
2350 if (page != ref_page)
2354 * Mark the VMA as having unmapped its page so that
2355 * future faults in this VMA will fail rather than
2356 * looking like data was lost
2358 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2361 pte = huge_ptep_get_and_clear(mm, address, ptep);
2362 if (huge_pte_none(pte))
2366 * HWPoisoned hugepage is already unmapped and dropped reference
2368 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2371 page = pte_page(pte);
2373 set_page_dirty(page);
2374 list_add(&page->lru, &page_list);
2376 spin_unlock(&mm->page_table_lock);
2377 flush_tlb_range(vma, start, end);
2378 mmu_notifier_invalidate_range_end(mm, start, end);
2379 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2380 page_remove_rmap(page);
2381 list_del(&page->lru);
2386 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2387 unsigned long start, unsigned long end,
2388 struct page *ref_page)
2390 __unmap_hugepage_range(vma, start, end, ref_page);
2393 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2394 * test will fail on a vma being torn down, and not grab a page table
2395 * on its way out. We're lucky that the flag has such an appropriate
2396 * name, and can in fact be safely cleared here. We could clear it
2397 * before the __unmap_hugepage_range above, but all that's necessary
2398 * is to clear it before releasing the i_mmap_mutex. This works
2399 * because in the context this is called, the VMA is about to be
2400 * destroyed and the i_mmap_mutex is held.
2402 vma->vm_flags &= ~VM_MAYSHARE;
2405 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2406 unsigned long end, struct page *ref_page)
2408 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2409 __unmap_hugepage_range(vma, start, end, ref_page);
2410 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2414 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2415 * mappping it owns the reserve page for. The intention is to unmap the page
2416 * from other VMAs and let the children be SIGKILLed if they are faulting the
2419 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2420 struct page *page, unsigned long address)
2422 struct hstate *h = hstate_vma(vma);
2423 struct vm_area_struct *iter_vma;
2424 struct address_space *mapping;
2425 struct prio_tree_iter iter;
2429 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2430 * from page cache lookup which is in HPAGE_SIZE units.
2432 address = address & huge_page_mask(h);
2433 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2435 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2438 * Take the mapping lock for the duration of the table walk. As
2439 * this mapping should be shared between all the VMAs,
2440 * __unmap_hugepage_range() is called as the lock is already held
2442 mutex_lock(&mapping->i_mmap_mutex);
2443 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2444 /* Do not unmap the current VMA */
2445 if (iter_vma == vma)
2449 * Unmap the page from other VMAs without their own reserves.
2450 * They get marked to be SIGKILLed if they fault in these
2451 * areas. This is because a future no-page fault on this VMA
2452 * could insert a zeroed page instead of the data existing
2453 * from the time of fork. This would look like data corruption
2455 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2456 __unmap_hugepage_range(iter_vma,
2457 address, address + huge_page_size(h),
2460 mutex_unlock(&mapping->i_mmap_mutex);
2466 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2468 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2469 unsigned long address, pte_t *ptep, pte_t pte,
2470 struct page *pagecache_page)
2472 struct hstate *h = hstate_vma(vma);
2473 struct page *old_page, *new_page;
2475 int outside_reserve = 0;
2477 old_page = pte_page(pte);
2480 /* If no-one else is actually using this page, avoid the copy
2481 * and just make the page writable */
2482 avoidcopy = (page_mapcount(old_page) == 1);
2484 if (PageAnon(old_page))
2485 page_move_anon_rmap(old_page, vma, address);
2486 set_huge_ptep_writable(vma, address, ptep);
2491 * If the process that created a MAP_PRIVATE mapping is about to
2492 * perform a COW due to a shared page count, attempt to satisfy
2493 * the allocation without using the existing reserves. The pagecache
2494 * page is used to determine if the reserve at this address was
2495 * consumed or not. If reserves were used, a partial faulted mapping
2496 * at the time of fork() could consume its reserves on COW instead
2497 * of the full address range.
2499 if (!(vma->vm_flags & VM_MAYSHARE) &&
2500 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2501 old_page != pagecache_page)
2502 outside_reserve = 1;
2504 page_cache_get(old_page);
2506 /* Drop page_table_lock as buddy allocator may be called */
2507 spin_unlock(&mm->page_table_lock);
2508 new_page = alloc_huge_page(vma, address, outside_reserve);
2510 if (IS_ERR(new_page)) {
2511 page_cache_release(old_page);
2514 * If a process owning a MAP_PRIVATE mapping fails to COW,
2515 * it is due to references held by a child and an insufficient
2516 * huge page pool. To guarantee the original mappers
2517 * reliability, unmap the page from child processes. The child
2518 * may get SIGKILLed if it later faults.
2520 if (outside_reserve) {
2521 BUG_ON(huge_pte_none(pte));
2522 if (unmap_ref_private(mm, vma, old_page, address)) {
2523 BUG_ON(huge_pte_none(pte));
2524 spin_lock(&mm->page_table_lock);
2525 goto retry_avoidcopy;
2530 /* Caller expects lock to be held */
2531 spin_lock(&mm->page_table_lock);
2532 return -PTR_ERR(new_page);
2536 * When the original hugepage is shared one, it does not have
2537 * anon_vma prepared.
2539 if (unlikely(anon_vma_prepare(vma))) {
2540 page_cache_release(new_page);
2541 page_cache_release(old_page);
2542 /* Caller expects lock to be held */
2543 spin_lock(&mm->page_table_lock);
2544 return VM_FAULT_OOM;
2547 copy_user_huge_page(new_page, old_page, address, vma,
2548 pages_per_huge_page(h));
2549 __SetPageUptodate(new_page);
2552 * Retake the page_table_lock to check for racing updates
2553 * before the page tables are altered
2555 spin_lock(&mm->page_table_lock);
2556 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2557 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2559 mmu_notifier_invalidate_range_start(mm,
2560 address & huge_page_mask(h),
2561 (address & huge_page_mask(h)) + huge_page_size(h));
2562 huge_ptep_clear_flush(vma, address, ptep);
2563 set_huge_pte_at(mm, address, ptep,
2564 make_huge_pte(vma, new_page, 1));
2565 page_remove_rmap(old_page);
2566 hugepage_add_new_anon_rmap(new_page, vma, address);
2567 /* Make the old page be freed below */
2568 new_page = old_page;
2569 mmu_notifier_invalidate_range_end(mm,
2570 address & huge_page_mask(h),
2571 (address & huge_page_mask(h)) + huge_page_size(h));
2573 page_cache_release(new_page);
2574 page_cache_release(old_page);
2578 /* Return the pagecache page at a given address within a VMA */
2579 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2580 struct vm_area_struct *vma, unsigned long address)
2582 struct address_space *mapping;
2585 mapping = vma->vm_file->f_mapping;
2586 idx = vma_hugecache_offset(h, vma, address);
2588 return find_lock_page(mapping, idx);
2592 * Return whether there is a pagecache page to back given address within VMA.
2593 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2595 static bool hugetlbfs_pagecache_present(struct hstate *h,
2596 struct vm_area_struct *vma, unsigned long address)
2598 struct address_space *mapping;
2602 mapping = vma->vm_file->f_mapping;
2603 idx = vma_hugecache_offset(h, vma, address);
2605 page = find_get_page(mapping, idx);
2608 return page != NULL;
2611 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2612 unsigned long address, pte_t *ptep, unsigned int flags)
2614 struct hstate *h = hstate_vma(vma);
2615 int ret = VM_FAULT_SIGBUS;
2619 struct address_space *mapping;
2623 * Currently, we are forced to kill the process in the event the
2624 * original mapper has unmapped pages from the child due to a failed
2625 * COW. Warn that such a situation has occurred as it may not be obvious
2627 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2629 "PID %d killed due to inadequate hugepage pool\n",
2634 mapping = vma->vm_file->f_mapping;
2635 idx = vma_hugecache_offset(h, vma, address);
2638 * Use page lock to guard against racing truncation
2639 * before we get page_table_lock.
2642 page = find_lock_page(mapping, idx);
2644 size = i_size_read(mapping->host) >> huge_page_shift(h);
2647 page = alloc_huge_page(vma, address, 0);
2649 ret = -PTR_ERR(page);
2652 clear_huge_page(page, address, pages_per_huge_page(h));
2653 __SetPageUptodate(page);
2655 if (vma->vm_flags & VM_MAYSHARE) {
2657 struct inode *inode = mapping->host;
2659 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2667 spin_lock(&inode->i_lock);
2668 inode->i_blocks += blocks_per_huge_page(h);
2669 spin_unlock(&inode->i_lock);
2670 page_dup_rmap(page);
2673 if (unlikely(anon_vma_prepare(vma))) {
2675 goto backout_unlocked;
2677 hugepage_add_new_anon_rmap(page, vma, address);
2681 * If memory error occurs between mmap() and fault, some process
2682 * don't have hwpoisoned swap entry for errored virtual address.
2683 * So we need to block hugepage fault by PG_hwpoison bit check.
2685 if (unlikely(PageHWPoison(page))) {
2686 ret = VM_FAULT_HWPOISON |
2687 VM_FAULT_SET_HINDEX(h - hstates);
2688 goto backout_unlocked;
2690 page_dup_rmap(page);
2694 * If we are going to COW a private mapping later, we examine the
2695 * pending reservations for this page now. This will ensure that
2696 * any allocations necessary to record that reservation occur outside
2699 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2700 if (vma_needs_reservation(h, vma, address) < 0) {
2702 goto backout_unlocked;
2705 spin_lock(&mm->page_table_lock);
2706 size = i_size_read(mapping->host) >> huge_page_shift(h);
2711 if (!huge_pte_none(huge_ptep_get(ptep)))
2714 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2715 && (vma->vm_flags & VM_SHARED)));
2716 set_huge_pte_at(mm, address, ptep, new_pte);
2718 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2719 /* Optimization, do the COW without a second fault */
2720 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2723 spin_unlock(&mm->page_table_lock);
2729 spin_unlock(&mm->page_table_lock);
2736 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2737 unsigned long address, unsigned int flags)
2742 struct page *page = NULL;
2743 struct page *pagecache_page = NULL;
2744 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2745 struct hstate *h = hstate_vma(vma);
2747 ptep = huge_pte_offset(mm, address);
2749 entry = huge_ptep_get(ptep);
2750 if (unlikely(is_hugetlb_entry_migration(entry))) {
2751 migration_entry_wait(mm, (pmd_t *)ptep, address);
2753 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2754 return VM_FAULT_HWPOISON_LARGE |
2755 VM_FAULT_SET_HINDEX(h - hstates);
2758 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2760 return VM_FAULT_OOM;
2763 * Serialize hugepage allocation and instantiation, so that we don't
2764 * get spurious allocation failures if two CPUs race to instantiate
2765 * the same page in the page cache.
2767 mutex_lock(&hugetlb_instantiation_mutex);
2768 entry = huge_ptep_get(ptep);
2769 if (huge_pte_none(entry)) {
2770 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2777 * If we are going to COW the mapping later, we examine the pending
2778 * reservations for this page now. This will ensure that any
2779 * allocations necessary to record that reservation occur outside the
2780 * spinlock. For private mappings, we also lookup the pagecache
2781 * page now as it is used to determine if a reservation has been
2784 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2785 if (vma_needs_reservation(h, vma, address) < 0) {
2790 if (!(vma->vm_flags & VM_MAYSHARE))
2791 pagecache_page = hugetlbfs_pagecache_page(h,
2796 * hugetlb_cow() requires page locks of pte_page(entry) and
2797 * pagecache_page, so here we need take the former one
2798 * when page != pagecache_page or !pagecache_page.
2799 * Note that locking order is always pagecache_page -> page,
2800 * so no worry about deadlock.
2802 page = pte_page(entry);
2804 if (page != pagecache_page)
2807 spin_lock(&mm->page_table_lock);
2808 /* Check for a racing update before calling hugetlb_cow */
2809 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2810 goto out_page_table_lock;
2813 if (flags & FAULT_FLAG_WRITE) {
2814 if (!pte_write(entry)) {
2815 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2817 goto out_page_table_lock;
2819 entry = pte_mkdirty(entry);
2821 entry = pte_mkyoung(entry);
2822 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2823 flags & FAULT_FLAG_WRITE))
2824 update_mmu_cache(vma, address, ptep);
2826 out_page_table_lock:
2827 spin_unlock(&mm->page_table_lock);
2829 if (pagecache_page) {
2830 unlock_page(pagecache_page);
2831 put_page(pagecache_page);
2833 if (page != pagecache_page)
2838 mutex_unlock(&hugetlb_instantiation_mutex);
2843 /* Can be overriden by architectures */
2844 __attribute__((weak)) struct page *
2845 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2846 pud_t *pud, int write)
2852 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2853 struct page **pages, struct vm_area_struct **vmas,
2854 unsigned long *position, int *length, int i,
2857 unsigned long pfn_offset;
2858 unsigned long vaddr = *position;
2859 int remainder = *length;
2860 struct hstate *h = hstate_vma(vma);
2862 spin_lock(&mm->page_table_lock);
2863 while (vaddr < vma->vm_end && remainder) {
2869 * Some archs (sparc64, sh*) have multiple pte_ts to
2870 * each hugepage. We have to make sure we get the
2871 * first, for the page indexing below to work.
2873 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2874 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2877 * When coredumping, it suits get_dump_page if we just return
2878 * an error where there's an empty slot with no huge pagecache
2879 * to back it. This way, we avoid allocating a hugepage, and
2880 * the sparse dumpfile avoids allocating disk blocks, but its
2881 * huge holes still show up with zeroes where they need to be.
2883 if (absent && (flags & FOLL_DUMP) &&
2884 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2890 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2893 spin_unlock(&mm->page_table_lock);
2894 ret = hugetlb_fault(mm, vma, vaddr,
2895 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2896 spin_lock(&mm->page_table_lock);
2897 if (!(ret & VM_FAULT_ERROR))
2904 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2905 page = pte_page(huge_ptep_get(pte));
2908 pages[i] = mem_map_offset(page, pfn_offset);
2919 if (vaddr < vma->vm_end && remainder &&
2920 pfn_offset < pages_per_huge_page(h)) {
2922 * We use pfn_offset to avoid touching the pageframes
2923 * of this compound page.
2928 spin_unlock(&mm->page_table_lock);
2929 *length = remainder;
2932 return i ? i : -EFAULT;
2935 void hugetlb_change_protection(struct vm_area_struct *vma,
2936 unsigned long address, unsigned long end, pgprot_t newprot)
2938 struct mm_struct *mm = vma->vm_mm;
2939 unsigned long start = address;
2942 struct hstate *h = hstate_vma(vma);
2944 BUG_ON(address >= end);
2945 flush_cache_range(vma, address, end);
2947 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2948 spin_lock(&mm->page_table_lock);
2949 for (; address < end; address += huge_page_size(h)) {
2950 ptep = huge_pte_offset(mm, address);
2953 if (huge_pmd_unshare(mm, &address, ptep))
2955 if (!huge_pte_none(huge_ptep_get(ptep))) {
2956 pte = huge_ptep_get_and_clear(mm, address, ptep);
2957 pte = pte_mkhuge(pte_modify(pte, newprot));
2958 set_huge_pte_at(mm, address, ptep, pte);
2961 spin_unlock(&mm->page_table_lock);
2963 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2964 * may have cleared our pud entry and done put_page on the page table:
2965 * once we release i_mmap_mutex, another task can do the final put_page
2966 * and that page table be reused and filled with junk.
2968 flush_tlb_range(vma, start, end);
2969 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2972 int hugetlb_reserve_pages(struct inode *inode,
2974 struct vm_area_struct *vma,
2975 vm_flags_t vm_flags)
2978 struct hstate *h = hstate_inode(inode);
2979 struct hugepage_subpool *spool = subpool_inode(inode);
2982 * Only apply hugepage reservation if asked. At fault time, an
2983 * attempt will be made for VM_NORESERVE to allocate a page
2984 * without using reserves
2986 if (vm_flags & VM_NORESERVE)
2990 * Shared mappings base their reservation on the number of pages that
2991 * are already allocated on behalf of the file. Private mappings need
2992 * to reserve the full area even if read-only as mprotect() may be
2993 * called to make the mapping read-write. Assume !vma is a shm mapping
2995 if (!vma || vma->vm_flags & VM_MAYSHARE)
2996 chg = region_chg(&inode->i_mapping->private_list, from, to);
2998 struct resv_map *resv_map = resv_map_alloc();
3004 set_vma_resv_map(vma, resv_map);
3005 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3013 /* There must be enough pages in the subpool for the mapping */
3014 if (hugepage_subpool_get_pages(spool, chg)) {
3020 * Check enough hugepages are available for the reservation.
3021 * Hand the pages back to the subpool if there are not
3023 ret = hugetlb_acct_memory(h, chg);
3025 hugepage_subpool_put_pages(spool, chg);
3030 * Account for the reservations made. Shared mappings record regions
3031 * that have reservations as they are shared by multiple VMAs.
3032 * When the last VMA disappears, the region map says how much
3033 * the reservation was and the page cache tells how much of
3034 * the reservation was consumed. Private mappings are per-VMA and
3035 * only the consumed reservations are tracked. When the VMA
3036 * disappears, the original reservation is the VMA size and the
3037 * consumed reservations are stored in the map. Hence, nothing
3038 * else has to be done for private mappings here
3040 if (!vma || vma->vm_flags & VM_MAYSHARE)
3041 region_add(&inode->i_mapping->private_list, from, to);
3049 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3051 struct hstate *h = hstate_inode(inode);
3052 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3053 struct hugepage_subpool *spool = subpool_inode(inode);
3055 spin_lock(&inode->i_lock);
3056 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3057 spin_unlock(&inode->i_lock);
3059 hugepage_subpool_put_pages(spool, (chg - freed));
3060 hugetlb_acct_memory(h, -(chg - freed));
3063 #ifdef CONFIG_MEMORY_FAILURE
3065 /* Should be called in hugetlb_lock */
3066 static int is_hugepage_on_freelist(struct page *hpage)
3070 struct hstate *h = page_hstate(hpage);
3071 int nid = page_to_nid(hpage);
3073 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3080 * This function is called from memory failure code.
3081 * Assume the caller holds page lock of the head page.
3083 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3085 struct hstate *h = page_hstate(hpage);
3086 int nid = page_to_nid(hpage);
3089 spin_lock(&hugetlb_lock);
3090 if (is_hugepage_on_freelist(hpage)) {
3091 list_del(&hpage->lru);
3092 set_page_refcounted(hpage);
3093 h->free_huge_pages--;
3094 h->free_huge_pages_node[nid]--;
3097 spin_unlock(&hugetlb_lock);