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);
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);
1421 ret = alloc_fresh_huge_page(h, nodes_allowed);
1422 spin_lock(&hugetlb_lock);
1426 /* Bail for signals. Probably ctrl-c from user */
1427 if (signal_pending(current))
1432 * Decrease the pool size
1433 * First return free pages to the buddy allocator (being careful
1434 * to keep enough around to satisfy reservations). Then place
1435 * pages into surplus state as needed so the pool will shrink
1436 * to the desired size as pages become free.
1438 * By placing pages into the surplus state independent of the
1439 * overcommit value, we are allowing the surplus pool size to
1440 * exceed overcommit. There are few sane options here. Since
1441 * alloc_buddy_huge_page() is checking the global counter,
1442 * though, we'll note that we're not allowed to exceed surplus
1443 * and won't grow the pool anywhere else. Not until one of the
1444 * sysctls are changed, or the surplus pages go out of use.
1446 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1447 min_count = max(count, min_count);
1448 try_to_free_low(h, min_count, nodes_allowed);
1449 while (min_count < persistent_huge_pages(h)) {
1450 if (!free_pool_huge_page(h, nodes_allowed, 0))
1452 cond_resched_lock(&hugetlb_lock);
1454 while (count < persistent_huge_pages(h)) {
1455 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1459 ret = persistent_huge_pages(h);
1460 spin_unlock(&hugetlb_lock);
1464 #define HSTATE_ATTR_RO(_name) \
1465 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1467 #define HSTATE_ATTR(_name) \
1468 static struct kobj_attribute _name##_attr = \
1469 __ATTR(_name, 0644, _name##_show, _name##_store)
1471 static struct kobject *hugepages_kobj;
1472 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1474 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1476 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1480 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1481 if (hstate_kobjs[i] == kobj) {
1483 *nidp = NUMA_NO_NODE;
1487 return kobj_to_node_hstate(kobj, nidp);
1490 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1491 struct kobj_attribute *attr, char *buf)
1494 unsigned long nr_huge_pages;
1497 h = kobj_to_hstate(kobj, &nid);
1498 if (nid == NUMA_NO_NODE)
1499 nr_huge_pages = h->nr_huge_pages;
1501 nr_huge_pages = h->nr_huge_pages_node[nid];
1503 return sprintf(buf, "%lu\n", nr_huge_pages);
1506 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1507 struct kobject *kobj, struct kobj_attribute *attr,
1508 const char *buf, size_t len)
1512 unsigned long count;
1514 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1516 err = strict_strtoul(buf, 10, &count);
1520 h = kobj_to_hstate(kobj, &nid);
1521 if (h->order >= MAX_ORDER) {
1526 if (nid == NUMA_NO_NODE) {
1528 * global hstate attribute
1530 if (!(obey_mempolicy &&
1531 init_nodemask_of_mempolicy(nodes_allowed))) {
1532 NODEMASK_FREE(nodes_allowed);
1533 nodes_allowed = &node_states[N_HIGH_MEMORY];
1535 } else if (nodes_allowed) {
1537 * per node hstate attribute: adjust count to global,
1538 * but restrict alloc/free to the specified node.
1540 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1541 init_nodemask_of_node(nodes_allowed, nid);
1543 nodes_allowed = &node_states[N_HIGH_MEMORY];
1545 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1547 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1548 NODEMASK_FREE(nodes_allowed);
1552 NODEMASK_FREE(nodes_allowed);
1556 static ssize_t nr_hugepages_show(struct kobject *kobj,
1557 struct kobj_attribute *attr, char *buf)
1559 return nr_hugepages_show_common(kobj, attr, buf);
1562 static ssize_t nr_hugepages_store(struct kobject *kobj,
1563 struct kobj_attribute *attr, const char *buf, size_t len)
1565 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1567 HSTATE_ATTR(nr_hugepages);
1572 * hstate attribute for optionally mempolicy-based constraint on persistent
1573 * huge page alloc/free.
1575 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1576 struct kobj_attribute *attr, char *buf)
1578 return nr_hugepages_show_common(kobj, attr, buf);
1581 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1582 struct kobj_attribute *attr, const char *buf, size_t len)
1584 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1586 HSTATE_ATTR(nr_hugepages_mempolicy);
1590 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1591 struct kobj_attribute *attr, char *buf)
1593 struct hstate *h = kobj_to_hstate(kobj, NULL);
1594 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1597 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1598 struct kobj_attribute *attr, const char *buf, size_t count)
1601 unsigned long input;
1602 struct hstate *h = kobj_to_hstate(kobj, NULL);
1604 if (h->order >= MAX_ORDER)
1607 err = strict_strtoul(buf, 10, &input);
1611 spin_lock(&hugetlb_lock);
1612 h->nr_overcommit_huge_pages = input;
1613 spin_unlock(&hugetlb_lock);
1617 HSTATE_ATTR(nr_overcommit_hugepages);
1619 static ssize_t free_hugepages_show(struct kobject *kobj,
1620 struct kobj_attribute *attr, char *buf)
1623 unsigned long free_huge_pages;
1626 h = kobj_to_hstate(kobj, &nid);
1627 if (nid == NUMA_NO_NODE)
1628 free_huge_pages = h->free_huge_pages;
1630 free_huge_pages = h->free_huge_pages_node[nid];
1632 return sprintf(buf, "%lu\n", free_huge_pages);
1634 HSTATE_ATTR_RO(free_hugepages);
1636 static ssize_t resv_hugepages_show(struct kobject *kobj,
1637 struct kobj_attribute *attr, char *buf)
1639 struct hstate *h = kobj_to_hstate(kobj, NULL);
1640 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1642 HSTATE_ATTR_RO(resv_hugepages);
1644 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1645 struct kobj_attribute *attr, char *buf)
1648 unsigned long surplus_huge_pages;
1651 h = kobj_to_hstate(kobj, &nid);
1652 if (nid == NUMA_NO_NODE)
1653 surplus_huge_pages = h->surplus_huge_pages;
1655 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1657 return sprintf(buf, "%lu\n", surplus_huge_pages);
1659 HSTATE_ATTR_RO(surplus_hugepages);
1661 static struct attribute *hstate_attrs[] = {
1662 &nr_hugepages_attr.attr,
1663 &nr_overcommit_hugepages_attr.attr,
1664 &free_hugepages_attr.attr,
1665 &resv_hugepages_attr.attr,
1666 &surplus_hugepages_attr.attr,
1668 &nr_hugepages_mempolicy_attr.attr,
1673 static struct attribute_group hstate_attr_group = {
1674 .attrs = hstate_attrs,
1677 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1678 struct kobject **hstate_kobjs,
1679 struct attribute_group *hstate_attr_group)
1682 int hi = h - hstates;
1684 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1685 if (!hstate_kobjs[hi])
1688 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1690 kobject_put(hstate_kobjs[hi]);
1695 static void __init hugetlb_sysfs_init(void)
1700 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1701 if (!hugepages_kobj)
1704 for_each_hstate(h) {
1705 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1706 hstate_kobjs, &hstate_attr_group);
1708 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1716 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1717 * with node devices in node_devices[] using a parallel array. The array
1718 * index of a node device or _hstate == node id.
1719 * This is here to avoid any static dependency of the node device driver, in
1720 * the base kernel, on the hugetlb module.
1722 struct node_hstate {
1723 struct kobject *hugepages_kobj;
1724 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1726 struct node_hstate node_hstates[MAX_NUMNODES];
1729 * A subset of global hstate attributes for node devices
1731 static struct attribute *per_node_hstate_attrs[] = {
1732 &nr_hugepages_attr.attr,
1733 &free_hugepages_attr.attr,
1734 &surplus_hugepages_attr.attr,
1738 static struct attribute_group per_node_hstate_attr_group = {
1739 .attrs = per_node_hstate_attrs,
1743 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1744 * Returns node id via non-NULL nidp.
1746 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1750 for (nid = 0; nid < nr_node_ids; nid++) {
1751 struct node_hstate *nhs = &node_hstates[nid];
1753 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1754 if (nhs->hstate_kobjs[i] == kobj) {
1766 * Unregister hstate attributes from a single node device.
1767 * No-op if no hstate attributes attached.
1769 void hugetlb_unregister_node(struct node *node)
1772 struct node_hstate *nhs = &node_hstates[node->dev.id];
1774 if (!nhs->hugepages_kobj)
1775 return; /* no hstate attributes */
1778 if (nhs->hstate_kobjs[h - hstates]) {
1779 kobject_put(nhs->hstate_kobjs[h - hstates]);
1780 nhs->hstate_kobjs[h - hstates] = NULL;
1783 kobject_put(nhs->hugepages_kobj);
1784 nhs->hugepages_kobj = NULL;
1788 * hugetlb module exit: unregister hstate attributes from node devices
1791 static void hugetlb_unregister_all_nodes(void)
1796 * disable node device registrations.
1798 register_hugetlbfs_with_node(NULL, NULL);
1801 * remove hstate attributes from any nodes that have them.
1803 for (nid = 0; nid < nr_node_ids; nid++)
1804 hugetlb_unregister_node(&node_devices[nid]);
1808 * Register hstate attributes for a single node device.
1809 * No-op if attributes already registered.
1811 void hugetlb_register_node(struct node *node)
1814 struct node_hstate *nhs = &node_hstates[node->dev.id];
1817 if (nhs->hugepages_kobj)
1818 return; /* already allocated */
1820 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1822 if (!nhs->hugepages_kobj)
1825 for_each_hstate(h) {
1826 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1828 &per_node_hstate_attr_group);
1830 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1832 h->name, node->dev.id);
1833 hugetlb_unregister_node(node);
1840 * hugetlb init time: register hstate attributes for all registered node
1841 * devices of nodes that have memory. All on-line nodes should have
1842 * registered their associated device by this time.
1844 static void hugetlb_register_all_nodes(void)
1848 for_each_node_state(nid, N_HIGH_MEMORY) {
1849 struct node *node = &node_devices[nid];
1850 if (node->dev.id == nid)
1851 hugetlb_register_node(node);
1855 * Let the node device driver know we're here so it can
1856 * [un]register hstate attributes on node hotplug.
1858 register_hugetlbfs_with_node(hugetlb_register_node,
1859 hugetlb_unregister_node);
1861 #else /* !CONFIG_NUMA */
1863 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1871 static void hugetlb_unregister_all_nodes(void) { }
1873 static void hugetlb_register_all_nodes(void) { }
1877 static void __exit hugetlb_exit(void)
1881 hugetlb_unregister_all_nodes();
1883 for_each_hstate(h) {
1884 kobject_put(hstate_kobjs[h - hstates]);
1887 kobject_put(hugepages_kobj);
1889 module_exit(hugetlb_exit);
1891 static int __init hugetlb_init(void)
1893 /* Some platform decide whether they support huge pages at boot
1894 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1895 * there is no such support
1897 if (HPAGE_SHIFT == 0)
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 tmp = h->max_huge_pages;
2016 if (write && h->order >= MAX_ORDER)
2020 table->maxlen = sizeof(unsigned long);
2021 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2026 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2027 GFP_KERNEL | __GFP_NORETRY);
2028 if (!(obey_mempolicy &&
2029 init_nodemask_of_mempolicy(nodes_allowed))) {
2030 NODEMASK_FREE(nodes_allowed);
2031 nodes_allowed = &node_states[N_HIGH_MEMORY];
2033 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2035 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2036 NODEMASK_FREE(nodes_allowed);
2042 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2043 void __user *buffer, size_t *length, loff_t *ppos)
2046 return hugetlb_sysctl_handler_common(false, table, write,
2047 buffer, length, ppos);
2051 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2052 void __user *buffer, size_t *length, loff_t *ppos)
2054 return hugetlb_sysctl_handler_common(true, table, write,
2055 buffer, length, ppos);
2057 #endif /* CONFIG_NUMA */
2059 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2060 void __user *buffer,
2061 size_t *length, loff_t *ppos)
2063 proc_dointvec(table, write, buffer, length, ppos);
2064 if (hugepages_treat_as_movable)
2065 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2067 htlb_alloc_mask = GFP_HIGHUSER;
2071 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2072 void __user *buffer,
2073 size_t *length, loff_t *ppos)
2075 struct hstate *h = &default_hstate;
2079 tmp = h->nr_overcommit_huge_pages;
2081 if (write && h->order >= MAX_ORDER)
2085 table->maxlen = sizeof(unsigned long);
2086 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2091 spin_lock(&hugetlb_lock);
2092 h->nr_overcommit_huge_pages = tmp;
2093 spin_unlock(&hugetlb_lock);
2099 #endif /* CONFIG_SYSCTL */
2101 void hugetlb_report_meminfo(struct seq_file *m)
2103 struct hstate *h = &default_hstate;
2105 "HugePages_Total: %5lu\n"
2106 "HugePages_Free: %5lu\n"
2107 "HugePages_Rsvd: %5lu\n"
2108 "HugePages_Surp: %5lu\n"
2109 "Hugepagesize: %8lu kB\n",
2113 h->surplus_huge_pages,
2114 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2117 int hugetlb_report_node_meminfo(int nid, char *buf)
2119 struct hstate *h = &default_hstate;
2121 "Node %d HugePages_Total: %5u\n"
2122 "Node %d HugePages_Free: %5u\n"
2123 "Node %d HugePages_Surp: %5u\n",
2124 nid, h->nr_huge_pages_node[nid],
2125 nid, h->free_huge_pages_node[nid],
2126 nid, h->surplus_huge_pages_node[nid]);
2129 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2130 unsigned long hugetlb_total_pages(void)
2133 unsigned long nr_total_pages = 0;
2136 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2137 return nr_total_pages;
2140 static int hugetlb_acct_memory(struct hstate *h, long delta)
2144 spin_lock(&hugetlb_lock);
2146 * When cpuset is configured, it breaks the strict hugetlb page
2147 * reservation as the accounting is done on a global variable. Such
2148 * reservation is completely rubbish in the presence of cpuset because
2149 * the reservation is not checked against page availability for the
2150 * current cpuset. Application can still potentially OOM'ed by kernel
2151 * with lack of free htlb page in cpuset that the task is in.
2152 * Attempt to enforce strict accounting with cpuset is almost
2153 * impossible (or too ugly) because cpuset is too fluid that
2154 * task or memory node can be dynamically moved between cpusets.
2156 * The change of semantics for shared hugetlb mapping with cpuset is
2157 * undesirable. However, in order to preserve some of the semantics,
2158 * we fall back to check against current free page availability as
2159 * a best attempt and hopefully to minimize the impact of changing
2160 * semantics that cpuset has.
2163 if (gather_surplus_pages(h, delta) < 0)
2166 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2167 return_unused_surplus_pages(h, delta);
2174 return_unused_surplus_pages(h, (unsigned long) -delta);
2177 spin_unlock(&hugetlb_lock);
2181 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2183 struct resv_map *reservations = vma_resv_map(vma);
2186 * This new VMA should share its siblings reservation map if present.
2187 * The VMA will only ever have a valid reservation map pointer where
2188 * it is being copied for another still existing VMA. As that VMA
2189 * has a reference to the reservation map it cannot disappear until
2190 * after this open call completes. It is therefore safe to take a
2191 * new reference here without additional locking.
2194 kref_get(&reservations->refs);
2197 static void resv_map_put(struct vm_area_struct *vma)
2199 struct resv_map *reservations = vma_resv_map(vma);
2203 kref_put(&reservations->refs, resv_map_release);
2206 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2208 struct hstate *h = hstate_vma(vma);
2209 struct resv_map *reservations = vma_resv_map(vma);
2210 struct hugepage_subpool *spool = subpool_vma(vma);
2211 unsigned long reserve;
2212 unsigned long start;
2216 start = vma_hugecache_offset(h, vma, vma->vm_start);
2217 end = vma_hugecache_offset(h, vma, vma->vm_end);
2219 reserve = (end - start) -
2220 region_count(&reservations->regions, start, end);
2225 hugetlb_acct_memory(h, -reserve);
2226 hugepage_subpool_put_pages(spool, reserve);
2232 * We cannot handle pagefaults against hugetlb pages at all. They cause
2233 * handle_mm_fault() to try to instantiate regular-sized pages in the
2234 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2237 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2243 const struct vm_operations_struct hugetlb_vm_ops = {
2244 .fault = hugetlb_vm_op_fault,
2245 .open = hugetlb_vm_op_open,
2246 .close = hugetlb_vm_op_close,
2249 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2256 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2258 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2260 entry = pte_mkyoung(entry);
2261 entry = pte_mkhuge(entry);
2266 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2267 unsigned long address, pte_t *ptep)
2271 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2272 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2273 update_mmu_cache(vma, address, ptep);
2276 static int is_hugetlb_entry_migration(pte_t pte)
2280 if (huge_pte_none(pte) || pte_present(pte))
2282 swp = pte_to_swp_entry(pte);
2283 if (non_swap_entry(swp) && is_migration_entry(swp))
2289 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2293 if (huge_pte_none(pte) || pte_present(pte))
2295 swp = pte_to_swp_entry(pte);
2296 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2302 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2303 struct vm_area_struct *vma)
2305 pte_t *src_pte, *dst_pte, entry;
2306 struct page *ptepage;
2309 struct hstate *h = hstate_vma(vma);
2310 unsigned long sz = huge_page_size(h);
2312 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2314 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2315 src_pte = huge_pte_offset(src, addr);
2318 dst_pte = huge_pte_alloc(dst, addr, sz);
2322 /* If the pagetables are shared don't copy or take references */
2323 if (dst_pte == src_pte)
2326 spin_lock(&dst->page_table_lock);
2327 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2328 entry = huge_ptep_get(src_pte);
2329 if (huge_pte_none(entry)) { /* skip none entry */
2331 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2332 is_hugetlb_entry_hwpoisoned(entry))) {
2333 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2335 if (is_write_migration_entry(swp_entry) && cow) {
2337 * COW mappings require pages in both
2338 * parent and child to be set to read.
2340 make_migration_entry_read(&swp_entry);
2341 entry = swp_entry_to_pte(swp_entry);
2342 set_huge_pte_at(src, addr, src_pte, entry);
2344 set_huge_pte_at(dst, addr, dst_pte, entry);
2347 huge_ptep_set_wrprotect(src, addr, src_pte);
2348 entry = huge_ptep_get(src_pte);
2349 ptepage = pte_page(entry);
2351 page_dup_rmap(ptepage);
2352 set_huge_pte_at(dst, addr, dst_pte, entry);
2354 spin_unlock(&src->page_table_lock);
2355 spin_unlock(&dst->page_table_lock);
2363 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2364 unsigned long end, struct page *ref_page)
2366 struct mm_struct *mm = vma->vm_mm;
2367 unsigned long address;
2372 struct hstate *h = hstate_vma(vma);
2373 unsigned long sz = huge_page_size(h);
2376 * A page gathering list, protected by per file i_mmap_mutex. The
2377 * lock is used to avoid list corruption from multiple unmapping
2378 * of the same page since we are using page->lru.
2380 LIST_HEAD(page_list);
2382 WARN_ON(!is_vm_hugetlb_page(vma));
2383 BUG_ON(start & ~huge_page_mask(h));
2384 BUG_ON(end & ~huge_page_mask(h));
2386 mmu_notifier_invalidate_range_start(mm, start, end);
2387 spin_lock(&mm->page_table_lock);
2388 for (address = start; address < end; address += sz) {
2389 ptep = huge_pte_offset(mm, address);
2393 if (huge_pmd_unshare(mm, &address, ptep))
2397 * If a reference page is supplied, it is because a specific
2398 * page is being unmapped, not a range. Ensure the page we
2399 * are about to unmap is the actual page of interest.
2402 pte = huge_ptep_get(ptep);
2403 if (huge_pte_none(pte))
2405 page = pte_page(pte);
2406 if (page != ref_page)
2410 * Mark the VMA as having unmapped its page so that
2411 * future faults in this VMA will fail rather than
2412 * looking like data was lost
2414 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2417 pte = huge_ptep_get_and_clear(mm, address, ptep);
2418 if (huge_pte_none(pte))
2422 * Migrating hugepage or HWPoisoned hugepage is already
2423 * unmapped and its refcount is dropped
2425 if (unlikely(!pte_present(pte)))
2428 page = pte_page(pte);
2430 set_page_dirty(page);
2431 list_add(&page->lru, &page_list);
2433 spin_unlock(&mm->page_table_lock);
2434 flush_tlb_range(vma, start, end);
2435 mmu_notifier_invalidate_range_end(mm, start, end);
2436 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2437 page_remove_rmap(page);
2438 list_del(&page->lru);
2443 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2444 unsigned long start, unsigned long end,
2445 struct page *ref_page)
2447 __unmap_hugepage_range(vma, start, end, ref_page);
2450 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2451 * test will fail on a vma being torn down, and not grab a page table
2452 * on its way out. We're lucky that the flag has such an appropriate
2453 * name, and can in fact be safely cleared here. We could clear it
2454 * before the __unmap_hugepage_range above, but all that's necessary
2455 * is to clear it before releasing the i_mmap_mutex. This works
2456 * because in the context this is called, the VMA is about to be
2457 * destroyed and the i_mmap_mutex is held.
2459 vma->vm_flags &= ~VM_MAYSHARE;
2462 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2463 unsigned long end, struct page *ref_page)
2465 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2466 __unmap_hugepage_range(vma, start, end, ref_page);
2467 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2471 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2472 * mappping it owns the reserve page for. The intention is to unmap the page
2473 * from other VMAs and let the children be SIGKILLed if they are faulting the
2476 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2477 struct page *page, unsigned long address)
2479 struct hstate *h = hstate_vma(vma);
2480 struct vm_area_struct *iter_vma;
2481 struct address_space *mapping;
2482 struct prio_tree_iter iter;
2486 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2487 * from page cache lookup which is in HPAGE_SIZE units.
2489 address = address & huge_page_mask(h);
2490 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2492 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2495 * Take the mapping lock for the duration of the table walk. As
2496 * this mapping should be shared between all the VMAs,
2497 * __unmap_hugepage_range() is called as the lock is already held
2499 mutex_lock(&mapping->i_mmap_mutex);
2500 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2501 /* Do not unmap the current VMA */
2502 if (iter_vma == vma)
2506 * Shared VMAs have their own reserves and do not affect
2507 * MAP_PRIVATE accounting but it is possible that a shared
2508 * VMA is using the same page so check and skip such VMAs.
2510 if (iter_vma->vm_flags & VM_MAYSHARE)
2514 * Unmap the page from other VMAs without their own reserves.
2515 * They get marked to be SIGKILLed if they fault in these
2516 * areas. This is because a future no-page fault on this VMA
2517 * could insert a zeroed page instead of the data existing
2518 * from the time of fork. This would look like data corruption
2520 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2521 __unmap_hugepage_range(iter_vma,
2522 address, address + huge_page_size(h),
2525 mutex_unlock(&mapping->i_mmap_mutex);
2531 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2533 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2534 unsigned long address, pte_t *ptep, pte_t pte,
2535 struct page *pagecache_page)
2537 struct hstate *h = hstate_vma(vma);
2538 struct page *old_page, *new_page;
2540 int outside_reserve = 0;
2542 old_page = pte_page(pte);
2545 /* If no-one else is actually using this page, avoid the copy
2546 * and just make the page writable */
2547 avoidcopy = (page_mapcount(old_page) == 1);
2549 if (PageAnon(old_page))
2550 page_move_anon_rmap(old_page, vma, address);
2551 set_huge_ptep_writable(vma, address, ptep);
2556 * If the process that created a MAP_PRIVATE mapping is about to
2557 * perform a COW due to a shared page count, attempt to satisfy
2558 * the allocation without using the existing reserves. The pagecache
2559 * page is used to determine if the reserve at this address was
2560 * consumed or not. If reserves were used, a partial faulted mapping
2561 * at the time of fork() could consume its reserves on COW instead
2562 * of the full address range.
2564 if (!(vma->vm_flags & VM_MAYSHARE) &&
2565 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2566 old_page != pagecache_page)
2567 outside_reserve = 1;
2569 page_cache_get(old_page);
2571 /* Drop page_table_lock as buddy allocator may be called */
2572 spin_unlock(&mm->page_table_lock);
2573 new_page = alloc_huge_page(vma, address, outside_reserve);
2575 if (IS_ERR(new_page)) {
2576 page_cache_release(old_page);
2579 * If a process owning a MAP_PRIVATE mapping fails to COW,
2580 * it is due to references held by a child and an insufficient
2581 * huge page pool. To guarantee the original mappers
2582 * reliability, unmap the page from child processes. The child
2583 * may get SIGKILLed if it later faults.
2585 if (outside_reserve) {
2586 BUG_ON(huge_pte_none(pte));
2587 if (unmap_ref_private(mm, vma, old_page, address)) {
2588 BUG_ON(huge_pte_none(pte));
2589 spin_lock(&mm->page_table_lock);
2590 goto retry_avoidcopy;
2595 /* Caller expects lock to be held */
2596 spin_lock(&mm->page_table_lock);
2597 return -PTR_ERR(new_page);
2601 * When the original hugepage is shared one, it does not have
2602 * anon_vma prepared.
2604 if (unlikely(anon_vma_prepare(vma))) {
2605 page_cache_release(new_page);
2606 page_cache_release(old_page);
2607 /* Caller expects lock to be held */
2608 spin_lock(&mm->page_table_lock);
2609 return VM_FAULT_OOM;
2612 copy_user_huge_page(new_page, old_page, address, vma,
2613 pages_per_huge_page(h));
2614 __SetPageUptodate(new_page);
2617 * Retake the page_table_lock to check for racing updates
2618 * before the page tables are altered
2620 spin_lock(&mm->page_table_lock);
2621 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2622 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2624 mmu_notifier_invalidate_range_start(mm,
2625 address & huge_page_mask(h),
2626 (address & huge_page_mask(h)) + huge_page_size(h));
2627 huge_ptep_clear_flush(vma, address, ptep);
2628 set_huge_pte_at(mm, address, ptep,
2629 make_huge_pte(vma, new_page, 1));
2630 page_remove_rmap(old_page);
2631 hugepage_add_new_anon_rmap(new_page, vma, address);
2632 /* Make the old page be freed below */
2633 new_page = old_page;
2634 mmu_notifier_invalidate_range_end(mm,
2635 address & huge_page_mask(h),
2636 (address & huge_page_mask(h)) + huge_page_size(h));
2638 page_cache_release(new_page);
2639 page_cache_release(old_page);
2643 /* Return the pagecache page at a given address within a VMA */
2644 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2645 struct vm_area_struct *vma, unsigned long address)
2647 struct address_space *mapping;
2650 mapping = vma->vm_file->f_mapping;
2651 idx = vma_hugecache_offset(h, vma, address);
2653 return find_lock_page(mapping, idx);
2657 * Return whether there is a pagecache page to back given address within VMA.
2658 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2660 static bool hugetlbfs_pagecache_present(struct hstate *h,
2661 struct vm_area_struct *vma, unsigned long address)
2663 struct address_space *mapping;
2667 mapping = vma->vm_file->f_mapping;
2668 idx = vma_hugecache_offset(h, vma, address);
2670 page = find_get_page(mapping, idx);
2673 return page != NULL;
2676 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2677 unsigned long address, pte_t *ptep, unsigned int flags)
2679 struct hstate *h = hstate_vma(vma);
2680 int ret = VM_FAULT_SIGBUS;
2684 struct address_space *mapping;
2688 * Currently, we are forced to kill the process in the event the
2689 * original mapper has unmapped pages from the child due to a failed
2690 * COW. Warn that such a situation has occurred as it may not be obvious
2692 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2694 "PID %d killed due to inadequate hugepage pool\n",
2699 mapping = vma->vm_file->f_mapping;
2700 idx = vma_hugecache_offset(h, vma, address);
2703 * Use page lock to guard against racing truncation
2704 * before we get page_table_lock.
2707 page = find_lock_page(mapping, idx);
2709 size = i_size_read(mapping->host) >> huge_page_shift(h);
2712 page = alloc_huge_page(vma, address, 0);
2714 ret = -PTR_ERR(page);
2717 clear_huge_page(page, address, pages_per_huge_page(h));
2718 __SetPageUptodate(page);
2720 if (vma->vm_flags & VM_MAYSHARE) {
2722 struct inode *inode = mapping->host;
2724 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2732 spin_lock(&inode->i_lock);
2733 inode->i_blocks += blocks_per_huge_page(h);
2734 spin_unlock(&inode->i_lock);
2735 page_dup_rmap(page);
2738 if (unlikely(anon_vma_prepare(vma))) {
2740 goto backout_unlocked;
2742 hugepage_add_new_anon_rmap(page, vma, address);
2746 * If memory error occurs between mmap() and fault, some process
2747 * don't have hwpoisoned swap entry for errored virtual address.
2748 * So we need to block hugepage fault by PG_hwpoison bit check.
2750 if (unlikely(PageHWPoison(page))) {
2751 ret = VM_FAULT_HWPOISON |
2752 VM_FAULT_SET_HINDEX(h - hstates);
2753 goto backout_unlocked;
2755 page_dup_rmap(page);
2759 * If we are going to COW a private mapping later, we examine the
2760 * pending reservations for this page now. This will ensure that
2761 * any allocations necessary to record that reservation occur outside
2764 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2765 if (vma_needs_reservation(h, vma, address) < 0) {
2767 goto backout_unlocked;
2770 spin_lock(&mm->page_table_lock);
2771 size = i_size_read(mapping->host) >> huge_page_shift(h);
2776 if (!huge_pte_none(huge_ptep_get(ptep)))
2779 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2780 && (vma->vm_flags & VM_SHARED)));
2781 set_huge_pte_at(mm, address, ptep, new_pte);
2783 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2784 /* Optimization, do the COW without a second fault */
2785 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2788 spin_unlock(&mm->page_table_lock);
2794 spin_unlock(&mm->page_table_lock);
2801 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2802 unsigned long address, unsigned int flags)
2807 struct page *page = NULL;
2808 struct page *pagecache_page = NULL;
2809 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2810 struct hstate *h = hstate_vma(vma);
2811 int need_wait_lock = 0;
2813 ptep = huge_pte_offset(mm, address);
2815 entry = huge_ptep_get(ptep);
2816 if (unlikely(is_hugetlb_entry_migration(entry))) {
2817 migration_entry_wait_huge(mm, ptep);
2819 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2820 return VM_FAULT_HWPOISON_LARGE |
2821 VM_FAULT_SET_HINDEX(h - hstates);
2823 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2825 return VM_FAULT_OOM;
2829 * Serialize hugepage allocation and instantiation, so that we don't
2830 * get spurious allocation failures if two CPUs race to instantiate
2831 * the same page in the page cache.
2833 mutex_lock(&hugetlb_instantiation_mutex);
2834 entry = huge_ptep_get(ptep);
2835 if (huge_pte_none(entry)) {
2836 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2843 * entry could be a migration/hwpoison entry at this point, so this
2844 * check prevents the kernel from going below assuming that we have
2845 * a active hugepage in pagecache. This goto expects the 2nd page fault,
2846 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
2849 if (!pte_present(entry))
2853 * If we are going to COW the mapping later, we examine the pending
2854 * reservations for this page now. This will ensure that any
2855 * allocations necessary to record that reservation occur outside the
2856 * spinlock. For private mappings, we also lookup the pagecache
2857 * page now as it is used to determine if a reservation has been
2860 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2861 if (vma_needs_reservation(h, vma, address) < 0) {
2866 if (!(vma->vm_flags & VM_MAYSHARE))
2867 pagecache_page = hugetlbfs_pagecache_page(h,
2871 spin_lock(&mm->page_table_lock);
2872 /* Check for a racing update before calling hugetlb_cow */
2873 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2874 goto out_page_table_lock;
2877 * hugetlb_cow() requires page locks of pte_page(entry) and
2878 * pagecache_page, so here we need take the former one
2879 * when page != pagecache_page or !pagecache_page.
2881 page = pte_page(entry);
2882 if (page != pagecache_page)
2883 if (!trylock_page(page)) {
2885 goto out_page_table_lock;
2890 if (flags & FAULT_FLAG_WRITE) {
2891 if (!pte_write(entry)) {
2892 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2896 entry = pte_mkdirty(entry);
2898 entry = pte_mkyoung(entry);
2899 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2900 flags & FAULT_FLAG_WRITE))
2901 update_mmu_cache(vma, address, ptep);
2903 if (page != pagecache_page)
2906 out_page_table_lock:
2907 spin_unlock(&mm->page_table_lock);
2909 if (pagecache_page) {
2910 unlock_page(pagecache_page);
2911 put_page(pagecache_page);
2914 mutex_unlock(&hugetlb_instantiation_mutex);
2917 * Generally it's safe to hold refcount during waiting page lock. But
2918 * here we just wait to defer the next page fault to avoid busy loop and
2919 * the page is not used after unlocked before returning from the current
2920 * page fault. So we are safe from accessing freed page, even if we wait
2921 * here without taking refcount.
2924 wait_on_page_locked(page);
2928 /* Can be overriden by architectures */
2929 __attribute__((weak)) struct page *
2930 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2931 pud_t *pud, int write)
2937 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2938 struct page **pages, struct vm_area_struct **vmas,
2939 unsigned long *position, int *length, int i,
2942 unsigned long pfn_offset;
2943 unsigned long vaddr = *position;
2944 int remainder = *length;
2945 struct hstate *h = hstate_vma(vma);
2947 spin_lock(&mm->page_table_lock);
2948 while (vaddr < vma->vm_end && remainder) {
2954 * Some archs (sparc64, sh*) have multiple pte_ts to
2955 * each hugepage. We have to make sure we get the
2956 * first, for the page indexing below to work.
2958 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2959 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2962 * When coredumping, it suits get_dump_page if we just return
2963 * an error where there's an empty slot with no huge pagecache
2964 * to back it. This way, we avoid allocating a hugepage, and
2965 * the sparse dumpfile avoids allocating disk blocks, but its
2966 * huge holes still show up with zeroes where they need to be.
2968 if (absent && (flags & FOLL_DUMP) &&
2969 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2975 * We need call hugetlb_fault for both hugepages under migration
2976 * (in which case hugetlb_fault waits for the migration,) and
2977 * hwpoisoned hugepages (in which case we need to prevent the
2978 * caller from accessing to them.) In order to do this, we use
2979 * here is_swap_pte instead of is_hugetlb_entry_migration and
2980 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2981 * both cases, and because we can't follow correct pages
2982 * directly from any kind of swap entries.
2984 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2985 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2988 spin_unlock(&mm->page_table_lock);
2989 ret = hugetlb_fault(mm, vma, vaddr,
2990 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2991 spin_lock(&mm->page_table_lock);
2992 if (!(ret & VM_FAULT_ERROR))
2999 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3000 page = pte_page(huge_ptep_get(pte));
3003 pages[i] = mem_map_offset(page, pfn_offset);
3014 if (vaddr < vma->vm_end && remainder &&
3015 pfn_offset < pages_per_huge_page(h)) {
3017 * We use pfn_offset to avoid touching the pageframes
3018 * of this compound page.
3023 spin_unlock(&mm->page_table_lock);
3024 *length = remainder;
3027 return i ? i : -EFAULT;
3030 void hugetlb_change_protection(struct vm_area_struct *vma,
3031 unsigned long address, unsigned long end, pgprot_t newprot)
3033 struct mm_struct *mm = vma->vm_mm;
3034 unsigned long start = address;
3037 struct hstate *h = hstate_vma(vma);
3039 BUG_ON(address >= end);
3040 flush_cache_range(vma, address, end);
3042 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3043 spin_lock(&mm->page_table_lock);
3044 for (; address < end; address += huge_page_size(h)) {
3045 ptep = huge_pte_offset(mm, address);
3048 if (huge_pmd_unshare(mm, &address, ptep))
3050 pte = huge_ptep_get(ptep);
3051 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
3053 if (unlikely(is_hugetlb_entry_migration(pte))) {
3054 swp_entry_t entry = pte_to_swp_entry(pte);
3056 if (is_write_migration_entry(entry)) {
3059 make_migration_entry_read(&entry);
3060 newpte = swp_entry_to_pte(entry);
3061 set_huge_pte_at(mm, address, ptep, newpte);
3065 if (!huge_pte_none(pte)) {
3066 pte = huge_ptep_get_and_clear(mm, address, ptep);
3067 pte = pte_mkhuge(pte_modify(pte, newprot));
3068 set_huge_pte_at(mm, address, ptep, pte);
3071 spin_unlock(&mm->page_table_lock);
3073 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3074 * may have cleared our pud entry and done put_page on the page table:
3075 * once we release i_mmap_mutex, another task can do the final put_page
3076 * and that page table be reused and filled with junk.
3078 flush_tlb_range(vma, start, end);
3079 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3082 int hugetlb_reserve_pages(struct inode *inode,
3084 struct vm_area_struct *vma,
3085 vm_flags_t vm_flags)
3088 struct hstate *h = hstate_inode(inode);
3089 struct hugepage_subpool *spool = subpool_inode(inode);
3092 * Only apply hugepage reservation if asked. At fault time, an
3093 * attempt will be made for VM_NORESERVE to allocate a page
3094 * without using reserves
3096 if (vm_flags & VM_NORESERVE)
3100 * Shared mappings base their reservation on the number of pages that
3101 * are already allocated on behalf of the file. Private mappings need
3102 * to reserve the full area even if read-only as mprotect() may be
3103 * called to make the mapping read-write. Assume !vma is a shm mapping
3105 if (!vma || vma->vm_flags & VM_MAYSHARE)
3106 chg = region_chg(&inode->i_mapping->private_list, from, to);
3108 struct resv_map *resv_map = resv_map_alloc();
3114 set_vma_resv_map(vma, resv_map);
3115 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3123 /* There must be enough pages in the subpool for the mapping */
3124 if (hugepage_subpool_get_pages(spool, chg)) {
3130 * Check enough hugepages are available for the reservation.
3131 * Hand the pages back to the subpool if there are not
3133 ret = hugetlb_acct_memory(h, chg);
3135 hugepage_subpool_put_pages(spool, chg);
3140 * Account for the reservations made. Shared mappings record regions
3141 * that have reservations as they are shared by multiple VMAs.
3142 * When the last VMA disappears, the region map says how much
3143 * the reservation was and the page cache tells how much of
3144 * the reservation was consumed. Private mappings are per-VMA and
3145 * only the consumed reservations are tracked. When the VMA
3146 * disappears, the original reservation is the VMA size and the
3147 * consumed reservations are stored in the map. Hence, nothing
3148 * else has to be done for private mappings here
3150 if (!vma || vma->vm_flags & VM_MAYSHARE)
3151 region_add(&inode->i_mapping->private_list, from, to);
3159 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3161 struct hstate *h = hstate_inode(inode);
3162 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3163 struct hugepage_subpool *spool = subpool_inode(inode);
3165 spin_lock(&inode->i_lock);
3166 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3167 spin_unlock(&inode->i_lock);
3169 hugepage_subpool_put_pages(spool, (chg - freed));
3170 hugetlb_acct_memory(h, -(chg - freed));
3173 #ifdef CONFIG_MEMORY_FAILURE
3175 /* Should be called in hugetlb_lock */
3176 static int is_hugepage_on_freelist(struct page *hpage)
3180 struct hstate *h = page_hstate(hpage);
3181 int nid = page_to_nid(hpage);
3183 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3190 * This function is called from memory failure code.
3191 * Assume the caller holds page lock of the head page.
3193 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3195 struct hstate *h = page_hstate(hpage);
3196 int nid = page_to_nid(hpage);
3199 spin_lock(&hugetlb_lock);
3200 if (is_hugepage_on_freelist(hpage)) {
3201 list_del(&hpage->lru);
3202 set_page_refcounted(hpage);
3203 h->free_huge_pages--;
3204 h->free_huge_pages_node[nid]--;
3207 spin_unlock(&hugetlb_lock);