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/mmdebug.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
27 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/node.h>
34 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
35 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
36 unsigned long hugepages_treat_as_movable;
38 static int max_hstate;
39 unsigned int default_hstate_idx;
40 struct hstate hstates[HUGE_MAX_HSTATE];
42 __initdata LIST_HEAD(huge_boot_pages);
44 /* for command line parsing */
45 static struct hstate * __initdata parsed_hstate;
46 static unsigned long __initdata default_hstate_max_huge_pages;
47 static unsigned long __initdata default_hstate_size;
49 #define for_each_hstate(h) \
50 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
53 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
55 static DEFINE_SPINLOCK(hugetlb_lock);
57 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
59 bool free = (spool->count == 0) && (spool->used_hpages == 0);
61 spin_unlock(&spool->lock);
63 /* If no pages are used, and no other handles to the subpool
64 * remain, free the subpool the subpool remain */
69 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
71 struct hugepage_subpool *spool;
73 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
77 spin_lock_init(&spool->lock);
79 spool->max_hpages = nr_blocks;
80 spool->used_hpages = 0;
85 void hugepage_put_subpool(struct hugepage_subpool *spool)
87 spin_lock(&spool->lock);
88 BUG_ON(!spool->count);
90 unlock_or_release_subpool(spool);
93 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 spin_lock(&spool->lock);
102 if ((spool->used_hpages + delta) <= spool->max_hpages) {
103 spool->used_hpages += delta;
107 spin_unlock(&spool->lock);
112 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
118 spin_lock(&spool->lock);
119 spool->used_hpages -= delta;
120 /* If hugetlbfs_put_super couldn't free spool due to
121 * an outstanding quota reference, free it now. */
122 unlock_or_release_subpool(spool);
125 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
127 return HUGETLBFS_SB(inode->i_sb)->spool;
130 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
132 return subpool_inode(vma->vm_file->f_dentry->d_inode);
136 * Region tracking -- allows tracking of reservations and instantiated pages
137 * across the pages in a mapping.
139 * The region data structures are protected by a combination of the mmap_sem
140 * and the hugetlb_instantion_mutex. To access or modify a region the caller
141 * must either hold the mmap_sem for write, or the mmap_sem for read and
142 * the hugetlb_instantiation mutex:
144 * down_write(&mm->mmap_sem);
146 * down_read(&mm->mmap_sem);
147 * mutex_lock(&hugetlb_instantiation_mutex);
150 struct list_head link;
155 static long region_add(struct list_head *head, long f, long t)
157 struct file_region *rg, *nrg, *trg;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg, head, link)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
171 if (&rg->link == head)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct list_head *head, long f, long t)
193 struct file_region *rg, *nrg;
196 /* Locate the region we are before or in. */
197 list_for_each_entry(rg, head, link)
201 /* If we are below the current region then a new region is required.
202 * Subtle, allocate a new region at the position but make it zero
203 * size such that we can guarantee to record the reservation. */
204 if (&rg->link == head || t < rg->from) {
205 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
210 INIT_LIST_HEAD(&nrg->link);
211 list_add(&nrg->link, rg->link.prev);
216 /* Round our left edge to the current segment if it encloses us. */
221 /* Check for and consume any regions we now overlap with. */
222 list_for_each_entry(rg, rg->link.prev, link) {
223 if (&rg->link == head)
228 /* We overlap with this area, if it extends further than
229 * us then we must extend ourselves. Account for its
230 * existing reservation. */
235 chg -= rg->to - rg->from;
240 static long region_truncate(struct list_head *head, long end)
242 struct file_region *rg, *trg;
245 /* Locate the region we are either in or before. */
246 list_for_each_entry(rg, head, link)
249 if (&rg->link == head)
252 /* If we are in the middle of a region then adjust it. */
253 if (end > rg->from) {
256 rg = list_entry(rg->link.next, typeof(*rg), link);
259 /* Drop any remaining regions. */
260 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
261 if (&rg->link == head)
263 chg += rg->to - rg->from;
270 static long region_count(struct list_head *head, long f, long t)
272 struct file_region *rg;
275 /* Locate each segment we overlap with, and count that overlap. */
276 list_for_each_entry(rg, head, link) {
285 seg_from = max(rg->from, f);
286 seg_to = min(rg->to, t);
288 chg += seg_to - seg_from;
295 * Convert the address within this vma to the page offset within
296 * the mapping, in pagecache page units; huge pages here.
298 static pgoff_t vma_hugecache_offset(struct hstate *h,
299 struct vm_area_struct *vma, unsigned long address)
301 return ((address - vma->vm_start) >> huge_page_shift(h)) +
302 (vma->vm_pgoff >> huge_page_order(h));
305 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
306 unsigned long address)
308 return vma_hugecache_offset(hstate_vma(vma), vma, address);
312 * Return the size of the pages allocated when backing a VMA. In the majority
313 * cases this will be same size as used by the page table entries.
315 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
317 struct hstate *hstate;
319 if (!is_vm_hugetlb_page(vma))
322 hstate = hstate_vma(vma);
324 return 1UL << (hstate->order + PAGE_SHIFT);
326 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
329 * Return the page size being used by the MMU to back a VMA. In the majority
330 * of cases, the page size used by the kernel matches the MMU size. On
331 * architectures where it differs, an architecture-specific version of this
332 * function is required.
334 #ifndef vma_mmu_pagesize
335 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
337 return vma_kernel_pagesize(vma);
342 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
343 * bits of the reservation map pointer, which are always clear due to
346 #define HPAGE_RESV_OWNER (1UL << 0)
347 #define HPAGE_RESV_UNMAPPED (1UL << 1)
348 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
351 * These helpers are used to track how many pages are reserved for
352 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
353 * is guaranteed to have their future faults succeed.
355 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
356 * the reserve counters are updated with the hugetlb_lock held. It is safe
357 * to reset the VMA at fork() time as it is not in use yet and there is no
358 * chance of the global counters getting corrupted as a result of the values.
360 * The private mapping reservation is represented in a subtly different
361 * manner to a shared mapping. A shared mapping has a region map associated
362 * with the underlying file, this region map represents the backing file
363 * pages which have ever had a reservation assigned which this persists even
364 * after the page is instantiated. A private mapping has a region map
365 * associated with the original mmap which is attached to all VMAs which
366 * reference it, this region map represents those offsets which have consumed
367 * reservation ie. where pages have been instantiated.
369 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
371 return (unsigned long)vma->vm_private_data;
374 static void set_vma_private_data(struct vm_area_struct *vma,
377 vma->vm_private_data = (void *)value;
382 struct list_head regions;
385 static struct resv_map *resv_map_alloc(void)
387 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
391 kref_init(&resv_map->refs);
392 INIT_LIST_HEAD(&resv_map->regions);
397 static void resv_map_release(struct kref *ref)
399 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
401 /* Clear out any active regions before we release the map. */
402 region_truncate(&resv_map->regions, 0);
406 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
408 VM_BUG_ON(!is_vm_hugetlb_page(vma));
409 if (!(vma->vm_flags & VM_MAYSHARE))
410 return (struct resv_map *)(get_vma_private_data(vma) &
415 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
417 VM_BUG_ON(!is_vm_hugetlb_page(vma));
418 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
420 set_vma_private_data(vma, (get_vma_private_data(vma) &
421 HPAGE_RESV_MASK) | (unsigned long)map);
424 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
426 VM_BUG_ON(!is_vm_hugetlb_page(vma));
427 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
429 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
432 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
434 VM_BUG_ON(!is_vm_hugetlb_page(vma));
436 return (get_vma_private_data(vma) & flag) != 0;
439 /* Decrement the reserved pages in the hugepage pool by one */
440 static void decrement_hugepage_resv_vma(struct hstate *h,
441 struct vm_area_struct *vma)
443 if (vma->vm_flags & VM_NORESERVE)
446 if (vma->vm_flags & VM_MAYSHARE) {
447 /* Shared mappings always use reserves */
448 h->resv_huge_pages--;
449 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
451 * Only the process that called mmap() has reserves for
454 h->resv_huge_pages--;
458 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
459 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
462 if (!(vma->vm_flags & VM_MAYSHARE))
463 vma->vm_private_data = (void *)0;
466 /* Returns true if the VMA has associated reserve pages */
467 static int vma_has_reserves(struct vm_area_struct *vma)
469 if (vma->vm_flags & VM_MAYSHARE)
471 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
476 static void copy_gigantic_page(struct page *dst, struct page *src)
479 struct hstate *h = page_hstate(src);
480 struct page *dst_base = dst;
481 struct page *src_base = src;
483 for (i = 0; i < pages_per_huge_page(h); ) {
485 copy_highpage(dst, src);
488 dst = mem_map_next(dst, dst_base, i);
489 src = mem_map_next(src, src_base, i);
493 void copy_huge_page(struct page *dst, struct page *src)
496 struct hstate *h = page_hstate(src);
498 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
499 copy_gigantic_page(dst, src);
504 for (i = 0; i < pages_per_huge_page(h); i++) {
506 copy_highpage(dst + i, src + i);
510 static void enqueue_huge_page(struct hstate *h, struct page *page)
512 int nid = page_to_nid(page);
513 list_add(&page->lru, &h->hugepage_freelists[nid]);
514 h->free_huge_pages++;
515 h->free_huge_pages_node[nid]++;
518 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 if (list_empty(&h->hugepage_freelists[nid]))
524 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
525 list_del(&page->lru);
526 set_page_refcounted(page);
527 h->free_huge_pages--;
528 h->free_huge_pages_node[nid]--;
532 static struct page *dequeue_huge_page_vma(struct hstate *h,
533 struct vm_area_struct *vma,
534 unsigned long address, int avoid_reserve)
536 struct page *page = NULL;
537 struct mempolicy *mpol;
538 nodemask_t *nodemask;
539 struct zonelist *zonelist;
542 unsigned int cpuset_mems_cookie;
545 cpuset_mems_cookie = get_mems_allowed();
546 zonelist = huge_zonelist(vma, address,
547 htlb_alloc_mask, &mpol, &nodemask);
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma) &&
554 h->free_huge_pages - h->resv_huge_pages == 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 for_each_zone_zonelist_nodemask(zone, z, zonelist,
562 MAX_NR_ZONES - 1, nodemask) {
563 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
564 page = dequeue_huge_page_node(h, zone_to_nid(zone));
567 decrement_hugepage_resv_vma(h, vma);
574 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
583 static void update_and_free_page(struct hstate *h, struct page *page)
587 VM_BUG_ON(h->order >= MAX_ORDER);
590 h->nr_huge_pages_node[page_to_nid(page)]--;
591 for (i = 0; i < pages_per_huge_page(h); i++) {
592 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
593 1 << PG_referenced | 1 << PG_dirty |
594 1 << PG_active | 1 << PG_reserved |
595 1 << PG_private | 1 << PG_writeback);
597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
599 arch_release_hugepage(page);
600 __free_pages(page, huge_page_order(h));
603 struct hstate *size_to_hstate(unsigned long size)
608 if (huge_page_size(h) == size)
614 static void free_huge_page(struct page *page)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate *h = page_hstate(page);
621 int nid = page_to_nid(page);
622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
625 set_page_private(page, 0);
626 page->mapping = NULL;
627 BUG_ON(page_count(page));
628 BUG_ON(page_mapcount(page));
629 INIT_LIST_HEAD(&page->lru);
631 spin_lock(&hugetlb_lock);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 update_and_free_page(h, page);
634 h->surplus_huge_pages--;
635 h->surplus_huge_pages_node[nid]--;
637 enqueue_huge_page(h, page);
639 spin_unlock(&hugetlb_lock);
640 hugepage_subpool_put_pages(spool, 1);
643 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
645 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);
684 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
685 * normal or transparent huge pages.
687 int PageHeadHuge(struct page *page_head)
689 compound_page_dtor *dtor;
691 if (!PageHead(page_head))
694 dtor = get_compound_page_dtor(page_head);
696 return dtor == free_huge_page;
698 EXPORT_SYMBOL_GPL(PageHeadHuge);
700 pgoff_t __basepage_index(struct page *page)
702 struct page *page_head = compound_head(page);
703 pgoff_t index = page_index(page_head);
704 unsigned long compound_idx;
706 if (!PageHuge(page_head))
707 return page_index(page);
709 if (compound_order(page_head) >= MAX_ORDER)
710 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
712 compound_idx = page - page_head;
714 return (index << compound_order(page_head)) + compound_idx;
717 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
721 if (h->order >= MAX_ORDER)
724 page = alloc_pages_exact_node(nid,
725 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
726 __GFP_REPEAT|__GFP_NOWARN,
729 if (arch_prepare_hugepage(page)) {
730 __free_pages(page, huge_page_order(h));
733 prep_new_huge_page(h, page, nid);
740 * common helper functions for hstate_next_node_to_{alloc|free}.
741 * We may have allocated or freed a huge page based on a different
742 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
743 * be outside of *nodes_allowed. Ensure that we use an allowed
744 * node for alloc or free.
746 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
748 nid = next_node(nid, *nodes_allowed);
749 if (nid == MAX_NUMNODES)
750 nid = first_node(*nodes_allowed);
751 VM_BUG_ON(nid >= MAX_NUMNODES);
756 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
758 if (!node_isset(nid, *nodes_allowed))
759 nid = next_node_allowed(nid, nodes_allowed);
764 * returns the previously saved node ["this node"] from which to
765 * allocate a persistent huge page for the pool and advance the
766 * next node from which to allocate, handling wrap at end of node
769 static int hstate_next_node_to_alloc(struct hstate *h,
770 nodemask_t *nodes_allowed)
774 VM_BUG_ON(!nodes_allowed);
776 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
777 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
782 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
789 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
790 next_nid = start_nid;
793 page = alloc_fresh_huge_page_node(h, next_nid);
798 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
799 } while (next_nid != start_nid);
802 count_vm_event(HTLB_BUDDY_PGALLOC);
804 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
810 * helper for free_pool_huge_page() - return the previously saved
811 * node ["this node"] from which to free a huge page. Advance the
812 * next node id whether or not we find a free huge page to free so
813 * that the next attempt to free addresses the next node.
815 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
819 VM_BUG_ON(!nodes_allowed);
821 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
822 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
828 * Free huge page from pool from next node to free.
829 * Attempt to keep persistent huge pages more or less
830 * balanced over allowed nodes.
831 * Called with hugetlb_lock locked.
833 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
840 start_nid = hstate_next_node_to_free(h, nodes_allowed);
841 next_nid = start_nid;
845 * If we're returning unused surplus pages, only examine
846 * nodes with surplus pages.
848 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
849 !list_empty(&h->hugepage_freelists[next_nid])) {
851 list_entry(h->hugepage_freelists[next_nid].next,
853 list_del(&page->lru);
854 h->free_huge_pages--;
855 h->free_huge_pages_node[next_nid]--;
857 h->surplus_huge_pages--;
858 h->surplus_huge_pages_node[next_nid]--;
860 update_and_free_page(h, page);
864 next_nid = hstate_next_node_to_free(h, nodes_allowed);
865 } while (next_nid != start_nid);
870 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
875 if (h->order >= MAX_ORDER)
879 * Assume we will successfully allocate the surplus page to
880 * prevent racing processes from causing the surplus to exceed
883 * This however introduces a different race, where a process B
884 * tries to grow the static hugepage pool while alloc_pages() is
885 * called by process A. B will only examine the per-node
886 * counters in determining if surplus huge pages can be
887 * converted to normal huge pages in adjust_pool_surplus(). A
888 * won't be able to increment the per-node counter, until the
889 * lock is dropped by B, but B doesn't drop hugetlb_lock until
890 * no more huge pages can be converted from surplus to normal
891 * state (and doesn't try to convert again). Thus, we have a
892 * case where a surplus huge page exists, the pool is grown, and
893 * the surplus huge page still exists after, even though it
894 * should just have been converted to a normal huge page. This
895 * does not leak memory, though, as the hugepage will be freed
896 * once it is out of use. It also does not allow the counters to
897 * go out of whack in adjust_pool_surplus() as we don't modify
898 * the node values until we've gotten the hugepage and only the
899 * per-node value is checked there.
901 spin_lock(&hugetlb_lock);
902 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
903 spin_unlock(&hugetlb_lock);
907 h->surplus_huge_pages++;
909 spin_unlock(&hugetlb_lock);
911 if (nid == NUMA_NO_NODE)
912 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
913 __GFP_REPEAT|__GFP_NOWARN,
916 page = alloc_pages_exact_node(nid,
917 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
918 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
920 if (page && arch_prepare_hugepage(page)) {
921 __free_pages(page, huge_page_order(h));
925 spin_lock(&hugetlb_lock);
927 r_nid = page_to_nid(page);
928 set_compound_page_dtor(page, free_huge_page);
930 * We incremented the global counters already
932 h->nr_huge_pages_node[r_nid]++;
933 h->surplus_huge_pages_node[r_nid]++;
934 __count_vm_event(HTLB_BUDDY_PGALLOC);
937 h->surplus_huge_pages--;
938 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
940 spin_unlock(&hugetlb_lock);
946 * This allocation function is useful in the context where vma is irrelevant.
947 * E.g. soft-offlining uses this function because it only cares physical
948 * address of error page.
950 struct page *alloc_huge_page_node(struct hstate *h, int nid)
954 spin_lock(&hugetlb_lock);
955 page = dequeue_huge_page_node(h, nid);
956 spin_unlock(&hugetlb_lock);
959 page = alloc_buddy_huge_page(h, nid);
965 * Increase the hugetlb pool such that it can accommodate a reservation
968 static int gather_surplus_pages(struct hstate *h, int delta)
970 struct list_head surplus_list;
971 struct page *page, *tmp;
973 int needed, allocated;
975 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
977 h->resv_huge_pages += delta;
982 INIT_LIST_HEAD(&surplus_list);
986 spin_unlock(&hugetlb_lock);
987 for (i = 0; i < needed; i++) {
988 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
991 * We were not able to allocate enough pages to
992 * satisfy the entire reservation so we free what
993 * we've allocated so far.
997 list_add(&page->lru, &surplus_list);
1002 * After retaking hugetlb_lock, we need to recalculate 'needed'
1003 * because either resv_huge_pages or free_huge_pages may have changed.
1005 spin_lock(&hugetlb_lock);
1006 needed = (h->resv_huge_pages + delta) -
1007 (h->free_huge_pages + allocated);
1012 * The surplus_list now contains _at_least_ the number of extra pages
1013 * needed to accommodate the reservation. Add the appropriate number
1014 * of pages to the hugetlb pool and free the extras back to the buddy
1015 * allocator. Commit the entire reservation here to prevent another
1016 * process from stealing the pages as they are added to the pool but
1017 * before they are reserved.
1019 needed += allocated;
1020 h->resv_huge_pages += delta;
1023 /* Free the needed pages to the hugetlb pool */
1024 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1027 list_del(&page->lru);
1029 * This page is now managed by the hugetlb allocator and has
1030 * no users -- drop the buddy allocator's reference.
1032 put_page_testzero(page);
1033 VM_BUG_ON(page_count(page));
1034 enqueue_huge_page(h, page);
1036 spin_unlock(&hugetlb_lock);
1038 /* Free unnecessary surplus pages to the buddy allocator */
1040 if (!list_empty(&surplus_list)) {
1041 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1042 list_del(&page->lru);
1046 spin_lock(&hugetlb_lock);
1052 * When releasing a hugetlb pool reservation, any surplus pages that were
1053 * allocated to satisfy the reservation must be explicitly freed if they were
1055 * Called with hugetlb_lock held.
1057 static void return_unused_surplus_pages(struct hstate *h,
1058 unsigned long unused_resv_pages)
1060 unsigned long nr_pages;
1062 /* Uncommit the reservation */
1063 h->resv_huge_pages -= unused_resv_pages;
1065 /* Cannot return gigantic pages currently */
1066 if (h->order >= MAX_ORDER)
1069 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1072 * We want to release as many surplus pages as possible, spread
1073 * evenly across all nodes with memory. Iterate across these nodes
1074 * until we can no longer free unreserved surplus pages. This occurs
1075 * when the nodes with surplus pages have no free pages.
1076 * free_pool_huge_page() will balance the the freed pages across the
1077 * on-line nodes with memory and will handle the hstate accounting.
1079 while (nr_pages--) {
1080 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1082 cond_resched_lock(&hugetlb_lock);
1087 * Determine if the huge page at addr within the vma has an associated
1088 * reservation. Where it does not we will need to logically increase
1089 * reservation and actually increase subpool usage before an allocation
1090 * can occur. Where any new reservation would be required the
1091 * reservation change is prepared, but not committed. Once the page
1092 * has been allocated from the subpool and instantiated the change should
1093 * be committed via vma_commit_reservation. No action is required on
1096 static long vma_needs_reservation(struct hstate *h,
1097 struct vm_area_struct *vma, unsigned long addr)
1099 struct address_space *mapping = vma->vm_file->f_mapping;
1100 struct inode *inode = mapping->host;
1102 if (vma->vm_flags & VM_MAYSHARE) {
1103 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1104 return region_chg(&inode->i_mapping->private_list,
1107 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1112 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1113 struct resv_map *reservations = vma_resv_map(vma);
1115 err = region_chg(&reservations->regions, idx, idx + 1);
1121 static void vma_commit_reservation(struct hstate *h,
1122 struct vm_area_struct *vma, unsigned long addr)
1124 struct address_space *mapping = vma->vm_file->f_mapping;
1125 struct inode *inode = mapping->host;
1127 if (vma->vm_flags & VM_MAYSHARE) {
1128 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1129 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1131 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1132 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1133 struct resv_map *reservations = vma_resv_map(vma);
1135 /* Mark this page used in the map. */
1136 region_add(&reservations->regions, idx, idx + 1);
1140 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1141 unsigned long addr, int avoid_reserve)
1143 struct hugepage_subpool *spool = subpool_vma(vma);
1144 struct hstate *h = hstate_vma(vma);
1149 * Processes that did not create the mapping will have no
1150 * reserves and will not have accounted against subpool
1151 * limit. Check that the subpool limit can be made before
1152 * satisfying the allocation MAP_NORESERVE mappings may also
1153 * need pages and subpool limit allocated allocated if no reserve
1156 chg = vma_needs_reservation(h, vma, addr);
1158 return ERR_PTR(-VM_FAULT_OOM);
1160 if (hugepage_subpool_get_pages(spool, chg))
1161 return ERR_PTR(-VM_FAULT_SIGBUS);
1163 spin_lock(&hugetlb_lock);
1164 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1165 spin_unlock(&hugetlb_lock);
1168 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1170 hugepage_subpool_put_pages(spool, chg);
1171 return ERR_PTR(-VM_FAULT_SIGBUS);
1175 set_page_private(page, (unsigned long)spool);
1177 vma_commit_reservation(h, vma, addr);
1182 int __weak alloc_bootmem_huge_page(struct hstate *h)
1184 struct huge_bootmem_page *m;
1185 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1190 addr = __alloc_bootmem_node_nopanic(
1191 NODE_DATA(hstate_next_node_to_alloc(h,
1192 &node_states[N_HIGH_MEMORY])),
1193 huge_page_size(h), huge_page_size(h), 0);
1197 * Use the beginning of the huge page to store the
1198 * huge_bootmem_page struct (until gather_bootmem
1199 * puts them into the mem_map).
1209 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1210 /* Put them into a private list first because mem_map is not up yet */
1211 list_add(&m->list, &huge_boot_pages);
1216 static void prep_compound_huge_page(struct page *page, int order)
1218 if (unlikely(order > (MAX_ORDER - 1)))
1219 prep_compound_gigantic_page(page, order);
1221 prep_compound_page(page, order);
1224 /* Put bootmem huge pages into the standard lists after mem_map is up */
1225 static void __init gather_bootmem_prealloc(void)
1227 struct huge_bootmem_page *m;
1229 list_for_each_entry(m, &huge_boot_pages, list) {
1230 struct hstate *h = m->hstate;
1233 #ifdef CONFIG_HIGHMEM
1234 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1235 free_bootmem_late((unsigned long)m,
1236 sizeof(struct huge_bootmem_page));
1238 page = virt_to_page(m);
1240 __ClearPageReserved(page);
1241 WARN_ON(page_count(page) != 1);
1242 prep_compound_huge_page(page, h->order);
1243 prep_new_huge_page(h, page, page_to_nid(page));
1245 * If we had gigantic hugepages allocated at boot time, we need
1246 * to restore the 'stolen' pages to totalram_pages in order to
1247 * fix confusing memory reports from free(1) and another
1248 * side-effects, like CommitLimit going negative.
1250 if (h->order > (MAX_ORDER - 1))
1251 totalram_pages += 1 << h->order;
1255 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1259 for (i = 0; i < h->max_huge_pages; ++i) {
1260 if (h->order >= MAX_ORDER) {
1261 if (!alloc_bootmem_huge_page(h))
1263 } else if (!alloc_fresh_huge_page(h,
1264 &node_states[N_HIGH_MEMORY]))
1267 h->max_huge_pages = i;
1270 static void __init hugetlb_init_hstates(void)
1274 for_each_hstate(h) {
1275 /* oversize hugepages were init'ed in early boot */
1276 if (h->order < MAX_ORDER)
1277 hugetlb_hstate_alloc_pages(h);
1281 static char * __init memfmt(char *buf, unsigned long n)
1283 if (n >= (1UL << 30))
1284 sprintf(buf, "%lu GB", n >> 30);
1285 else if (n >= (1UL << 20))
1286 sprintf(buf, "%lu MB", n >> 20);
1288 sprintf(buf, "%lu KB", n >> 10);
1292 static void __init report_hugepages(void)
1296 for_each_hstate(h) {
1298 printk(KERN_INFO "HugeTLB registered %s page size, "
1299 "pre-allocated %ld pages\n",
1300 memfmt(buf, huge_page_size(h)),
1301 h->free_huge_pages);
1305 #ifdef CONFIG_HIGHMEM
1306 static void try_to_free_low(struct hstate *h, unsigned long count,
1307 nodemask_t *nodes_allowed)
1311 if (h->order >= MAX_ORDER)
1314 for_each_node_mask(i, *nodes_allowed) {
1315 struct page *page, *next;
1316 struct list_head *freel = &h->hugepage_freelists[i];
1317 list_for_each_entry_safe(page, next, freel, lru) {
1318 if (count >= h->nr_huge_pages)
1320 if (PageHighMem(page))
1322 list_del(&page->lru);
1323 update_and_free_page(h, page);
1324 h->free_huge_pages--;
1325 h->free_huge_pages_node[page_to_nid(page)]--;
1330 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1331 nodemask_t *nodes_allowed)
1337 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1338 * balanced by operating on them in a round-robin fashion.
1339 * Returns 1 if an adjustment was made.
1341 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1344 int start_nid, next_nid;
1347 VM_BUG_ON(delta != -1 && delta != 1);
1350 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1352 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1353 next_nid = start_nid;
1359 * To shrink on this node, there must be a surplus page
1361 if (!h->surplus_huge_pages_node[nid]) {
1362 next_nid = hstate_next_node_to_alloc(h,
1369 * Surplus cannot exceed the total number of pages
1371 if (h->surplus_huge_pages_node[nid] >=
1372 h->nr_huge_pages_node[nid]) {
1373 next_nid = hstate_next_node_to_free(h,
1379 h->surplus_huge_pages += delta;
1380 h->surplus_huge_pages_node[nid] += delta;
1383 } while (next_nid != start_nid);
1388 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1389 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1390 nodemask_t *nodes_allowed)
1392 unsigned long min_count, ret;
1394 if (h->order >= MAX_ORDER)
1395 return h->max_huge_pages;
1398 * Increase the pool size
1399 * First take pages out of surplus state. Then make up the
1400 * remaining difference by allocating fresh huge pages.
1402 * We might race with alloc_buddy_huge_page() here and be unable
1403 * to convert a surplus huge page to a normal huge page. That is
1404 * not critical, though, it just means the overall size of the
1405 * pool might be one hugepage larger than it needs to be, but
1406 * within all the constraints specified by the sysctls.
1408 spin_lock(&hugetlb_lock);
1409 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1410 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1414 while (count > persistent_huge_pages(h)) {
1416 * If this allocation races such that we no longer need the
1417 * page, free_huge_page will handle it by freeing the page
1418 * and reducing the surplus.
1420 spin_unlock(&hugetlb_lock);
1422 /* yield cpu to avoid soft lockup */
1425 ret = alloc_fresh_huge_page(h, nodes_allowed);
1426 spin_lock(&hugetlb_lock);
1430 /* Bail for signals. Probably ctrl-c from user */
1431 if (signal_pending(current))
1436 * Decrease the pool size
1437 * First return free pages to the buddy allocator (being careful
1438 * to keep enough around to satisfy reservations). Then place
1439 * pages into surplus state as needed so the pool will shrink
1440 * to the desired size as pages become free.
1442 * By placing pages into the surplus state independent of the
1443 * overcommit value, we are allowing the surplus pool size to
1444 * exceed overcommit. There are few sane options here. Since
1445 * alloc_buddy_huge_page() is checking the global counter,
1446 * though, we'll note that we're not allowed to exceed surplus
1447 * and won't grow the pool anywhere else. Not until one of the
1448 * sysctls are changed, or the surplus pages go out of use.
1450 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1451 min_count = max(count, min_count);
1452 try_to_free_low(h, min_count, nodes_allowed);
1453 while (min_count < persistent_huge_pages(h)) {
1454 if (!free_pool_huge_page(h, nodes_allowed, 0))
1456 cond_resched_lock(&hugetlb_lock);
1458 while (count < persistent_huge_pages(h)) {
1459 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1463 ret = persistent_huge_pages(h);
1464 spin_unlock(&hugetlb_lock);
1468 #define HSTATE_ATTR_RO(_name) \
1469 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1471 #define HSTATE_ATTR(_name) \
1472 static struct kobj_attribute _name##_attr = \
1473 __ATTR(_name, 0644, _name##_show, _name##_store)
1475 static struct kobject *hugepages_kobj;
1476 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1478 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1480 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1484 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1485 if (hstate_kobjs[i] == kobj) {
1487 *nidp = NUMA_NO_NODE;
1491 return kobj_to_node_hstate(kobj, nidp);
1494 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1495 struct kobj_attribute *attr, char *buf)
1498 unsigned long nr_huge_pages;
1501 h = kobj_to_hstate(kobj, &nid);
1502 if (nid == NUMA_NO_NODE)
1503 nr_huge_pages = h->nr_huge_pages;
1505 nr_huge_pages = h->nr_huge_pages_node[nid];
1507 return sprintf(buf, "%lu\n", nr_huge_pages);
1510 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1511 struct kobject *kobj, struct kobj_attribute *attr,
1512 const char *buf, size_t len)
1516 unsigned long count;
1518 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1520 err = strict_strtoul(buf, 10, &count);
1524 h = kobj_to_hstate(kobj, &nid);
1525 if (h->order >= MAX_ORDER) {
1530 if (nid == NUMA_NO_NODE) {
1532 * global hstate attribute
1534 if (!(obey_mempolicy &&
1535 init_nodemask_of_mempolicy(nodes_allowed))) {
1536 NODEMASK_FREE(nodes_allowed);
1537 nodes_allowed = &node_states[N_HIGH_MEMORY];
1539 } else if (nodes_allowed) {
1541 * per node hstate attribute: adjust count to global,
1542 * but restrict alloc/free to the specified node.
1544 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1545 init_nodemask_of_node(nodes_allowed, nid);
1547 nodes_allowed = &node_states[N_HIGH_MEMORY];
1549 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1551 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1552 NODEMASK_FREE(nodes_allowed);
1556 NODEMASK_FREE(nodes_allowed);
1560 static ssize_t nr_hugepages_show(struct kobject *kobj,
1561 struct kobj_attribute *attr, char *buf)
1563 return nr_hugepages_show_common(kobj, attr, buf);
1566 static ssize_t nr_hugepages_store(struct kobject *kobj,
1567 struct kobj_attribute *attr, const char *buf, size_t len)
1569 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1571 HSTATE_ATTR(nr_hugepages);
1576 * hstate attribute for optionally mempolicy-based constraint on persistent
1577 * huge page alloc/free.
1579 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1580 struct kobj_attribute *attr, char *buf)
1582 return nr_hugepages_show_common(kobj, attr, buf);
1585 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1586 struct kobj_attribute *attr, const char *buf, size_t len)
1588 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1590 HSTATE_ATTR(nr_hugepages_mempolicy);
1594 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1595 struct kobj_attribute *attr, char *buf)
1597 struct hstate *h = kobj_to_hstate(kobj, NULL);
1598 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1601 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1602 struct kobj_attribute *attr, const char *buf, size_t count)
1605 unsigned long input;
1606 struct hstate *h = kobj_to_hstate(kobj, NULL);
1608 if (h->order >= MAX_ORDER)
1611 err = strict_strtoul(buf, 10, &input);
1615 spin_lock(&hugetlb_lock);
1616 h->nr_overcommit_huge_pages = input;
1617 spin_unlock(&hugetlb_lock);
1621 HSTATE_ATTR(nr_overcommit_hugepages);
1623 static ssize_t free_hugepages_show(struct kobject *kobj,
1624 struct kobj_attribute *attr, char *buf)
1627 unsigned long free_huge_pages;
1630 h = kobj_to_hstate(kobj, &nid);
1631 if (nid == NUMA_NO_NODE)
1632 free_huge_pages = h->free_huge_pages;
1634 free_huge_pages = h->free_huge_pages_node[nid];
1636 return sprintf(buf, "%lu\n", free_huge_pages);
1638 HSTATE_ATTR_RO(free_hugepages);
1640 static ssize_t resv_hugepages_show(struct kobject *kobj,
1641 struct kobj_attribute *attr, char *buf)
1643 struct hstate *h = kobj_to_hstate(kobj, NULL);
1644 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1646 HSTATE_ATTR_RO(resv_hugepages);
1648 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1649 struct kobj_attribute *attr, char *buf)
1652 unsigned long surplus_huge_pages;
1655 h = kobj_to_hstate(kobj, &nid);
1656 if (nid == NUMA_NO_NODE)
1657 surplus_huge_pages = h->surplus_huge_pages;
1659 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1661 return sprintf(buf, "%lu\n", surplus_huge_pages);
1663 HSTATE_ATTR_RO(surplus_hugepages);
1665 static struct attribute *hstate_attrs[] = {
1666 &nr_hugepages_attr.attr,
1667 &nr_overcommit_hugepages_attr.attr,
1668 &free_hugepages_attr.attr,
1669 &resv_hugepages_attr.attr,
1670 &surplus_hugepages_attr.attr,
1672 &nr_hugepages_mempolicy_attr.attr,
1677 static struct attribute_group hstate_attr_group = {
1678 .attrs = hstate_attrs,
1681 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1682 struct kobject **hstate_kobjs,
1683 struct attribute_group *hstate_attr_group)
1686 int hi = h - hstates;
1688 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1689 if (!hstate_kobjs[hi])
1692 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1694 kobject_put(hstate_kobjs[hi]);
1699 static void __init hugetlb_sysfs_init(void)
1704 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1705 if (!hugepages_kobj)
1708 for_each_hstate(h) {
1709 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1710 hstate_kobjs, &hstate_attr_group);
1712 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1720 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1721 * with node sysdevs in node_devices[] using a parallel array. The array
1722 * index of a node sysdev or _hstate == node id.
1723 * This is here to avoid any static dependency of the node sysdev driver, in
1724 * the base kernel, on the hugetlb module.
1726 struct node_hstate {
1727 struct kobject *hugepages_kobj;
1728 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1730 struct node_hstate node_hstates[MAX_NUMNODES];
1733 * A subset of global hstate attributes for node sysdevs
1735 static struct attribute *per_node_hstate_attrs[] = {
1736 &nr_hugepages_attr.attr,
1737 &free_hugepages_attr.attr,
1738 &surplus_hugepages_attr.attr,
1742 static struct attribute_group per_node_hstate_attr_group = {
1743 .attrs = per_node_hstate_attrs,
1747 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1748 * Returns node id via non-NULL nidp.
1750 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1754 for (nid = 0; nid < nr_node_ids; nid++) {
1755 struct node_hstate *nhs = &node_hstates[nid];
1757 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1758 if (nhs->hstate_kobjs[i] == kobj) {
1770 * Unregister hstate attributes from a single node sysdev.
1771 * No-op if no hstate attributes attached.
1773 void hugetlb_unregister_node(struct node *node)
1776 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1778 if (!nhs->hugepages_kobj)
1779 return; /* no hstate attributes */
1782 if (nhs->hstate_kobjs[h - hstates]) {
1783 kobject_put(nhs->hstate_kobjs[h - hstates]);
1784 nhs->hstate_kobjs[h - hstates] = NULL;
1787 kobject_put(nhs->hugepages_kobj);
1788 nhs->hugepages_kobj = NULL;
1792 * hugetlb module exit: unregister hstate attributes from node sysdevs
1795 static void hugetlb_unregister_all_nodes(void)
1800 * disable node sysdev registrations.
1802 register_hugetlbfs_with_node(NULL, NULL);
1805 * remove hstate attributes from any nodes that have them.
1807 for (nid = 0; nid < nr_node_ids; nid++)
1808 hugetlb_unregister_node(&node_devices[nid]);
1812 * Register hstate attributes for a single node sysdev.
1813 * No-op if attributes already registered.
1815 void hugetlb_register_node(struct node *node)
1818 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1821 if (nhs->hugepages_kobj)
1822 return; /* already allocated */
1824 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1825 &node->sysdev.kobj);
1826 if (!nhs->hugepages_kobj)
1829 for_each_hstate(h) {
1830 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1832 &per_node_hstate_attr_group);
1834 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1836 h->name, node->sysdev.id);
1837 hugetlb_unregister_node(node);
1844 * hugetlb init time: register hstate attributes for all registered node
1845 * sysdevs of nodes that have memory. All on-line nodes should have
1846 * registered their associated sysdev by this time.
1848 static void hugetlb_register_all_nodes(void)
1852 for_each_node_state(nid, N_HIGH_MEMORY) {
1853 struct node *node = &node_devices[nid];
1854 if (node->sysdev.id == nid)
1855 hugetlb_register_node(node);
1859 * Let the node sysdev driver know we're here so it can
1860 * [un]register hstate attributes on node hotplug.
1862 register_hugetlbfs_with_node(hugetlb_register_node,
1863 hugetlb_unregister_node);
1865 #else /* !CONFIG_NUMA */
1867 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1875 static void hugetlb_unregister_all_nodes(void) { }
1877 static void hugetlb_register_all_nodes(void) { }
1881 static void __exit hugetlb_exit(void)
1885 hugetlb_unregister_all_nodes();
1887 for_each_hstate(h) {
1888 kobject_put(hstate_kobjs[h - hstates]);
1891 kobject_put(hugepages_kobj);
1893 module_exit(hugetlb_exit);
1895 static int __init hugetlb_init(void)
1897 if (!hugepages_supported())
1900 if (!size_to_hstate(default_hstate_size)) {
1901 default_hstate_size = HPAGE_SIZE;
1902 if (!size_to_hstate(default_hstate_size))
1903 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1905 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1906 if (default_hstate_max_huge_pages)
1907 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1909 hugetlb_init_hstates();
1911 gather_bootmem_prealloc();
1915 hugetlb_sysfs_init();
1917 hugetlb_register_all_nodes();
1921 module_init(hugetlb_init);
1923 /* Should be called on processing a hugepagesz=... option */
1924 void __init hugetlb_add_hstate(unsigned order)
1929 if (size_to_hstate(PAGE_SIZE << order)) {
1930 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1933 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1935 h = &hstates[max_hstate++];
1937 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1938 h->nr_huge_pages = 0;
1939 h->free_huge_pages = 0;
1940 for (i = 0; i < MAX_NUMNODES; ++i)
1941 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1942 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1943 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1944 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1945 huge_page_size(h)/1024);
1950 static int __init hugetlb_nrpages_setup(char *s)
1953 static unsigned long *last_mhp;
1956 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1957 * so this hugepages= parameter goes to the "default hstate".
1960 mhp = &default_hstate_max_huge_pages;
1962 mhp = &parsed_hstate->max_huge_pages;
1964 if (mhp == last_mhp) {
1965 printk(KERN_WARNING "hugepages= specified twice without "
1966 "interleaving hugepagesz=, ignoring\n");
1970 if (sscanf(s, "%lu", mhp) <= 0)
1974 * Global state is always initialized later in hugetlb_init.
1975 * But we need to allocate >= MAX_ORDER hstates here early to still
1976 * use the bootmem allocator.
1978 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1979 hugetlb_hstate_alloc_pages(parsed_hstate);
1985 __setup("hugepages=", hugetlb_nrpages_setup);
1987 static int __init hugetlb_default_setup(char *s)
1989 default_hstate_size = memparse(s, &s);
1992 __setup("default_hugepagesz=", hugetlb_default_setup);
1994 static unsigned int cpuset_mems_nr(unsigned int *array)
1997 unsigned int nr = 0;
1999 for_each_node_mask(node, cpuset_current_mems_allowed)
2005 #ifdef CONFIG_SYSCTL
2006 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2007 struct ctl_table *table, int write,
2008 void __user *buffer, size_t *length, loff_t *ppos)
2010 struct hstate *h = &default_hstate;
2014 if (!hugepages_supported())
2017 tmp = h->max_huge_pages;
2019 if (write && h->order >= MAX_ORDER)
2023 table->maxlen = sizeof(unsigned long);
2024 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2029 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2030 GFP_KERNEL | __GFP_NORETRY);
2031 if (!(obey_mempolicy &&
2032 init_nodemask_of_mempolicy(nodes_allowed))) {
2033 NODEMASK_FREE(nodes_allowed);
2034 nodes_allowed = &node_states[N_HIGH_MEMORY];
2036 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2038 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2039 NODEMASK_FREE(nodes_allowed);
2045 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2046 void __user *buffer, size_t *length, loff_t *ppos)
2049 return hugetlb_sysctl_handler_common(false, table, write,
2050 buffer, length, ppos);
2054 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2055 void __user *buffer, size_t *length, loff_t *ppos)
2057 return hugetlb_sysctl_handler_common(true, table, write,
2058 buffer, length, ppos);
2060 #endif /* CONFIG_NUMA */
2062 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2063 void __user *buffer,
2064 size_t *length, loff_t *ppos)
2066 proc_dointvec(table, write, buffer, length, ppos);
2067 if (hugepages_treat_as_movable)
2068 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2070 htlb_alloc_mask = GFP_HIGHUSER;
2074 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2075 void __user *buffer,
2076 size_t *length, loff_t *ppos)
2078 struct hstate *h = &default_hstate;
2082 if (!hugepages_supported())
2085 tmp = h->nr_overcommit_huge_pages;
2087 if (write && h->order >= MAX_ORDER)
2091 table->maxlen = sizeof(unsigned long);
2092 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2097 spin_lock(&hugetlb_lock);
2098 h->nr_overcommit_huge_pages = tmp;
2099 spin_unlock(&hugetlb_lock);
2105 #endif /* CONFIG_SYSCTL */
2107 void hugetlb_report_meminfo(struct seq_file *m)
2109 struct hstate *h = &default_hstate;
2110 if (!hugepages_supported())
2113 "HugePages_Total: %5lu\n"
2114 "HugePages_Free: %5lu\n"
2115 "HugePages_Rsvd: %5lu\n"
2116 "HugePages_Surp: %5lu\n"
2117 "Hugepagesize: %8lu kB\n",
2121 h->surplus_huge_pages,
2122 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2125 int hugetlb_report_node_meminfo(int nid, char *buf)
2127 struct hstate *h = &default_hstate;
2128 if (!hugepages_supported())
2131 "Node %d HugePages_Total: %5u\n"
2132 "Node %d HugePages_Free: %5u\n"
2133 "Node %d HugePages_Surp: %5u\n",
2134 nid, h->nr_huge_pages_node[nid],
2135 nid, h->free_huge_pages_node[nid],
2136 nid, h->surplus_huge_pages_node[nid]);
2139 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2140 unsigned long hugetlb_total_pages(void)
2143 unsigned long nr_total_pages = 0;
2146 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2147 return nr_total_pages;
2150 static int hugetlb_acct_memory(struct hstate *h, long delta)
2154 spin_lock(&hugetlb_lock);
2156 * When cpuset is configured, it breaks the strict hugetlb page
2157 * reservation as the accounting is done on a global variable. Such
2158 * reservation is completely rubbish in the presence of cpuset because
2159 * the reservation is not checked against page availability for the
2160 * current cpuset. Application can still potentially OOM'ed by kernel
2161 * with lack of free htlb page in cpuset that the task is in.
2162 * Attempt to enforce strict accounting with cpuset is almost
2163 * impossible (or too ugly) because cpuset is too fluid that
2164 * task or memory node can be dynamically moved between cpusets.
2166 * The change of semantics for shared hugetlb mapping with cpuset is
2167 * undesirable. However, in order to preserve some of the semantics,
2168 * we fall back to check against current free page availability as
2169 * a best attempt and hopefully to minimize the impact of changing
2170 * semantics that cpuset has.
2173 if (gather_surplus_pages(h, delta) < 0)
2176 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2177 return_unused_surplus_pages(h, delta);
2184 return_unused_surplus_pages(h, (unsigned long) -delta);
2187 spin_unlock(&hugetlb_lock);
2191 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2193 struct resv_map *reservations = vma_resv_map(vma);
2196 * This new VMA should share its siblings reservation map if present.
2197 * The VMA will only ever have a valid reservation map pointer where
2198 * it is being copied for another still existing VMA. As that VMA
2199 * has a reference to the reservation map it cannot disappear until
2200 * after this open call completes. It is therefore safe to take a
2201 * new reference here without additional locking.
2204 kref_get(&reservations->refs);
2207 static void resv_map_put(struct vm_area_struct *vma)
2209 struct resv_map *reservations = vma_resv_map(vma);
2213 kref_put(&reservations->refs, resv_map_release);
2216 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2218 struct hstate *h = hstate_vma(vma);
2219 struct resv_map *reservations = vma_resv_map(vma);
2220 struct hugepage_subpool *spool = subpool_vma(vma);
2221 unsigned long reserve;
2222 unsigned long start;
2226 start = vma_hugecache_offset(h, vma, vma->vm_start);
2227 end = vma_hugecache_offset(h, vma, vma->vm_end);
2229 reserve = (end - start) -
2230 region_count(&reservations->regions, start, end);
2235 hugetlb_acct_memory(h, -reserve);
2236 hugepage_subpool_put_pages(spool, reserve);
2242 * We cannot handle pagefaults against hugetlb pages at all. They cause
2243 * handle_mm_fault() to try to instantiate regular-sized pages in the
2244 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2247 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2253 const struct vm_operations_struct hugetlb_vm_ops = {
2254 .fault = hugetlb_vm_op_fault,
2255 .open = hugetlb_vm_op_open,
2256 .close = hugetlb_vm_op_close,
2259 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2266 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2268 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2270 entry = pte_mkyoung(entry);
2271 entry = pte_mkhuge(entry);
2276 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2277 unsigned long address, pte_t *ptep)
2281 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2282 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2283 update_mmu_cache(vma, address, ptep);
2286 static int is_hugetlb_entry_migration(pte_t pte)
2290 if (huge_pte_none(pte) || pte_present(pte))
2292 swp = pte_to_swp_entry(pte);
2293 if (non_swap_entry(swp) && is_migration_entry(swp))
2299 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2303 if (huge_pte_none(pte) || pte_present(pte))
2305 swp = pte_to_swp_entry(pte);
2306 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2312 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2313 struct vm_area_struct *vma)
2315 pte_t *src_pte, *dst_pte, entry;
2316 struct page *ptepage;
2319 struct hstate *h = hstate_vma(vma);
2320 unsigned long sz = huge_page_size(h);
2322 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2324 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2325 src_pte = huge_pte_offset(src, addr);
2328 dst_pte = huge_pte_alloc(dst, addr, sz);
2332 /* If the pagetables are shared don't copy or take references */
2333 if (dst_pte == src_pte)
2336 spin_lock(&dst->page_table_lock);
2337 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2338 entry = huge_ptep_get(src_pte);
2339 if (huge_pte_none(entry)) { /* skip none entry */
2341 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2342 is_hugetlb_entry_hwpoisoned(entry))) {
2343 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2345 if (is_write_migration_entry(swp_entry) && cow) {
2347 * COW mappings require pages in both
2348 * parent and child to be set to read.
2350 make_migration_entry_read(&swp_entry);
2351 entry = swp_entry_to_pte(swp_entry);
2352 set_huge_pte_at(src, addr, src_pte, entry);
2354 set_huge_pte_at(dst, addr, dst_pte, entry);
2357 huge_ptep_set_wrprotect(src, addr, src_pte);
2358 entry = huge_ptep_get(src_pte);
2359 ptepage = pte_page(entry);
2361 page_dup_rmap(ptepage);
2362 set_huge_pte_at(dst, addr, dst_pte, entry);
2364 spin_unlock(&src->page_table_lock);
2365 spin_unlock(&dst->page_table_lock);
2373 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2374 unsigned long end, struct page *ref_page)
2376 struct mm_struct *mm = vma->vm_mm;
2377 unsigned long address;
2382 struct hstate *h = hstate_vma(vma);
2383 unsigned long sz = huge_page_size(h);
2386 * A page gathering list, protected by per file i_mmap_mutex. The
2387 * lock is used to avoid list corruption from multiple unmapping
2388 * of the same page since we are using page->lru.
2390 LIST_HEAD(page_list);
2392 WARN_ON(!is_vm_hugetlb_page(vma));
2393 BUG_ON(start & ~huge_page_mask(h));
2394 BUG_ON(end & ~huge_page_mask(h));
2396 mmu_notifier_invalidate_range_start(mm, start, end);
2397 spin_lock(&mm->page_table_lock);
2398 for (address = start; address < end; address += sz) {
2399 ptep = huge_pte_offset(mm, address);
2403 if (huge_pmd_unshare(mm, &address, ptep))
2407 * If a reference page is supplied, it is because a specific
2408 * page is being unmapped, not a range. Ensure the page we
2409 * are about to unmap is the actual page of interest.
2412 pte = huge_ptep_get(ptep);
2413 if (huge_pte_none(pte))
2415 page = pte_page(pte);
2416 if (page != ref_page)
2420 * Mark the VMA as having unmapped its page so that
2421 * future faults in this VMA will fail rather than
2422 * looking like data was lost
2424 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2427 pte = huge_ptep_get_and_clear(mm, address, ptep);
2428 if (huge_pte_none(pte))
2432 * Migrating hugepage or HWPoisoned hugepage is already
2433 * unmapped and its refcount is dropped
2435 if (unlikely(!pte_present(pte)))
2438 page = pte_page(pte);
2440 set_page_dirty(page);
2441 list_add(&page->lru, &page_list);
2443 spin_unlock(&mm->page_table_lock);
2444 flush_tlb_range(vma, start, end);
2445 mmu_notifier_invalidate_range_end(mm, start, end);
2446 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2447 page_remove_rmap(page);
2448 list_del(&page->lru);
2453 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2454 unsigned long start, unsigned long end,
2455 struct page *ref_page)
2457 __unmap_hugepage_range(vma, start, end, ref_page);
2460 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2461 * test will fail on a vma being torn down, and not grab a page table
2462 * on its way out. We're lucky that the flag has such an appropriate
2463 * name, and can in fact be safely cleared here. We could clear it
2464 * before the __unmap_hugepage_range above, but all that's necessary
2465 * is to clear it before releasing the i_mmap_mutex. This works
2466 * because in the context this is called, the VMA is about to be
2467 * destroyed and the i_mmap_mutex is held.
2469 vma->vm_flags &= ~VM_MAYSHARE;
2472 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2473 unsigned long end, struct page *ref_page)
2475 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2476 __unmap_hugepage_range(vma, start, end, ref_page);
2477 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2481 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2482 * mappping it owns the reserve page for. The intention is to unmap the page
2483 * from other VMAs and let the children be SIGKILLed if they are faulting the
2486 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2487 struct page *page, unsigned long address)
2489 struct hstate *h = hstate_vma(vma);
2490 struct vm_area_struct *iter_vma;
2491 struct address_space *mapping;
2492 struct prio_tree_iter iter;
2496 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2497 * from page cache lookup which is in HPAGE_SIZE units.
2499 address = address & huge_page_mask(h);
2500 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2502 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2505 * Take the mapping lock for the duration of the table walk. As
2506 * this mapping should be shared between all the VMAs,
2507 * __unmap_hugepage_range() is called as the lock is already held
2509 mutex_lock(&mapping->i_mmap_mutex);
2510 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2511 /* Do not unmap the current VMA */
2512 if (iter_vma == vma)
2516 * Shared VMAs have their own reserves and do not affect
2517 * MAP_PRIVATE accounting but it is possible that a shared
2518 * VMA is using the same page so check and skip such VMAs.
2520 if (iter_vma->vm_flags & VM_MAYSHARE)
2524 * Unmap the page from other VMAs without their own reserves.
2525 * They get marked to be SIGKILLed if they fault in these
2526 * areas. This is because a future no-page fault on this VMA
2527 * could insert a zeroed page instead of the data existing
2528 * from the time of fork. This would look like data corruption
2530 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2531 __unmap_hugepage_range(iter_vma,
2532 address, address + huge_page_size(h),
2535 mutex_unlock(&mapping->i_mmap_mutex);
2541 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2543 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2544 unsigned long address, pte_t *ptep, pte_t pte,
2545 struct page *pagecache_page)
2547 struct hstate *h = hstate_vma(vma);
2548 struct page *old_page, *new_page;
2550 int outside_reserve = 0;
2552 old_page = pte_page(pte);
2555 /* If no-one else is actually using this page, avoid the copy
2556 * and just make the page writable */
2557 avoidcopy = (page_mapcount(old_page) == 1);
2559 if (PageAnon(old_page))
2560 page_move_anon_rmap(old_page, vma, address);
2561 set_huge_ptep_writable(vma, address, ptep);
2566 * If the process that created a MAP_PRIVATE mapping is about to
2567 * perform a COW due to a shared page count, attempt to satisfy
2568 * the allocation without using the existing reserves. The pagecache
2569 * page is used to determine if the reserve at this address was
2570 * consumed or not. If reserves were used, a partial faulted mapping
2571 * at the time of fork() could consume its reserves on COW instead
2572 * of the full address range.
2574 if (!(vma->vm_flags & VM_MAYSHARE) &&
2575 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2576 old_page != pagecache_page)
2577 outside_reserve = 1;
2579 page_cache_get(old_page);
2581 /* Drop page_table_lock as buddy allocator may be called */
2582 spin_unlock(&mm->page_table_lock);
2583 new_page = alloc_huge_page(vma, address, outside_reserve);
2585 if (IS_ERR(new_page)) {
2586 page_cache_release(old_page);
2589 * If a process owning a MAP_PRIVATE mapping fails to COW,
2590 * it is due to references held by a child and an insufficient
2591 * huge page pool. To guarantee the original mappers
2592 * reliability, unmap the page from child processes. The child
2593 * may get SIGKILLed if it later faults.
2595 if (outside_reserve) {
2596 BUG_ON(huge_pte_none(pte));
2597 if (unmap_ref_private(mm, vma, old_page, address)) {
2598 BUG_ON(huge_pte_none(pte));
2599 spin_lock(&mm->page_table_lock);
2600 goto retry_avoidcopy;
2605 /* Caller expects lock to be held */
2606 spin_lock(&mm->page_table_lock);
2607 return -PTR_ERR(new_page);
2611 * When the original hugepage is shared one, it does not have
2612 * anon_vma prepared.
2614 if (unlikely(anon_vma_prepare(vma))) {
2615 page_cache_release(new_page);
2616 page_cache_release(old_page);
2617 /* Caller expects lock to be held */
2618 spin_lock(&mm->page_table_lock);
2619 return VM_FAULT_OOM;
2622 copy_user_huge_page(new_page, old_page, address, vma,
2623 pages_per_huge_page(h));
2624 __SetPageUptodate(new_page);
2627 * Retake the page_table_lock to check for racing updates
2628 * before the page tables are altered
2630 spin_lock(&mm->page_table_lock);
2631 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2632 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2634 mmu_notifier_invalidate_range_start(mm,
2635 address & huge_page_mask(h),
2636 (address & huge_page_mask(h)) + huge_page_size(h));
2637 huge_ptep_clear_flush(vma, address, ptep);
2638 set_huge_pte_at(mm, address, ptep,
2639 make_huge_pte(vma, new_page, 1));
2640 page_remove_rmap(old_page);
2641 hugepage_add_new_anon_rmap(new_page, vma, address);
2642 /* Make the old page be freed below */
2643 new_page = old_page;
2644 mmu_notifier_invalidate_range_end(mm,
2645 address & huge_page_mask(h),
2646 (address & huge_page_mask(h)) + huge_page_size(h));
2648 page_cache_release(new_page);
2649 page_cache_release(old_page);
2653 /* Return the pagecache page at a given address within a VMA */
2654 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2655 struct vm_area_struct *vma, unsigned long address)
2657 struct address_space *mapping;
2660 mapping = vma->vm_file->f_mapping;
2661 idx = vma_hugecache_offset(h, vma, address);
2663 return find_lock_page(mapping, idx);
2667 * Return whether there is a pagecache page to back given address within VMA.
2668 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2670 static bool hugetlbfs_pagecache_present(struct hstate *h,
2671 struct vm_area_struct *vma, unsigned long address)
2673 struct address_space *mapping;
2677 mapping = vma->vm_file->f_mapping;
2678 idx = vma_hugecache_offset(h, vma, address);
2680 page = find_get_page(mapping, idx);
2683 return page != NULL;
2686 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2687 unsigned long address, pte_t *ptep, unsigned int flags)
2689 struct hstate *h = hstate_vma(vma);
2690 int ret = VM_FAULT_SIGBUS;
2694 struct address_space *mapping;
2698 * Currently, we are forced to kill the process in the event the
2699 * original mapper has unmapped pages from the child due to a failed
2700 * COW. Warn that such a situation has occurred as it may not be obvious
2702 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2704 "PID %d killed due to inadequate hugepage pool\n",
2709 mapping = vma->vm_file->f_mapping;
2710 idx = vma_hugecache_offset(h, vma, address);
2713 * Use page lock to guard against racing truncation
2714 * before we get page_table_lock.
2717 page = find_lock_page(mapping, idx);
2719 size = i_size_read(mapping->host) >> huge_page_shift(h);
2722 page = alloc_huge_page(vma, address, 0);
2724 ret = -PTR_ERR(page);
2727 clear_huge_page(page, address, pages_per_huge_page(h));
2728 __SetPageUptodate(page);
2730 if (vma->vm_flags & VM_MAYSHARE) {
2732 struct inode *inode = mapping->host;
2734 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2742 spin_lock(&inode->i_lock);
2743 inode->i_blocks += blocks_per_huge_page(h);
2744 spin_unlock(&inode->i_lock);
2745 page_dup_rmap(page);
2748 if (unlikely(anon_vma_prepare(vma))) {
2750 goto backout_unlocked;
2752 hugepage_add_new_anon_rmap(page, vma, address);
2756 * If memory error occurs between mmap() and fault, some process
2757 * don't have hwpoisoned swap entry for errored virtual address.
2758 * So we need to block hugepage fault by PG_hwpoison bit check.
2760 if (unlikely(PageHWPoison(page))) {
2761 ret = VM_FAULT_HWPOISON |
2762 VM_FAULT_SET_HINDEX(h - hstates);
2763 goto backout_unlocked;
2765 page_dup_rmap(page);
2769 * If we are going to COW a private mapping later, we examine the
2770 * pending reservations for this page now. This will ensure that
2771 * any allocations necessary to record that reservation occur outside
2774 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2775 if (vma_needs_reservation(h, vma, address) < 0) {
2777 goto backout_unlocked;
2780 spin_lock(&mm->page_table_lock);
2781 size = i_size_read(mapping->host) >> huge_page_shift(h);
2786 if (!huge_pte_none(huge_ptep_get(ptep)))
2789 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2790 && (vma->vm_flags & VM_SHARED)));
2791 set_huge_pte_at(mm, address, ptep, new_pte);
2793 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2794 /* Optimization, do the COW without a second fault */
2795 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2798 spin_unlock(&mm->page_table_lock);
2804 spin_unlock(&mm->page_table_lock);
2811 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2812 unsigned long address, unsigned int flags)
2817 struct page *page = NULL;
2818 struct page *pagecache_page = NULL;
2819 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2820 struct hstate *h = hstate_vma(vma);
2821 int need_wait_lock = 0;
2823 ptep = huge_pte_offset(mm, address);
2825 entry = huge_ptep_get(ptep);
2826 if (unlikely(is_hugetlb_entry_migration(entry))) {
2827 migration_entry_wait_huge(mm, ptep);
2829 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2830 return VM_FAULT_HWPOISON_LARGE |
2831 VM_FAULT_SET_HINDEX(h - hstates);
2833 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2835 return VM_FAULT_OOM;
2839 * Serialize hugepage allocation and instantiation, so that we don't
2840 * get spurious allocation failures if two CPUs race to instantiate
2841 * the same page in the page cache.
2843 mutex_lock(&hugetlb_instantiation_mutex);
2844 entry = huge_ptep_get(ptep);
2845 if (huge_pte_none(entry)) {
2846 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2853 * entry could be a migration/hwpoison entry at this point, so this
2854 * check prevents the kernel from going below assuming that we have
2855 * a active hugepage in pagecache. This goto expects the 2nd page fault,
2856 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
2859 if (!pte_present(entry))
2863 * If we are going to COW the mapping later, we examine the pending
2864 * reservations for this page now. This will ensure that any
2865 * allocations necessary to record that reservation occur outside the
2866 * spinlock. For private mappings, we also lookup the pagecache
2867 * page now as it is used to determine if a reservation has been
2870 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2871 if (vma_needs_reservation(h, vma, address) < 0) {
2876 if (!(vma->vm_flags & VM_MAYSHARE))
2877 pagecache_page = hugetlbfs_pagecache_page(h,
2881 spin_lock(&mm->page_table_lock);
2882 /* Check for a racing update before calling hugetlb_cow */
2883 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2884 goto out_page_table_lock;
2887 * hugetlb_cow() requires page locks of pte_page(entry) and
2888 * pagecache_page, so here we need take the former one
2889 * when page != pagecache_page or !pagecache_page.
2891 page = pte_page(entry);
2892 if (page != pagecache_page)
2893 if (!trylock_page(page)) {
2895 goto out_page_table_lock;
2900 if (flags & FAULT_FLAG_WRITE) {
2901 if (!pte_write(entry)) {
2902 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2906 entry = pte_mkdirty(entry);
2908 entry = pte_mkyoung(entry);
2909 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2910 flags & FAULT_FLAG_WRITE))
2911 update_mmu_cache(vma, address, ptep);
2913 if (page != pagecache_page)
2916 out_page_table_lock:
2917 spin_unlock(&mm->page_table_lock);
2919 if (pagecache_page) {
2920 unlock_page(pagecache_page);
2921 put_page(pagecache_page);
2924 mutex_unlock(&hugetlb_instantiation_mutex);
2927 * Generally it's safe to hold refcount during waiting page lock. But
2928 * here we just wait to defer the next page fault to avoid busy loop and
2929 * the page is not used after unlocked before returning from the current
2930 * page fault. So we are safe from accessing freed page, even if we wait
2931 * here without taking refcount.
2934 wait_on_page_locked(page);
2938 /* Can be overriden by architectures */
2939 __attribute__((weak)) struct page *
2940 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2941 pud_t *pud, int write)
2947 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2948 struct page **pages, struct vm_area_struct **vmas,
2949 unsigned long *position, int *length, int i,
2952 unsigned long pfn_offset;
2953 unsigned long vaddr = *position;
2954 int remainder = *length;
2955 struct hstate *h = hstate_vma(vma);
2957 spin_lock(&mm->page_table_lock);
2958 while (vaddr < vma->vm_end && remainder) {
2964 * Some archs (sparc64, sh*) have multiple pte_ts to
2965 * each hugepage. We have to make sure we get the
2966 * first, for the page indexing below to work.
2968 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2969 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2972 * When coredumping, it suits get_dump_page if we just return
2973 * an error where there's an empty slot with no huge pagecache
2974 * to back it. This way, we avoid allocating a hugepage, and
2975 * the sparse dumpfile avoids allocating disk blocks, but its
2976 * huge holes still show up with zeroes where they need to be.
2978 if (absent && (flags & FOLL_DUMP) &&
2979 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2985 * We need call hugetlb_fault for both hugepages under migration
2986 * (in which case hugetlb_fault waits for the migration,) and
2987 * hwpoisoned hugepages (in which case we need to prevent the
2988 * caller from accessing to them.) In order to do this, we use
2989 * here is_swap_pte instead of is_hugetlb_entry_migration and
2990 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2991 * both cases, and because we can't follow correct pages
2992 * directly from any kind of swap entries.
2994 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2995 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2998 spin_unlock(&mm->page_table_lock);
2999 ret = hugetlb_fault(mm, vma, vaddr,
3000 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3001 spin_lock(&mm->page_table_lock);
3002 if (!(ret & VM_FAULT_ERROR))
3009 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3010 page = pte_page(huge_ptep_get(pte));
3013 pages[i] = mem_map_offset(page, pfn_offset);
3024 if (vaddr < vma->vm_end && remainder &&
3025 pfn_offset < pages_per_huge_page(h)) {
3027 * We use pfn_offset to avoid touching the pageframes
3028 * of this compound page.
3033 spin_unlock(&mm->page_table_lock);
3034 *length = remainder;
3037 return i ? i : -EFAULT;
3040 void hugetlb_change_protection(struct vm_area_struct *vma,
3041 unsigned long address, unsigned long end, pgprot_t newprot)
3043 struct mm_struct *mm = vma->vm_mm;
3044 unsigned long start = address;
3047 struct hstate *h = hstate_vma(vma);
3049 BUG_ON(address >= end);
3050 flush_cache_range(vma, address, end);
3052 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3053 spin_lock(&mm->page_table_lock);
3054 for (; address < end; address += huge_page_size(h)) {
3055 ptep = huge_pte_offset(mm, address);
3058 if (huge_pmd_unshare(mm, &address, ptep))
3060 pte = huge_ptep_get(ptep);
3061 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
3063 if (unlikely(is_hugetlb_entry_migration(pte))) {
3064 swp_entry_t entry = pte_to_swp_entry(pte);
3066 if (is_write_migration_entry(entry)) {
3069 make_migration_entry_read(&entry);
3070 newpte = swp_entry_to_pte(entry);
3071 set_huge_pte_at(mm, address, ptep, newpte);
3075 if (!huge_pte_none(pte)) {
3076 pte = huge_ptep_get_and_clear(mm, address, ptep);
3077 pte = pte_mkhuge(pte_modify(pte, newprot));
3078 set_huge_pte_at(mm, address, ptep, pte);
3081 spin_unlock(&mm->page_table_lock);
3083 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3084 * may have cleared our pud entry and done put_page on the page table:
3085 * once we release i_mmap_mutex, another task can do the final put_page
3086 * and that page table be reused and filled with junk.
3088 flush_tlb_range(vma, start, end);
3089 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3092 int hugetlb_reserve_pages(struct inode *inode,
3094 struct vm_area_struct *vma,
3095 vm_flags_t vm_flags)
3098 struct hstate *h = hstate_inode(inode);
3099 struct hugepage_subpool *spool = subpool_inode(inode);
3101 /* This should never happen */
3103 #ifdef CONFIG_DEBUG_VM
3104 WARN(1, "%s called with a negative range\n", __func__);
3110 * Only apply hugepage reservation if asked. At fault time, an
3111 * attempt will be made for VM_NORESERVE to allocate a page
3112 * without using reserves
3114 if (vm_flags & VM_NORESERVE)
3118 * Shared mappings base their reservation on the number of pages that
3119 * are already allocated on behalf of the file. Private mappings need
3120 * to reserve the full area even if read-only as mprotect() may be
3121 * called to make the mapping read-write. Assume !vma is a shm mapping
3123 if (!vma || vma->vm_flags & VM_MAYSHARE)
3124 chg = region_chg(&inode->i_mapping->private_list, from, to);
3126 struct resv_map *resv_map = resv_map_alloc();
3132 set_vma_resv_map(vma, resv_map);
3133 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3141 /* There must be enough pages in the subpool for the mapping */
3142 if (hugepage_subpool_get_pages(spool, chg)) {
3148 * Check enough hugepages are available for the reservation.
3149 * Hand the pages back to the subpool if there are not
3151 ret = hugetlb_acct_memory(h, chg);
3153 hugepage_subpool_put_pages(spool, chg);
3158 * Account for the reservations made. Shared mappings record regions
3159 * that have reservations as they are shared by multiple VMAs.
3160 * When the last VMA disappears, the region map says how much
3161 * the reservation was and the page cache tells how much of
3162 * the reservation was consumed. Private mappings are per-VMA and
3163 * only the consumed reservations are tracked. When the VMA
3164 * disappears, the original reservation is the VMA size and the
3165 * consumed reservations are stored in the map. Hence, nothing
3166 * else has to be done for private mappings here
3168 if (!vma || vma->vm_flags & VM_MAYSHARE)
3169 region_add(&inode->i_mapping->private_list, from, to);
3177 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3179 struct hstate *h = hstate_inode(inode);
3180 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3181 struct hugepage_subpool *spool = subpool_inode(inode);
3183 spin_lock(&inode->i_lock);
3184 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3185 spin_unlock(&inode->i_lock);
3187 hugepage_subpool_put_pages(spool, (chg - freed));
3188 hugetlb_acct_memory(h, -(chg - freed));
3191 #ifdef CONFIG_MEMORY_FAILURE
3193 /* Should be called in hugetlb_lock */
3194 static int is_hugepage_on_freelist(struct page *hpage)
3198 struct hstate *h = page_hstate(hpage);
3199 int nid = page_to_nid(hpage);
3201 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3208 * This function is called from memory failure code.
3209 * Assume the caller holds page lock of the head page.
3211 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3213 struct hstate *h = page_hstate(hpage);
3214 int nid = page_to_nid(hpage);
3217 spin_lock(&hugetlb_lock);
3218 if (is_hugepage_on_freelist(hpage)) {
3219 list_del(&hpage->lru);
3220 set_page_refcounted(hpage);
3221 h->free_huge_pages--;
3222 h->free_huge_pages_node[nid]--;
3225 spin_unlock(&hugetlb_lock);