2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
41 __initdata LIST_HEAD(huge_boot_pages);
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static DEFINE_SPINLOCK(hugetlb_lock);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
60 spin_unlock(&spool->lock);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
70 struct hugepage_subpool *spool;
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
76 spin_lock_init(&spool->lock);
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
89 unlock_or_release_subpool(spool);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
106 spin_unlock(&spool->lock);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
126 return HUGETLBFS_SB(inode->i_sb)->spool;
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
131 return subpool_inode(vma->vm_file->f_dentry->d_inode);
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link;
154 static long region_add(struct list_head *head, long f, long t)
156 struct file_region *rg, *nrg, *trg;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head *head, long f, long t)
192 struct file_region *rg, *nrg;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg -= rg->to - rg->from;
239 static long region_truncate(struct list_head *head, long end)
241 struct file_region *rg, *trg;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
248 if (&rg->link == head)
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
255 rg = list_entry(rg->link.next, typeof(*rg), link);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
262 chg += rg->to - rg->from;
269 static long region_count(struct list_head *head, long f, long t)
271 struct file_region *rg;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
287 chg += seg_to - seg_from;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
316 struct hstate *hstate;
318 if (!is_vm_hugetlb_page(vma))
321 hstate = hstate_vma(vma);
323 return 1UL << (hstate->order + PAGE_SHIFT);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
336 return vma_kernel_pagesize(vma);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
370 return (unsigned long)vma->vm_private_data;
373 static void set_vma_private_data(struct vm_area_struct *vma,
376 vma->vm_private_data = (void *)value;
381 struct list_head regions;
384 static struct resv_map *resv_map_alloc(void)
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
396 static void resv_map_release(struct kref *ref)
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
435 return (get_vma_private_data(vma) & flag) != 0;
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
442 if (vma->vm_flags & VM_NORESERVE)
445 if (vma->vm_flags & VM_MAYSHARE) {
446 /* Shared mappings always use reserves */
447 h->resv_huge_pages--;
448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
450 * Only the process that called mmap() has reserves for
453 h->resv_huge_pages--;
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
461 if (!(vma->vm_flags & VM_MAYSHARE))
462 vma->vm_private_data = (void *)0;
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct *vma)
468 if (vma->vm_flags & VM_MAYSHARE)
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
475 static void copy_gigantic_page(struct page *dst, struct page *src)
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
482 for (i = 0; i < pages_per_huge_page(h); ) {
484 copy_highpage(dst, src);
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
492 void copy_huge_page(struct page *dst, struct page *src)
495 struct hstate *h = page_hstate(src);
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
503 for (i = 0; i < pages_per_huge_page(h); i++) {
505 copy_highpage(dst + i, src + i);
509 static void enqueue_huge_page(struct hstate *h, struct page *page)
511 int nid = page_to_nid(page);
512 list_add(&page->lru, &h->hugepage_freelists[nid]);
513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
517 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
521 if (list_empty(&h->hugepage_freelists[nid]))
523 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
524 list_del(&page->lru);
525 set_page_refcounted(page);
526 h->free_huge_pages--;
527 h->free_huge_pages_node[nid]--;
531 static struct page *dequeue_huge_page_vma(struct hstate *h,
532 struct vm_area_struct *vma,
533 unsigned long address, int avoid_reserve)
535 struct page *page = NULL;
536 struct mempolicy *mpol;
537 nodemask_t *nodemask;
538 struct zonelist *zonelist;
541 unsigned int cpuset_mems_cookie;
544 cpuset_mems_cookie = get_mems_allowed();
545 zonelist = huge_zonelist(vma, address,
546 htlb_alloc_mask, &mpol, &nodemask);
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma) &&
553 h->free_huge_pages - h->resv_huge_pages == 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
566 decrement_hugepage_resv_vma(h, vma);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
582 static void update_and_free_page(struct hstate *h, struct page *page)
586 VM_BUG_ON(h->order >= MAX_ORDER);
589 h->nr_huge_pages_node[page_to_nid(page)]--;
590 for (i = 0; i < pages_per_huge_page(h); i++) {
591 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
592 1 << PG_referenced | 1 << PG_dirty |
593 1 << PG_active | 1 << PG_reserved |
594 1 << PG_private | 1 << PG_writeback);
596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
598 arch_release_hugepage(page);
599 __free_pages(page, huge_page_order(h));
602 struct hstate *size_to_hstate(unsigned long size)
607 if (huge_page_size(h) == size)
613 static void free_huge_page(struct page *page)
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
619 struct hstate *h = page_hstate(page);
620 int nid = page_to_nid(page);
621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
624 set_page_private(page, 0);
625 page->mapping = NULL;
626 BUG_ON(page_count(page));
627 BUG_ON(page_mapcount(page));
628 INIT_LIST_HEAD(&page->lru);
630 spin_lock(&hugetlb_lock);
631 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
632 update_and_free_page(h, page);
633 h->surplus_huge_pages--;
634 h->surplus_huge_pages_node[nid]--;
636 enqueue_huge_page(h, page);
638 spin_unlock(&hugetlb_lock);
639 hugepage_subpool_put_pages(spool, 1);
642 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
644 set_compound_page_dtor(page, free_huge_page);
645 spin_lock(&hugetlb_lock);
647 h->nr_huge_pages_node[nid]++;
648 spin_unlock(&hugetlb_lock);
649 put_page(page); /* free it into the hugepage allocator */
652 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
655 int nr_pages = 1 << order;
656 struct page *p = page + 1;
658 /* we rely on prep_new_huge_page to set the destructor */
659 set_compound_order(page, order);
661 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
663 set_page_count(p, 0);
664 p->first_page = page;
668 int PageHuge(struct page *page)
670 compound_page_dtor *dtor;
672 if (!PageCompound(page))
675 page = compound_head(page);
676 dtor = get_compound_page_dtor(page);
678 return dtor == free_huge_page;
680 EXPORT_SYMBOL_GPL(PageHuge);
682 pgoff_t __basepage_index(struct page *page)
684 struct page *page_head = compound_head(page);
685 pgoff_t index = page_index(page_head);
686 unsigned long compound_idx;
688 if (!PageHuge(page_head))
689 return page_index(page);
691 if (compound_order(page_head) >= MAX_ORDER)
692 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
694 compound_idx = page - page_head;
696 return (index << compound_order(page_head)) + compound_idx;
699 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
703 if (h->order >= MAX_ORDER)
706 page = alloc_pages_exact_node(nid,
707 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
708 __GFP_REPEAT|__GFP_NOWARN,
711 if (arch_prepare_hugepage(page)) {
712 __free_pages(page, huge_page_order(h));
715 prep_new_huge_page(h, page, nid);
722 * common helper functions for hstate_next_node_to_{alloc|free}.
723 * We may have allocated or freed a huge page based on a different
724 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
725 * be outside of *nodes_allowed. Ensure that we use an allowed
726 * node for alloc or free.
728 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
730 nid = next_node(nid, *nodes_allowed);
731 if (nid == MAX_NUMNODES)
732 nid = first_node(*nodes_allowed);
733 VM_BUG_ON(nid >= MAX_NUMNODES);
738 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
740 if (!node_isset(nid, *nodes_allowed))
741 nid = next_node_allowed(nid, nodes_allowed);
746 * returns the previously saved node ["this node"] from which to
747 * allocate a persistent huge page for the pool and advance the
748 * next node from which to allocate, handling wrap at end of node
751 static int hstate_next_node_to_alloc(struct hstate *h,
752 nodemask_t *nodes_allowed)
756 VM_BUG_ON(!nodes_allowed);
758 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
759 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
764 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
771 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
772 next_nid = start_nid;
775 page = alloc_fresh_huge_page_node(h, next_nid);
780 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
781 } while (next_nid != start_nid);
784 count_vm_event(HTLB_BUDDY_PGALLOC);
786 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
792 * helper for free_pool_huge_page() - return the previously saved
793 * node ["this node"] from which to free a huge page. Advance the
794 * next node id whether or not we find a free huge page to free so
795 * that the next attempt to free addresses the next node.
797 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
801 VM_BUG_ON(!nodes_allowed);
803 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
804 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
810 * Free huge page from pool from next node to free.
811 * Attempt to keep persistent huge pages more or less
812 * balanced over allowed nodes.
813 * Called with hugetlb_lock locked.
815 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
822 start_nid = hstate_next_node_to_free(h, nodes_allowed);
823 next_nid = start_nid;
827 * If we're returning unused surplus pages, only examine
828 * nodes with surplus pages.
830 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
831 !list_empty(&h->hugepage_freelists[next_nid])) {
833 list_entry(h->hugepage_freelists[next_nid].next,
835 list_del(&page->lru);
836 h->free_huge_pages--;
837 h->free_huge_pages_node[next_nid]--;
839 h->surplus_huge_pages--;
840 h->surplus_huge_pages_node[next_nid]--;
842 update_and_free_page(h, page);
846 next_nid = hstate_next_node_to_free(h, nodes_allowed);
847 } while (next_nid != start_nid);
852 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
857 if (h->order >= MAX_ORDER)
861 * Assume we will successfully allocate the surplus page to
862 * prevent racing processes from causing the surplus to exceed
865 * This however introduces a different race, where a process B
866 * tries to grow the static hugepage pool while alloc_pages() is
867 * called by process A. B will only examine the per-node
868 * counters in determining if surplus huge pages can be
869 * converted to normal huge pages in adjust_pool_surplus(). A
870 * won't be able to increment the per-node counter, until the
871 * lock is dropped by B, but B doesn't drop hugetlb_lock until
872 * no more huge pages can be converted from surplus to normal
873 * state (and doesn't try to convert again). Thus, we have a
874 * case where a surplus huge page exists, the pool is grown, and
875 * the surplus huge page still exists after, even though it
876 * should just have been converted to a normal huge page. This
877 * does not leak memory, though, as the hugepage will be freed
878 * once it is out of use. It also does not allow the counters to
879 * go out of whack in adjust_pool_surplus() as we don't modify
880 * the node values until we've gotten the hugepage and only the
881 * per-node value is checked there.
883 spin_lock(&hugetlb_lock);
884 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
885 spin_unlock(&hugetlb_lock);
889 h->surplus_huge_pages++;
891 spin_unlock(&hugetlb_lock);
893 if (nid == NUMA_NO_NODE)
894 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
895 __GFP_REPEAT|__GFP_NOWARN,
898 page = alloc_pages_exact_node(nid,
899 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
900 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
902 if (page && arch_prepare_hugepage(page)) {
903 __free_pages(page, huge_page_order(h));
907 spin_lock(&hugetlb_lock);
909 r_nid = page_to_nid(page);
910 set_compound_page_dtor(page, free_huge_page);
912 * We incremented the global counters already
914 h->nr_huge_pages_node[r_nid]++;
915 h->surplus_huge_pages_node[r_nid]++;
916 __count_vm_event(HTLB_BUDDY_PGALLOC);
919 h->surplus_huge_pages--;
920 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
922 spin_unlock(&hugetlb_lock);
928 * This allocation function is useful in the context where vma is irrelevant.
929 * E.g. soft-offlining uses this function because it only cares physical
930 * address of error page.
932 struct page *alloc_huge_page_node(struct hstate *h, int nid)
936 spin_lock(&hugetlb_lock);
937 page = dequeue_huge_page_node(h, nid);
938 spin_unlock(&hugetlb_lock);
941 page = alloc_buddy_huge_page(h, nid);
947 * Increase the hugetlb pool such that it can accommodate a reservation
950 static int gather_surplus_pages(struct hstate *h, int delta)
952 struct list_head surplus_list;
953 struct page *page, *tmp;
955 int needed, allocated;
957 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
959 h->resv_huge_pages += delta;
964 INIT_LIST_HEAD(&surplus_list);
968 spin_unlock(&hugetlb_lock);
969 for (i = 0; i < needed; i++) {
970 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
973 * We were not able to allocate enough pages to
974 * satisfy the entire reservation so we free what
975 * we've allocated so far.
979 list_add(&page->lru, &surplus_list);
984 * After retaking hugetlb_lock, we need to recalculate 'needed'
985 * because either resv_huge_pages or free_huge_pages may have changed.
987 spin_lock(&hugetlb_lock);
988 needed = (h->resv_huge_pages + delta) -
989 (h->free_huge_pages + allocated);
994 * The surplus_list now contains _at_least_ the number of extra pages
995 * needed to accommodate the reservation. Add the appropriate number
996 * of pages to the hugetlb pool and free the extras back to the buddy
997 * allocator. Commit the entire reservation here to prevent another
998 * process from stealing the pages as they are added to the pool but
999 * before they are reserved.
1001 needed += allocated;
1002 h->resv_huge_pages += delta;
1005 /* Free the needed pages to the hugetlb pool */
1006 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1009 list_del(&page->lru);
1011 * This page is now managed by the hugetlb allocator and has
1012 * no users -- drop the buddy allocator's reference.
1014 put_page_testzero(page);
1015 VM_BUG_ON(page_count(page));
1016 enqueue_huge_page(h, page);
1018 spin_unlock(&hugetlb_lock);
1020 /* Free unnecessary surplus pages to the buddy allocator */
1022 if (!list_empty(&surplus_list)) {
1023 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1024 list_del(&page->lru);
1028 spin_lock(&hugetlb_lock);
1034 * When releasing a hugetlb pool reservation, any surplus pages that were
1035 * allocated to satisfy the reservation must be explicitly freed if they were
1037 * Called with hugetlb_lock held.
1039 static void return_unused_surplus_pages(struct hstate *h,
1040 unsigned long unused_resv_pages)
1042 unsigned long nr_pages;
1044 /* Uncommit the reservation */
1045 h->resv_huge_pages -= unused_resv_pages;
1047 /* Cannot return gigantic pages currently */
1048 if (h->order >= MAX_ORDER)
1051 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1054 * We want to release as many surplus pages as possible, spread
1055 * evenly across all nodes with memory. Iterate across these nodes
1056 * until we can no longer free unreserved surplus pages. This occurs
1057 * when the nodes with surplus pages have no free pages.
1058 * free_pool_huge_page() will balance the the freed pages across the
1059 * on-line nodes with memory and will handle the hstate accounting.
1061 while (nr_pages--) {
1062 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1068 * Determine if the huge page at addr within the vma has an associated
1069 * reservation. Where it does not we will need to logically increase
1070 * reservation and actually increase subpool usage before an allocation
1071 * can occur. Where any new reservation would be required the
1072 * reservation change is prepared, but not committed. Once the page
1073 * has been allocated from the subpool and instantiated the change should
1074 * be committed via vma_commit_reservation. No action is required on
1077 static long vma_needs_reservation(struct hstate *h,
1078 struct vm_area_struct *vma, unsigned long addr)
1080 struct address_space *mapping = vma->vm_file->f_mapping;
1081 struct inode *inode = mapping->host;
1083 if (vma->vm_flags & VM_MAYSHARE) {
1084 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1085 return region_chg(&inode->i_mapping->private_list,
1088 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1093 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1094 struct resv_map *reservations = vma_resv_map(vma);
1096 err = region_chg(&reservations->regions, idx, idx + 1);
1102 static void vma_commit_reservation(struct hstate *h,
1103 struct vm_area_struct *vma, unsigned long addr)
1105 struct address_space *mapping = vma->vm_file->f_mapping;
1106 struct inode *inode = mapping->host;
1108 if (vma->vm_flags & VM_MAYSHARE) {
1109 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1110 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1112 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1113 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1114 struct resv_map *reservations = vma_resv_map(vma);
1116 /* Mark this page used in the map. */
1117 region_add(&reservations->regions, idx, idx + 1);
1121 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1122 unsigned long addr, int avoid_reserve)
1124 struct hugepage_subpool *spool = subpool_vma(vma);
1125 struct hstate *h = hstate_vma(vma);
1130 * Processes that did not create the mapping will have no
1131 * reserves and will not have accounted against subpool
1132 * limit. Check that the subpool limit can be made before
1133 * satisfying the allocation MAP_NORESERVE mappings may also
1134 * need pages and subpool limit allocated allocated if no reserve
1137 chg = vma_needs_reservation(h, vma, addr);
1139 return ERR_PTR(-VM_FAULT_OOM);
1141 if (hugepage_subpool_get_pages(spool, chg))
1142 return ERR_PTR(-VM_FAULT_SIGBUS);
1144 spin_lock(&hugetlb_lock);
1145 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1146 spin_unlock(&hugetlb_lock);
1149 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1151 hugepage_subpool_put_pages(spool, chg);
1152 return ERR_PTR(-VM_FAULT_SIGBUS);
1156 set_page_private(page, (unsigned long)spool);
1158 vma_commit_reservation(h, vma, addr);
1163 int __weak alloc_bootmem_huge_page(struct hstate *h)
1165 struct huge_bootmem_page *m;
1166 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1171 addr = __alloc_bootmem_node_nopanic(
1172 NODE_DATA(hstate_next_node_to_alloc(h,
1173 &node_states[N_HIGH_MEMORY])),
1174 huge_page_size(h), huge_page_size(h), 0);
1178 * Use the beginning of the huge page to store the
1179 * huge_bootmem_page struct (until gather_bootmem
1180 * puts them into the mem_map).
1190 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1191 /* Put them into a private list first because mem_map is not up yet */
1192 list_add(&m->list, &huge_boot_pages);
1197 static void prep_compound_huge_page(struct page *page, int order)
1199 if (unlikely(order > (MAX_ORDER - 1)))
1200 prep_compound_gigantic_page(page, order);
1202 prep_compound_page(page, order);
1205 /* Put bootmem huge pages into the standard lists after mem_map is up */
1206 static void __init gather_bootmem_prealloc(void)
1208 struct huge_bootmem_page *m;
1210 list_for_each_entry(m, &huge_boot_pages, list) {
1211 struct hstate *h = m->hstate;
1214 #ifdef CONFIG_HIGHMEM
1215 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1216 free_bootmem_late((unsigned long)m,
1217 sizeof(struct huge_bootmem_page));
1219 page = virt_to_page(m);
1221 __ClearPageReserved(page);
1222 WARN_ON(page_count(page) != 1);
1223 prep_compound_huge_page(page, h->order);
1224 prep_new_huge_page(h, page, page_to_nid(page));
1226 * If we had gigantic hugepages allocated at boot time, we need
1227 * to restore the 'stolen' pages to totalram_pages in order to
1228 * fix confusing memory reports from free(1) and another
1229 * side-effects, like CommitLimit going negative.
1231 if (h->order > (MAX_ORDER - 1))
1232 totalram_pages += 1 << h->order;
1236 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1240 for (i = 0; i < h->max_huge_pages; ++i) {
1241 if (h->order >= MAX_ORDER) {
1242 if (!alloc_bootmem_huge_page(h))
1244 } else if (!alloc_fresh_huge_page(h,
1245 &node_states[N_HIGH_MEMORY]))
1248 h->max_huge_pages = i;
1251 static void __init hugetlb_init_hstates(void)
1255 for_each_hstate(h) {
1256 /* oversize hugepages were init'ed in early boot */
1257 if (h->order < MAX_ORDER)
1258 hugetlb_hstate_alloc_pages(h);
1262 static char * __init memfmt(char *buf, unsigned long n)
1264 if (n >= (1UL << 30))
1265 sprintf(buf, "%lu GB", n >> 30);
1266 else if (n >= (1UL << 20))
1267 sprintf(buf, "%lu MB", n >> 20);
1269 sprintf(buf, "%lu KB", n >> 10);
1273 static void __init report_hugepages(void)
1277 for_each_hstate(h) {
1279 printk(KERN_INFO "HugeTLB registered %s page size, "
1280 "pre-allocated %ld pages\n",
1281 memfmt(buf, huge_page_size(h)),
1282 h->free_huge_pages);
1286 #ifdef CONFIG_HIGHMEM
1287 static void try_to_free_low(struct hstate *h, unsigned long count,
1288 nodemask_t *nodes_allowed)
1292 if (h->order >= MAX_ORDER)
1295 for_each_node_mask(i, *nodes_allowed) {
1296 struct page *page, *next;
1297 struct list_head *freel = &h->hugepage_freelists[i];
1298 list_for_each_entry_safe(page, next, freel, lru) {
1299 if (count >= h->nr_huge_pages)
1301 if (PageHighMem(page))
1303 list_del(&page->lru);
1304 update_and_free_page(h, page);
1305 h->free_huge_pages--;
1306 h->free_huge_pages_node[page_to_nid(page)]--;
1311 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1312 nodemask_t *nodes_allowed)
1318 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1319 * balanced by operating on them in a round-robin fashion.
1320 * Returns 1 if an adjustment was made.
1322 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1325 int start_nid, next_nid;
1328 VM_BUG_ON(delta != -1 && delta != 1);
1331 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1333 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1334 next_nid = start_nid;
1340 * To shrink on this node, there must be a surplus page
1342 if (!h->surplus_huge_pages_node[nid]) {
1343 next_nid = hstate_next_node_to_alloc(h,
1350 * Surplus cannot exceed the total number of pages
1352 if (h->surplus_huge_pages_node[nid] >=
1353 h->nr_huge_pages_node[nid]) {
1354 next_nid = hstate_next_node_to_free(h,
1360 h->surplus_huge_pages += delta;
1361 h->surplus_huge_pages_node[nid] += delta;
1364 } while (next_nid != start_nid);
1369 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1370 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1371 nodemask_t *nodes_allowed)
1373 unsigned long min_count, ret;
1375 if (h->order >= MAX_ORDER)
1376 return h->max_huge_pages;
1379 * Increase the pool size
1380 * First take pages out of surplus state. Then make up the
1381 * remaining difference by allocating fresh huge pages.
1383 * We might race with alloc_buddy_huge_page() here and be unable
1384 * to convert a surplus huge page to a normal huge page. That is
1385 * not critical, though, it just means the overall size of the
1386 * pool might be one hugepage larger than it needs to be, but
1387 * within all the constraints specified by the sysctls.
1389 spin_lock(&hugetlb_lock);
1390 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1391 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1395 while (count > persistent_huge_pages(h)) {
1397 * If this allocation races such that we no longer need the
1398 * page, free_huge_page will handle it by freeing the page
1399 * and reducing the surplus.
1401 spin_unlock(&hugetlb_lock);
1402 ret = alloc_fresh_huge_page(h, nodes_allowed);
1403 spin_lock(&hugetlb_lock);
1407 /* Bail for signals. Probably ctrl-c from user */
1408 if (signal_pending(current))
1413 * Decrease the pool size
1414 * First return free pages to the buddy allocator (being careful
1415 * to keep enough around to satisfy reservations). Then place
1416 * pages into surplus state as needed so the pool will shrink
1417 * to the desired size as pages become free.
1419 * By placing pages into the surplus state independent of the
1420 * overcommit value, we are allowing the surplus pool size to
1421 * exceed overcommit. There are few sane options here. Since
1422 * alloc_buddy_huge_page() is checking the global counter,
1423 * though, we'll note that we're not allowed to exceed surplus
1424 * and won't grow the pool anywhere else. Not until one of the
1425 * sysctls are changed, or the surplus pages go out of use.
1427 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1428 min_count = max(count, min_count);
1429 try_to_free_low(h, min_count, nodes_allowed);
1430 while (min_count < persistent_huge_pages(h)) {
1431 if (!free_pool_huge_page(h, nodes_allowed, 0))
1434 while (count < persistent_huge_pages(h)) {
1435 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1439 ret = persistent_huge_pages(h);
1440 spin_unlock(&hugetlb_lock);
1444 #define HSTATE_ATTR_RO(_name) \
1445 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1447 #define HSTATE_ATTR(_name) \
1448 static struct kobj_attribute _name##_attr = \
1449 __ATTR(_name, 0644, _name##_show, _name##_store)
1451 static struct kobject *hugepages_kobj;
1452 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1454 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1456 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1460 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1461 if (hstate_kobjs[i] == kobj) {
1463 *nidp = NUMA_NO_NODE;
1467 return kobj_to_node_hstate(kobj, nidp);
1470 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1471 struct kobj_attribute *attr, char *buf)
1474 unsigned long nr_huge_pages;
1477 h = kobj_to_hstate(kobj, &nid);
1478 if (nid == NUMA_NO_NODE)
1479 nr_huge_pages = h->nr_huge_pages;
1481 nr_huge_pages = h->nr_huge_pages_node[nid];
1483 return sprintf(buf, "%lu\n", nr_huge_pages);
1486 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1487 struct kobject *kobj, struct kobj_attribute *attr,
1488 const char *buf, size_t len)
1492 unsigned long count;
1494 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1496 err = strict_strtoul(buf, 10, &count);
1500 h = kobj_to_hstate(kobj, &nid);
1501 if (h->order >= MAX_ORDER) {
1506 if (nid == NUMA_NO_NODE) {
1508 * global hstate attribute
1510 if (!(obey_mempolicy &&
1511 init_nodemask_of_mempolicy(nodes_allowed))) {
1512 NODEMASK_FREE(nodes_allowed);
1513 nodes_allowed = &node_states[N_HIGH_MEMORY];
1515 } else if (nodes_allowed) {
1517 * per node hstate attribute: adjust count to global,
1518 * but restrict alloc/free to the specified node.
1520 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1521 init_nodemask_of_node(nodes_allowed, nid);
1523 nodes_allowed = &node_states[N_HIGH_MEMORY];
1525 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1527 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1528 NODEMASK_FREE(nodes_allowed);
1532 NODEMASK_FREE(nodes_allowed);
1536 static ssize_t nr_hugepages_show(struct kobject *kobj,
1537 struct kobj_attribute *attr, char *buf)
1539 return nr_hugepages_show_common(kobj, attr, buf);
1542 static ssize_t nr_hugepages_store(struct kobject *kobj,
1543 struct kobj_attribute *attr, const char *buf, size_t len)
1545 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1547 HSTATE_ATTR(nr_hugepages);
1552 * hstate attribute for optionally mempolicy-based constraint on persistent
1553 * huge page alloc/free.
1555 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1556 struct kobj_attribute *attr, char *buf)
1558 return nr_hugepages_show_common(kobj, attr, buf);
1561 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1562 struct kobj_attribute *attr, const char *buf, size_t len)
1564 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1566 HSTATE_ATTR(nr_hugepages_mempolicy);
1570 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1571 struct kobj_attribute *attr, char *buf)
1573 struct hstate *h = kobj_to_hstate(kobj, NULL);
1574 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1577 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1578 struct kobj_attribute *attr, const char *buf, size_t count)
1581 unsigned long input;
1582 struct hstate *h = kobj_to_hstate(kobj, NULL);
1584 if (h->order >= MAX_ORDER)
1587 err = strict_strtoul(buf, 10, &input);
1591 spin_lock(&hugetlb_lock);
1592 h->nr_overcommit_huge_pages = input;
1593 spin_unlock(&hugetlb_lock);
1597 HSTATE_ATTR(nr_overcommit_hugepages);
1599 static ssize_t free_hugepages_show(struct kobject *kobj,
1600 struct kobj_attribute *attr, char *buf)
1603 unsigned long free_huge_pages;
1606 h = kobj_to_hstate(kobj, &nid);
1607 if (nid == NUMA_NO_NODE)
1608 free_huge_pages = h->free_huge_pages;
1610 free_huge_pages = h->free_huge_pages_node[nid];
1612 return sprintf(buf, "%lu\n", free_huge_pages);
1614 HSTATE_ATTR_RO(free_hugepages);
1616 static ssize_t resv_hugepages_show(struct kobject *kobj,
1617 struct kobj_attribute *attr, char *buf)
1619 struct hstate *h = kobj_to_hstate(kobj, NULL);
1620 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1622 HSTATE_ATTR_RO(resv_hugepages);
1624 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1625 struct kobj_attribute *attr, char *buf)
1628 unsigned long surplus_huge_pages;
1631 h = kobj_to_hstate(kobj, &nid);
1632 if (nid == NUMA_NO_NODE)
1633 surplus_huge_pages = h->surplus_huge_pages;
1635 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1637 return sprintf(buf, "%lu\n", surplus_huge_pages);
1639 HSTATE_ATTR_RO(surplus_hugepages);
1641 static struct attribute *hstate_attrs[] = {
1642 &nr_hugepages_attr.attr,
1643 &nr_overcommit_hugepages_attr.attr,
1644 &free_hugepages_attr.attr,
1645 &resv_hugepages_attr.attr,
1646 &surplus_hugepages_attr.attr,
1648 &nr_hugepages_mempolicy_attr.attr,
1653 static struct attribute_group hstate_attr_group = {
1654 .attrs = hstate_attrs,
1657 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1658 struct kobject **hstate_kobjs,
1659 struct attribute_group *hstate_attr_group)
1662 int hi = h - hstates;
1664 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1665 if (!hstate_kobjs[hi])
1668 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1670 kobject_put(hstate_kobjs[hi]);
1675 static void __init hugetlb_sysfs_init(void)
1680 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1681 if (!hugepages_kobj)
1684 for_each_hstate(h) {
1685 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1686 hstate_kobjs, &hstate_attr_group);
1688 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1696 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1697 * with node sysdevs in node_devices[] using a parallel array. The array
1698 * index of a node sysdev or _hstate == node id.
1699 * This is here to avoid any static dependency of the node sysdev driver, in
1700 * the base kernel, on the hugetlb module.
1702 struct node_hstate {
1703 struct kobject *hugepages_kobj;
1704 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1706 struct node_hstate node_hstates[MAX_NUMNODES];
1709 * A subset of global hstate attributes for node sysdevs
1711 static struct attribute *per_node_hstate_attrs[] = {
1712 &nr_hugepages_attr.attr,
1713 &free_hugepages_attr.attr,
1714 &surplus_hugepages_attr.attr,
1718 static struct attribute_group per_node_hstate_attr_group = {
1719 .attrs = per_node_hstate_attrs,
1723 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1724 * Returns node id via non-NULL nidp.
1726 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1730 for (nid = 0; nid < nr_node_ids; nid++) {
1731 struct node_hstate *nhs = &node_hstates[nid];
1733 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1734 if (nhs->hstate_kobjs[i] == kobj) {
1746 * Unregister hstate attributes from a single node sysdev.
1747 * No-op if no hstate attributes attached.
1749 void hugetlb_unregister_node(struct node *node)
1752 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1754 if (!nhs->hugepages_kobj)
1755 return; /* no hstate attributes */
1758 if (nhs->hstate_kobjs[h - hstates]) {
1759 kobject_put(nhs->hstate_kobjs[h - hstates]);
1760 nhs->hstate_kobjs[h - hstates] = NULL;
1763 kobject_put(nhs->hugepages_kobj);
1764 nhs->hugepages_kobj = NULL;
1768 * hugetlb module exit: unregister hstate attributes from node sysdevs
1771 static void hugetlb_unregister_all_nodes(void)
1776 * disable node sysdev registrations.
1778 register_hugetlbfs_with_node(NULL, NULL);
1781 * remove hstate attributes from any nodes that have them.
1783 for (nid = 0; nid < nr_node_ids; nid++)
1784 hugetlb_unregister_node(&node_devices[nid]);
1788 * Register hstate attributes for a single node sysdev.
1789 * No-op if attributes already registered.
1791 void hugetlb_register_node(struct node *node)
1794 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1797 if (nhs->hugepages_kobj)
1798 return; /* already allocated */
1800 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1801 &node->sysdev.kobj);
1802 if (!nhs->hugepages_kobj)
1805 for_each_hstate(h) {
1806 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1808 &per_node_hstate_attr_group);
1810 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1812 h->name, node->sysdev.id);
1813 hugetlb_unregister_node(node);
1820 * hugetlb init time: register hstate attributes for all registered node
1821 * sysdevs of nodes that have memory. All on-line nodes should have
1822 * registered their associated sysdev by this time.
1824 static void hugetlb_register_all_nodes(void)
1828 for_each_node_state(nid, N_HIGH_MEMORY) {
1829 struct node *node = &node_devices[nid];
1830 if (node->sysdev.id == nid)
1831 hugetlb_register_node(node);
1835 * Let the node sysdev driver know we're here so it can
1836 * [un]register hstate attributes on node hotplug.
1838 register_hugetlbfs_with_node(hugetlb_register_node,
1839 hugetlb_unregister_node);
1841 #else /* !CONFIG_NUMA */
1843 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1851 static void hugetlb_unregister_all_nodes(void) { }
1853 static void hugetlb_register_all_nodes(void) { }
1857 static void __exit hugetlb_exit(void)
1861 hugetlb_unregister_all_nodes();
1863 for_each_hstate(h) {
1864 kobject_put(hstate_kobjs[h - hstates]);
1867 kobject_put(hugepages_kobj);
1869 module_exit(hugetlb_exit);
1871 static int __init hugetlb_init(void)
1873 /* Some platform decide whether they support huge pages at boot
1874 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1875 * there is no such support
1877 if (HPAGE_SHIFT == 0)
1880 if (!size_to_hstate(default_hstate_size)) {
1881 default_hstate_size = HPAGE_SIZE;
1882 if (!size_to_hstate(default_hstate_size))
1883 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1885 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1886 if (default_hstate_max_huge_pages)
1887 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1889 hugetlb_init_hstates();
1891 gather_bootmem_prealloc();
1895 hugetlb_sysfs_init();
1897 hugetlb_register_all_nodes();
1901 module_init(hugetlb_init);
1903 /* Should be called on processing a hugepagesz=... option */
1904 void __init hugetlb_add_hstate(unsigned order)
1909 if (size_to_hstate(PAGE_SIZE << order)) {
1910 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1913 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1915 h = &hstates[max_hstate++];
1917 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1918 h->nr_huge_pages = 0;
1919 h->free_huge_pages = 0;
1920 for (i = 0; i < MAX_NUMNODES; ++i)
1921 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1922 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1923 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1924 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1925 huge_page_size(h)/1024);
1930 static int __init hugetlb_nrpages_setup(char *s)
1933 static unsigned long *last_mhp;
1936 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1937 * so this hugepages= parameter goes to the "default hstate".
1940 mhp = &default_hstate_max_huge_pages;
1942 mhp = &parsed_hstate->max_huge_pages;
1944 if (mhp == last_mhp) {
1945 printk(KERN_WARNING "hugepages= specified twice without "
1946 "interleaving hugepagesz=, ignoring\n");
1950 if (sscanf(s, "%lu", mhp) <= 0)
1954 * Global state is always initialized later in hugetlb_init.
1955 * But we need to allocate >= MAX_ORDER hstates here early to still
1956 * use the bootmem allocator.
1958 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1959 hugetlb_hstate_alloc_pages(parsed_hstate);
1965 __setup("hugepages=", hugetlb_nrpages_setup);
1967 static int __init hugetlb_default_setup(char *s)
1969 default_hstate_size = memparse(s, &s);
1972 __setup("default_hugepagesz=", hugetlb_default_setup);
1974 static unsigned int cpuset_mems_nr(unsigned int *array)
1977 unsigned int nr = 0;
1979 for_each_node_mask(node, cpuset_current_mems_allowed)
1985 #ifdef CONFIG_SYSCTL
1986 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1987 struct ctl_table *table, int write,
1988 void __user *buffer, size_t *length, loff_t *ppos)
1990 struct hstate *h = &default_hstate;
1994 tmp = h->max_huge_pages;
1996 if (write && h->order >= MAX_ORDER)
2000 table->maxlen = sizeof(unsigned long);
2001 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2006 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2007 GFP_KERNEL | __GFP_NORETRY);
2008 if (!(obey_mempolicy &&
2009 init_nodemask_of_mempolicy(nodes_allowed))) {
2010 NODEMASK_FREE(nodes_allowed);
2011 nodes_allowed = &node_states[N_HIGH_MEMORY];
2013 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2015 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2016 NODEMASK_FREE(nodes_allowed);
2022 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2023 void __user *buffer, size_t *length, loff_t *ppos)
2026 return hugetlb_sysctl_handler_common(false, table, write,
2027 buffer, length, ppos);
2031 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2032 void __user *buffer, size_t *length, loff_t *ppos)
2034 return hugetlb_sysctl_handler_common(true, table, write,
2035 buffer, length, ppos);
2037 #endif /* CONFIG_NUMA */
2039 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2040 void __user *buffer,
2041 size_t *length, loff_t *ppos)
2043 proc_dointvec(table, write, buffer, length, ppos);
2044 if (hugepages_treat_as_movable)
2045 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2047 htlb_alloc_mask = GFP_HIGHUSER;
2051 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2052 void __user *buffer,
2053 size_t *length, loff_t *ppos)
2055 struct hstate *h = &default_hstate;
2059 tmp = h->nr_overcommit_huge_pages;
2061 if (write && h->order >= MAX_ORDER)
2065 table->maxlen = sizeof(unsigned long);
2066 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2071 spin_lock(&hugetlb_lock);
2072 h->nr_overcommit_huge_pages = tmp;
2073 spin_unlock(&hugetlb_lock);
2079 #endif /* CONFIG_SYSCTL */
2081 void hugetlb_report_meminfo(struct seq_file *m)
2083 struct hstate *h = &default_hstate;
2085 "HugePages_Total: %5lu\n"
2086 "HugePages_Free: %5lu\n"
2087 "HugePages_Rsvd: %5lu\n"
2088 "HugePages_Surp: %5lu\n"
2089 "Hugepagesize: %8lu kB\n",
2093 h->surplus_huge_pages,
2094 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2097 int hugetlb_report_node_meminfo(int nid, char *buf)
2099 struct hstate *h = &default_hstate;
2101 "Node %d HugePages_Total: %5u\n"
2102 "Node %d HugePages_Free: %5u\n"
2103 "Node %d HugePages_Surp: %5u\n",
2104 nid, h->nr_huge_pages_node[nid],
2105 nid, h->free_huge_pages_node[nid],
2106 nid, h->surplus_huge_pages_node[nid]);
2109 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2110 unsigned long hugetlb_total_pages(void)
2113 unsigned long nr_total_pages = 0;
2116 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2117 return nr_total_pages;
2120 static int hugetlb_acct_memory(struct hstate *h, long delta)
2124 spin_lock(&hugetlb_lock);
2126 * When cpuset is configured, it breaks the strict hugetlb page
2127 * reservation as the accounting is done on a global variable. Such
2128 * reservation is completely rubbish in the presence of cpuset because
2129 * the reservation is not checked against page availability for the
2130 * current cpuset. Application can still potentially OOM'ed by kernel
2131 * with lack of free htlb page in cpuset that the task is in.
2132 * Attempt to enforce strict accounting with cpuset is almost
2133 * impossible (or too ugly) because cpuset is too fluid that
2134 * task or memory node can be dynamically moved between cpusets.
2136 * The change of semantics for shared hugetlb mapping with cpuset is
2137 * undesirable. However, in order to preserve some of the semantics,
2138 * we fall back to check against current free page availability as
2139 * a best attempt and hopefully to minimize the impact of changing
2140 * semantics that cpuset has.
2143 if (gather_surplus_pages(h, delta) < 0)
2146 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2147 return_unused_surplus_pages(h, delta);
2154 return_unused_surplus_pages(h, (unsigned long) -delta);
2157 spin_unlock(&hugetlb_lock);
2161 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2163 struct resv_map *reservations = vma_resv_map(vma);
2166 * This new VMA should share its siblings reservation map if present.
2167 * The VMA will only ever have a valid reservation map pointer where
2168 * it is being copied for another still existing VMA. As that VMA
2169 * has a reference to the reservation map it cannot disappear until
2170 * after this open call completes. It is therefore safe to take a
2171 * new reference here without additional locking.
2174 kref_get(&reservations->refs);
2177 static void resv_map_put(struct vm_area_struct *vma)
2179 struct resv_map *reservations = vma_resv_map(vma);
2183 kref_put(&reservations->refs, resv_map_release);
2186 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2188 struct hstate *h = hstate_vma(vma);
2189 struct resv_map *reservations = vma_resv_map(vma);
2190 struct hugepage_subpool *spool = subpool_vma(vma);
2191 unsigned long reserve;
2192 unsigned long start;
2196 start = vma_hugecache_offset(h, vma, vma->vm_start);
2197 end = vma_hugecache_offset(h, vma, vma->vm_end);
2199 reserve = (end - start) -
2200 region_count(&reservations->regions, start, end);
2205 hugetlb_acct_memory(h, -reserve);
2206 hugepage_subpool_put_pages(spool, reserve);
2212 * We cannot handle pagefaults against hugetlb pages at all. They cause
2213 * handle_mm_fault() to try to instantiate regular-sized pages in the
2214 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2217 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2223 const struct vm_operations_struct hugetlb_vm_ops = {
2224 .fault = hugetlb_vm_op_fault,
2225 .open = hugetlb_vm_op_open,
2226 .close = hugetlb_vm_op_close,
2229 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2236 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2238 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2240 entry = pte_mkyoung(entry);
2241 entry = pte_mkhuge(entry);
2246 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2247 unsigned long address, pte_t *ptep)
2251 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2252 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2253 update_mmu_cache(vma, address, ptep);
2257 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2258 struct vm_area_struct *vma)
2260 pte_t *src_pte, *dst_pte, entry;
2261 struct page *ptepage;
2264 struct hstate *h = hstate_vma(vma);
2265 unsigned long sz = huge_page_size(h);
2267 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2269 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2270 src_pte = huge_pte_offset(src, addr);
2273 dst_pte = huge_pte_alloc(dst, addr, sz);
2277 /* If the pagetables are shared don't copy or take references */
2278 if (dst_pte == src_pte)
2281 spin_lock(&dst->page_table_lock);
2282 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2283 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2285 huge_ptep_set_wrprotect(src, addr, src_pte);
2286 entry = huge_ptep_get(src_pte);
2287 ptepage = pte_page(entry);
2289 page_dup_rmap(ptepage);
2290 set_huge_pte_at(dst, addr, dst_pte, entry);
2292 spin_unlock(&src->page_table_lock);
2293 spin_unlock(&dst->page_table_lock);
2301 static int is_hugetlb_entry_migration(pte_t pte)
2305 if (huge_pte_none(pte) || pte_present(pte))
2307 swp = pte_to_swp_entry(pte);
2308 if (non_swap_entry(swp) && is_migration_entry(swp))
2314 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2318 if (huge_pte_none(pte) || pte_present(pte))
2320 swp = pte_to_swp_entry(pte);
2321 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2327 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2328 unsigned long end, struct page *ref_page)
2330 struct mm_struct *mm = vma->vm_mm;
2331 unsigned long address;
2336 struct hstate *h = hstate_vma(vma);
2337 unsigned long sz = huge_page_size(h);
2340 * A page gathering list, protected by per file i_mmap_mutex. The
2341 * lock is used to avoid list corruption from multiple unmapping
2342 * of the same page since we are using page->lru.
2344 LIST_HEAD(page_list);
2346 WARN_ON(!is_vm_hugetlb_page(vma));
2347 BUG_ON(start & ~huge_page_mask(h));
2348 BUG_ON(end & ~huge_page_mask(h));
2350 mmu_notifier_invalidate_range_start(mm, start, end);
2351 spin_lock(&mm->page_table_lock);
2352 for (address = start; address < end; address += sz) {
2353 ptep = huge_pte_offset(mm, address);
2357 if (huge_pmd_unshare(mm, &address, ptep))
2361 * If a reference page is supplied, it is because a specific
2362 * page is being unmapped, not a range. Ensure the page we
2363 * are about to unmap is the actual page of interest.
2366 pte = huge_ptep_get(ptep);
2367 if (huge_pte_none(pte))
2369 page = pte_page(pte);
2370 if (page != ref_page)
2374 * Mark the VMA as having unmapped its page so that
2375 * future faults in this VMA will fail rather than
2376 * looking like data was lost
2378 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2381 pte = huge_ptep_get_and_clear(mm, address, ptep);
2382 if (huge_pte_none(pte))
2386 * HWPoisoned hugepage is already unmapped and dropped reference
2388 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2391 page = pte_page(pte);
2393 set_page_dirty(page);
2394 list_add(&page->lru, &page_list);
2396 spin_unlock(&mm->page_table_lock);
2397 flush_tlb_range(vma, start, end);
2398 mmu_notifier_invalidate_range_end(mm, start, end);
2399 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2400 page_remove_rmap(page);
2401 list_del(&page->lru);
2406 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2407 unsigned long start, unsigned long end,
2408 struct page *ref_page)
2410 __unmap_hugepage_range(vma, start, end, ref_page);
2413 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2414 * test will fail on a vma being torn down, and not grab a page table
2415 * on its way out. We're lucky that the flag has such an appropriate
2416 * name, and can in fact be safely cleared here. We could clear it
2417 * before the __unmap_hugepage_range above, but all that's necessary
2418 * is to clear it before releasing the i_mmap_mutex. This works
2419 * because in the context this is called, the VMA is about to be
2420 * destroyed and the i_mmap_mutex is held.
2422 vma->vm_flags &= ~VM_MAYSHARE;
2425 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2426 unsigned long end, struct page *ref_page)
2428 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2429 __unmap_hugepage_range(vma, start, end, ref_page);
2430 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2434 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2435 * mappping it owns the reserve page for. The intention is to unmap the page
2436 * from other VMAs and let the children be SIGKILLed if they are faulting the
2439 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2440 struct page *page, unsigned long address)
2442 struct hstate *h = hstate_vma(vma);
2443 struct vm_area_struct *iter_vma;
2444 struct address_space *mapping;
2445 struct prio_tree_iter iter;
2449 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2450 * from page cache lookup which is in HPAGE_SIZE units.
2452 address = address & huge_page_mask(h);
2453 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2455 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2458 * Take the mapping lock for the duration of the table walk. As
2459 * this mapping should be shared between all the VMAs,
2460 * __unmap_hugepage_range() is called as the lock is already held
2462 mutex_lock(&mapping->i_mmap_mutex);
2463 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2464 /* Do not unmap the current VMA */
2465 if (iter_vma == vma)
2469 * Unmap the page from other VMAs without their own reserves.
2470 * They get marked to be SIGKILLed if they fault in these
2471 * areas. This is because a future no-page fault on this VMA
2472 * could insert a zeroed page instead of the data existing
2473 * from the time of fork. This would look like data corruption
2475 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2476 __unmap_hugepage_range(iter_vma,
2477 address, address + huge_page_size(h),
2480 mutex_unlock(&mapping->i_mmap_mutex);
2486 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2488 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2489 unsigned long address, pte_t *ptep, pte_t pte,
2490 struct page *pagecache_page)
2492 struct hstate *h = hstate_vma(vma);
2493 struct page *old_page, *new_page;
2495 int outside_reserve = 0;
2497 old_page = pte_page(pte);
2500 /* If no-one else is actually using this page, avoid the copy
2501 * and just make the page writable */
2502 avoidcopy = (page_mapcount(old_page) == 1);
2504 if (PageAnon(old_page))
2505 page_move_anon_rmap(old_page, vma, address);
2506 set_huge_ptep_writable(vma, address, ptep);
2511 * If the process that created a MAP_PRIVATE mapping is about to
2512 * perform a COW due to a shared page count, attempt to satisfy
2513 * the allocation without using the existing reserves. The pagecache
2514 * page is used to determine if the reserve at this address was
2515 * consumed or not. If reserves were used, a partial faulted mapping
2516 * at the time of fork() could consume its reserves on COW instead
2517 * of the full address range.
2519 if (!(vma->vm_flags & VM_MAYSHARE) &&
2520 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2521 old_page != pagecache_page)
2522 outside_reserve = 1;
2524 page_cache_get(old_page);
2526 /* Drop page_table_lock as buddy allocator may be called */
2527 spin_unlock(&mm->page_table_lock);
2528 new_page = alloc_huge_page(vma, address, outside_reserve);
2530 if (IS_ERR(new_page)) {
2531 page_cache_release(old_page);
2534 * If a process owning a MAP_PRIVATE mapping fails to COW,
2535 * it is due to references held by a child and an insufficient
2536 * huge page pool. To guarantee the original mappers
2537 * reliability, unmap the page from child processes. The child
2538 * may get SIGKILLed if it later faults.
2540 if (outside_reserve) {
2541 BUG_ON(huge_pte_none(pte));
2542 if (unmap_ref_private(mm, vma, old_page, address)) {
2543 BUG_ON(huge_pte_none(pte));
2544 spin_lock(&mm->page_table_lock);
2545 goto retry_avoidcopy;
2550 /* Caller expects lock to be held */
2551 spin_lock(&mm->page_table_lock);
2552 return -PTR_ERR(new_page);
2556 * When the original hugepage is shared one, it does not have
2557 * anon_vma prepared.
2559 if (unlikely(anon_vma_prepare(vma))) {
2560 page_cache_release(new_page);
2561 page_cache_release(old_page);
2562 /* Caller expects lock to be held */
2563 spin_lock(&mm->page_table_lock);
2564 return VM_FAULT_OOM;
2567 copy_user_huge_page(new_page, old_page, address, vma,
2568 pages_per_huge_page(h));
2569 __SetPageUptodate(new_page);
2572 * Retake the page_table_lock to check for racing updates
2573 * before the page tables are altered
2575 spin_lock(&mm->page_table_lock);
2576 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2577 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2579 mmu_notifier_invalidate_range_start(mm,
2580 address & huge_page_mask(h),
2581 (address & huge_page_mask(h)) + huge_page_size(h));
2582 huge_ptep_clear_flush(vma, address, ptep);
2583 set_huge_pte_at(mm, address, ptep,
2584 make_huge_pte(vma, new_page, 1));
2585 page_remove_rmap(old_page);
2586 hugepage_add_new_anon_rmap(new_page, vma, address);
2587 /* Make the old page be freed below */
2588 new_page = old_page;
2589 mmu_notifier_invalidate_range_end(mm,
2590 address & huge_page_mask(h),
2591 (address & huge_page_mask(h)) + huge_page_size(h));
2593 page_cache_release(new_page);
2594 page_cache_release(old_page);
2598 /* Return the pagecache page at a given address within a VMA */
2599 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2600 struct vm_area_struct *vma, unsigned long address)
2602 struct address_space *mapping;
2605 mapping = vma->vm_file->f_mapping;
2606 idx = vma_hugecache_offset(h, vma, address);
2608 return find_lock_page(mapping, idx);
2612 * Return whether there is a pagecache page to back given address within VMA.
2613 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2615 static bool hugetlbfs_pagecache_present(struct hstate *h,
2616 struct vm_area_struct *vma, unsigned long address)
2618 struct address_space *mapping;
2622 mapping = vma->vm_file->f_mapping;
2623 idx = vma_hugecache_offset(h, vma, address);
2625 page = find_get_page(mapping, idx);
2628 return page != NULL;
2631 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2632 unsigned long address, pte_t *ptep, unsigned int flags)
2634 struct hstate *h = hstate_vma(vma);
2635 int ret = VM_FAULT_SIGBUS;
2639 struct address_space *mapping;
2643 * Currently, we are forced to kill the process in the event the
2644 * original mapper has unmapped pages from the child due to a failed
2645 * COW. Warn that such a situation has occurred as it may not be obvious
2647 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2649 "PID %d killed due to inadequate hugepage pool\n",
2654 mapping = vma->vm_file->f_mapping;
2655 idx = vma_hugecache_offset(h, vma, address);
2658 * Use page lock to guard against racing truncation
2659 * before we get page_table_lock.
2662 page = find_lock_page(mapping, idx);
2664 size = i_size_read(mapping->host) >> huge_page_shift(h);
2667 page = alloc_huge_page(vma, address, 0);
2669 ret = -PTR_ERR(page);
2672 clear_huge_page(page, address, pages_per_huge_page(h));
2673 __SetPageUptodate(page);
2675 if (vma->vm_flags & VM_MAYSHARE) {
2677 struct inode *inode = mapping->host;
2679 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2687 spin_lock(&inode->i_lock);
2688 inode->i_blocks += blocks_per_huge_page(h);
2689 spin_unlock(&inode->i_lock);
2690 page_dup_rmap(page);
2693 if (unlikely(anon_vma_prepare(vma))) {
2695 goto backout_unlocked;
2697 hugepage_add_new_anon_rmap(page, vma, address);
2701 * If memory error occurs between mmap() and fault, some process
2702 * don't have hwpoisoned swap entry for errored virtual address.
2703 * So we need to block hugepage fault by PG_hwpoison bit check.
2705 if (unlikely(PageHWPoison(page))) {
2706 ret = VM_FAULT_HWPOISON |
2707 VM_FAULT_SET_HINDEX(h - hstates);
2708 goto backout_unlocked;
2710 page_dup_rmap(page);
2714 * If we are going to COW a private mapping later, we examine the
2715 * pending reservations for this page now. This will ensure that
2716 * any allocations necessary to record that reservation occur outside
2719 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2720 if (vma_needs_reservation(h, vma, address) < 0) {
2722 goto backout_unlocked;
2725 spin_lock(&mm->page_table_lock);
2726 size = i_size_read(mapping->host) >> huge_page_shift(h);
2731 if (!huge_pte_none(huge_ptep_get(ptep)))
2734 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2735 && (vma->vm_flags & VM_SHARED)));
2736 set_huge_pte_at(mm, address, ptep, new_pte);
2738 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2739 /* Optimization, do the COW without a second fault */
2740 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2743 spin_unlock(&mm->page_table_lock);
2749 spin_unlock(&mm->page_table_lock);
2756 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2757 unsigned long address, unsigned int flags)
2762 struct page *page = NULL;
2763 struct page *pagecache_page = NULL;
2764 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2765 struct hstate *h = hstate_vma(vma);
2767 ptep = huge_pte_offset(mm, address);
2769 entry = huge_ptep_get(ptep);
2770 if (unlikely(is_hugetlb_entry_migration(entry))) {
2771 migration_entry_wait_huge(mm, ptep);
2773 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2774 return VM_FAULT_HWPOISON_LARGE |
2775 VM_FAULT_SET_HINDEX(h - hstates);
2778 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2780 return VM_FAULT_OOM;
2783 * Serialize hugepage allocation and instantiation, so that we don't
2784 * get spurious allocation failures if two CPUs race to instantiate
2785 * the same page in the page cache.
2787 mutex_lock(&hugetlb_instantiation_mutex);
2788 entry = huge_ptep_get(ptep);
2789 if (huge_pte_none(entry)) {
2790 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2797 * If we are going to COW the mapping later, we examine the pending
2798 * reservations for this page now. This will ensure that any
2799 * allocations necessary to record that reservation occur outside the
2800 * spinlock. For private mappings, we also lookup the pagecache
2801 * page now as it is used to determine if a reservation has been
2804 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2805 if (vma_needs_reservation(h, vma, address) < 0) {
2810 if (!(vma->vm_flags & VM_MAYSHARE))
2811 pagecache_page = hugetlbfs_pagecache_page(h,
2816 * hugetlb_cow() requires page locks of pte_page(entry) and
2817 * pagecache_page, so here we need take the former one
2818 * when page != pagecache_page or !pagecache_page.
2819 * Note that locking order is always pagecache_page -> page,
2820 * so no worry about deadlock.
2822 page = pte_page(entry);
2824 if (page != pagecache_page)
2827 spin_lock(&mm->page_table_lock);
2828 /* Check for a racing update before calling hugetlb_cow */
2829 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2830 goto out_page_table_lock;
2833 if (flags & FAULT_FLAG_WRITE) {
2834 if (!pte_write(entry)) {
2835 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2837 goto out_page_table_lock;
2839 entry = pte_mkdirty(entry);
2841 entry = pte_mkyoung(entry);
2842 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2843 flags & FAULT_FLAG_WRITE))
2844 update_mmu_cache(vma, address, ptep);
2846 out_page_table_lock:
2847 spin_unlock(&mm->page_table_lock);
2849 if (pagecache_page) {
2850 unlock_page(pagecache_page);
2851 put_page(pagecache_page);
2853 if (page != pagecache_page)
2858 mutex_unlock(&hugetlb_instantiation_mutex);
2863 /* Can be overriden by architectures */
2864 __attribute__((weak)) struct page *
2865 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2866 pud_t *pud, int write)
2872 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2873 struct page **pages, struct vm_area_struct **vmas,
2874 unsigned long *position, int *length, int i,
2877 unsigned long pfn_offset;
2878 unsigned long vaddr = *position;
2879 int remainder = *length;
2880 struct hstate *h = hstate_vma(vma);
2882 spin_lock(&mm->page_table_lock);
2883 while (vaddr < vma->vm_end && remainder) {
2889 * Some archs (sparc64, sh*) have multiple pte_ts to
2890 * each hugepage. We have to make sure we get the
2891 * first, for the page indexing below to work.
2893 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2894 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2897 * When coredumping, it suits get_dump_page if we just return
2898 * an error where there's an empty slot with no huge pagecache
2899 * to back it. This way, we avoid allocating a hugepage, and
2900 * the sparse dumpfile avoids allocating disk blocks, but its
2901 * huge holes still show up with zeroes where they need to be.
2903 if (absent && (flags & FOLL_DUMP) &&
2904 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2910 * We need call hugetlb_fault for both hugepages under migration
2911 * (in which case hugetlb_fault waits for the migration,) and
2912 * hwpoisoned hugepages (in which case we need to prevent the
2913 * caller from accessing to them.) In order to do this, we use
2914 * here is_swap_pte instead of is_hugetlb_entry_migration and
2915 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2916 * both cases, and because we can't follow correct pages
2917 * directly from any kind of swap entries.
2919 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2920 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2923 spin_unlock(&mm->page_table_lock);
2924 ret = hugetlb_fault(mm, vma, vaddr,
2925 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2926 spin_lock(&mm->page_table_lock);
2927 if (!(ret & VM_FAULT_ERROR))
2934 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2935 page = pte_page(huge_ptep_get(pte));
2938 pages[i] = mem_map_offset(page, pfn_offset);
2949 if (vaddr < vma->vm_end && remainder &&
2950 pfn_offset < pages_per_huge_page(h)) {
2952 * We use pfn_offset to avoid touching the pageframes
2953 * of this compound page.
2958 spin_unlock(&mm->page_table_lock);
2959 *length = remainder;
2962 return i ? i : -EFAULT;
2965 void hugetlb_change_protection(struct vm_area_struct *vma,
2966 unsigned long address, unsigned long end, pgprot_t newprot)
2968 struct mm_struct *mm = vma->vm_mm;
2969 unsigned long start = address;
2972 struct hstate *h = hstate_vma(vma);
2974 BUG_ON(address >= end);
2975 flush_cache_range(vma, address, end);
2977 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2978 spin_lock(&mm->page_table_lock);
2979 for (; address < end; address += huge_page_size(h)) {
2980 ptep = huge_pte_offset(mm, address);
2983 if (huge_pmd_unshare(mm, &address, ptep))
2985 if (!huge_pte_none(huge_ptep_get(ptep))) {
2986 pte = huge_ptep_get_and_clear(mm, address, ptep);
2987 pte = pte_mkhuge(pte_modify(pte, newprot));
2988 set_huge_pte_at(mm, address, ptep, pte);
2991 spin_unlock(&mm->page_table_lock);
2993 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2994 * may have cleared our pud entry and done put_page on the page table:
2995 * once we release i_mmap_mutex, another task can do the final put_page
2996 * and that page table be reused and filled with junk.
2998 flush_tlb_range(vma, start, end);
2999 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3002 int hugetlb_reserve_pages(struct inode *inode,
3004 struct vm_area_struct *vma,
3005 vm_flags_t vm_flags)
3008 struct hstate *h = hstate_inode(inode);
3009 struct hugepage_subpool *spool = subpool_inode(inode);
3012 * Only apply hugepage reservation if asked. At fault time, an
3013 * attempt will be made for VM_NORESERVE to allocate a page
3014 * without using reserves
3016 if (vm_flags & VM_NORESERVE)
3020 * Shared mappings base their reservation on the number of pages that
3021 * are already allocated on behalf of the file. Private mappings need
3022 * to reserve the full area even if read-only as mprotect() may be
3023 * called to make the mapping read-write. Assume !vma is a shm mapping
3025 if (!vma || vma->vm_flags & VM_MAYSHARE)
3026 chg = region_chg(&inode->i_mapping->private_list, from, to);
3028 struct resv_map *resv_map = resv_map_alloc();
3034 set_vma_resv_map(vma, resv_map);
3035 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3043 /* There must be enough pages in the subpool for the mapping */
3044 if (hugepage_subpool_get_pages(spool, chg)) {
3050 * Check enough hugepages are available for the reservation.
3051 * Hand the pages back to the subpool if there are not
3053 ret = hugetlb_acct_memory(h, chg);
3055 hugepage_subpool_put_pages(spool, chg);
3060 * Account for the reservations made. Shared mappings record regions
3061 * that have reservations as they are shared by multiple VMAs.
3062 * When the last VMA disappears, the region map says how much
3063 * the reservation was and the page cache tells how much of
3064 * the reservation was consumed. Private mappings are per-VMA and
3065 * only the consumed reservations are tracked. When the VMA
3066 * disappears, the original reservation is the VMA size and the
3067 * consumed reservations are stored in the map. Hence, nothing
3068 * else has to be done for private mappings here
3070 if (!vma || vma->vm_flags & VM_MAYSHARE)
3071 region_add(&inode->i_mapping->private_list, from, to);
3079 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3081 struct hstate *h = hstate_inode(inode);
3082 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3083 struct hugepage_subpool *spool = subpool_inode(inode);
3085 spin_lock(&inode->i_lock);
3086 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3087 spin_unlock(&inode->i_lock);
3089 hugepage_subpool_put_pages(spool, (chg - freed));
3090 hugetlb_acct_memory(h, -(chg - freed));
3093 #ifdef CONFIG_MEMORY_FAILURE
3095 /* Should be called in hugetlb_lock */
3096 static int is_hugepage_on_freelist(struct page *hpage)
3100 struct hstate *h = page_hstate(hpage);
3101 int nid = page_to_nid(hpage);
3103 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3110 * This function is called from memory failure code.
3111 * Assume the caller holds page lock of the head page.
3113 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3115 struct hstate *h = page_hstate(hpage);
3116 int nid = page_to_nid(hpage);
3119 spin_lock(&hugetlb_lock);
3120 if (is_hugepage_on_freelist(hpage)) {
3121 list_del(&hpage->lru);
3122 set_page_refcounted(hpage);
3123 h->free_huge_pages--;
3124 h->free_huge_pages_node[nid]--;
3127 spin_unlock(&hugetlb_lock);