2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
27 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/node.h>
34 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
35 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
36 unsigned long hugepages_treat_as_movable;
38 static int max_hstate;
39 unsigned int default_hstate_idx;
40 struct hstate hstates[HUGE_MAX_HSTATE];
42 __initdata LIST_HEAD(huge_boot_pages);
44 /* for command line parsing */
45 static struct hstate * __initdata parsed_hstate;
46 static unsigned long __initdata default_hstate_max_huge_pages;
47 static unsigned long __initdata default_hstate_size;
49 #define for_each_hstate(h) \
50 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
53 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
55 static DEFINE_SPINLOCK(hugetlb_lock);
57 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
59 bool free = (spool->count == 0) && (spool->used_hpages == 0);
61 spin_unlock(&spool->lock);
63 /* If no pages are used, and no other handles to the subpool
64 * remain, free the subpool the subpool remain */
69 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
71 struct hugepage_subpool *spool;
73 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
77 spin_lock_init(&spool->lock);
79 spool->max_hpages = nr_blocks;
80 spool->used_hpages = 0;
85 void hugepage_put_subpool(struct hugepage_subpool *spool)
87 spin_lock(&spool->lock);
88 BUG_ON(!spool->count);
90 unlock_or_release_subpool(spool);
93 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 spin_lock(&spool->lock);
102 if ((spool->used_hpages + delta) <= spool->max_hpages) {
103 spool->used_hpages += delta;
107 spin_unlock(&spool->lock);
112 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
118 spin_lock(&spool->lock);
119 spool->used_hpages -= delta;
120 /* If hugetlbfs_put_super couldn't free spool due to
121 * an outstanding quota reference, free it now. */
122 unlock_or_release_subpool(spool);
125 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
127 return HUGETLBFS_SB(inode->i_sb)->spool;
130 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
132 return subpool_inode(vma->vm_file->f_dentry->d_inode);
136 * Region tracking -- allows tracking of reservations and instantiated pages
137 * across the pages in a mapping.
139 * The region data structures are protected by a combination of the mmap_sem
140 * and the hugetlb_instantion_mutex. To access or modify a region the caller
141 * must either hold the mmap_sem for write, or the mmap_sem for read and
142 * the hugetlb_instantiation mutex:
144 * down_write(&mm->mmap_sem);
146 * down_read(&mm->mmap_sem);
147 * mutex_lock(&hugetlb_instantiation_mutex);
150 struct list_head link;
155 static long region_add(struct list_head *head, long f, long t)
157 struct file_region *rg, *nrg, *trg;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg, head, link)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
171 if (&rg->link == head)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct list_head *head, long f, long t)
193 struct file_region *rg, *nrg;
196 /* Locate the region we are before or in. */
197 list_for_each_entry(rg, head, link)
201 /* If we are below the current region then a new region is required.
202 * Subtle, allocate a new region at the position but make it zero
203 * size such that we can guarantee to record the reservation. */
204 if (&rg->link == head || t < rg->from) {
205 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
210 INIT_LIST_HEAD(&nrg->link);
211 list_add(&nrg->link, rg->link.prev);
216 /* Round our left edge to the current segment if it encloses us. */
221 /* Check for and consume any regions we now overlap with. */
222 list_for_each_entry(rg, rg->link.prev, link) {
223 if (&rg->link == head)
228 /* We overlap with this area, if it extends further than
229 * us then we must extend ourselves. Account for its
230 * existing reservation. */
235 chg -= rg->to - rg->from;
240 static long region_truncate(struct list_head *head, long end)
242 struct file_region *rg, *trg;
245 /* Locate the region we are either in or before. */
246 list_for_each_entry(rg, head, link)
249 if (&rg->link == head)
252 /* If we are in the middle of a region then adjust it. */
253 if (end > rg->from) {
256 rg = list_entry(rg->link.next, typeof(*rg), link);
259 /* Drop any remaining regions. */
260 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
261 if (&rg->link == head)
263 chg += rg->to - rg->from;
270 static long region_count(struct list_head *head, long f, long t)
272 struct file_region *rg;
275 /* Locate each segment we overlap with, and count that overlap. */
276 list_for_each_entry(rg, head, link) {
285 seg_from = max(rg->from, f);
286 seg_to = min(rg->to, t);
288 chg += seg_to - seg_from;
295 * Convert the address within this vma to the page offset within
296 * the mapping, in pagecache page units; huge pages here.
298 static pgoff_t vma_hugecache_offset(struct hstate *h,
299 struct vm_area_struct *vma, unsigned long address)
301 return ((address - vma->vm_start) >> huge_page_shift(h)) +
302 (vma->vm_pgoff >> huge_page_order(h));
305 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
306 unsigned long address)
308 return vma_hugecache_offset(hstate_vma(vma), vma, address);
312 * Return the size of the pages allocated when backing a VMA. In the majority
313 * cases this will be same size as used by the page table entries.
315 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
317 struct hstate *hstate;
319 if (!is_vm_hugetlb_page(vma))
322 hstate = hstate_vma(vma);
324 return 1UL << (hstate->order + PAGE_SHIFT);
326 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
329 * Return the page size being used by the MMU to back a VMA. In the majority
330 * of cases, the page size used by the kernel matches the MMU size. On
331 * architectures where it differs, an architecture-specific version of this
332 * function is required.
334 #ifndef vma_mmu_pagesize
335 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
337 return vma_kernel_pagesize(vma);
342 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
343 * bits of the reservation map pointer, which are always clear due to
346 #define HPAGE_RESV_OWNER (1UL << 0)
347 #define HPAGE_RESV_UNMAPPED (1UL << 1)
348 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
351 * These helpers are used to track how many pages are reserved for
352 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
353 * is guaranteed to have their future faults succeed.
355 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
356 * the reserve counters are updated with the hugetlb_lock held. It is safe
357 * to reset the VMA at fork() time as it is not in use yet and there is no
358 * chance of the global counters getting corrupted as a result of the values.
360 * The private mapping reservation is represented in a subtly different
361 * manner to a shared mapping. A shared mapping has a region map associated
362 * with the underlying file, this region map represents the backing file
363 * pages which have ever had a reservation assigned which this persists even
364 * after the page is instantiated. A private mapping has a region map
365 * associated with the original mmap which is attached to all VMAs which
366 * reference it, this region map represents those offsets which have consumed
367 * reservation ie. where pages have been instantiated.
369 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
371 return (unsigned long)vma->vm_private_data;
374 static void set_vma_private_data(struct vm_area_struct *vma,
377 vma->vm_private_data = (void *)value;
382 struct list_head regions;
385 static struct resv_map *resv_map_alloc(void)
387 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
391 kref_init(&resv_map->refs);
392 INIT_LIST_HEAD(&resv_map->regions);
397 static void resv_map_release(struct kref *ref)
399 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
401 /* Clear out any active regions before we release the map. */
402 region_truncate(&resv_map->regions, 0);
406 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
408 VM_BUG_ON(!is_vm_hugetlb_page(vma));
409 if (!(vma->vm_flags & VM_MAYSHARE))
410 return (struct resv_map *)(get_vma_private_data(vma) &
415 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
417 VM_BUG_ON(!is_vm_hugetlb_page(vma));
418 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
420 set_vma_private_data(vma, (get_vma_private_data(vma) &
421 HPAGE_RESV_MASK) | (unsigned long)map);
424 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
426 VM_BUG_ON(!is_vm_hugetlb_page(vma));
427 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
429 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
432 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
434 VM_BUG_ON(!is_vm_hugetlb_page(vma));
436 return (get_vma_private_data(vma) & flag) != 0;
439 /* Decrement the reserved pages in the hugepage pool by one */
440 static void decrement_hugepage_resv_vma(struct hstate *h,
441 struct vm_area_struct *vma)
443 if (vma->vm_flags & VM_NORESERVE)
446 if (vma->vm_flags & VM_MAYSHARE) {
447 /* Shared mappings always use reserves */
448 h->resv_huge_pages--;
449 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
451 * Only the process that called mmap() has reserves for
454 h->resv_huge_pages--;
458 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
459 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
462 if (!(vma->vm_flags & VM_MAYSHARE))
463 vma->vm_private_data = (void *)0;
466 /* Returns true if the VMA has associated reserve pages */
467 static int vma_has_reserves(struct vm_area_struct *vma)
469 if (vma->vm_flags & VM_MAYSHARE)
471 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
476 static void copy_gigantic_page(struct page *dst, struct page *src)
479 struct hstate *h = page_hstate(src);
480 struct page *dst_base = dst;
481 struct page *src_base = src;
483 for (i = 0; i < pages_per_huge_page(h); ) {
485 copy_highpage(dst, src);
488 dst = mem_map_next(dst, dst_base, i);
489 src = mem_map_next(src, src_base, i);
493 void copy_huge_page(struct page *dst, struct page *src)
496 struct hstate *h = page_hstate(src);
498 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
499 copy_gigantic_page(dst, src);
504 for (i = 0; i < pages_per_huge_page(h); i++) {
506 copy_highpage(dst + i, src + i);
510 static void enqueue_huge_page(struct hstate *h, struct page *page)
512 int nid = page_to_nid(page);
513 list_add(&page->lru, &h->hugepage_freelists[nid]);
514 h->free_huge_pages++;
515 h->free_huge_pages_node[nid]++;
518 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 if (list_empty(&h->hugepage_freelists[nid]))
524 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
525 list_del(&page->lru);
526 set_page_refcounted(page);
527 h->free_huge_pages--;
528 h->free_huge_pages_node[nid]--;
532 static struct page *dequeue_huge_page_vma(struct hstate *h,
533 struct vm_area_struct *vma,
534 unsigned long address, int avoid_reserve)
536 struct page *page = NULL;
537 struct mempolicy *mpol;
538 nodemask_t *nodemask;
539 struct zonelist *zonelist;
542 unsigned int cpuset_mems_cookie;
545 cpuset_mems_cookie = get_mems_allowed();
546 zonelist = huge_zonelist(vma, address,
547 htlb_alloc_mask, &mpol, &nodemask);
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma) &&
554 h->free_huge_pages - h->resv_huge_pages == 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 for_each_zone_zonelist_nodemask(zone, z, zonelist,
562 MAX_NR_ZONES - 1, nodemask) {
563 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
564 page = dequeue_huge_page_node(h, zone_to_nid(zone));
567 decrement_hugepage_resv_vma(h, vma);
574 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
583 static void update_and_free_page(struct hstate *h, struct page *page)
587 VM_BUG_ON(h->order >= MAX_ORDER);
590 h->nr_huge_pages_node[page_to_nid(page)]--;
591 for (i = 0; i < pages_per_huge_page(h); i++) {
592 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
593 1 << PG_referenced | 1 << PG_dirty |
594 1 << PG_active | 1 << PG_reserved |
595 1 << PG_private | 1 << PG_writeback);
597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
599 arch_release_hugepage(page);
600 __free_pages(page, huge_page_order(h));
603 struct hstate *size_to_hstate(unsigned long size)
608 if (huge_page_size(h) == size)
614 static void free_huge_page(struct page *page)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate *h = page_hstate(page);
621 int nid = page_to_nid(page);
622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
625 set_page_private(page, 0);
626 page->mapping = NULL;
627 BUG_ON(page_count(page));
628 BUG_ON(page_mapcount(page));
629 INIT_LIST_HEAD(&page->lru);
631 spin_lock(&hugetlb_lock);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 update_and_free_page(h, page);
634 h->surplus_huge_pages--;
635 h->surplus_huge_pages_node[nid]--;
637 arch_clear_hugepage_flags(page);
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 set_compound_page_dtor(page, free_huge_page);
647 spin_lock(&hugetlb_lock);
649 h->nr_huge_pages_node[nid]++;
650 spin_unlock(&hugetlb_lock);
651 put_page(page); /* free it into the hugepage allocator */
654 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
657 int nr_pages = 1 << order;
658 struct page *p = page + 1;
660 /* we rely on prep_new_huge_page to set the destructor */
661 set_compound_order(page, order);
663 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
665 set_page_count(p, 0);
666 p->first_page = page;
670 int PageHuge(struct page *page)
672 compound_page_dtor *dtor;
674 if (!PageCompound(page))
677 page = compound_head(page);
678 dtor = get_compound_page_dtor(page);
680 return dtor == free_huge_page;
682 EXPORT_SYMBOL_GPL(PageHuge);
685 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
686 * normal or transparent huge pages.
688 int PageHeadHuge(struct page *page_head)
690 compound_page_dtor *dtor;
692 if (!PageHead(page_head))
695 dtor = get_compound_page_dtor(page_head);
697 return dtor == free_huge_page;
699 EXPORT_SYMBOL_GPL(PageHeadHuge);
701 pgoff_t __basepage_index(struct page *page)
703 struct page *page_head = compound_head(page);
704 pgoff_t index = page_index(page_head);
705 unsigned long compound_idx;
707 if (!PageHuge(page_head))
708 return page_index(page);
710 if (compound_order(page_head) >= MAX_ORDER)
711 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
713 compound_idx = page - page_head;
715 return (index << compound_order(page_head)) + compound_idx;
718 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
722 if (h->order >= MAX_ORDER)
725 page = alloc_pages_exact_node(nid,
726 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
727 __GFP_REPEAT|__GFP_NOWARN,
730 if (arch_prepare_hugepage(page)) {
731 __free_pages(page, huge_page_order(h));
734 prep_new_huge_page(h, page, nid);
741 * common helper functions for hstate_next_node_to_{alloc|free}.
742 * We may have allocated or freed a huge page based on a different
743 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
744 * be outside of *nodes_allowed. Ensure that we use an allowed
745 * node for alloc or free.
747 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
749 nid = next_node(nid, *nodes_allowed);
750 if (nid == MAX_NUMNODES)
751 nid = first_node(*nodes_allowed);
752 VM_BUG_ON(nid >= MAX_NUMNODES);
757 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
759 if (!node_isset(nid, *nodes_allowed))
760 nid = next_node_allowed(nid, nodes_allowed);
765 * returns the previously saved node ["this node"] from which to
766 * allocate a persistent huge page for the pool and advance the
767 * next node from which to allocate, handling wrap at end of node
770 static int hstate_next_node_to_alloc(struct hstate *h,
771 nodemask_t *nodes_allowed)
775 VM_BUG_ON(!nodes_allowed);
777 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
778 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
783 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
790 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
791 next_nid = start_nid;
794 page = alloc_fresh_huge_page_node(h, next_nid);
799 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
800 } while (next_nid != start_nid);
803 count_vm_event(HTLB_BUDDY_PGALLOC);
805 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
811 * helper for free_pool_huge_page() - return the previously saved
812 * node ["this node"] from which to free a huge page. Advance the
813 * next node id whether or not we find a free huge page to free so
814 * that the next attempt to free addresses the next node.
816 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
820 VM_BUG_ON(!nodes_allowed);
822 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
823 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
829 * Free huge page from pool from next node to free.
830 * Attempt to keep persistent huge pages more or less
831 * balanced over allowed nodes.
832 * Called with hugetlb_lock locked.
834 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
841 start_nid = hstate_next_node_to_free(h, nodes_allowed);
842 next_nid = start_nid;
846 * If we're returning unused surplus pages, only examine
847 * nodes with surplus pages.
849 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
850 !list_empty(&h->hugepage_freelists[next_nid])) {
852 list_entry(h->hugepage_freelists[next_nid].next,
854 list_del(&page->lru);
855 h->free_huge_pages--;
856 h->free_huge_pages_node[next_nid]--;
858 h->surplus_huge_pages--;
859 h->surplus_huge_pages_node[next_nid]--;
861 update_and_free_page(h, page);
865 next_nid = hstate_next_node_to_free(h, nodes_allowed);
866 } while (next_nid != start_nid);
871 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
876 if (h->order >= MAX_ORDER)
880 * Assume we will successfully allocate the surplus page to
881 * prevent racing processes from causing the surplus to exceed
884 * This however introduces a different race, where a process B
885 * tries to grow the static hugepage pool while alloc_pages() is
886 * called by process A. B will only examine the per-node
887 * counters in determining if surplus huge pages can be
888 * converted to normal huge pages in adjust_pool_surplus(). A
889 * won't be able to increment the per-node counter, until the
890 * lock is dropped by B, but B doesn't drop hugetlb_lock until
891 * no more huge pages can be converted from surplus to normal
892 * state (and doesn't try to convert again). Thus, we have a
893 * case where a surplus huge page exists, the pool is grown, and
894 * the surplus huge page still exists after, even though it
895 * should just have been converted to a normal huge page. This
896 * does not leak memory, though, as the hugepage will be freed
897 * once it is out of use. It also does not allow the counters to
898 * go out of whack in adjust_pool_surplus() as we don't modify
899 * the node values until we've gotten the hugepage and only the
900 * per-node value is checked there.
902 spin_lock(&hugetlb_lock);
903 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
904 spin_unlock(&hugetlb_lock);
908 h->surplus_huge_pages++;
910 spin_unlock(&hugetlb_lock);
912 if (nid == NUMA_NO_NODE)
913 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
914 __GFP_REPEAT|__GFP_NOWARN,
917 page = alloc_pages_exact_node(nid,
918 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
919 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
921 if (page && arch_prepare_hugepage(page)) {
922 __free_pages(page, huge_page_order(h));
926 spin_lock(&hugetlb_lock);
928 r_nid = page_to_nid(page);
929 set_compound_page_dtor(page, free_huge_page);
931 * We incremented the global counters already
933 h->nr_huge_pages_node[r_nid]++;
934 h->surplus_huge_pages_node[r_nid]++;
935 __count_vm_event(HTLB_BUDDY_PGALLOC);
938 h->surplus_huge_pages--;
939 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
941 spin_unlock(&hugetlb_lock);
947 * This allocation function is useful in the context where vma is irrelevant.
948 * E.g. soft-offlining uses this function because it only cares physical
949 * address of error page.
951 struct page *alloc_huge_page_node(struct hstate *h, int nid)
955 spin_lock(&hugetlb_lock);
956 page = dequeue_huge_page_node(h, nid);
957 spin_unlock(&hugetlb_lock);
960 page = alloc_buddy_huge_page(h, nid);
966 * Increase the hugetlb pool such that it can accommodate a reservation
969 static int gather_surplus_pages(struct hstate *h, int delta)
971 struct list_head surplus_list;
972 struct page *page, *tmp;
974 int needed, allocated;
976 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
978 h->resv_huge_pages += delta;
983 INIT_LIST_HEAD(&surplus_list);
987 spin_unlock(&hugetlb_lock);
988 for (i = 0; i < needed; i++) {
989 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
992 * We were not able to allocate enough pages to
993 * satisfy the entire reservation so we free what
994 * we've allocated so far.
998 list_add(&page->lru, &surplus_list);
1000 allocated += needed;
1003 * After retaking hugetlb_lock, we need to recalculate 'needed'
1004 * because either resv_huge_pages or free_huge_pages may have changed.
1006 spin_lock(&hugetlb_lock);
1007 needed = (h->resv_huge_pages + delta) -
1008 (h->free_huge_pages + allocated);
1013 * The surplus_list now contains _at_least_ the number of extra pages
1014 * needed to accommodate the reservation. Add the appropriate number
1015 * of pages to the hugetlb pool and free the extras back to the buddy
1016 * allocator. Commit the entire reservation here to prevent another
1017 * process from stealing the pages as they are added to the pool but
1018 * before they are reserved.
1020 needed += allocated;
1021 h->resv_huge_pages += delta;
1024 /* Free the needed pages to the hugetlb pool */
1025 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1028 list_del(&page->lru);
1030 * This page is now managed by the hugetlb allocator and has
1031 * no users -- drop the buddy allocator's reference.
1033 put_page_testzero(page);
1034 VM_BUG_ON(page_count(page));
1035 enqueue_huge_page(h, page);
1037 spin_unlock(&hugetlb_lock);
1039 /* Free unnecessary surplus pages to the buddy allocator */
1041 if (!list_empty(&surplus_list)) {
1042 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1043 list_del(&page->lru);
1047 spin_lock(&hugetlb_lock);
1053 * When releasing a hugetlb pool reservation, any surplus pages that were
1054 * allocated to satisfy the reservation must be explicitly freed if they were
1056 * Called with hugetlb_lock held.
1058 static void return_unused_surplus_pages(struct hstate *h,
1059 unsigned long unused_resv_pages)
1061 unsigned long nr_pages;
1063 /* Uncommit the reservation */
1064 h->resv_huge_pages -= unused_resv_pages;
1066 /* Cannot return gigantic pages currently */
1067 if (h->order >= MAX_ORDER)
1070 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1073 * We want to release as many surplus pages as possible, spread
1074 * evenly across all nodes with memory. Iterate across these nodes
1075 * until we can no longer free unreserved surplus pages. This occurs
1076 * when the nodes with surplus pages have no free pages.
1077 * free_pool_huge_page() will balance the the freed pages across the
1078 * on-line nodes with memory and will handle the hstate accounting.
1080 while (nr_pages--) {
1081 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1083 cond_resched_lock(&hugetlb_lock);
1088 * Determine if the huge page at addr within the vma has an associated
1089 * reservation. Where it does not we will need to logically increase
1090 * reservation and actually increase subpool usage before an allocation
1091 * can occur. Where any new reservation would be required the
1092 * reservation change is prepared, but not committed. Once the page
1093 * has been allocated from the subpool and instantiated the change should
1094 * be committed via vma_commit_reservation. No action is required on
1097 static long vma_needs_reservation(struct hstate *h,
1098 struct vm_area_struct *vma, unsigned long addr)
1100 struct address_space *mapping = vma->vm_file->f_mapping;
1101 struct inode *inode = mapping->host;
1103 if (vma->vm_flags & VM_MAYSHARE) {
1104 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1105 return region_chg(&inode->i_mapping->private_list,
1108 } 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 err = region_chg(&reservations->regions, idx, idx + 1);
1122 static void vma_commit_reservation(struct hstate *h,
1123 struct vm_area_struct *vma, unsigned long addr)
1125 struct address_space *mapping = vma->vm_file->f_mapping;
1126 struct inode *inode = mapping->host;
1128 if (vma->vm_flags & VM_MAYSHARE) {
1129 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1130 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1132 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1133 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1134 struct resv_map *reservations = vma_resv_map(vma);
1136 /* Mark this page used in the map. */
1137 region_add(&reservations->regions, idx, idx + 1);
1141 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1142 unsigned long addr, int avoid_reserve)
1144 struct hugepage_subpool *spool = subpool_vma(vma);
1145 struct hstate *h = hstate_vma(vma);
1150 * Processes that did not create the mapping will have no
1151 * reserves and will not have accounted against subpool
1152 * limit. Check that the subpool limit can be made before
1153 * satisfying the allocation MAP_NORESERVE mappings may also
1154 * need pages and subpool limit allocated allocated if no reserve
1157 chg = vma_needs_reservation(h, vma, addr);
1159 return ERR_PTR(-VM_FAULT_OOM);
1161 if (hugepage_subpool_get_pages(spool, chg))
1162 return ERR_PTR(-VM_FAULT_SIGBUS);
1164 spin_lock(&hugetlb_lock);
1165 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1166 spin_unlock(&hugetlb_lock);
1169 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1171 hugepage_subpool_put_pages(spool, chg);
1172 return ERR_PTR(-VM_FAULT_SIGBUS);
1176 set_page_private(page, (unsigned long)spool);
1178 vma_commit_reservation(h, vma, addr);
1183 int __weak alloc_bootmem_huge_page(struct hstate *h)
1185 struct huge_bootmem_page *m;
1186 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1191 addr = __alloc_bootmem_node_nopanic(
1192 NODE_DATA(hstate_next_node_to_alloc(h,
1193 &node_states[N_HIGH_MEMORY])),
1194 huge_page_size(h), huge_page_size(h), 0);
1198 * Use the beginning of the huge page to store the
1199 * huge_bootmem_page struct (until gather_bootmem
1200 * puts them into the mem_map).
1210 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1211 /* Put them into a private list first because mem_map is not up yet */
1212 list_add(&m->list, &huge_boot_pages);
1217 static void prep_compound_huge_page(struct page *page, int order)
1219 if (unlikely(order > (MAX_ORDER - 1)))
1220 prep_compound_gigantic_page(page, order);
1222 prep_compound_page(page, order);
1225 /* Put bootmem huge pages into the standard lists after mem_map is up */
1226 static void __init gather_bootmem_prealloc(void)
1228 struct huge_bootmem_page *m;
1230 list_for_each_entry(m, &huge_boot_pages, list) {
1231 struct hstate *h = m->hstate;
1234 #ifdef CONFIG_HIGHMEM
1235 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1236 free_bootmem_late((unsigned long)m,
1237 sizeof(struct huge_bootmem_page));
1239 page = virt_to_page(m);
1241 __ClearPageReserved(page);
1242 WARN_ON(page_count(page) != 1);
1243 prep_compound_huge_page(page, h->order);
1244 prep_new_huge_page(h, page, page_to_nid(page));
1246 * If we had gigantic hugepages allocated at boot time, we need
1247 * to restore the 'stolen' pages to totalram_pages in order to
1248 * fix confusing memory reports from free(1) and another
1249 * side-effects, like CommitLimit going negative.
1251 if (h->order > (MAX_ORDER - 1))
1252 totalram_pages += 1 << h->order;
1256 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1260 for (i = 0; i < h->max_huge_pages; ++i) {
1261 if (h->order >= MAX_ORDER) {
1262 if (!alloc_bootmem_huge_page(h))
1264 } else if (!alloc_fresh_huge_page(h,
1265 &node_states[N_HIGH_MEMORY]))
1268 h->max_huge_pages = i;
1271 static void __init hugetlb_init_hstates(void)
1275 for_each_hstate(h) {
1276 /* oversize hugepages were init'ed in early boot */
1277 if (h->order < MAX_ORDER)
1278 hugetlb_hstate_alloc_pages(h);
1282 static char * __init memfmt(char *buf, unsigned long n)
1284 if (n >= (1UL << 30))
1285 sprintf(buf, "%lu GB", n >> 30);
1286 else if (n >= (1UL << 20))
1287 sprintf(buf, "%lu MB", n >> 20);
1289 sprintf(buf, "%lu KB", n >> 10);
1293 static void __init report_hugepages(void)
1297 for_each_hstate(h) {
1299 printk(KERN_INFO "HugeTLB registered %s page size, "
1300 "pre-allocated %ld pages\n",
1301 memfmt(buf, huge_page_size(h)),
1302 h->free_huge_pages);
1306 #ifdef CONFIG_HIGHMEM
1307 static void try_to_free_low(struct hstate *h, unsigned long count,
1308 nodemask_t *nodes_allowed)
1312 if (h->order >= MAX_ORDER)
1315 for_each_node_mask(i, *nodes_allowed) {
1316 struct page *page, *next;
1317 struct list_head *freel = &h->hugepage_freelists[i];
1318 list_for_each_entry_safe(page, next, freel, lru) {
1319 if (count >= h->nr_huge_pages)
1321 if (PageHighMem(page))
1323 list_del(&page->lru);
1324 update_and_free_page(h, page);
1325 h->free_huge_pages--;
1326 h->free_huge_pages_node[page_to_nid(page)]--;
1331 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1332 nodemask_t *nodes_allowed)
1338 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1339 * balanced by operating on them in a round-robin fashion.
1340 * Returns 1 if an adjustment was made.
1342 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1345 int start_nid, next_nid;
1348 VM_BUG_ON(delta != -1 && delta != 1);
1351 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1353 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1354 next_nid = start_nid;
1360 * To shrink on this node, there must be a surplus page
1362 if (!h->surplus_huge_pages_node[nid]) {
1363 next_nid = hstate_next_node_to_alloc(h,
1370 * Surplus cannot exceed the total number of pages
1372 if (h->surplus_huge_pages_node[nid] >=
1373 h->nr_huge_pages_node[nid]) {
1374 next_nid = hstate_next_node_to_free(h,
1380 h->surplus_huge_pages += delta;
1381 h->surplus_huge_pages_node[nid] += delta;
1384 } while (next_nid != start_nid);
1389 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1390 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1391 nodemask_t *nodes_allowed)
1393 unsigned long min_count, ret;
1395 if (h->order >= MAX_ORDER)
1396 return h->max_huge_pages;
1399 * Increase the pool size
1400 * First take pages out of surplus state. Then make up the
1401 * remaining difference by allocating fresh huge pages.
1403 * We might race with alloc_buddy_huge_page() here and be unable
1404 * to convert a surplus huge page to a normal huge page. That is
1405 * not critical, though, it just means the overall size of the
1406 * pool might be one hugepage larger than it needs to be, but
1407 * within all the constraints specified by the sysctls.
1409 spin_lock(&hugetlb_lock);
1410 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1411 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1415 while (count > persistent_huge_pages(h)) {
1417 * If this allocation races such that we no longer need the
1418 * page, free_huge_page will handle it by freeing the page
1419 * and reducing the surplus.
1421 spin_unlock(&hugetlb_lock);
1423 /* yield cpu to avoid soft lockup */
1426 ret = alloc_fresh_huge_page(h, nodes_allowed);
1427 spin_lock(&hugetlb_lock);
1431 /* Bail for signals. Probably ctrl-c from user */
1432 if (signal_pending(current))
1437 * Decrease the pool size
1438 * First return free pages to the buddy allocator (being careful
1439 * to keep enough around to satisfy reservations). Then place
1440 * pages into surplus state as needed so the pool will shrink
1441 * to the desired size as pages become free.
1443 * By placing pages into the surplus state independent of the
1444 * overcommit value, we are allowing the surplus pool size to
1445 * exceed overcommit. There are few sane options here. Since
1446 * alloc_buddy_huge_page() is checking the global counter,
1447 * though, we'll note that we're not allowed to exceed surplus
1448 * and won't grow the pool anywhere else. Not until one of the
1449 * sysctls are changed, or the surplus pages go out of use.
1451 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1452 min_count = max(count, min_count);
1453 try_to_free_low(h, min_count, nodes_allowed);
1454 while (min_count < persistent_huge_pages(h)) {
1455 if (!free_pool_huge_page(h, nodes_allowed, 0))
1457 cond_resched_lock(&hugetlb_lock);
1459 while (count < persistent_huge_pages(h)) {
1460 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1464 ret = persistent_huge_pages(h);
1465 spin_unlock(&hugetlb_lock);
1469 #define HSTATE_ATTR_RO(_name) \
1470 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1472 #define HSTATE_ATTR(_name) \
1473 static struct kobj_attribute _name##_attr = \
1474 __ATTR(_name, 0644, _name##_show, _name##_store)
1476 static struct kobject *hugepages_kobj;
1477 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1479 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1481 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1485 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1486 if (hstate_kobjs[i] == kobj) {
1488 *nidp = NUMA_NO_NODE;
1492 return kobj_to_node_hstate(kobj, nidp);
1495 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1496 struct kobj_attribute *attr, char *buf)
1499 unsigned long nr_huge_pages;
1502 h = kobj_to_hstate(kobj, &nid);
1503 if (nid == NUMA_NO_NODE)
1504 nr_huge_pages = h->nr_huge_pages;
1506 nr_huge_pages = h->nr_huge_pages_node[nid];
1508 return sprintf(buf, "%lu\n", nr_huge_pages);
1511 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1512 struct kobject *kobj, struct kobj_attribute *attr,
1513 const char *buf, size_t len)
1517 unsigned long count;
1519 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1521 err = strict_strtoul(buf, 10, &count);
1525 h = kobj_to_hstate(kobj, &nid);
1526 if (h->order >= MAX_ORDER) {
1531 if (nid == NUMA_NO_NODE) {
1533 * global hstate attribute
1535 if (!(obey_mempolicy &&
1536 init_nodemask_of_mempolicy(nodes_allowed))) {
1537 NODEMASK_FREE(nodes_allowed);
1538 nodes_allowed = &node_states[N_HIGH_MEMORY];
1540 } else if (nodes_allowed) {
1542 * per node hstate attribute: adjust count to global,
1543 * but restrict alloc/free to the specified node.
1545 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1546 init_nodemask_of_node(nodes_allowed, nid);
1548 nodes_allowed = &node_states[N_HIGH_MEMORY];
1550 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1552 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1553 NODEMASK_FREE(nodes_allowed);
1557 NODEMASK_FREE(nodes_allowed);
1561 static ssize_t nr_hugepages_show(struct kobject *kobj,
1562 struct kobj_attribute *attr, char *buf)
1564 return nr_hugepages_show_common(kobj, attr, buf);
1567 static ssize_t nr_hugepages_store(struct kobject *kobj,
1568 struct kobj_attribute *attr, const char *buf, size_t len)
1570 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1572 HSTATE_ATTR(nr_hugepages);
1577 * hstate attribute for optionally mempolicy-based constraint on persistent
1578 * huge page alloc/free.
1580 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1581 struct kobj_attribute *attr, char *buf)
1583 return nr_hugepages_show_common(kobj, attr, buf);
1586 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1587 struct kobj_attribute *attr, const char *buf, size_t len)
1589 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1591 HSTATE_ATTR(nr_hugepages_mempolicy);
1595 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1596 struct kobj_attribute *attr, char *buf)
1598 struct hstate *h = kobj_to_hstate(kobj, NULL);
1599 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1602 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1603 struct kobj_attribute *attr, const char *buf, size_t count)
1606 unsigned long input;
1607 struct hstate *h = kobj_to_hstate(kobj, NULL);
1609 if (h->order >= MAX_ORDER)
1612 err = strict_strtoul(buf, 10, &input);
1616 spin_lock(&hugetlb_lock);
1617 h->nr_overcommit_huge_pages = input;
1618 spin_unlock(&hugetlb_lock);
1622 HSTATE_ATTR(nr_overcommit_hugepages);
1624 static ssize_t free_hugepages_show(struct kobject *kobj,
1625 struct kobj_attribute *attr, char *buf)
1628 unsigned long free_huge_pages;
1631 h = kobj_to_hstate(kobj, &nid);
1632 if (nid == NUMA_NO_NODE)
1633 free_huge_pages = h->free_huge_pages;
1635 free_huge_pages = h->free_huge_pages_node[nid];
1637 return sprintf(buf, "%lu\n", free_huge_pages);
1639 HSTATE_ATTR_RO(free_hugepages);
1641 static ssize_t resv_hugepages_show(struct kobject *kobj,
1642 struct kobj_attribute *attr, char *buf)
1644 struct hstate *h = kobj_to_hstate(kobj, NULL);
1645 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1647 HSTATE_ATTR_RO(resv_hugepages);
1649 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1650 struct kobj_attribute *attr, char *buf)
1653 unsigned long surplus_huge_pages;
1656 h = kobj_to_hstate(kobj, &nid);
1657 if (nid == NUMA_NO_NODE)
1658 surplus_huge_pages = h->surplus_huge_pages;
1660 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1662 return sprintf(buf, "%lu\n", surplus_huge_pages);
1664 HSTATE_ATTR_RO(surplus_hugepages);
1666 static struct attribute *hstate_attrs[] = {
1667 &nr_hugepages_attr.attr,
1668 &nr_overcommit_hugepages_attr.attr,
1669 &free_hugepages_attr.attr,
1670 &resv_hugepages_attr.attr,
1671 &surplus_hugepages_attr.attr,
1673 &nr_hugepages_mempolicy_attr.attr,
1678 static struct attribute_group hstate_attr_group = {
1679 .attrs = hstate_attrs,
1682 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1683 struct kobject **hstate_kobjs,
1684 struct attribute_group *hstate_attr_group)
1687 int hi = h - hstates;
1689 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1690 if (!hstate_kobjs[hi])
1693 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1695 kobject_put(hstate_kobjs[hi]);
1700 static void __init hugetlb_sysfs_init(void)
1705 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1706 if (!hugepages_kobj)
1709 for_each_hstate(h) {
1710 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1711 hstate_kobjs, &hstate_attr_group);
1713 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1721 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1722 * with node devices in node_devices[] using a parallel array. The array
1723 * index of a node device or _hstate == node id.
1724 * This is here to avoid any static dependency of the node device driver, in
1725 * the base kernel, on the hugetlb module.
1727 struct node_hstate {
1728 struct kobject *hugepages_kobj;
1729 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1731 struct node_hstate node_hstates[MAX_NUMNODES];
1734 * A subset of global hstate attributes for node devices
1736 static struct attribute *per_node_hstate_attrs[] = {
1737 &nr_hugepages_attr.attr,
1738 &free_hugepages_attr.attr,
1739 &surplus_hugepages_attr.attr,
1743 static struct attribute_group per_node_hstate_attr_group = {
1744 .attrs = per_node_hstate_attrs,
1748 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1749 * Returns node id via non-NULL nidp.
1751 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1755 for (nid = 0; nid < nr_node_ids; nid++) {
1756 struct node_hstate *nhs = &node_hstates[nid];
1758 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1759 if (nhs->hstate_kobjs[i] == kobj) {
1771 * Unregister hstate attributes from a single node device.
1772 * No-op if no hstate attributes attached.
1774 void hugetlb_unregister_node(struct node *node)
1777 struct node_hstate *nhs = &node_hstates[node->dev.id];
1779 if (!nhs->hugepages_kobj)
1780 return; /* no hstate attributes */
1783 if (nhs->hstate_kobjs[h - hstates]) {
1784 kobject_put(nhs->hstate_kobjs[h - hstates]);
1785 nhs->hstate_kobjs[h - hstates] = NULL;
1788 kobject_put(nhs->hugepages_kobj);
1789 nhs->hugepages_kobj = NULL;
1793 * hugetlb module exit: unregister hstate attributes from node devices
1796 static void hugetlb_unregister_all_nodes(void)
1801 * disable node device registrations.
1803 register_hugetlbfs_with_node(NULL, NULL);
1806 * remove hstate attributes from any nodes that have them.
1808 for (nid = 0; nid < nr_node_ids; nid++)
1809 hugetlb_unregister_node(&node_devices[nid]);
1813 * Register hstate attributes for a single node device.
1814 * No-op if attributes already registered.
1816 void hugetlb_register_node(struct node *node)
1819 struct node_hstate *nhs = &node_hstates[node->dev.id];
1822 if (nhs->hugepages_kobj)
1823 return; /* already allocated */
1825 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1827 if (!nhs->hugepages_kobj)
1830 for_each_hstate(h) {
1831 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1833 &per_node_hstate_attr_group);
1835 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1837 h->name, node->dev.id);
1838 hugetlb_unregister_node(node);
1845 * hugetlb init time: register hstate attributes for all registered node
1846 * devices of nodes that have memory. All on-line nodes should have
1847 * registered their associated device by this time.
1849 static void hugetlb_register_all_nodes(void)
1853 for_each_node_state(nid, N_HIGH_MEMORY) {
1854 struct node *node = &node_devices[nid];
1855 if (node->dev.id == nid)
1856 hugetlb_register_node(node);
1860 * Let the node device driver know we're here so it can
1861 * [un]register hstate attributes on node hotplug.
1863 register_hugetlbfs_with_node(hugetlb_register_node,
1864 hugetlb_unregister_node);
1866 #else /* !CONFIG_NUMA */
1868 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1876 static void hugetlb_unregister_all_nodes(void) { }
1878 static void hugetlb_register_all_nodes(void) { }
1882 static void __exit hugetlb_exit(void)
1886 hugetlb_unregister_all_nodes();
1888 for_each_hstate(h) {
1889 kobject_put(hstate_kobjs[h - hstates]);
1892 kobject_put(hugepages_kobj);
1894 module_exit(hugetlb_exit);
1896 static int __init hugetlb_init(void)
1898 if (!hugepages_supported())
1901 if (!size_to_hstate(default_hstate_size)) {
1902 default_hstate_size = HPAGE_SIZE;
1903 if (!size_to_hstate(default_hstate_size))
1904 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1906 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1907 if (default_hstate_max_huge_pages)
1908 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1910 hugetlb_init_hstates();
1912 gather_bootmem_prealloc();
1916 hugetlb_sysfs_init();
1918 hugetlb_register_all_nodes();
1922 module_init(hugetlb_init);
1924 /* Should be called on processing a hugepagesz=... option */
1925 void __init hugetlb_add_hstate(unsigned order)
1930 if (size_to_hstate(PAGE_SIZE << order)) {
1931 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1934 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1936 h = &hstates[max_hstate++];
1938 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1939 h->nr_huge_pages = 0;
1940 h->free_huge_pages = 0;
1941 for (i = 0; i < MAX_NUMNODES; ++i)
1942 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1943 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1944 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1945 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1946 huge_page_size(h)/1024);
1951 static int __init hugetlb_nrpages_setup(char *s)
1954 static unsigned long *last_mhp;
1957 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1958 * so this hugepages= parameter goes to the "default hstate".
1961 mhp = &default_hstate_max_huge_pages;
1963 mhp = &parsed_hstate->max_huge_pages;
1965 if (mhp == last_mhp) {
1966 printk(KERN_WARNING "hugepages= specified twice without "
1967 "interleaving hugepagesz=, ignoring\n");
1971 if (sscanf(s, "%lu", mhp) <= 0)
1975 * Global state is always initialized later in hugetlb_init.
1976 * But we need to allocate >= MAX_ORDER hstates here early to still
1977 * use the bootmem allocator.
1979 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1980 hugetlb_hstate_alloc_pages(parsed_hstate);
1986 __setup("hugepages=", hugetlb_nrpages_setup);
1988 static int __init hugetlb_default_setup(char *s)
1990 default_hstate_size = memparse(s, &s);
1993 __setup("default_hugepagesz=", hugetlb_default_setup);
1995 static unsigned int cpuset_mems_nr(unsigned int *array)
1998 unsigned int nr = 0;
2000 for_each_node_mask(node, cpuset_current_mems_allowed)
2006 #ifdef CONFIG_SYSCTL
2007 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2008 struct ctl_table *table, int write,
2009 void __user *buffer, size_t *length, loff_t *ppos)
2011 struct hstate *h = &default_hstate;
2015 if (!hugepages_supported())
2018 tmp = h->max_huge_pages;
2020 if (write && h->order >= MAX_ORDER)
2024 table->maxlen = sizeof(unsigned long);
2025 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2030 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2031 GFP_KERNEL | __GFP_NORETRY);
2032 if (!(obey_mempolicy &&
2033 init_nodemask_of_mempolicy(nodes_allowed))) {
2034 NODEMASK_FREE(nodes_allowed);
2035 nodes_allowed = &node_states[N_HIGH_MEMORY];
2037 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2039 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2040 NODEMASK_FREE(nodes_allowed);
2046 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2047 void __user *buffer, size_t *length, loff_t *ppos)
2050 return hugetlb_sysctl_handler_common(false, table, write,
2051 buffer, length, ppos);
2055 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2056 void __user *buffer, size_t *length, loff_t *ppos)
2058 return hugetlb_sysctl_handler_common(true, table, write,
2059 buffer, length, ppos);
2061 #endif /* CONFIG_NUMA */
2063 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2064 void __user *buffer,
2065 size_t *length, loff_t *ppos)
2067 proc_dointvec(table, write, buffer, length, ppos);
2068 if (hugepages_treat_as_movable)
2069 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2071 htlb_alloc_mask = GFP_HIGHUSER;
2075 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2076 void __user *buffer,
2077 size_t *length, loff_t *ppos)
2079 struct hstate *h = &default_hstate;
2083 if (!hugepages_supported())
2086 tmp = h->nr_overcommit_huge_pages;
2088 if (write && h->order >= MAX_ORDER)
2092 table->maxlen = sizeof(unsigned long);
2093 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2098 spin_lock(&hugetlb_lock);
2099 h->nr_overcommit_huge_pages = tmp;
2100 spin_unlock(&hugetlb_lock);
2106 #endif /* CONFIG_SYSCTL */
2108 void hugetlb_report_meminfo(struct seq_file *m)
2110 struct hstate *h = &default_hstate;
2111 if (!hugepages_supported())
2114 "HugePages_Total: %5lu\n"
2115 "HugePages_Free: %5lu\n"
2116 "HugePages_Rsvd: %5lu\n"
2117 "HugePages_Surp: %5lu\n"
2118 "Hugepagesize: %8lu kB\n",
2122 h->surplus_huge_pages,
2123 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2126 int hugetlb_report_node_meminfo(int nid, char *buf)
2128 struct hstate *h = &default_hstate;
2129 if (!hugepages_supported())
2132 "Node %d HugePages_Total: %5u\n"
2133 "Node %d HugePages_Free: %5u\n"
2134 "Node %d HugePages_Surp: %5u\n",
2135 nid, h->nr_huge_pages_node[nid],
2136 nid, h->free_huge_pages_node[nid],
2137 nid, h->surplus_huge_pages_node[nid]);
2140 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2141 unsigned long hugetlb_total_pages(void)
2144 unsigned long nr_total_pages = 0;
2147 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2148 return nr_total_pages;
2151 static int hugetlb_acct_memory(struct hstate *h, long delta)
2155 spin_lock(&hugetlb_lock);
2157 * When cpuset is configured, it breaks the strict hugetlb page
2158 * reservation as the accounting is done on a global variable. Such
2159 * reservation is completely rubbish in the presence of cpuset because
2160 * the reservation is not checked against page availability for the
2161 * current cpuset. Application can still potentially OOM'ed by kernel
2162 * with lack of free htlb page in cpuset that the task is in.
2163 * Attempt to enforce strict accounting with cpuset is almost
2164 * impossible (or too ugly) because cpuset is too fluid that
2165 * task or memory node can be dynamically moved between cpusets.
2167 * The change of semantics for shared hugetlb mapping with cpuset is
2168 * undesirable. However, in order to preserve some of the semantics,
2169 * we fall back to check against current free page availability as
2170 * a best attempt and hopefully to minimize the impact of changing
2171 * semantics that cpuset has.
2174 if (gather_surplus_pages(h, delta) < 0)
2177 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2178 return_unused_surplus_pages(h, delta);
2185 return_unused_surplus_pages(h, (unsigned long) -delta);
2188 spin_unlock(&hugetlb_lock);
2192 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2194 struct resv_map *reservations = vma_resv_map(vma);
2197 * This new VMA should share its siblings reservation map if present.
2198 * The VMA will only ever have a valid reservation map pointer where
2199 * it is being copied for another still existing VMA. As that VMA
2200 * has a reference to the reservation map it cannot disappear until
2201 * after this open call completes. It is therefore safe to take a
2202 * new reference here without additional locking.
2205 kref_get(&reservations->refs);
2208 static void resv_map_put(struct vm_area_struct *vma)
2210 struct resv_map *reservations = vma_resv_map(vma);
2214 kref_put(&reservations->refs, resv_map_release);
2217 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2219 struct hstate *h = hstate_vma(vma);
2220 struct resv_map *reservations = vma_resv_map(vma);
2221 struct hugepage_subpool *spool = subpool_vma(vma);
2222 unsigned long reserve;
2223 unsigned long start;
2227 start = vma_hugecache_offset(h, vma, vma->vm_start);
2228 end = vma_hugecache_offset(h, vma, vma->vm_end);
2230 reserve = (end - start) -
2231 region_count(&reservations->regions, start, end);
2236 hugetlb_acct_memory(h, -reserve);
2237 hugepage_subpool_put_pages(spool, reserve);
2243 * We cannot handle pagefaults against hugetlb pages at all. They cause
2244 * handle_mm_fault() to try to instantiate regular-sized pages in the
2245 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2248 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2254 const struct vm_operations_struct hugetlb_vm_ops = {
2255 .fault = hugetlb_vm_op_fault,
2256 .open = hugetlb_vm_op_open,
2257 .close = hugetlb_vm_op_close,
2260 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2267 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2269 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2271 entry = pte_mkyoung(entry);
2272 entry = pte_mkhuge(entry);
2277 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2278 unsigned long address, pte_t *ptep)
2282 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2283 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2284 update_mmu_cache(vma, address, ptep);
2287 static int is_hugetlb_entry_migration(pte_t pte)
2291 if (huge_pte_none(pte) || pte_present(pte))
2293 swp = pte_to_swp_entry(pte);
2294 if (non_swap_entry(swp) && is_migration_entry(swp))
2300 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2304 if (huge_pte_none(pte) || pte_present(pte))
2306 swp = pte_to_swp_entry(pte);
2307 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2313 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2314 struct vm_area_struct *vma)
2316 pte_t *src_pte, *dst_pte, entry;
2317 struct page *ptepage;
2320 struct hstate *h = hstate_vma(vma);
2321 unsigned long sz = huge_page_size(h);
2323 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2325 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2326 src_pte = huge_pte_offset(src, addr);
2329 dst_pte = huge_pte_alloc(dst, addr, sz);
2333 /* If the pagetables are shared don't copy or take references */
2334 if (dst_pte == src_pte)
2337 spin_lock(&dst->page_table_lock);
2338 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2339 entry = huge_ptep_get(src_pte);
2340 if (huge_pte_none(entry)) { /* skip none entry */
2342 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2343 is_hugetlb_entry_hwpoisoned(entry))) {
2344 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2346 if (is_write_migration_entry(swp_entry) && cow) {
2348 * COW mappings require pages in both
2349 * parent and child to be set to read.
2351 make_migration_entry_read(&swp_entry);
2352 entry = swp_entry_to_pte(swp_entry);
2353 set_huge_pte_at(src, addr, src_pte, entry);
2355 set_huge_pte_at(dst, addr, dst_pte, entry);
2358 huge_ptep_set_wrprotect(src, addr, src_pte);
2359 entry = huge_ptep_get(src_pte);
2360 ptepage = pte_page(entry);
2362 page_dup_rmap(ptepage);
2363 set_huge_pte_at(dst, addr, dst_pte, entry);
2365 spin_unlock(&src->page_table_lock);
2366 spin_unlock(&dst->page_table_lock);
2374 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2375 unsigned long end, struct page *ref_page)
2377 struct mm_struct *mm = vma->vm_mm;
2378 unsigned long address;
2383 struct hstate *h = hstate_vma(vma);
2384 unsigned long sz = huge_page_size(h);
2387 * A page gathering list, protected by per file i_mmap_mutex. The
2388 * lock is used to avoid list corruption from multiple unmapping
2389 * of the same page since we are using page->lru.
2391 LIST_HEAD(page_list);
2393 WARN_ON(!is_vm_hugetlb_page(vma));
2394 BUG_ON(start & ~huge_page_mask(h));
2395 BUG_ON(end & ~huge_page_mask(h));
2397 mmu_notifier_invalidate_range_start(mm, start, end);
2398 spin_lock(&mm->page_table_lock);
2399 for (address = start; address < end; address += sz) {
2400 ptep = huge_pte_offset(mm, address);
2404 if (huge_pmd_unshare(mm, &address, ptep))
2408 * If a reference page is supplied, it is because a specific
2409 * page is being unmapped, not a range. Ensure the page we
2410 * are about to unmap is the actual page of interest.
2413 pte = huge_ptep_get(ptep);
2414 if (huge_pte_none(pte))
2416 page = pte_page(pte);
2417 if (page != ref_page)
2421 * Mark the VMA as having unmapped its page so that
2422 * future faults in this VMA will fail rather than
2423 * looking like data was lost
2425 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2428 pte = huge_ptep_get_and_clear(mm, address, ptep);
2429 if (huge_pte_none(pte))
2433 * Migrating hugepage or HWPoisoned hugepage is already
2434 * unmapped and its refcount is dropped
2436 if (unlikely(!pte_present(pte)))
2439 page = pte_page(pte);
2441 set_page_dirty(page);
2442 list_add(&page->lru, &page_list);
2444 spin_unlock(&mm->page_table_lock);
2445 flush_tlb_range(vma, start, end);
2446 mmu_notifier_invalidate_range_end(mm, start, end);
2447 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2448 page_remove_rmap(page);
2449 list_del(&page->lru);
2454 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2455 unsigned long start, unsigned long end,
2456 struct page *ref_page)
2458 __unmap_hugepage_range(vma, start, end, ref_page);
2461 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2462 * test will fail on a vma being torn down, and not grab a page table
2463 * on its way out. We're lucky that the flag has such an appropriate
2464 * name, and can in fact be safely cleared here. We could clear it
2465 * before the __unmap_hugepage_range above, but all that's necessary
2466 * is to clear it before releasing the i_mmap_mutex. This works
2467 * because in the context this is called, the VMA is about to be
2468 * destroyed and the i_mmap_mutex is held.
2470 vma->vm_flags &= ~VM_MAYSHARE;
2473 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2474 unsigned long end, struct page *ref_page)
2476 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2477 __unmap_hugepage_range(vma, start, end, ref_page);
2478 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2482 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2483 * mappping it owns the reserve page for. The intention is to unmap the page
2484 * from other VMAs and let the children be SIGKILLed if they are faulting the
2487 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2488 struct page *page, unsigned long address)
2490 struct hstate *h = hstate_vma(vma);
2491 struct vm_area_struct *iter_vma;
2492 struct address_space *mapping;
2493 struct prio_tree_iter iter;
2497 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2498 * from page cache lookup which is in HPAGE_SIZE units.
2500 address = address & huge_page_mask(h);
2501 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2503 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2506 * Take the mapping lock for the duration of the table walk. As
2507 * this mapping should be shared between all the VMAs,
2508 * __unmap_hugepage_range() is called as the lock is already held
2510 mutex_lock(&mapping->i_mmap_mutex);
2511 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2512 /* Do not unmap the current VMA */
2513 if (iter_vma == vma)
2517 * Shared VMAs have their own reserves and do not affect
2518 * MAP_PRIVATE accounting but it is possible that a shared
2519 * VMA is using the same page so check and skip such VMAs.
2521 if (iter_vma->vm_flags & VM_MAYSHARE)
2525 * Unmap the page from other VMAs without their own reserves.
2526 * They get marked to be SIGKILLed if they fault in these
2527 * areas. This is because a future no-page fault on this VMA
2528 * could insert a zeroed page instead of the data existing
2529 * from the time of fork. This would look like data corruption
2531 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2532 __unmap_hugepage_range(iter_vma,
2533 address, address + huge_page_size(h),
2536 mutex_unlock(&mapping->i_mmap_mutex);
2542 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2544 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2545 unsigned long address, pte_t *ptep, pte_t pte,
2546 struct page *pagecache_page)
2548 struct hstate *h = hstate_vma(vma);
2549 struct page *old_page, *new_page;
2551 int outside_reserve = 0;
2553 old_page = pte_page(pte);
2556 /* If no-one else is actually using this page, avoid the copy
2557 * and just make the page writable */
2558 avoidcopy = (page_mapcount(old_page) == 1);
2560 if (PageAnon(old_page))
2561 page_move_anon_rmap(old_page, vma, address);
2562 set_huge_ptep_writable(vma, address, ptep);
2567 * If the process that created a MAP_PRIVATE mapping is about to
2568 * perform a COW due to a shared page count, attempt to satisfy
2569 * the allocation without using the existing reserves. The pagecache
2570 * page is used to determine if the reserve at this address was
2571 * consumed or not. If reserves were used, a partial faulted mapping
2572 * at the time of fork() could consume its reserves on COW instead
2573 * of the full address range.
2575 if (!(vma->vm_flags & VM_MAYSHARE) &&
2576 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2577 old_page != pagecache_page)
2578 outside_reserve = 1;
2580 page_cache_get(old_page);
2582 /* Drop page_table_lock as buddy allocator may be called */
2583 spin_unlock(&mm->page_table_lock);
2584 new_page = alloc_huge_page(vma, address, outside_reserve);
2586 if (IS_ERR(new_page)) {
2587 page_cache_release(old_page);
2590 * If a process owning a MAP_PRIVATE mapping fails to COW,
2591 * it is due to references held by a child and an insufficient
2592 * huge page pool. To guarantee the original mappers
2593 * reliability, unmap the page from child processes. The child
2594 * may get SIGKILLed if it later faults.
2596 if (outside_reserve) {
2597 BUG_ON(huge_pte_none(pte));
2598 if (unmap_ref_private(mm, vma, old_page, address)) {
2599 BUG_ON(huge_pte_none(pte));
2600 spin_lock(&mm->page_table_lock);
2601 goto retry_avoidcopy;
2606 /* Caller expects lock to be held */
2607 spin_lock(&mm->page_table_lock);
2608 return -PTR_ERR(new_page);
2612 * When the original hugepage is shared one, it does not have
2613 * anon_vma prepared.
2615 if (unlikely(anon_vma_prepare(vma))) {
2616 page_cache_release(new_page);
2617 page_cache_release(old_page);
2618 /* Caller expects lock to be held */
2619 spin_lock(&mm->page_table_lock);
2620 return VM_FAULT_OOM;
2623 copy_user_huge_page(new_page, old_page, address, vma,
2624 pages_per_huge_page(h));
2625 __SetPageUptodate(new_page);
2628 * Retake the page_table_lock to check for racing updates
2629 * before the page tables are altered
2631 spin_lock(&mm->page_table_lock);
2632 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2633 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2635 mmu_notifier_invalidate_range_start(mm,
2636 address & huge_page_mask(h),
2637 (address & huge_page_mask(h)) + huge_page_size(h));
2638 huge_ptep_clear_flush(vma, address, ptep);
2639 set_huge_pte_at(mm, address, ptep,
2640 make_huge_pte(vma, new_page, 1));
2641 page_remove_rmap(old_page);
2642 hugepage_add_new_anon_rmap(new_page, vma, address);
2643 /* Make the old page be freed below */
2644 new_page = old_page;
2645 mmu_notifier_invalidate_range_end(mm,
2646 address & huge_page_mask(h),
2647 (address & huge_page_mask(h)) + huge_page_size(h));
2649 page_cache_release(new_page);
2650 page_cache_release(old_page);
2654 /* Return the pagecache page at a given address within a VMA */
2655 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2656 struct vm_area_struct *vma, unsigned long address)
2658 struct address_space *mapping;
2661 mapping = vma->vm_file->f_mapping;
2662 idx = vma_hugecache_offset(h, vma, address);
2664 return find_lock_page(mapping, idx);
2668 * Return whether there is a pagecache page to back given address within VMA.
2669 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2671 static bool hugetlbfs_pagecache_present(struct hstate *h,
2672 struct vm_area_struct *vma, unsigned long address)
2674 struct address_space *mapping;
2678 mapping = vma->vm_file->f_mapping;
2679 idx = vma_hugecache_offset(h, vma, address);
2681 page = find_get_page(mapping, idx);
2684 return page != NULL;
2687 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2688 unsigned long address, pte_t *ptep, unsigned int flags)
2690 struct hstate *h = hstate_vma(vma);
2691 int ret = VM_FAULT_SIGBUS;
2695 struct address_space *mapping;
2699 * Currently, we are forced to kill the process in the event the
2700 * original mapper has unmapped pages from the child due to a failed
2701 * COW. Warn that such a situation has occurred as it may not be obvious
2703 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2705 "PID %d killed due to inadequate hugepage pool\n",
2710 mapping = vma->vm_file->f_mapping;
2711 idx = vma_hugecache_offset(h, vma, address);
2714 * Use page lock to guard against racing truncation
2715 * before we get page_table_lock.
2718 page = find_lock_page(mapping, idx);
2720 size = i_size_read(mapping->host) >> huge_page_shift(h);
2723 page = alloc_huge_page(vma, address, 0);
2725 ret = -PTR_ERR(page);
2728 clear_huge_page(page, address, pages_per_huge_page(h));
2729 __SetPageUptodate(page);
2731 if (vma->vm_flags & VM_MAYSHARE) {
2733 struct inode *inode = mapping->host;
2735 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2743 spin_lock(&inode->i_lock);
2744 inode->i_blocks += blocks_per_huge_page(h);
2745 spin_unlock(&inode->i_lock);
2746 page_dup_rmap(page);
2749 if (unlikely(anon_vma_prepare(vma))) {
2751 goto backout_unlocked;
2753 hugepage_add_new_anon_rmap(page, vma, address);
2757 * If memory error occurs between mmap() and fault, some process
2758 * don't have hwpoisoned swap entry for errored virtual address.
2759 * So we need to block hugepage fault by PG_hwpoison bit check.
2761 if (unlikely(PageHWPoison(page))) {
2762 ret = VM_FAULT_HWPOISON |
2763 VM_FAULT_SET_HINDEX(h - hstates);
2764 goto backout_unlocked;
2766 page_dup_rmap(page);
2770 * If we are going to COW a private mapping later, we examine the
2771 * pending reservations for this page now. This will ensure that
2772 * any allocations necessary to record that reservation occur outside
2775 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2776 if (vma_needs_reservation(h, vma, address) < 0) {
2778 goto backout_unlocked;
2781 spin_lock(&mm->page_table_lock);
2782 size = i_size_read(mapping->host) >> huge_page_shift(h);
2787 if (!huge_pte_none(huge_ptep_get(ptep)))
2790 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2791 && (vma->vm_flags & VM_SHARED)));
2792 set_huge_pte_at(mm, address, ptep, new_pte);
2794 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2795 /* Optimization, do the COW without a second fault */
2796 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2799 spin_unlock(&mm->page_table_lock);
2805 spin_unlock(&mm->page_table_lock);
2812 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2813 unsigned long address, unsigned int flags)
2818 struct page *page = NULL;
2819 struct page *pagecache_page = NULL;
2820 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2821 struct hstate *h = hstate_vma(vma);
2822 int need_wait_lock = 0;
2824 ptep = huge_pte_offset(mm, address);
2826 entry = huge_ptep_get(ptep);
2827 if (unlikely(is_hugetlb_entry_migration(entry))) {
2828 migration_entry_wait_huge(mm, ptep);
2830 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2831 return VM_FAULT_HWPOISON_LARGE |
2832 VM_FAULT_SET_HINDEX(h - hstates);
2834 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2836 return VM_FAULT_OOM;
2840 * Serialize hugepage allocation and instantiation, so that we don't
2841 * get spurious allocation failures if two CPUs race to instantiate
2842 * the same page in the page cache.
2844 mutex_lock(&hugetlb_instantiation_mutex);
2845 entry = huge_ptep_get(ptep);
2846 if (huge_pte_none(entry)) {
2847 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2854 * entry could be a migration/hwpoison entry at this point, so this
2855 * check prevents the kernel from going below assuming that we have
2856 * a active hugepage in pagecache. This goto expects the 2nd page fault,
2857 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
2860 if (!pte_present(entry))
2864 * If we are going to COW the mapping later, we examine the pending
2865 * reservations for this page now. This will ensure that any
2866 * allocations necessary to record that reservation occur outside the
2867 * spinlock. For private mappings, we also lookup the pagecache
2868 * page now as it is used to determine if a reservation has been
2871 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2872 if (vma_needs_reservation(h, vma, address) < 0) {
2877 if (!(vma->vm_flags & VM_MAYSHARE))
2878 pagecache_page = hugetlbfs_pagecache_page(h,
2882 spin_lock(&mm->page_table_lock);
2883 /* Check for a racing update before calling hugetlb_cow */
2884 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2885 goto out_page_table_lock;
2888 * hugetlb_cow() requires page locks of pte_page(entry) and
2889 * pagecache_page, so here we need take the former one
2890 * when page != pagecache_page or !pagecache_page.
2892 page = pte_page(entry);
2893 if (page != pagecache_page)
2894 if (!trylock_page(page)) {
2896 goto out_page_table_lock;
2901 if (flags & FAULT_FLAG_WRITE) {
2902 if (!pte_write(entry)) {
2903 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2907 entry = pte_mkdirty(entry);
2909 entry = pte_mkyoung(entry);
2910 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2911 flags & FAULT_FLAG_WRITE))
2912 update_mmu_cache(vma, address, ptep);
2914 if (page != pagecache_page)
2917 out_page_table_lock:
2918 spin_unlock(&mm->page_table_lock);
2920 if (pagecache_page) {
2921 unlock_page(pagecache_page);
2922 put_page(pagecache_page);
2925 mutex_unlock(&hugetlb_instantiation_mutex);
2928 * Generally it's safe to hold refcount during waiting page lock. But
2929 * here we just wait to defer the next page fault to avoid busy loop and
2930 * the page is not used after unlocked before returning from the current
2931 * page fault. So we are safe from accessing freed page, even if we wait
2932 * here without taking refcount.
2935 wait_on_page_locked(page);
2939 /* Can be overriden by architectures */
2940 __attribute__((weak)) struct page *
2941 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2942 pud_t *pud, int write)
2948 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2949 struct page **pages, struct vm_area_struct **vmas,
2950 unsigned long *position, int *length, int i,
2953 unsigned long pfn_offset;
2954 unsigned long vaddr = *position;
2955 int remainder = *length;
2956 struct hstate *h = hstate_vma(vma);
2958 spin_lock(&mm->page_table_lock);
2959 while (vaddr < vma->vm_end && remainder) {
2965 * Some archs (sparc64, sh*) have multiple pte_ts to
2966 * each hugepage. We have to make sure we get the
2967 * first, for the page indexing below to work.
2969 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2970 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2973 * When coredumping, it suits get_dump_page if we just return
2974 * an error where there's an empty slot with no huge pagecache
2975 * to back it. This way, we avoid allocating a hugepage, and
2976 * the sparse dumpfile avoids allocating disk blocks, but its
2977 * huge holes still show up with zeroes where they need to be.
2979 if (absent && (flags & FOLL_DUMP) &&
2980 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2986 * We need call hugetlb_fault for both hugepages under migration
2987 * (in which case hugetlb_fault waits for the migration,) and
2988 * hwpoisoned hugepages (in which case we need to prevent the
2989 * caller from accessing to them.) In order to do this, we use
2990 * here is_swap_pte instead of is_hugetlb_entry_migration and
2991 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2992 * both cases, and because we can't follow correct pages
2993 * directly from any kind of swap entries.
2995 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2996 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2999 spin_unlock(&mm->page_table_lock);
3000 ret = hugetlb_fault(mm, vma, vaddr,
3001 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3002 spin_lock(&mm->page_table_lock);
3003 if (!(ret & VM_FAULT_ERROR))
3010 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3011 page = pte_page(huge_ptep_get(pte));
3014 pages[i] = mem_map_offset(page, pfn_offset);
3025 if (vaddr < vma->vm_end && remainder &&
3026 pfn_offset < pages_per_huge_page(h)) {
3028 * We use pfn_offset to avoid touching the pageframes
3029 * of this compound page.
3034 spin_unlock(&mm->page_table_lock);
3035 *length = remainder;
3038 return i ? i : -EFAULT;
3041 void hugetlb_change_protection(struct vm_area_struct *vma,
3042 unsigned long address, unsigned long end, pgprot_t newprot)
3044 struct mm_struct *mm = vma->vm_mm;
3045 unsigned long start = address;
3048 struct hstate *h = hstate_vma(vma);
3050 BUG_ON(address >= end);
3051 flush_cache_range(vma, address, end);
3053 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3054 spin_lock(&mm->page_table_lock);
3055 for (; address < end; address += huge_page_size(h)) {
3056 ptep = huge_pte_offset(mm, address);
3059 if (huge_pmd_unshare(mm, &address, ptep))
3061 pte = huge_ptep_get(ptep);
3062 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
3064 if (unlikely(is_hugetlb_entry_migration(pte))) {
3065 swp_entry_t entry = pte_to_swp_entry(pte);
3067 if (is_write_migration_entry(entry)) {
3070 make_migration_entry_read(&entry);
3071 newpte = swp_entry_to_pte(entry);
3072 set_huge_pte_at(mm, address, ptep, newpte);
3076 if (!huge_pte_none(pte)) {
3077 pte = huge_ptep_get_and_clear(mm, address, ptep);
3078 pte = pte_mkhuge(pte_modify(pte, newprot));
3079 set_huge_pte_at(mm, address, ptep, pte);
3082 spin_unlock(&mm->page_table_lock);
3084 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3085 * may have cleared our pud entry and done put_page on the page table:
3086 * once we release i_mmap_mutex, another task can do the final put_page
3087 * and that page table be reused and filled with junk.
3089 flush_tlb_range(vma, start, end);
3090 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3093 int hugetlb_reserve_pages(struct inode *inode,
3095 struct vm_area_struct *vma,
3096 vm_flags_t vm_flags)
3099 struct hstate *h = hstate_inode(inode);
3100 struct hugepage_subpool *spool = subpool_inode(inode);
3102 /* This should never happen */
3104 #ifdef CONFIG_DEBUG_VM
3105 WARN(1, "%s called with a negative range\n", __func__);
3111 * Only apply hugepage reservation if asked. At fault time, an
3112 * attempt will be made for VM_NORESERVE to allocate a page
3113 * without using reserves
3115 if (vm_flags & VM_NORESERVE)
3119 * Shared mappings base their reservation on the number of pages that
3120 * are already allocated on behalf of the file. Private mappings need
3121 * to reserve the full area even if read-only as mprotect() may be
3122 * called to make the mapping read-write. Assume !vma is a shm mapping
3124 if (!vma || vma->vm_flags & VM_MAYSHARE)
3125 chg = region_chg(&inode->i_mapping->private_list, from, to);
3127 struct resv_map *resv_map = resv_map_alloc();
3133 set_vma_resv_map(vma, resv_map);
3134 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3142 /* There must be enough pages in the subpool for the mapping */
3143 if (hugepage_subpool_get_pages(spool, chg)) {
3149 * Check enough hugepages are available for the reservation.
3150 * Hand the pages back to the subpool if there are not
3152 ret = hugetlb_acct_memory(h, chg);
3154 hugepage_subpool_put_pages(spool, chg);
3159 * Account for the reservations made. Shared mappings record regions
3160 * that have reservations as they are shared by multiple VMAs.
3161 * When the last VMA disappears, the region map says how much
3162 * the reservation was and the page cache tells how much of
3163 * the reservation was consumed. Private mappings are per-VMA and
3164 * only the consumed reservations are tracked. When the VMA
3165 * disappears, the original reservation is the VMA size and the
3166 * consumed reservations are stored in the map. Hence, nothing
3167 * else has to be done for private mappings here
3169 if (!vma || vma->vm_flags & VM_MAYSHARE)
3170 region_add(&inode->i_mapping->private_list, from, to);
3178 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3180 struct hstate *h = hstate_inode(inode);
3181 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3182 struct hugepage_subpool *spool = subpool_inode(inode);
3184 spin_lock(&inode->i_lock);
3185 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3186 spin_unlock(&inode->i_lock);
3188 hugepage_subpool_put_pages(spool, (chg - freed));
3189 hugetlb_acct_memory(h, -(chg - freed));
3192 #ifdef CONFIG_MEMORY_FAILURE
3194 /* Should be called in hugetlb_lock */
3195 static int is_hugepage_on_freelist(struct page *hpage)
3199 struct hstate *h = page_hstate(hpage);
3200 int nid = page_to_nid(hpage);
3202 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3209 * This function is called from memory failure code.
3210 * Assume the caller holds page lock of the head page.
3212 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3214 struct hstate *h = page_hstate(hpage);
3215 int nid = page_to_nid(hpage);
3218 spin_lock(&hugetlb_lock);
3219 if (is_hugepage_on_freelist(hpage)) {
3220 list_del(&hpage->lru);
3221 set_page_refcounted(hpage);
3222 h->free_huge_pages--;
3223 h->free_huge_pages_node[nid]--;
3226 spin_unlock(&hugetlb_lock);