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>
30 #include <linux/hugetlb.h>
31 #include <linux/node.h>
34 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
35 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
36 unsigned long hugepages_treat_as_movable;
38 static int hugetlb_max_hstate;
39 unsigned int default_hstate_idx;
40 struct hstate hstates[HUGE_MAX_HSTATE];
42 __initdata LIST_HEAD(huge_boot_pages);
44 /* for command line parsing */
45 static struct hstate * __initdata parsed_hstate;
46 static unsigned long __initdata default_hstate_max_huge_pages;
47 static unsigned long __initdata default_hstate_size;
49 #define for_each_hstate(h) \
50 for ((h) = hstates; (h) < &hstates[hugetlb_max_hstate]; (h)++)
53 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
55 static DEFINE_SPINLOCK(hugetlb_lock);
57 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
59 bool free = (spool->count == 0) && (spool->used_hpages == 0);
61 spin_unlock(&spool->lock);
63 /* If no pages are used, and no other handles to the subpool
64 * remain, free the subpool the subpool remain */
69 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
71 struct hugepage_subpool *spool;
73 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
77 spin_lock_init(&spool->lock);
79 spool->max_hpages = nr_blocks;
80 spool->used_hpages = 0;
85 void hugepage_put_subpool(struct hugepage_subpool *spool)
87 spin_lock(&spool->lock);
88 BUG_ON(!spool->count);
90 unlock_or_release_subpool(spool);
93 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 spin_lock(&spool->lock);
102 if ((spool->used_hpages + delta) <= spool->max_hpages) {
103 spool->used_hpages += delta;
107 spin_unlock(&spool->lock);
112 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
118 spin_lock(&spool->lock);
119 spool->used_hpages -= delta;
120 /* If hugetlbfs_put_super couldn't free spool due to
121 * an outstanding quota reference, free it now. */
122 unlock_or_release_subpool(spool);
125 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
127 return HUGETLBFS_SB(inode->i_sb)->spool;
130 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
132 return subpool_inode(vma->vm_file->f_dentry->d_inode);
136 * Region tracking -- allows tracking of reservations and instantiated pages
137 * across the pages in a mapping.
139 * The region data structures are protected by a combination of the mmap_sem
140 * and the hugetlb_instantion_mutex. To access or modify a region the caller
141 * must either hold the mmap_sem for write, or the mmap_sem for read and
142 * the hugetlb_instantiation mutex:
144 * down_write(&mm->mmap_sem);
146 * down_read(&mm->mmap_sem);
147 * mutex_lock(&hugetlb_instantiation_mutex);
150 struct list_head link;
155 static long region_add(struct list_head *head, long f, long t)
157 struct file_region *rg, *nrg, *trg;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg, head, link)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
171 if (&rg->link == head)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct list_head *head, long f, long t)
193 struct file_region *rg, *nrg;
196 /* Locate the region we are before or in. */
197 list_for_each_entry(rg, head, link)
201 /* If we are below the current region then a new region is required.
202 * Subtle, allocate a new region at the position but make it zero
203 * size such that we can guarantee to record the reservation. */
204 if (&rg->link == head || t < rg->from) {
205 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
210 INIT_LIST_HEAD(&nrg->link);
211 list_add(&nrg->link, rg->link.prev);
216 /* Round our left edge to the current segment if it encloses us. */
221 /* Check for and consume any regions we now overlap with. */
222 list_for_each_entry(rg, rg->link.prev, link) {
223 if (&rg->link == head)
228 /* We overlap with this area, if it extends further than
229 * us then we must extend ourselves. Account for its
230 * existing reservation. */
235 chg -= rg->to - rg->from;
240 static long region_truncate(struct list_head *head, long end)
242 struct file_region *rg, *trg;
245 /* Locate the region we are either in or before. */
246 list_for_each_entry(rg, head, link)
249 if (&rg->link == head)
252 /* If we are in the middle of a region then adjust it. */
253 if (end > rg->from) {
256 rg = list_entry(rg->link.next, typeof(*rg), link);
259 /* Drop any remaining regions. */
260 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
261 if (&rg->link == head)
263 chg += rg->to - rg->from;
270 static long region_count(struct list_head *head, long f, long t)
272 struct file_region *rg;
275 /* Locate each segment we overlap with, and count that overlap. */
276 list_for_each_entry(rg, head, link) {
285 seg_from = max(rg->from, f);
286 seg_to = min(rg->to, t);
288 chg += seg_to - seg_from;
295 * Convert the address within this vma to the page offset within
296 * the mapping, in pagecache page units; huge pages here.
298 static pgoff_t vma_hugecache_offset(struct hstate *h,
299 struct vm_area_struct *vma, unsigned long address)
301 return ((address - vma->vm_start) >> huge_page_shift(h)) +
302 (vma->vm_pgoff >> huge_page_order(h));
305 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
306 unsigned long address)
308 return vma_hugecache_offset(hstate_vma(vma), vma, address);
312 * Return the size of the pages allocated when backing a VMA. In the majority
313 * cases this will be same size as used by the page table entries.
315 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
317 struct hstate *hstate;
319 if (!is_vm_hugetlb_page(vma))
322 hstate = hstate_vma(vma);
324 return 1UL << (hstate->order + PAGE_SHIFT);
326 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
329 * Return the page size being used by the MMU to back a VMA. In the majority
330 * of cases, the page size used by the kernel matches the MMU size. On
331 * architectures where it differs, an architecture-specific version of this
332 * function is required.
334 #ifndef vma_mmu_pagesize
335 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
337 return vma_kernel_pagesize(vma);
342 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
343 * bits of the reservation map pointer, which are always clear due to
346 #define HPAGE_RESV_OWNER (1UL << 0)
347 #define HPAGE_RESV_UNMAPPED (1UL << 1)
348 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
351 * These helpers are used to track how many pages are reserved for
352 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
353 * is guaranteed to have their future faults succeed.
355 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
356 * the reserve counters are updated with the hugetlb_lock held. It is safe
357 * to reset the VMA at fork() time as it is not in use yet and there is no
358 * chance of the global counters getting corrupted as a result of the values.
360 * The private mapping reservation is represented in a subtly different
361 * manner to a shared mapping. A shared mapping has a region map associated
362 * with the underlying file, this region map represents the backing file
363 * pages which have ever had a reservation assigned which this persists even
364 * after the page is instantiated. A private mapping has a region map
365 * associated with the original mmap which is attached to all VMAs which
366 * reference it, this region map represents those offsets which have consumed
367 * reservation ie. where pages have been instantiated.
369 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
371 return (unsigned long)vma->vm_private_data;
374 static void set_vma_private_data(struct vm_area_struct *vma,
377 vma->vm_private_data = (void *)value;
382 struct list_head regions;
385 static struct resv_map *resv_map_alloc(void)
387 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
391 kref_init(&resv_map->refs);
392 INIT_LIST_HEAD(&resv_map->regions);
397 static void resv_map_release(struct kref *ref)
399 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
401 /* Clear out any active regions before we release the map. */
402 region_truncate(&resv_map->regions, 0);
406 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
408 VM_BUG_ON(!is_vm_hugetlb_page(vma));
409 if (!(vma->vm_flags & VM_MAYSHARE))
410 return (struct resv_map *)(get_vma_private_data(vma) &
415 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
417 VM_BUG_ON(!is_vm_hugetlb_page(vma));
418 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
420 set_vma_private_data(vma, (get_vma_private_data(vma) &
421 HPAGE_RESV_MASK) | (unsigned long)map);
424 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
426 VM_BUG_ON(!is_vm_hugetlb_page(vma));
427 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
429 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
432 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
434 VM_BUG_ON(!is_vm_hugetlb_page(vma));
436 return (get_vma_private_data(vma) & flag) != 0;
439 /* Decrement the reserved pages in the hugepage pool by one */
440 static void decrement_hugepage_resv_vma(struct hstate *h,
441 struct vm_area_struct *vma)
443 if (vma->vm_flags & VM_NORESERVE)
446 if (vma->vm_flags & VM_MAYSHARE) {
447 /* Shared mappings always use reserves */
448 h->resv_huge_pages--;
449 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
451 * Only the process that called mmap() has reserves for
454 h->resv_huge_pages--;
458 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
459 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
462 if (!(vma->vm_flags & VM_MAYSHARE))
463 vma->vm_private_data = (void *)0;
466 /* Returns true if the VMA has associated reserve pages */
467 static int vma_has_reserves(struct vm_area_struct *vma)
469 if (vma->vm_flags & VM_MAYSHARE)
471 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
476 static void copy_gigantic_page(struct page *dst, struct page *src)
479 struct hstate *h = page_hstate(src);
480 struct page *dst_base = dst;
481 struct page *src_base = src;
483 for (i = 0; i < pages_per_huge_page(h); ) {
485 copy_highpage(dst, src);
488 dst = mem_map_next(dst, dst_base, i);
489 src = mem_map_next(src, src_base, i);
493 void copy_huge_page(struct page *dst, struct page *src)
496 struct hstate *h = page_hstate(src);
498 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
499 copy_gigantic_page(dst, src);
504 for (i = 0; i < pages_per_huge_page(h); i++) {
506 copy_highpage(dst + i, src + i);
510 static void enqueue_huge_page(struct hstate *h, struct page *page)
512 int nid = page_to_nid(page);
513 list_move(&page->lru, &h->hugepage_freelists[nid]);
514 h->free_huge_pages++;
515 h->free_huge_pages_node[nid]++;
518 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 if (list_empty(&h->hugepage_freelists[nid]))
524 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
525 list_move(&page->lru, &h->hugepage_activelist);
526 set_page_refcounted(page);
527 h->free_huge_pages--;
528 h->free_huge_pages_node[nid]--;
532 static struct page *dequeue_huge_page_vma(struct hstate *h,
533 struct vm_area_struct *vma,
534 unsigned long address, int avoid_reserve)
536 struct page *page = NULL;
537 struct mempolicy *mpol;
538 nodemask_t *nodemask;
539 struct zonelist *zonelist;
542 unsigned int cpuset_mems_cookie;
545 cpuset_mems_cookie = get_mems_allowed();
546 zonelist = huge_zonelist(vma, address,
547 htlb_alloc_mask, &mpol, &nodemask);
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma) &&
554 h->free_huge_pages - h->resv_huge_pages == 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 for_each_zone_zonelist_nodemask(zone, z, zonelist,
562 MAX_NR_ZONES - 1, nodemask) {
563 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
564 page = dequeue_huge_page_node(h, zone_to_nid(zone));
567 decrement_hugepage_resv_vma(h, vma);
574 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
583 static void update_and_free_page(struct hstate *h, struct page *page)
587 VM_BUG_ON(h->order >= MAX_ORDER);
590 h->nr_huge_pages_node[page_to_nid(page)]--;
591 for (i = 0; i < pages_per_huge_page(h); i++) {
592 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
593 1 << PG_referenced | 1 << PG_dirty |
594 1 << PG_active | 1 << PG_reserved |
595 1 << PG_private | 1 << PG_writeback);
597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
599 arch_release_hugepage(page);
600 __free_pages(page, huge_page_order(h));
603 struct hstate *size_to_hstate(unsigned long size)
608 if (huge_page_size(h) == size)
614 static void free_huge_page(struct page *page)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate *h = page_hstate(page);
621 int nid = page_to_nid(page);
622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
625 set_page_private(page, 0);
626 page->mapping = NULL;
627 BUG_ON(page_count(page));
628 BUG_ON(page_mapcount(page));
630 spin_lock(&hugetlb_lock);
631 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
632 /* remove the page from active list */
633 list_del(&page->lru);
634 update_and_free_page(h, page);
635 h->surplus_huge_pages--;
636 h->surplus_huge_pages_node[nid]--;
638 enqueue_huge_page(h, page);
640 spin_unlock(&hugetlb_lock);
641 hugepage_subpool_put_pages(spool, 1);
644 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
646 INIT_LIST_HEAD(&page->lru);
647 set_compound_page_dtor(page, free_huge_page);
648 spin_lock(&hugetlb_lock);
650 h->nr_huge_pages_node[nid]++;
651 spin_unlock(&hugetlb_lock);
652 put_page(page); /* free it into the hugepage allocator */
655 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
658 int nr_pages = 1 << order;
659 struct page *p = page + 1;
661 /* we rely on prep_new_huge_page to set the destructor */
662 set_compound_order(page, order);
664 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
666 set_page_count(p, 0);
667 p->first_page = page;
671 int PageHuge(struct page *page)
673 compound_page_dtor *dtor;
675 if (!PageCompound(page))
678 page = compound_head(page);
679 dtor = get_compound_page_dtor(page);
681 return dtor == free_huge_page;
683 EXPORT_SYMBOL_GPL(PageHuge);
685 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
689 if (h->order >= MAX_ORDER)
692 page = alloc_pages_exact_node(nid,
693 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
694 __GFP_REPEAT|__GFP_NOWARN,
697 if (arch_prepare_hugepage(page)) {
698 __free_pages(page, huge_page_order(h));
701 prep_new_huge_page(h, page, nid);
708 * common helper functions for hstate_next_node_to_{alloc|free}.
709 * We may have allocated or freed a huge page based on a different
710 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
711 * be outside of *nodes_allowed. Ensure that we use an allowed
712 * node for alloc or free.
714 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
716 nid = next_node(nid, *nodes_allowed);
717 if (nid == MAX_NUMNODES)
718 nid = first_node(*nodes_allowed);
719 VM_BUG_ON(nid >= MAX_NUMNODES);
724 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
726 if (!node_isset(nid, *nodes_allowed))
727 nid = next_node_allowed(nid, nodes_allowed);
732 * returns the previously saved node ["this node"] from which to
733 * allocate a persistent huge page for the pool and advance the
734 * next node from which to allocate, handling wrap at end of node
737 static int hstate_next_node_to_alloc(struct hstate *h,
738 nodemask_t *nodes_allowed)
742 VM_BUG_ON(!nodes_allowed);
744 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
745 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
750 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
757 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
758 next_nid = start_nid;
761 page = alloc_fresh_huge_page_node(h, next_nid);
766 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
767 } while (next_nid != start_nid);
770 count_vm_event(HTLB_BUDDY_PGALLOC);
772 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
778 * helper for free_pool_huge_page() - return the previously saved
779 * node ["this node"] from which to free a huge page. Advance the
780 * next node id whether or not we find a free huge page to free so
781 * that the next attempt to free addresses the next node.
783 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
787 VM_BUG_ON(!nodes_allowed);
789 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
790 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
796 * Free huge page from pool from next node to free.
797 * Attempt to keep persistent huge pages more or less
798 * balanced over allowed nodes.
799 * Called with hugetlb_lock locked.
801 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
808 start_nid = hstate_next_node_to_free(h, nodes_allowed);
809 next_nid = start_nid;
813 * If we're returning unused surplus pages, only examine
814 * nodes with surplus pages.
816 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
817 !list_empty(&h->hugepage_freelists[next_nid])) {
819 list_entry(h->hugepage_freelists[next_nid].next,
821 list_del(&page->lru);
822 h->free_huge_pages--;
823 h->free_huge_pages_node[next_nid]--;
825 h->surplus_huge_pages--;
826 h->surplus_huge_pages_node[next_nid]--;
828 update_and_free_page(h, page);
832 next_nid = hstate_next_node_to_free(h, nodes_allowed);
833 } while (next_nid != start_nid);
838 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
843 if (h->order >= MAX_ORDER)
847 * Assume we will successfully allocate the surplus page to
848 * prevent racing processes from causing the surplus to exceed
851 * This however introduces a different race, where a process B
852 * tries to grow the static hugepage pool while alloc_pages() is
853 * called by process A. B will only examine the per-node
854 * counters in determining if surplus huge pages can be
855 * converted to normal huge pages in adjust_pool_surplus(). A
856 * won't be able to increment the per-node counter, until the
857 * lock is dropped by B, but B doesn't drop hugetlb_lock until
858 * no more huge pages can be converted from surplus to normal
859 * state (and doesn't try to convert again). Thus, we have a
860 * case where a surplus huge page exists, the pool is grown, and
861 * the surplus huge page still exists after, even though it
862 * should just have been converted to a normal huge page. This
863 * does not leak memory, though, as the hugepage will be freed
864 * once it is out of use. It also does not allow the counters to
865 * go out of whack in adjust_pool_surplus() as we don't modify
866 * the node values until we've gotten the hugepage and only the
867 * per-node value is checked there.
869 spin_lock(&hugetlb_lock);
870 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
871 spin_unlock(&hugetlb_lock);
875 h->surplus_huge_pages++;
877 spin_unlock(&hugetlb_lock);
879 if (nid == NUMA_NO_NODE)
880 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
881 __GFP_REPEAT|__GFP_NOWARN,
884 page = alloc_pages_exact_node(nid,
885 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
886 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
888 if (page && arch_prepare_hugepage(page)) {
889 __free_pages(page, huge_page_order(h));
893 spin_lock(&hugetlb_lock);
895 INIT_LIST_HEAD(&page->lru);
896 r_nid = page_to_nid(page);
897 set_compound_page_dtor(page, free_huge_page);
899 * We incremented the global counters already
901 h->nr_huge_pages_node[r_nid]++;
902 h->surplus_huge_pages_node[r_nid]++;
903 __count_vm_event(HTLB_BUDDY_PGALLOC);
906 h->surplus_huge_pages--;
907 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
909 spin_unlock(&hugetlb_lock);
915 * This allocation function is useful in the context where vma is irrelevant.
916 * E.g. soft-offlining uses this function because it only cares physical
917 * address of error page.
919 struct page *alloc_huge_page_node(struct hstate *h, int nid)
923 spin_lock(&hugetlb_lock);
924 page = dequeue_huge_page_node(h, nid);
925 spin_unlock(&hugetlb_lock);
928 page = alloc_buddy_huge_page(h, nid);
934 * Increase the hugetlb pool such that it can accommodate a reservation
937 static int gather_surplus_pages(struct hstate *h, int delta)
939 struct list_head surplus_list;
940 struct page *page, *tmp;
942 int needed, allocated;
943 bool alloc_ok = true;
945 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
947 h->resv_huge_pages += delta;
952 INIT_LIST_HEAD(&surplus_list);
956 spin_unlock(&hugetlb_lock);
957 for (i = 0; i < needed; i++) {
958 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
963 list_add(&page->lru, &surplus_list);
968 * After retaking hugetlb_lock, we need to recalculate 'needed'
969 * because either resv_huge_pages or free_huge_pages may have changed.
971 spin_lock(&hugetlb_lock);
972 needed = (h->resv_huge_pages + delta) -
973 (h->free_huge_pages + allocated);
978 * We were not able to allocate enough pages to
979 * satisfy the entire reservation so we free what
980 * we've allocated so far.
985 * The surplus_list now contains _at_least_ the number of extra pages
986 * needed to accommodate the reservation. Add the appropriate number
987 * of pages to the hugetlb pool and free the extras back to the buddy
988 * allocator. Commit the entire reservation here to prevent another
989 * process from stealing the pages as they are added to the pool but
990 * before they are reserved.
993 h->resv_huge_pages += delta;
996 /* Free the needed pages to the hugetlb pool */
997 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1001 * This page is now managed by the hugetlb allocator and has
1002 * no users -- drop the buddy allocator's reference.
1004 put_page_testzero(page);
1005 VM_BUG_ON(page_count(page));
1006 enqueue_huge_page(h, page);
1009 spin_unlock(&hugetlb_lock);
1011 /* Free unnecessary surplus pages to the buddy allocator */
1012 if (!list_empty(&surplus_list)) {
1013 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1017 spin_lock(&hugetlb_lock);
1023 * When releasing a hugetlb pool reservation, any surplus pages that were
1024 * allocated to satisfy the reservation must be explicitly freed if they were
1026 * Called with hugetlb_lock held.
1028 static void return_unused_surplus_pages(struct hstate *h,
1029 unsigned long unused_resv_pages)
1031 unsigned long nr_pages;
1033 /* Uncommit the reservation */
1034 h->resv_huge_pages -= unused_resv_pages;
1036 /* Cannot return gigantic pages currently */
1037 if (h->order >= MAX_ORDER)
1040 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1043 * We want to release as many surplus pages as possible, spread
1044 * evenly across all nodes with memory. Iterate across these nodes
1045 * until we can no longer free unreserved surplus pages. This occurs
1046 * when the nodes with surplus pages have no free pages.
1047 * free_pool_huge_page() will balance the the freed pages across the
1048 * on-line nodes with memory and will handle the hstate accounting.
1050 while (nr_pages--) {
1051 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1057 * Determine if the huge page at addr within the vma has an associated
1058 * reservation. Where it does not we will need to logically increase
1059 * reservation and actually increase subpool usage before an allocation
1060 * can occur. Where any new reservation would be required the
1061 * reservation change is prepared, but not committed. Once the page
1062 * has been allocated from the subpool and instantiated the change should
1063 * be committed via vma_commit_reservation. No action is required on
1066 static long vma_needs_reservation(struct hstate *h,
1067 struct vm_area_struct *vma, unsigned long addr)
1069 struct address_space *mapping = vma->vm_file->f_mapping;
1070 struct inode *inode = mapping->host;
1072 if (vma->vm_flags & VM_MAYSHARE) {
1073 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1074 return region_chg(&inode->i_mapping->private_list,
1077 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1082 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1083 struct resv_map *reservations = vma_resv_map(vma);
1085 err = region_chg(&reservations->regions, idx, idx + 1);
1091 static void vma_commit_reservation(struct hstate *h,
1092 struct vm_area_struct *vma, unsigned long addr)
1094 struct address_space *mapping = vma->vm_file->f_mapping;
1095 struct inode *inode = mapping->host;
1097 if (vma->vm_flags & VM_MAYSHARE) {
1098 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1099 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1101 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1102 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1103 struct resv_map *reservations = vma_resv_map(vma);
1105 /* Mark this page used in the map. */
1106 region_add(&reservations->regions, idx, idx + 1);
1110 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1111 unsigned long addr, int avoid_reserve)
1113 struct hugepage_subpool *spool = subpool_vma(vma);
1114 struct hstate *h = hstate_vma(vma);
1119 * Processes that did not create the mapping will have no
1120 * reserves and will not have accounted against subpool
1121 * limit. Check that the subpool limit can be made before
1122 * satisfying the allocation MAP_NORESERVE mappings may also
1123 * need pages and subpool limit allocated allocated if no reserve
1126 chg = vma_needs_reservation(h, vma, addr);
1128 return ERR_PTR(-ENOMEM);
1130 if (hugepage_subpool_get_pages(spool, chg))
1131 return ERR_PTR(-ENOSPC);
1133 spin_lock(&hugetlb_lock);
1134 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1135 spin_unlock(&hugetlb_lock);
1138 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1140 hugepage_subpool_put_pages(spool, chg);
1141 return ERR_PTR(-ENOSPC);
1145 set_page_private(page, (unsigned long)spool);
1147 vma_commit_reservation(h, vma, addr);
1152 int __weak alloc_bootmem_huge_page(struct hstate *h)
1154 struct huge_bootmem_page *m;
1155 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1160 addr = __alloc_bootmem_node_nopanic(
1161 NODE_DATA(hstate_next_node_to_alloc(h,
1162 &node_states[N_HIGH_MEMORY])),
1163 huge_page_size(h), huge_page_size(h), 0);
1167 * Use the beginning of the huge page to store the
1168 * huge_bootmem_page struct (until gather_bootmem
1169 * puts them into the mem_map).
1179 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1180 /* Put them into a private list first because mem_map is not up yet */
1181 list_add(&m->list, &huge_boot_pages);
1186 static void prep_compound_huge_page(struct page *page, int order)
1188 if (unlikely(order > (MAX_ORDER - 1)))
1189 prep_compound_gigantic_page(page, order);
1191 prep_compound_page(page, order);
1194 /* Put bootmem huge pages into the standard lists after mem_map is up */
1195 static void __init gather_bootmem_prealloc(void)
1197 struct huge_bootmem_page *m;
1199 list_for_each_entry(m, &huge_boot_pages, list) {
1200 struct hstate *h = m->hstate;
1203 #ifdef CONFIG_HIGHMEM
1204 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1205 free_bootmem_late((unsigned long)m,
1206 sizeof(struct huge_bootmem_page));
1208 page = virt_to_page(m);
1210 __ClearPageReserved(page);
1211 WARN_ON(page_count(page) != 1);
1212 prep_compound_huge_page(page, h->order);
1213 prep_new_huge_page(h, page, page_to_nid(page));
1215 * If we had gigantic hugepages allocated at boot time, we need
1216 * to restore the 'stolen' pages to totalram_pages in order to
1217 * fix confusing memory reports from free(1) and another
1218 * side-effects, like CommitLimit going negative.
1220 if (h->order > (MAX_ORDER - 1))
1221 totalram_pages += 1 << h->order;
1225 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1229 for (i = 0; i < h->max_huge_pages; ++i) {
1230 if (h->order >= MAX_ORDER) {
1231 if (!alloc_bootmem_huge_page(h))
1233 } else if (!alloc_fresh_huge_page(h,
1234 &node_states[N_HIGH_MEMORY]))
1237 h->max_huge_pages = i;
1240 static void __init hugetlb_init_hstates(void)
1244 for_each_hstate(h) {
1245 /* oversize hugepages were init'ed in early boot */
1246 if (h->order < MAX_ORDER)
1247 hugetlb_hstate_alloc_pages(h);
1251 static char * __init memfmt(char *buf, unsigned long n)
1253 if (n >= (1UL << 30))
1254 sprintf(buf, "%lu GB", n >> 30);
1255 else if (n >= (1UL << 20))
1256 sprintf(buf, "%lu MB", n >> 20);
1258 sprintf(buf, "%lu KB", n >> 10);
1262 static void __init report_hugepages(void)
1266 for_each_hstate(h) {
1268 printk(KERN_INFO "HugeTLB registered %s page size, "
1269 "pre-allocated %ld pages\n",
1270 memfmt(buf, huge_page_size(h)),
1271 h->free_huge_pages);
1275 #ifdef CONFIG_HIGHMEM
1276 static void try_to_free_low(struct hstate *h, unsigned long count,
1277 nodemask_t *nodes_allowed)
1281 if (h->order >= MAX_ORDER)
1284 for_each_node_mask(i, *nodes_allowed) {
1285 struct page *page, *next;
1286 struct list_head *freel = &h->hugepage_freelists[i];
1287 list_for_each_entry_safe(page, next, freel, lru) {
1288 if (count >= h->nr_huge_pages)
1290 if (PageHighMem(page))
1292 list_del(&page->lru);
1293 update_and_free_page(h, page);
1294 h->free_huge_pages--;
1295 h->free_huge_pages_node[page_to_nid(page)]--;
1300 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1301 nodemask_t *nodes_allowed)
1307 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1308 * balanced by operating on them in a round-robin fashion.
1309 * Returns 1 if an adjustment was made.
1311 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1314 int start_nid, next_nid;
1317 VM_BUG_ON(delta != -1 && delta != 1);
1320 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1322 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1323 next_nid = start_nid;
1329 * To shrink on this node, there must be a surplus page
1331 if (!h->surplus_huge_pages_node[nid]) {
1332 next_nid = hstate_next_node_to_alloc(h,
1339 * Surplus cannot exceed the total number of pages
1341 if (h->surplus_huge_pages_node[nid] >=
1342 h->nr_huge_pages_node[nid]) {
1343 next_nid = hstate_next_node_to_free(h,
1349 h->surplus_huge_pages += delta;
1350 h->surplus_huge_pages_node[nid] += delta;
1353 } while (next_nid != start_nid);
1358 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1359 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1360 nodemask_t *nodes_allowed)
1362 unsigned long min_count, ret;
1364 if (h->order >= MAX_ORDER)
1365 return h->max_huge_pages;
1368 * Increase the pool size
1369 * First take pages out of surplus state. Then make up the
1370 * remaining difference by allocating fresh huge pages.
1372 * We might race with alloc_buddy_huge_page() here and be unable
1373 * to convert a surplus huge page to a normal huge page. That is
1374 * not critical, though, it just means the overall size of the
1375 * pool might be one hugepage larger than it needs to be, but
1376 * within all the constraints specified by the sysctls.
1378 spin_lock(&hugetlb_lock);
1379 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1380 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1384 while (count > persistent_huge_pages(h)) {
1386 * If this allocation races such that we no longer need the
1387 * page, free_huge_page will handle it by freeing the page
1388 * and reducing the surplus.
1390 spin_unlock(&hugetlb_lock);
1391 ret = alloc_fresh_huge_page(h, nodes_allowed);
1392 spin_lock(&hugetlb_lock);
1396 /* Bail for signals. Probably ctrl-c from user */
1397 if (signal_pending(current))
1402 * Decrease the pool size
1403 * First return free pages to the buddy allocator (being careful
1404 * to keep enough around to satisfy reservations). Then place
1405 * pages into surplus state as needed so the pool will shrink
1406 * to the desired size as pages become free.
1408 * By placing pages into the surplus state independent of the
1409 * overcommit value, we are allowing the surplus pool size to
1410 * exceed overcommit. There are few sane options here. Since
1411 * alloc_buddy_huge_page() is checking the global counter,
1412 * though, we'll note that we're not allowed to exceed surplus
1413 * and won't grow the pool anywhere else. Not until one of the
1414 * sysctls are changed, or the surplus pages go out of use.
1416 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1417 min_count = max(count, min_count);
1418 try_to_free_low(h, min_count, nodes_allowed);
1419 while (min_count < persistent_huge_pages(h)) {
1420 if (!free_pool_huge_page(h, nodes_allowed, 0))
1423 while (count < persistent_huge_pages(h)) {
1424 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1428 ret = persistent_huge_pages(h);
1429 spin_unlock(&hugetlb_lock);
1433 #define HSTATE_ATTR_RO(_name) \
1434 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1436 #define HSTATE_ATTR(_name) \
1437 static struct kobj_attribute _name##_attr = \
1438 __ATTR(_name, 0644, _name##_show, _name##_store)
1440 static struct kobject *hugepages_kobj;
1441 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1443 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1445 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1449 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1450 if (hstate_kobjs[i] == kobj) {
1452 *nidp = NUMA_NO_NODE;
1456 return kobj_to_node_hstate(kobj, nidp);
1459 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1460 struct kobj_attribute *attr, char *buf)
1463 unsigned long nr_huge_pages;
1466 h = kobj_to_hstate(kobj, &nid);
1467 if (nid == NUMA_NO_NODE)
1468 nr_huge_pages = h->nr_huge_pages;
1470 nr_huge_pages = h->nr_huge_pages_node[nid];
1472 return sprintf(buf, "%lu\n", nr_huge_pages);
1475 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1476 struct kobject *kobj, struct kobj_attribute *attr,
1477 const char *buf, size_t len)
1481 unsigned long count;
1483 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1485 err = strict_strtoul(buf, 10, &count);
1489 h = kobj_to_hstate(kobj, &nid);
1490 if (h->order >= MAX_ORDER) {
1495 if (nid == NUMA_NO_NODE) {
1497 * global hstate attribute
1499 if (!(obey_mempolicy &&
1500 init_nodemask_of_mempolicy(nodes_allowed))) {
1501 NODEMASK_FREE(nodes_allowed);
1502 nodes_allowed = &node_states[N_HIGH_MEMORY];
1504 } else if (nodes_allowed) {
1506 * per node hstate attribute: adjust count to global,
1507 * but restrict alloc/free to the specified node.
1509 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1510 init_nodemask_of_node(nodes_allowed, nid);
1512 nodes_allowed = &node_states[N_HIGH_MEMORY];
1514 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1516 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1517 NODEMASK_FREE(nodes_allowed);
1521 NODEMASK_FREE(nodes_allowed);
1525 static ssize_t nr_hugepages_show(struct kobject *kobj,
1526 struct kobj_attribute *attr, char *buf)
1528 return nr_hugepages_show_common(kobj, attr, buf);
1531 static ssize_t nr_hugepages_store(struct kobject *kobj,
1532 struct kobj_attribute *attr, const char *buf, size_t len)
1534 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1536 HSTATE_ATTR(nr_hugepages);
1541 * hstate attribute for optionally mempolicy-based constraint on persistent
1542 * huge page alloc/free.
1544 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1545 struct kobj_attribute *attr, char *buf)
1547 return nr_hugepages_show_common(kobj, attr, buf);
1550 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1551 struct kobj_attribute *attr, const char *buf, size_t len)
1553 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1555 HSTATE_ATTR(nr_hugepages_mempolicy);
1559 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1560 struct kobj_attribute *attr, char *buf)
1562 struct hstate *h = kobj_to_hstate(kobj, NULL);
1563 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1566 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1567 struct kobj_attribute *attr, const char *buf, size_t count)
1570 unsigned long input;
1571 struct hstate *h = kobj_to_hstate(kobj, NULL);
1573 if (h->order >= MAX_ORDER)
1576 err = strict_strtoul(buf, 10, &input);
1580 spin_lock(&hugetlb_lock);
1581 h->nr_overcommit_huge_pages = input;
1582 spin_unlock(&hugetlb_lock);
1586 HSTATE_ATTR(nr_overcommit_hugepages);
1588 static ssize_t free_hugepages_show(struct kobject *kobj,
1589 struct kobj_attribute *attr, char *buf)
1592 unsigned long free_huge_pages;
1595 h = kobj_to_hstate(kobj, &nid);
1596 if (nid == NUMA_NO_NODE)
1597 free_huge_pages = h->free_huge_pages;
1599 free_huge_pages = h->free_huge_pages_node[nid];
1601 return sprintf(buf, "%lu\n", free_huge_pages);
1603 HSTATE_ATTR_RO(free_hugepages);
1605 static ssize_t resv_hugepages_show(struct kobject *kobj,
1606 struct kobj_attribute *attr, char *buf)
1608 struct hstate *h = kobj_to_hstate(kobj, NULL);
1609 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1611 HSTATE_ATTR_RO(resv_hugepages);
1613 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1614 struct kobj_attribute *attr, char *buf)
1617 unsigned long surplus_huge_pages;
1620 h = kobj_to_hstate(kobj, &nid);
1621 if (nid == NUMA_NO_NODE)
1622 surplus_huge_pages = h->surplus_huge_pages;
1624 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1626 return sprintf(buf, "%lu\n", surplus_huge_pages);
1628 HSTATE_ATTR_RO(surplus_hugepages);
1630 static struct attribute *hstate_attrs[] = {
1631 &nr_hugepages_attr.attr,
1632 &nr_overcommit_hugepages_attr.attr,
1633 &free_hugepages_attr.attr,
1634 &resv_hugepages_attr.attr,
1635 &surplus_hugepages_attr.attr,
1637 &nr_hugepages_mempolicy_attr.attr,
1642 static struct attribute_group hstate_attr_group = {
1643 .attrs = hstate_attrs,
1646 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1647 struct kobject **hstate_kobjs,
1648 struct attribute_group *hstate_attr_group)
1651 int hi = hstate_index(h);
1653 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1654 if (!hstate_kobjs[hi])
1657 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1659 kobject_put(hstate_kobjs[hi]);
1664 static void __init hugetlb_sysfs_init(void)
1669 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1670 if (!hugepages_kobj)
1673 for_each_hstate(h) {
1674 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1675 hstate_kobjs, &hstate_attr_group);
1677 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1685 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1686 * with node devices in node_devices[] using a parallel array. The array
1687 * index of a node device or _hstate == node id.
1688 * This is here to avoid any static dependency of the node device driver, in
1689 * the base kernel, on the hugetlb module.
1691 struct node_hstate {
1692 struct kobject *hugepages_kobj;
1693 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1695 struct node_hstate node_hstates[MAX_NUMNODES];
1698 * A subset of global hstate attributes for node devices
1700 static struct attribute *per_node_hstate_attrs[] = {
1701 &nr_hugepages_attr.attr,
1702 &free_hugepages_attr.attr,
1703 &surplus_hugepages_attr.attr,
1707 static struct attribute_group per_node_hstate_attr_group = {
1708 .attrs = per_node_hstate_attrs,
1712 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1713 * Returns node id via non-NULL nidp.
1715 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1719 for (nid = 0; nid < nr_node_ids; nid++) {
1720 struct node_hstate *nhs = &node_hstates[nid];
1722 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1723 if (nhs->hstate_kobjs[i] == kobj) {
1735 * Unregister hstate attributes from a single node device.
1736 * No-op if no hstate attributes attached.
1738 void hugetlb_unregister_node(struct node *node)
1741 struct node_hstate *nhs = &node_hstates[node->dev.id];
1743 if (!nhs->hugepages_kobj)
1744 return; /* no hstate attributes */
1746 for_each_hstate(h) {
1747 int idx = hstate_index(h);
1748 if (nhs->hstate_kobjs[idx]) {
1749 kobject_put(nhs->hstate_kobjs[idx]);
1750 nhs->hstate_kobjs[idx] = NULL;
1754 kobject_put(nhs->hugepages_kobj);
1755 nhs->hugepages_kobj = NULL;
1759 * hugetlb module exit: unregister hstate attributes from node devices
1762 static void hugetlb_unregister_all_nodes(void)
1767 * disable node device registrations.
1769 register_hugetlbfs_with_node(NULL, NULL);
1772 * remove hstate attributes from any nodes that have them.
1774 for (nid = 0; nid < nr_node_ids; nid++)
1775 hugetlb_unregister_node(&node_devices[nid]);
1779 * Register hstate attributes for a single node device.
1780 * No-op if attributes already registered.
1782 void hugetlb_register_node(struct node *node)
1785 struct node_hstate *nhs = &node_hstates[node->dev.id];
1788 if (nhs->hugepages_kobj)
1789 return; /* already allocated */
1791 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1793 if (!nhs->hugepages_kobj)
1796 for_each_hstate(h) {
1797 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1799 &per_node_hstate_attr_group);
1801 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1803 h->name, node->dev.id);
1804 hugetlb_unregister_node(node);
1811 * hugetlb init time: register hstate attributes for all registered node
1812 * devices of nodes that have memory. All on-line nodes should have
1813 * registered their associated device by this time.
1815 static void hugetlb_register_all_nodes(void)
1819 for_each_node_state(nid, N_HIGH_MEMORY) {
1820 struct node *node = &node_devices[nid];
1821 if (node->dev.id == nid)
1822 hugetlb_register_node(node);
1826 * Let the node device driver know we're here so it can
1827 * [un]register hstate attributes on node hotplug.
1829 register_hugetlbfs_with_node(hugetlb_register_node,
1830 hugetlb_unregister_node);
1832 #else /* !CONFIG_NUMA */
1834 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1842 static void hugetlb_unregister_all_nodes(void) { }
1844 static void hugetlb_register_all_nodes(void) { }
1848 static void __exit hugetlb_exit(void)
1852 hugetlb_unregister_all_nodes();
1854 for_each_hstate(h) {
1855 kobject_put(hstate_kobjs[hstate_index(h)]);
1858 kobject_put(hugepages_kobj);
1860 module_exit(hugetlb_exit);
1862 static int __init hugetlb_init(void)
1864 /* Some platform decide whether they support huge pages at boot
1865 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1866 * there is no such support
1868 if (HPAGE_SHIFT == 0)
1871 if (!size_to_hstate(default_hstate_size)) {
1872 default_hstate_size = HPAGE_SIZE;
1873 if (!size_to_hstate(default_hstate_size))
1874 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1876 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1877 if (default_hstate_max_huge_pages)
1878 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1880 hugetlb_init_hstates();
1882 gather_bootmem_prealloc();
1886 hugetlb_sysfs_init();
1888 hugetlb_register_all_nodes();
1892 module_init(hugetlb_init);
1894 /* Should be called on processing a hugepagesz=... option */
1895 void __init hugetlb_add_hstate(unsigned order)
1900 if (size_to_hstate(PAGE_SIZE << order)) {
1901 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1904 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1906 h = &hstates[hugetlb_max_hstate++];
1908 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1909 h->nr_huge_pages = 0;
1910 h->free_huge_pages = 0;
1911 for (i = 0; i < MAX_NUMNODES; ++i)
1912 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1913 INIT_LIST_HEAD(&h->hugepage_activelist);
1914 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1915 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1916 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1917 huge_page_size(h)/1024);
1922 static int __init hugetlb_nrpages_setup(char *s)
1925 static unsigned long *last_mhp;
1928 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1929 * so this hugepages= parameter goes to the "default hstate".
1931 if (!hugetlb_max_hstate)
1932 mhp = &default_hstate_max_huge_pages;
1934 mhp = &parsed_hstate->max_huge_pages;
1936 if (mhp == last_mhp) {
1937 printk(KERN_WARNING "hugepages= specified twice without "
1938 "interleaving hugepagesz=, ignoring\n");
1942 if (sscanf(s, "%lu", mhp) <= 0)
1946 * Global state is always initialized later in hugetlb_init.
1947 * But we need to allocate >= MAX_ORDER hstates here early to still
1948 * use the bootmem allocator.
1950 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1951 hugetlb_hstate_alloc_pages(parsed_hstate);
1957 __setup("hugepages=", hugetlb_nrpages_setup);
1959 static int __init hugetlb_default_setup(char *s)
1961 default_hstate_size = memparse(s, &s);
1964 __setup("default_hugepagesz=", hugetlb_default_setup);
1966 static unsigned int cpuset_mems_nr(unsigned int *array)
1969 unsigned int nr = 0;
1971 for_each_node_mask(node, cpuset_current_mems_allowed)
1977 #ifdef CONFIG_SYSCTL
1978 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1979 struct ctl_table *table, int write,
1980 void __user *buffer, size_t *length, loff_t *ppos)
1982 struct hstate *h = &default_hstate;
1986 tmp = h->max_huge_pages;
1988 if (write && h->order >= MAX_ORDER)
1992 table->maxlen = sizeof(unsigned long);
1993 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1998 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1999 GFP_KERNEL | __GFP_NORETRY);
2000 if (!(obey_mempolicy &&
2001 init_nodemask_of_mempolicy(nodes_allowed))) {
2002 NODEMASK_FREE(nodes_allowed);
2003 nodes_allowed = &node_states[N_HIGH_MEMORY];
2005 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2007 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2008 NODEMASK_FREE(nodes_allowed);
2014 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2015 void __user *buffer, size_t *length, loff_t *ppos)
2018 return hugetlb_sysctl_handler_common(false, table, write,
2019 buffer, length, ppos);
2023 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2024 void __user *buffer, size_t *length, loff_t *ppos)
2026 return hugetlb_sysctl_handler_common(true, table, write,
2027 buffer, length, ppos);
2029 #endif /* CONFIG_NUMA */
2031 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2032 void __user *buffer,
2033 size_t *length, loff_t *ppos)
2035 proc_dointvec(table, write, buffer, length, ppos);
2036 if (hugepages_treat_as_movable)
2037 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2039 htlb_alloc_mask = GFP_HIGHUSER;
2043 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2044 void __user *buffer,
2045 size_t *length, loff_t *ppos)
2047 struct hstate *h = &default_hstate;
2051 tmp = h->nr_overcommit_huge_pages;
2053 if (write && h->order >= MAX_ORDER)
2057 table->maxlen = sizeof(unsigned long);
2058 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2063 spin_lock(&hugetlb_lock);
2064 h->nr_overcommit_huge_pages = tmp;
2065 spin_unlock(&hugetlb_lock);
2071 #endif /* CONFIG_SYSCTL */
2073 void hugetlb_report_meminfo(struct seq_file *m)
2075 struct hstate *h = &default_hstate;
2077 "HugePages_Total: %5lu\n"
2078 "HugePages_Free: %5lu\n"
2079 "HugePages_Rsvd: %5lu\n"
2080 "HugePages_Surp: %5lu\n"
2081 "Hugepagesize: %8lu kB\n",
2085 h->surplus_huge_pages,
2086 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2089 int hugetlb_report_node_meminfo(int nid, char *buf)
2091 struct hstate *h = &default_hstate;
2093 "Node %d HugePages_Total: %5u\n"
2094 "Node %d HugePages_Free: %5u\n"
2095 "Node %d HugePages_Surp: %5u\n",
2096 nid, h->nr_huge_pages_node[nid],
2097 nid, h->free_huge_pages_node[nid],
2098 nid, h->surplus_huge_pages_node[nid]);
2101 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2102 unsigned long hugetlb_total_pages(void)
2104 struct hstate *h = &default_hstate;
2105 return h->nr_huge_pages * pages_per_huge_page(h);
2108 static int hugetlb_acct_memory(struct hstate *h, long delta)
2112 spin_lock(&hugetlb_lock);
2114 * When cpuset is configured, it breaks the strict hugetlb page
2115 * reservation as the accounting is done on a global variable. Such
2116 * reservation is completely rubbish in the presence of cpuset because
2117 * the reservation is not checked against page availability for the
2118 * current cpuset. Application can still potentially OOM'ed by kernel
2119 * with lack of free htlb page in cpuset that the task is in.
2120 * Attempt to enforce strict accounting with cpuset is almost
2121 * impossible (or too ugly) because cpuset is too fluid that
2122 * task or memory node can be dynamically moved between cpusets.
2124 * The change of semantics for shared hugetlb mapping with cpuset is
2125 * undesirable. However, in order to preserve some of the semantics,
2126 * we fall back to check against current free page availability as
2127 * a best attempt and hopefully to minimize the impact of changing
2128 * semantics that cpuset has.
2131 if (gather_surplus_pages(h, delta) < 0)
2134 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2135 return_unused_surplus_pages(h, delta);
2142 return_unused_surplus_pages(h, (unsigned long) -delta);
2145 spin_unlock(&hugetlb_lock);
2149 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2151 struct resv_map *reservations = vma_resv_map(vma);
2154 * This new VMA should share its siblings reservation map if present.
2155 * The VMA will only ever have a valid reservation map pointer where
2156 * it is being copied for another still existing VMA. As that VMA
2157 * has a reference to the reservation map it cannot disappear until
2158 * after this open call completes. It is therefore safe to take a
2159 * new reference here without additional locking.
2162 kref_get(&reservations->refs);
2165 static void resv_map_put(struct vm_area_struct *vma)
2167 struct resv_map *reservations = vma_resv_map(vma);
2171 kref_put(&reservations->refs, resv_map_release);
2174 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2176 struct hstate *h = hstate_vma(vma);
2177 struct resv_map *reservations = vma_resv_map(vma);
2178 struct hugepage_subpool *spool = subpool_vma(vma);
2179 unsigned long reserve;
2180 unsigned long start;
2184 start = vma_hugecache_offset(h, vma, vma->vm_start);
2185 end = vma_hugecache_offset(h, vma, vma->vm_end);
2187 reserve = (end - start) -
2188 region_count(&reservations->regions, start, end);
2193 hugetlb_acct_memory(h, -reserve);
2194 hugepage_subpool_put_pages(spool, reserve);
2200 * We cannot handle pagefaults against hugetlb pages at all. They cause
2201 * handle_mm_fault() to try to instantiate regular-sized pages in the
2202 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2205 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2211 const struct vm_operations_struct hugetlb_vm_ops = {
2212 .fault = hugetlb_vm_op_fault,
2213 .open = hugetlb_vm_op_open,
2214 .close = hugetlb_vm_op_close,
2217 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2224 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2226 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2228 entry = pte_mkyoung(entry);
2229 entry = pte_mkhuge(entry);
2230 entry = arch_make_huge_pte(entry, vma, page, writable);
2235 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2236 unsigned long address, pte_t *ptep)
2240 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2241 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2242 update_mmu_cache(vma, address, ptep);
2246 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2247 struct vm_area_struct *vma)
2249 pte_t *src_pte, *dst_pte, entry;
2250 struct page *ptepage;
2253 struct hstate *h = hstate_vma(vma);
2254 unsigned long sz = huge_page_size(h);
2256 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2258 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2259 src_pte = huge_pte_offset(src, addr);
2262 dst_pte = huge_pte_alloc(dst, addr, sz);
2266 /* If the pagetables are shared don't copy or take references */
2267 if (dst_pte == src_pte)
2270 spin_lock(&dst->page_table_lock);
2271 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2272 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2274 huge_ptep_set_wrprotect(src, addr, src_pte);
2275 entry = huge_ptep_get(src_pte);
2276 ptepage = pte_page(entry);
2278 page_dup_rmap(ptepage);
2279 set_huge_pte_at(dst, addr, dst_pte, entry);
2281 spin_unlock(&src->page_table_lock);
2282 spin_unlock(&dst->page_table_lock);
2290 static int is_hugetlb_entry_migration(pte_t pte)
2294 if (huge_pte_none(pte) || pte_present(pte))
2296 swp = pte_to_swp_entry(pte);
2297 if (non_swap_entry(swp) && is_migration_entry(swp))
2303 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2307 if (huge_pte_none(pte) || pte_present(pte))
2309 swp = pte_to_swp_entry(pte);
2310 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2316 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2317 unsigned long start, unsigned long end,
2318 struct page *ref_page)
2320 int force_flush = 0;
2321 struct mm_struct *mm = vma->vm_mm;
2322 unsigned long address;
2326 struct hstate *h = hstate_vma(vma);
2327 unsigned long sz = huge_page_size(h);
2329 WARN_ON(!is_vm_hugetlb_page(vma));
2330 BUG_ON(start & ~huge_page_mask(h));
2331 BUG_ON(end & ~huge_page_mask(h));
2333 tlb_start_vma(tlb, vma);
2334 mmu_notifier_invalidate_range_start(mm, start, end);
2336 spin_lock(&mm->page_table_lock);
2337 for (address = start; address < end; address += sz) {
2338 ptep = huge_pte_offset(mm, address);
2342 if (huge_pmd_unshare(mm, &address, ptep))
2345 pte = huge_ptep_get(ptep);
2346 if (huge_pte_none(pte))
2350 * HWPoisoned hugepage is already unmapped and dropped reference
2352 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2355 page = pte_page(pte);
2357 * If a reference page is supplied, it is because a specific
2358 * page is being unmapped, not a range. Ensure the page we
2359 * are about to unmap is the actual page of interest.
2362 if (page != ref_page)
2366 * Mark the VMA as having unmapped its page so that
2367 * future faults in this VMA will fail rather than
2368 * looking like data was lost
2370 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2373 pte = huge_ptep_get_and_clear(mm, address, ptep);
2374 tlb_remove_tlb_entry(tlb, ptep, address);
2376 set_page_dirty(page);
2378 page_remove_rmap(page);
2379 force_flush = !__tlb_remove_page(tlb, page);
2382 /* Bail out after unmapping reference page if supplied */
2386 spin_unlock(&mm->page_table_lock);
2388 * mmu_gather ran out of room to batch pages, we break out of
2389 * the PTE lock to avoid doing the potential expensive TLB invalidate
2390 * and page-free while holding it.
2395 if (address < end && !ref_page)
2398 mmu_notifier_invalidate_range_end(mm, start, end);
2399 tlb_end_vma(tlb, vma);
2402 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2403 unsigned long end, struct page *ref_page)
2405 struct mm_struct *mm;
2406 struct mmu_gather tlb;
2410 tlb_gather_mmu(&tlb, mm, 0);
2411 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2412 tlb_finish_mmu(&tlb, start, end);
2416 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2417 * mappping it owns the reserve page for. The intention is to unmap the page
2418 * from other VMAs and let the children be SIGKILLed if they are faulting the
2421 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2422 struct page *page, unsigned long address)
2424 struct hstate *h = hstate_vma(vma);
2425 struct vm_area_struct *iter_vma;
2426 struct address_space *mapping;
2427 struct prio_tree_iter iter;
2431 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2432 * from page cache lookup which is in HPAGE_SIZE units.
2434 address = address & huge_page_mask(h);
2435 pgoff = vma_hugecache_offset(h, vma, address);
2436 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2439 * Take the mapping lock for the duration of the table walk. As
2440 * this mapping should be shared between all the VMAs,
2441 * __unmap_hugepage_range() is called as the lock is already held
2443 mutex_lock(&mapping->i_mmap_mutex);
2444 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2445 /* Do not unmap the current VMA */
2446 if (iter_vma == vma)
2450 * Unmap the page from other VMAs without their own reserves.
2451 * They get marked to be SIGKILLed if they fault in these
2452 * areas. This is because a future no-page fault on this VMA
2453 * could insert a zeroed page instead of the data existing
2454 * from the time of fork. This would look like data corruption
2456 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2457 unmap_hugepage_range(iter_vma, address,
2458 address + huge_page_size(h), page);
2460 mutex_unlock(&mapping->i_mmap_mutex);
2466 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2467 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2468 * cannot race with other handlers or page migration.
2469 * Keep the pte_same checks anyway to make transition from the mutex easier.
2471 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2472 unsigned long address, pte_t *ptep, pte_t pte,
2473 struct page *pagecache_page)
2475 struct hstate *h = hstate_vma(vma);
2476 struct page *old_page, *new_page;
2478 int outside_reserve = 0;
2480 old_page = pte_page(pte);
2483 /* If no-one else is actually using this page, avoid the copy
2484 * and just make the page writable */
2485 avoidcopy = (page_mapcount(old_page) == 1);
2487 if (PageAnon(old_page))
2488 page_move_anon_rmap(old_page, vma, address);
2489 set_huge_ptep_writable(vma, address, ptep);
2494 * If the process that created a MAP_PRIVATE mapping is about to
2495 * perform a COW due to a shared page count, attempt to satisfy
2496 * the allocation without using the existing reserves. The pagecache
2497 * page is used to determine if the reserve at this address was
2498 * consumed or not. If reserves were used, a partial faulted mapping
2499 * at the time of fork() could consume its reserves on COW instead
2500 * of the full address range.
2502 if (!(vma->vm_flags & VM_MAYSHARE) &&
2503 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2504 old_page != pagecache_page)
2505 outside_reserve = 1;
2507 page_cache_get(old_page);
2509 /* Drop page_table_lock as buddy allocator may be called */
2510 spin_unlock(&mm->page_table_lock);
2511 new_page = alloc_huge_page(vma, address, outside_reserve);
2513 if (IS_ERR(new_page)) {
2514 long err = PTR_ERR(new_page);
2515 page_cache_release(old_page);
2518 * If a process owning a MAP_PRIVATE mapping fails to COW,
2519 * it is due to references held by a child and an insufficient
2520 * huge page pool. To guarantee the original mappers
2521 * reliability, unmap the page from child processes. The child
2522 * may get SIGKILLed if it later faults.
2524 if (outside_reserve) {
2525 BUG_ON(huge_pte_none(pte));
2526 if (unmap_ref_private(mm, vma, old_page, address)) {
2527 BUG_ON(huge_pte_none(pte));
2528 spin_lock(&mm->page_table_lock);
2529 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2530 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2531 goto retry_avoidcopy;
2533 * race occurs while re-acquiring page_table_lock, and
2541 /* Caller expects lock to be held */
2542 spin_lock(&mm->page_table_lock);
2544 return VM_FAULT_OOM;
2546 return VM_FAULT_SIGBUS;
2550 * When the original hugepage is shared one, it does not have
2551 * anon_vma prepared.
2553 if (unlikely(anon_vma_prepare(vma))) {
2554 page_cache_release(new_page);
2555 page_cache_release(old_page);
2556 /* Caller expects lock to be held */
2557 spin_lock(&mm->page_table_lock);
2558 return VM_FAULT_OOM;
2561 copy_user_huge_page(new_page, old_page, address, vma,
2562 pages_per_huge_page(h));
2563 __SetPageUptodate(new_page);
2566 * Retake the page_table_lock to check for racing updates
2567 * before the page tables are altered
2569 spin_lock(&mm->page_table_lock);
2570 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2571 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2573 mmu_notifier_invalidate_range_start(mm,
2574 address & huge_page_mask(h),
2575 (address & huge_page_mask(h)) + huge_page_size(h));
2576 huge_ptep_clear_flush(vma, address, ptep);
2577 set_huge_pte_at(mm, address, ptep,
2578 make_huge_pte(vma, new_page, 1));
2579 page_remove_rmap(old_page);
2580 hugepage_add_new_anon_rmap(new_page, vma, address);
2581 /* Make the old page be freed below */
2582 new_page = old_page;
2583 mmu_notifier_invalidate_range_end(mm,
2584 address & huge_page_mask(h),
2585 (address & huge_page_mask(h)) + huge_page_size(h));
2587 page_cache_release(new_page);
2588 page_cache_release(old_page);
2592 /* Return the pagecache page at a given address within a VMA */
2593 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2594 struct vm_area_struct *vma, unsigned long address)
2596 struct address_space *mapping;
2599 mapping = vma->vm_file->f_mapping;
2600 idx = vma_hugecache_offset(h, vma, address);
2602 return find_lock_page(mapping, idx);
2606 * Return whether there is a pagecache page to back given address within VMA.
2607 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2609 static bool hugetlbfs_pagecache_present(struct hstate *h,
2610 struct vm_area_struct *vma, unsigned long address)
2612 struct address_space *mapping;
2616 mapping = vma->vm_file->f_mapping;
2617 idx = vma_hugecache_offset(h, vma, address);
2619 page = find_get_page(mapping, idx);
2622 return page != NULL;
2625 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2626 unsigned long address, pte_t *ptep, unsigned int flags)
2628 struct hstate *h = hstate_vma(vma);
2629 int ret = VM_FAULT_SIGBUS;
2634 struct address_space *mapping;
2638 * Currently, we are forced to kill the process in the event the
2639 * original mapper has unmapped pages from the child due to a failed
2640 * COW. Warn that such a situation has occurred as it may not be obvious
2642 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2644 "PID %d killed due to inadequate hugepage pool\n",
2649 mapping = vma->vm_file->f_mapping;
2650 idx = vma_hugecache_offset(h, vma, address);
2653 * Use page lock to guard against racing truncation
2654 * before we get page_table_lock.
2657 page = find_lock_page(mapping, idx);
2659 size = i_size_read(mapping->host) >> huge_page_shift(h);
2662 page = alloc_huge_page(vma, address, 0);
2664 ret = PTR_ERR(page);
2668 ret = VM_FAULT_SIGBUS;
2671 clear_huge_page(page, address, pages_per_huge_page(h));
2672 __SetPageUptodate(page);
2674 if (vma->vm_flags & VM_MAYSHARE) {
2676 struct inode *inode = mapping->host;
2678 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2686 spin_lock(&inode->i_lock);
2687 inode->i_blocks += blocks_per_huge_page(h);
2688 spin_unlock(&inode->i_lock);
2691 if (unlikely(anon_vma_prepare(vma))) {
2693 goto backout_unlocked;
2699 * If memory error occurs between mmap() and fault, some process
2700 * don't have hwpoisoned swap entry for errored virtual address.
2701 * So we need to block hugepage fault by PG_hwpoison bit check.
2703 if (unlikely(PageHWPoison(page))) {
2704 ret = VM_FAULT_HWPOISON |
2705 VM_FAULT_SET_HINDEX(hstate_index(h));
2706 goto backout_unlocked;
2711 * If we are going to COW a private mapping later, we examine the
2712 * pending reservations for this page now. This will ensure that
2713 * any allocations necessary to record that reservation occur outside
2716 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2717 if (vma_needs_reservation(h, vma, address) < 0) {
2719 goto backout_unlocked;
2722 spin_lock(&mm->page_table_lock);
2723 size = i_size_read(mapping->host) >> huge_page_shift(h);
2728 if (!huge_pte_none(huge_ptep_get(ptep)))
2732 hugepage_add_new_anon_rmap(page, vma, address);
2734 page_dup_rmap(page);
2735 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2736 && (vma->vm_flags & VM_SHARED)));
2737 set_huge_pte_at(mm, address, ptep, new_pte);
2739 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2740 /* Optimization, do the COW without a second fault */
2741 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2744 spin_unlock(&mm->page_table_lock);
2750 spin_unlock(&mm->page_table_lock);
2757 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2758 unsigned long address, unsigned int flags)
2763 struct page *page = NULL;
2764 struct page *pagecache_page = NULL;
2765 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2766 struct hstate *h = hstate_vma(vma);
2768 address &= huge_page_mask(h);
2770 ptep = huge_pte_offset(mm, address);
2772 entry = huge_ptep_get(ptep);
2773 if (unlikely(is_hugetlb_entry_migration(entry))) {
2774 migration_entry_wait(mm, (pmd_t *)ptep, address);
2776 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2777 return VM_FAULT_HWPOISON_LARGE |
2778 VM_FAULT_SET_HINDEX(hstate_index(h));
2781 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2783 return VM_FAULT_OOM;
2786 * Serialize hugepage allocation and instantiation, so that we don't
2787 * get spurious allocation failures if two CPUs race to instantiate
2788 * the same page in the page cache.
2790 mutex_lock(&hugetlb_instantiation_mutex);
2791 entry = huge_ptep_get(ptep);
2792 if (huge_pte_none(entry)) {
2793 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2800 * If we are going to COW the mapping later, we examine the pending
2801 * reservations for this page now. This will ensure that any
2802 * allocations necessary to record that reservation occur outside the
2803 * spinlock. For private mappings, we also lookup the pagecache
2804 * page now as it is used to determine if a reservation has been
2807 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2808 if (vma_needs_reservation(h, vma, address) < 0) {
2813 if (!(vma->vm_flags & VM_MAYSHARE))
2814 pagecache_page = hugetlbfs_pagecache_page(h,
2819 * hugetlb_cow() requires page locks of pte_page(entry) and
2820 * pagecache_page, so here we need take the former one
2821 * when page != pagecache_page or !pagecache_page.
2822 * Note that locking order is always pagecache_page -> page,
2823 * so no worry about deadlock.
2825 page = pte_page(entry);
2827 if (page != pagecache_page)
2830 spin_lock(&mm->page_table_lock);
2831 /* Check for a racing update before calling hugetlb_cow */
2832 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2833 goto out_page_table_lock;
2836 if (flags & FAULT_FLAG_WRITE) {
2837 if (!pte_write(entry)) {
2838 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2840 goto out_page_table_lock;
2842 entry = pte_mkdirty(entry);
2844 entry = pte_mkyoung(entry);
2845 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2846 flags & FAULT_FLAG_WRITE))
2847 update_mmu_cache(vma, address, ptep);
2849 out_page_table_lock:
2850 spin_unlock(&mm->page_table_lock);
2852 if (pagecache_page) {
2853 unlock_page(pagecache_page);
2854 put_page(pagecache_page);
2856 if (page != pagecache_page)
2861 mutex_unlock(&hugetlb_instantiation_mutex);
2866 /* Can be overriden by architectures */
2867 __attribute__((weak)) struct page *
2868 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2869 pud_t *pud, int write)
2875 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2876 struct page **pages, struct vm_area_struct **vmas,
2877 unsigned long *position, int *length, int i,
2880 unsigned long pfn_offset;
2881 unsigned long vaddr = *position;
2882 int remainder = *length;
2883 struct hstate *h = hstate_vma(vma);
2885 spin_lock(&mm->page_table_lock);
2886 while (vaddr < vma->vm_end && remainder) {
2892 * Some archs (sparc64, sh*) have multiple pte_ts to
2893 * each hugepage. We have to make sure we get the
2894 * first, for the page indexing below to work.
2896 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2897 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2900 * When coredumping, it suits get_dump_page if we just return
2901 * an error where there's an empty slot with no huge pagecache
2902 * to back it. This way, we avoid allocating a hugepage, and
2903 * the sparse dumpfile avoids allocating disk blocks, but its
2904 * huge holes still show up with zeroes where they need to be.
2906 if (absent && (flags & FOLL_DUMP) &&
2907 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2913 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2916 spin_unlock(&mm->page_table_lock);
2917 ret = hugetlb_fault(mm, vma, vaddr,
2918 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2919 spin_lock(&mm->page_table_lock);
2920 if (!(ret & VM_FAULT_ERROR))
2927 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2928 page = pte_page(huge_ptep_get(pte));
2931 pages[i] = mem_map_offset(page, pfn_offset);
2942 if (vaddr < vma->vm_end && remainder &&
2943 pfn_offset < pages_per_huge_page(h)) {
2945 * We use pfn_offset to avoid touching the pageframes
2946 * of this compound page.
2951 spin_unlock(&mm->page_table_lock);
2952 *length = remainder;
2955 return i ? i : -EFAULT;
2958 void hugetlb_change_protection(struct vm_area_struct *vma,
2959 unsigned long address, unsigned long end, pgprot_t newprot)
2961 struct mm_struct *mm = vma->vm_mm;
2962 unsigned long start = address;
2965 struct hstate *h = hstate_vma(vma);
2967 BUG_ON(address >= end);
2968 flush_cache_range(vma, address, end);
2970 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2971 spin_lock(&mm->page_table_lock);
2972 for (; address < end; address += huge_page_size(h)) {
2973 ptep = huge_pte_offset(mm, address);
2976 if (huge_pmd_unshare(mm, &address, ptep))
2978 if (!huge_pte_none(huge_ptep_get(ptep))) {
2979 pte = huge_ptep_get_and_clear(mm, address, ptep);
2980 pte = pte_mkhuge(pte_modify(pte, newprot));
2981 set_huge_pte_at(mm, address, ptep, pte);
2984 spin_unlock(&mm->page_table_lock);
2985 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2987 flush_tlb_range(vma, start, end);
2990 int hugetlb_reserve_pages(struct inode *inode,
2992 struct vm_area_struct *vma,
2993 vm_flags_t vm_flags)
2996 struct hstate *h = hstate_inode(inode);
2997 struct hugepage_subpool *spool = subpool_inode(inode);
3000 * Only apply hugepage reservation if asked. At fault time, an
3001 * attempt will be made for VM_NORESERVE to allocate a page
3002 * without using reserves
3004 if (vm_flags & VM_NORESERVE)
3008 * Shared mappings base their reservation on the number of pages that
3009 * are already allocated on behalf of the file. Private mappings need
3010 * to reserve the full area even if read-only as mprotect() may be
3011 * called to make the mapping read-write. Assume !vma is a shm mapping
3013 if (!vma || vma->vm_flags & VM_MAYSHARE)
3014 chg = region_chg(&inode->i_mapping->private_list, from, to);
3016 struct resv_map *resv_map = resv_map_alloc();
3022 set_vma_resv_map(vma, resv_map);
3023 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3031 /* There must be enough pages in the subpool for the mapping */
3032 if (hugepage_subpool_get_pages(spool, chg)) {
3038 * Check enough hugepages are available for the reservation.
3039 * Hand the pages back to the subpool if there are not
3041 ret = hugetlb_acct_memory(h, chg);
3043 hugepage_subpool_put_pages(spool, chg);
3048 * Account for the reservations made. Shared mappings record regions
3049 * that have reservations as they are shared by multiple VMAs.
3050 * When the last VMA disappears, the region map says how much
3051 * the reservation was and the page cache tells how much of
3052 * the reservation was consumed. Private mappings are per-VMA and
3053 * only the consumed reservations are tracked. When the VMA
3054 * disappears, the original reservation is the VMA size and the
3055 * consumed reservations are stored in the map. Hence, nothing
3056 * else has to be done for private mappings here
3058 if (!vma || vma->vm_flags & VM_MAYSHARE)
3059 region_add(&inode->i_mapping->private_list, from, to);
3067 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3069 struct hstate *h = hstate_inode(inode);
3070 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3071 struct hugepage_subpool *spool = subpool_inode(inode);
3073 spin_lock(&inode->i_lock);
3074 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3075 spin_unlock(&inode->i_lock);
3077 hugepage_subpool_put_pages(spool, (chg - freed));
3078 hugetlb_acct_memory(h, -(chg - freed));
3081 #ifdef CONFIG_MEMORY_FAILURE
3083 /* Should be called in hugetlb_lock */
3084 static int is_hugepage_on_freelist(struct page *hpage)
3088 struct hstate *h = page_hstate(hpage);
3089 int nid = page_to_nid(hpage);
3091 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3098 * This function is called from memory failure code.
3099 * Assume the caller holds page lock of the head page.
3101 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3103 struct hstate *h = page_hstate(hpage);
3104 int nid = page_to_nid(hpage);
3107 spin_lock(&hugetlb_lock);
3108 if (is_hugepage_on_freelist(hpage)) {
3109 list_del(&hpage->lru);
3110 set_page_refcounted(hpage);
3111 h->free_huge_pages--;
3112 h->free_huge_pages_node[nid]--;
3115 spin_unlock(&hugetlb_lock);