2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, 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>
24 #include <linux/page-isolation.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
37 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
38 unsigned long hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 __initdata LIST_HEAD(huge_boot_pages);
46 /* for command line parsing */
47 static struct hstate * __initdata parsed_hstate;
48 static unsigned long __initdata default_hstate_max_huge_pages;
49 static unsigned long __initdata default_hstate_size;
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
55 DEFINE_SPINLOCK(hugetlb_lock);
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
61 static int num_fault_mutexes;
62 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
64 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
66 bool free = (spool->count == 0) && (spool->used_hpages == 0);
68 spin_unlock(&spool->lock);
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
76 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
78 struct hugepage_subpool *spool;
80 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
84 spin_lock_init(&spool->lock);
86 spool->max_hpages = nr_blocks;
87 spool->used_hpages = 0;
92 void hugepage_put_subpool(struct hugepage_subpool *spool)
94 spin_lock(&spool->lock);
95 BUG_ON(!spool->count);
97 unlock_or_release_subpool(spool);
100 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
108 spin_lock(&spool->lock);
109 if ((spool->used_hpages + delta) <= spool->max_hpages) {
110 spool->used_hpages += delta;
114 spin_unlock(&spool->lock);
119 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
125 spin_lock(&spool->lock);
126 spool->used_hpages -= delta;
127 /* If hugetlbfs_put_super couldn't free spool due to
128 * an outstanding quota reference, free it now. */
129 unlock_or_release_subpool(spool);
132 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
134 return HUGETLBFS_SB(inode->i_sb)->spool;
137 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
139 return subpool_inode(file_inode(vma->vm_file));
143 * Region tracking -- allows tracking of reservations and instantiated pages
144 * across the pages in a mapping.
146 * The region data structures are embedded into a resv_map and
147 * protected by a resv_map's lock
150 struct list_head link;
155 static long region_add(struct resv_map *resv, long f, long t)
157 struct list_head *head = &resv->regions;
158 struct file_region *rg, *nrg, *trg;
160 spin_lock(&resv->lock);
161 /* Locate the region we are either in or before. */
162 list_for_each_entry(rg, head, link)
166 /* Round our left edge to the current segment if it encloses us. */
170 /* Check for and consume any regions we now overlap with. */
172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
173 if (&rg->link == head)
178 /* If this area reaches higher then extend our area to
179 * include it completely. If this is not the first area
180 * which we intend to reuse, free it. */
190 spin_unlock(&resv->lock);
194 static long region_chg(struct resv_map *resv, long f, long t)
196 struct list_head *head = &resv->regions;
197 struct file_region *rg, *nrg = NULL;
201 spin_lock(&resv->lock);
202 /* Locate the region we are before or in. */
203 list_for_each_entry(rg, head, link)
207 /* If we are below the current region then a new region is required.
208 * Subtle, allocate a new region at the position but make it zero
209 * size such that we can guarantee to record the reservation. */
210 if (&rg->link == head || t < rg->from) {
212 spin_unlock(&resv->lock);
213 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
219 INIT_LIST_HEAD(&nrg->link);
223 list_add(&nrg->link, rg->link.prev);
228 /* Round our left edge to the current segment if it encloses us. */
233 /* Check for and consume any regions we now overlap with. */
234 list_for_each_entry(rg, rg->link.prev, link) {
235 if (&rg->link == head)
240 /* We overlap with this area, if it extends further than
241 * us then we must extend ourselves. Account for its
242 * existing reservation. */
247 chg -= rg->to - rg->from;
251 spin_unlock(&resv->lock);
252 /* We already know we raced and no longer need the new region */
256 spin_unlock(&resv->lock);
260 static long region_truncate(struct resv_map *resv, long end)
262 struct list_head *head = &resv->regions;
263 struct file_region *rg, *trg;
266 spin_lock(&resv->lock);
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
271 if (&rg->link == head)
274 /* If we are in the middle of a region then adjust it. */
275 if (end > rg->from) {
278 rg = list_entry(rg->link.next, typeof(*rg), link);
281 /* Drop any remaining regions. */
282 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
283 if (&rg->link == head)
285 chg += rg->to - rg->from;
291 spin_unlock(&resv->lock);
295 static long region_count(struct resv_map *resv, long f, long t)
297 struct list_head *head = &resv->regions;
298 struct file_region *rg;
301 spin_lock(&resv->lock);
302 /* Locate each segment we overlap with, and count that overlap. */
303 list_for_each_entry(rg, head, link) {
312 seg_from = max(rg->from, f);
313 seg_to = min(rg->to, t);
315 chg += seg_to - seg_from;
317 spin_unlock(&resv->lock);
323 * Convert the address within this vma to the page offset within
324 * the mapping, in pagecache page units; huge pages here.
326 static pgoff_t vma_hugecache_offset(struct hstate *h,
327 struct vm_area_struct *vma, unsigned long address)
329 return ((address - vma->vm_start) >> huge_page_shift(h)) +
330 (vma->vm_pgoff >> huge_page_order(h));
333 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
334 unsigned long address)
336 return vma_hugecache_offset(hstate_vma(vma), vma, address);
340 * Return the size of the pages allocated when backing a VMA. In the majority
341 * cases this will be same size as used by the page table entries.
343 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
345 struct hstate *hstate;
347 if (!is_vm_hugetlb_page(vma))
350 hstate = hstate_vma(vma);
352 return 1UL << huge_page_shift(hstate);
354 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
357 * Return the page size being used by the MMU to back a VMA. In the majority
358 * of cases, the page size used by the kernel matches the MMU size. On
359 * architectures where it differs, an architecture-specific version of this
360 * function is required.
362 #ifndef vma_mmu_pagesize
363 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
365 return vma_kernel_pagesize(vma);
370 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
371 * bits of the reservation map pointer, which are always clear due to
374 #define HPAGE_RESV_OWNER (1UL << 0)
375 #define HPAGE_RESV_UNMAPPED (1UL << 1)
376 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
379 * These helpers are used to track how many pages are reserved for
380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
381 * is guaranteed to have their future faults succeed.
383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
384 * the reserve counters are updated with the hugetlb_lock held. It is safe
385 * to reset the VMA at fork() time as it is not in use yet and there is no
386 * chance of the global counters getting corrupted as a result of the values.
388 * The private mapping reservation is represented in a subtly different
389 * manner to a shared mapping. A shared mapping has a region map associated
390 * with the underlying file, this region map represents the backing file
391 * pages which have ever had a reservation assigned which this persists even
392 * after the page is instantiated. A private mapping has a region map
393 * associated with the original mmap which is attached to all VMAs which
394 * reference it, this region map represents those offsets which have consumed
395 * reservation ie. where pages have been instantiated.
397 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
399 return (unsigned long)vma->vm_private_data;
402 static void set_vma_private_data(struct vm_area_struct *vma,
405 vma->vm_private_data = (void *)value;
408 struct resv_map *resv_map_alloc(void)
410 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
414 kref_init(&resv_map->refs);
415 spin_lock_init(&resv_map->lock);
416 INIT_LIST_HEAD(&resv_map->regions);
421 void resv_map_release(struct kref *ref)
423 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
425 /* Clear out any active regions before we release the map. */
426 region_truncate(resv_map, 0);
430 static inline struct resv_map *inode_resv_map(struct inode *inode)
432 return inode->i_mapping->private_data;
435 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
437 VM_BUG_ON(!is_vm_hugetlb_page(vma));
438 if (vma->vm_flags & VM_MAYSHARE) {
439 struct address_space *mapping = vma->vm_file->f_mapping;
440 struct inode *inode = mapping->host;
442 return inode_resv_map(inode);
445 return (struct resv_map *)(get_vma_private_data(vma) &
450 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
452 VM_BUG_ON(!is_vm_hugetlb_page(vma));
453 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
455 set_vma_private_data(vma, (get_vma_private_data(vma) &
456 HPAGE_RESV_MASK) | (unsigned long)map);
459 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma));
462 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
464 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
467 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
469 VM_BUG_ON(!is_vm_hugetlb_page(vma));
471 return (get_vma_private_data(vma) & flag) != 0;
474 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
475 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
477 VM_BUG_ON(!is_vm_hugetlb_page(vma));
478 if (!(vma->vm_flags & VM_MAYSHARE))
479 vma->vm_private_data = (void *)0;
482 /* Returns true if the VMA has associated reserve pages */
483 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
485 if (vma->vm_flags & VM_NORESERVE) {
487 * This address is already reserved by other process(chg == 0),
488 * so, we should decrement reserved count. Without decrementing,
489 * reserve count remains after releasing inode, because this
490 * allocated page will go into page cache and is regarded as
491 * coming from reserved pool in releasing step. Currently, we
492 * don't have any other solution to deal with this situation
493 * properly, so add work-around here.
495 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
501 /* Shared mappings always use reserves */
502 if (vma->vm_flags & VM_MAYSHARE)
506 * Only the process that called mmap() has reserves for
509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
515 static void enqueue_huge_page(struct hstate *h, struct page *page)
517 int nid = page_to_nid(page);
518 list_move(&page->lru, &h->hugepage_freelists[nid]);
519 h->free_huge_pages++;
520 h->free_huge_pages_node[nid]++;
523 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
527 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
528 if (!is_migrate_isolate_page(page))
531 * if 'non-isolated free hugepage' not found on the list,
532 * the allocation fails.
534 if (&h->hugepage_freelists[nid] == &page->lru)
536 list_move(&page->lru, &h->hugepage_activelist);
537 set_page_refcounted(page);
538 h->free_huge_pages--;
539 h->free_huge_pages_node[nid]--;
543 /* Movability of hugepages depends on migration support. */
544 static inline gfp_t htlb_alloc_mask(struct hstate *h)
546 if (hugepages_treat_as_movable || hugepage_migration_support(h))
547 return GFP_HIGHUSER_MOVABLE;
552 static struct page *dequeue_huge_page_vma(struct hstate *h,
553 struct vm_area_struct *vma,
554 unsigned long address, int avoid_reserve,
557 struct page *page = NULL;
558 struct mempolicy *mpol;
559 nodemask_t *nodemask;
560 struct zonelist *zonelist;
563 unsigned int cpuset_mems_cookie;
566 * A child process with MAP_PRIVATE mappings created by their parent
567 * have no page reserves. This check ensures that reservations are
568 * not "stolen". The child may still get SIGKILLed
570 if (!vma_has_reserves(vma, chg) &&
571 h->free_huge_pages - h->resv_huge_pages == 0)
574 /* If reserves cannot be used, ensure enough pages are in the pool */
575 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
579 cpuset_mems_cookie = read_mems_allowed_begin();
580 zonelist = huge_zonelist(vma, address,
581 htlb_alloc_mask(h), &mpol, &nodemask);
583 for_each_zone_zonelist_nodemask(zone, z, zonelist,
584 MAX_NR_ZONES - 1, nodemask) {
585 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
586 page = dequeue_huge_page_node(h, zone_to_nid(zone));
590 if (!vma_has_reserves(vma, chg))
593 SetPagePrivate(page);
594 h->resv_huge_pages--;
601 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
609 static void update_and_free_page(struct hstate *h, struct page *page)
613 VM_BUG_ON(h->order >= MAX_ORDER);
616 h->nr_huge_pages_node[page_to_nid(page)]--;
617 for (i = 0; i < pages_per_huge_page(h); i++) {
618 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
619 1 << PG_referenced | 1 << PG_dirty |
620 1 << PG_active | 1 << PG_reserved |
621 1 << PG_private | 1 << PG_writeback);
623 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
624 set_compound_page_dtor(page, NULL);
625 set_page_refcounted(page);
626 arch_release_hugepage(page);
627 __free_pages(page, huge_page_order(h));
630 struct hstate *size_to_hstate(unsigned long size)
635 if (huge_page_size(h) == size)
641 static void free_huge_page(struct page *page)
644 * Can't pass hstate in here because it is called from the
645 * compound page destructor.
647 struct hstate *h = page_hstate(page);
648 int nid = page_to_nid(page);
649 struct hugepage_subpool *spool =
650 (struct hugepage_subpool *)page_private(page);
651 bool restore_reserve;
653 set_page_private(page, 0);
654 page->mapping = NULL;
655 BUG_ON(page_count(page));
656 BUG_ON(page_mapcount(page));
657 restore_reserve = PagePrivate(page);
658 ClearPagePrivate(page);
660 spin_lock(&hugetlb_lock);
661 hugetlb_cgroup_uncharge_page(hstate_index(h),
662 pages_per_huge_page(h), page);
664 h->resv_huge_pages++;
666 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
667 /* remove the page from active list */
668 list_del(&page->lru);
669 update_and_free_page(h, page);
670 h->surplus_huge_pages--;
671 h->surplus_huge_pages_node[nid]--;
673 arch_clear_hugepage_flags(page);
674 enqueue_huge_page(h, page);
676 spin_unlock(&hugetlb_lock);
677 hugepage_subpool_put_pages(spool, 1);
680 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
682 INIT_LIST_HEAD(&page->lru);
683 set_compound_page_dtor(page, free_huge_page);
684 spin_lock(&hugetlb_lock);
685 set_hugetlb_cgroup(page, NULL);
687 h->nr_huge_pages_node[nid]++;
688 spin_unlock(&hugetlb_lock);
689 put_page(page); /* free it into the hugepage allocator */
692 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
695 int nr_pages = 1 << order;
696 struct page *p = page + 1;
698 /* we rely on prep_new_huge_page to set the destructor */
699 set_compound_order(page, order);
701 __ClearPageReserved(page);
702 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
705 * For gigantic hugepages allocated through bootmem at
706 * boot, it's safer to be consistent with the not-gigantic
707 * hugepages and clear the PG_reserved bit from all tail pages
708 * too. Otherwse drivers using get_user_pages() to access tail
709 * pages may get the reference counting wrong if they see
710 * PG_reserved set on a tail page (despite the head page not
711 * having PG_reserved set). Enforcing this consistency between
712 * head and tail pages allows drivers to optimize away a check
713 * on the head page when they need know if put_page() is needed
714 * after get_user_pages().
716 __ClearPageReserved(p);
717 set_page_count(p, 0);
718 p->first_page = page;
723 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
724 * transparent huge pages. See the PageTransHuge() documentation for more
727 int PageHuge(struct page *page)
729 if (!PageCompound(page))
732 page = compound_head(page);
733 return get_compound_page_dtor(page) == free_huge_page;
735 EXPORT_SYMBOL_GPL(PageHuge);
738 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
739 * normal or transparent huge pages.
741 int PageHeadHuge(struct page *page_head)
743 if (!PageHead(page_head))
746 return get_compound_page_dtor(page_head) == free_huge_page;
749 pgoff_t __basepage_index(struct page *page)
751 struct page *page_head = compound_head(page);
752 pgoff_t index = page_index(page_head);
753 unsigned long compound_idx;
755 if (!PageHuge(page_head))
756 return page_index(page);
758 if (compound_order(page_head) >= MAX_ORDER)
759 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
761 compound_idx = page - page_head;
763 return (index << compound_order(page_head)) + compound_idx;
766 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
770 if (h->order >= MAX_ORDER)
773 page = alloc_pages_exact_node(nid,
774 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
775 __GFP_REPEAT|__GFP_NOWARN,
778 if (arch_prepare_hugepage(page)) {
779 __free_pages(page, huge_page_order(h));
782 prep_new_huge_page(h, page, nid);
789 * common helper functions for hstate_next_node_to_{alloc|free}.
790 * We may have allocated or freed a huge page based on a different
791 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
792 * be outside of *nodes_allowed. Ensure that we use an allowed
793 * node for alloc or free.
795 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
797 nid = next_node(nid, *nodes_allowed);
798 if (nid == MAX_NUMNODES)
799 nid = first_node(*nodes_allowed);
800 VM_BUG_ON(nid >= MAX_NUMNODES);
805 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
807 if (!node_isset(nid, *nodes_allowed))
808 nid = next_node_allowed(nid, nodes_allowed);
813 * returns the previously saved node ["this node"] from which to
814 * allocate a persistent huge page for the pool and advance the
815 * next node from which to allocate, handling wrap at end of node
818 static int hstate_next_node_to_alloc(struct hstate *h,
819 nodemask_t *nodes_allowed)
823 VM_BUG_ON(!nodes_allowed);
825 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
826 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
832 * helper for free_pool_huge_page() - return the previously saved
833 * node ["this node"] from which to free a huge page. Advance the
834 * next node id whether or not we find a free huge page to free so
835 * that the next attempt to free addresses the next node.
837 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
841 VM_BUG_ON(!nodes_allowed);
843 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
844 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
849 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
850 for (nr_nodes = nodes_weight(*mask); \
852 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
855 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
856 for (nr_nodes = nodes_weight(*mask); \
858 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
861 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
867 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
868 page = alloc_fresh_huge_page_node(h, node);
876 count_vm_event(HTLB_BUDDY_PGALLOC);
878 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
884 * Free huge page from pool from next node to free.
885 * Attempt to keep persistent huge pages more or less
886 * balanced over allowed nodes.
887 * Called with hugetlb_lock locked.
889 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
895 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
897 * If we're returning unused surplus pages, only examine
898 * nodes with surplus pages.
900 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
901 !list_empty(&h->hugepage_freelists[node])) {
903 list_entry(h->hugepage_freelists[node].next,
905 list_del(&page->lru);
906 h->free_huge_pages--;
907 h->free_huge_pages_node[node]--;
909 h->surplus_huge_pages--;
910 h->surplus_huge_pages_node[node]--;
912 update_and_free_page(h, page);
922 * Dissolve a given free hugepage into free buddy pages. This function does
923 * nothing for in-use (including surplus) hugepages.
925 static void dissolve_free_huge_page(struct page *page)
927 spin_lock(&hugetlb_lock);
928 if (PageHuge(page) && !page_count(page)) {
929 struct hstate *h = page_hstate(page);
930 int nid = page_to_nid(page);
931 list_del(&page->lru);
932 h->free_huge_pages--;
933 h->free_huge_pages_node[nid]--;
934 update_and_free_page(h, page);
936 spin_unlock(&hugetlb_lock);
940 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
941 * make specified memory blocks removable from the system.
942 * Note that start_pfn should aligned with (minimum) hugepage size.
944 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
946 unsigned int order = 8 * sizeof(void *);
950 /* Set scan step to minimum hugepage size */
952 if (order > huge_page_order(h))
953 order = huge_page_order(h);
954 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
955 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
956 dissolve_free_huge_page(pfn_to_page(pfn));
959 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
964 if (h->order >= MAX_ORDER)
968 * Assume we will successfully allocate the surplus page to
969 * prevent racing processes from causing the surplus to exceed
972 * This however introduces a different race, where a process B
973 * tries to grow the static hugepage pool while alloc_pages() is
974 * called by process A. B will only examine the per-node
975 * counters in determining if surplus huge pages can be
976 * converted to normal huge pages in adjust_pool_surplus(). A
977 * won't be able to increment the per-node counter, until the
978 * lock is dropped by B, but B doesn't drop hugetlb_lock until
979 * no more huge pages can be converted from surplus to normal
980 * state (and doesn't try to convert again). Thus, we have a
981 * case where a surplus huge page exists, the pool is grown, and
982 * the surplus huge page still exists after, even though it
983 * should just have been converted to a normal huge page. This
984 * does not leak memory, though, as the hugepage will be freed
985 * once it is out of use. It also does not allow the counters to
986 * go out of whack in adjust_pool_surplus() as we don't modify
987 * the node values until we've gotten the hugepage and only the
988 * per-node value is checked there.
990 spin_lock(&hugetlb_lock);
991 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
992 spin_unlock(&hugetlb_lock);
996 h->surplus_huge_pages++;
998 spin_unlock(&hugetlb_lock);
1000 if (nid == NUMA_NO_NODE)
1001 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1002 __GFP_REPEAT|__GFP_NOWARN,
1003 huge_page_order(h));
1005 page = alloc_pages_exact_node(nid,
1006 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1007 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1009 if (page && arch_prepare_hugepage(page)) {
1010 __free_pages(page, huge_page_order(h));
1014 spin_lock(&hugetlb_lock);
1016 INIT_LIST_HEAD(&page->lru);
1017 r_nid = page_to_nid(page);
1018 set_compound_page_dtor(page, free_huge_page);
1019 set_hugetlb_cgroup(page, NULL);
1021 * We incremented the global counters already
1023 h->nr_huge_pages_node[r_nid]++;
1024 h->surplus_huge_pages_node[r_nid]++;
1025 __count_vm_event(HTLB_BUDDY_PGALLOC);
1028 h->surplus_huge_pages--;
1029 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1031 spin_unlock(&hugetlb_lock);
1037 * This allocation function is useful in the context where vma is irrelevant.
1038 * E.g. soft-offlining uses this function because it only cares physical
1039 * address of error page.
1041 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1043 struct page *page = NULL;
1045 spin_lock(&hugetlb_lock);
1046 if (h->free_huge_pages - h->resv_huge_pages > 0)
1047 page = dequeue_huge_page_node(h, nid);
1048 spin_unlock(&hugetlb_lock);
1051 page = alloc_buddy_huge_page(h, nid);
1057 * Increase the hugetlb pool such that it can accommodate a reservation
1060 static int gather_surplus_pages(struct hstate *h, int delta)
1062 struct list_head surplus_list;
1063 struct page *page, *tmp;
1065 int needed, allocated;
1066 bool alloc_ok = true;
1068 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1070 h->resv_huge_pages += delta;
1075 INIT_LIST_HEAD(&surplus_list);
1079 spin_unlock(&hugetlb_lock);
1080 for (i = 0; i < needed; i++) {
1081 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1086 list_add(&page->lru, &surplus_list);
1091 * After retaking hugetlb_lock, we need to recalculate 'needed'
1092 * because either resv_huge_pages or free_huge_pages may have changed.
1094 spin_lock(&hugetlb_lock);
1095 needed = (h->resv_huge_pages + delta) -
1096 (h->free_huge_pages + allocated);
1101 * We were not able to allocate enough pages to
1102 * satisfy the entire reservation so we free what
1103 * we've allocated so far.
1108 * The surplus_list now contains _at_least_ the number of extra pages
1109 * needed to accommodate the reservation. Add the appropriate number
1110 * of pages to the hugetlb pool and free the extras back to the buddy
1111 * allocator. Commit the entire reservation here to prevent another
1112 * process from stealing the pages as they are added to the pool but
1113 * before they are reserved.
1115 needed += allocated;
1116 h->resv_huge_pages += delta;
1119 /* Free the needed pages to the hugetlb pool */
1120 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1124 * This page is now managed by the hugetlb allocator and has
1125 * no users -- drop the buddy allocator's reference.
1127 put_page_testzero(page);
1128 VM_BUG_ON_PAGE(page_count(page), page);
1129 enqueue_huge_page(h, page);
1132 spin_unlock(&hugetlb_lock);
1134 /* Free unnecessary surplus pages to the buddy allocator */
1135 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1137 spin_lock(&hugetlb_lock);
1143 * When releasing a hugetlb pool reservation, any surplus pages that were
1144 * allocated to satisfy the reservation must be explicitly freed if they were
1146 * Called with hugetlb_lock held.
1148 static void return_unused_surplus_pages(struct hstate *h,
1149 unsigned long unused_resv_pages)
1151 unsigned long nr_pages;
1153 /* Uncommit the reservation */
1154 h->resv_huge_pages -= unused_resv_pages;
1156 /* Cannot return gigantic pages currently */
1157 if (h->order >= MAX_ORDER)
1160 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1163 * We want to release as many surplus pages as possible, spread
1164 * evenly across all nodes with memory. Iterate across these nodes
1165 * until we can no longer free unreserved surplus pages. This occurs
1166 * when the nodes with surplus pages have no free pages.
1167 * free_pool_huge_page() will balance the the freed pages across the
1168 * on-line nodes with memory and will handle the hstate accounting.
1170 while (nr_pages--) {
1171 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1177 * Determine if the huge page at addr within the vma has an associated
1178 * reservation. Where it does not we will need to logically increase
1179 * reservation and actually increase subpool usage before an allocation
1180 * can occur. Where any new reservation would be required the
1181 * reservation change is prepared, but not committed. Once the page
1182 * has been allocated from the subpool and instantiated the change should
1183 * be committed via vma_commit_reservation. No action is required on
1186 static long vma_needs_reservation(struct hstate *h,
1187 struct vm_area_struct *vma, unsigned long addr)
1189 struct resv_map *resv;
1193 resv = vma_resv_map(vma);
1197 idx = vma_hugecache_offset(h, vma, addr);
1198 chg = region_chg(resv, idx, idx + 1);
1200 if (vma->vm_flags & VM_MAYSHARE)
1203 return chg < 0 ? chg : 0;
1205 static void vma_commit_reservation(struct hstate *h,
1206 struct vm_area_struct *vma, unsigned long addr)
1208 struct resv_map *resv;
1211 resv = vma_resv_map(vma);
1215 idx = vma_hugecache_offset(h, vma, addr);
1216 region_add(resv, idx, idx + 1);
1219 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1220 unsigned long addr, int avoid_reserve)
1222 struct hugepage_subpool *spool = subpool_vma(vma);
1223 struct hstate *h = hstate_vma(vma);
1227 struct hugetlb_cgroup *h_cg;
1229 idx = hstate_index(h);
1231 * Processes that did not create the mapping will have no
1232 * reserves and will not have accounted against subpool
1233 * limit. Check that the subpool limit can be made before
1234 * satisfying the allocation MAP_NORESERVE mappings may also
1235 * need pages and subpool limit allocated allocated if no reserve
1238 chg = vma_needs_reservation(h, vma, addr);
1240 return ERR_PTR(-ENOMEM);
1241 if (chg || avoid_reserve)
1242 if (hugepage_subpool_get_pages(spool, 1))
1243 return ERR_PTR(-ENOSPC);
1245 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1247 if (chg || avoid_reserve)
1248 hugepage_subpool_put_pages(spool, 1);
1249 return ERR_PTR(-ENOSPC);
1251 spin_lock(&hugetlb_lock);
1252 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1254 spin_unlock(&hugetlb_lock);
1255 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1257 hugetlb_cgroup_uncharge_cgroup(idx,
1258 pages_per_huge_page(h),
1260 if (chg || avoid_reserve)
1261 hugepage_subpool_put_pages(spool, 1);
1262 return ERR_PTR(-ENOSPC);
1264 spin_lock(&hugetlb_lock);
1265 list_move(&page->lru, &h->hugepage_activelist);
1268 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1269 spin_unlock(&hugetlb_lock);
1271 set_page_private(page, (unsigned long)spool);
1273 vma_commit_reservation(h, vma, addr);
1278 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1279 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1280 * where no ERR_VALUE is expected to be returned.
1282 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1283 unsigned long addr, int avoid_reserve)
1285 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1291 int __weak alloc_bootmem_huge_page(struct hstate *h)
1293 struct huge_bootmem_page *m;
1296 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1299 addr = memblock_virt_alloc_try_nid_nopanic(
1300 huge_page_size(h), huge_page_size(h),
1301 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1304 * Use the beginning of the huge page to store the
1305 * huge_bootmem_page struct (until gather_bootmem
1306 * puts them into the mem_map).
1315 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1316 /* Put them into a private list first because mem_map is not up yet */
1317 list_add(&m->list, &huge_boot_pages);
1322 static void prep_compound_huge_page(struct page *page, int order)
1324 if (unlikely(order > (MAX_ORDER - 1)))
1325 prep_compound_gigantic_page(page, order);
1327 prep_compound_page(page, order);
1330 /* Put bootmem huge pages into the standard lists after mem_map is up */
1331 static void __init gather_bootmem_prealloc(void)
1333 struct huge_bootmem_page *m;
1335 list_for_each_entry(m, &huge_boot_pages, list) {
1336 struct hstate *h = m->hstate;
1339 #ifdef CONFIG_HIGHMEM
1340 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1341 memblock_free_late(__pa(m),
1342 sizeof(struct huge_bootmem_page));
1344 page = virt_to_page(m);
1346 WARN_ON(page_count(page) != 1);
1347 prep_compound_huge_page(page, h->order);
1348 WARN_ON(PageReserved(page));
1349 prep_new_huge_page(h, page, page_to_nid(page));
1351 * If we had gigantic hugepages allocated at boot time, we need
1352 * to restore the 'stolen' pages to totalram_pages in order to
1353 * fix confusing memory reports from free(1) and another
1354 * side-effects, like CommitLimit going negative.
1356 if (h->order > (MAX_ORDER - 1))
1357 adjust_managed_page_count(page, 1 << h->order);
1361 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1365 for (i = 0; i < h->max_huge_pages; ++i) {
1366 if (h->order >= MAX_ORDER) {
1367 if (!alloc_bootmem_huge_page(h))
1369 } else if (!alloc_fresh_huge_page(h,
1370 &node_states[N_MEMORY]))
1373 h->max_huge_pages = i;
1376 static void __init hugetlb_init_hstates(void)
1380 for_each_hstate(h) {
1381 /* oversize hugepages were init'ed in early boot */
1382 if (h->order < MAX_ORDER)
1383 hugetlb_hstate_alloc_pages(h);
1387 static char * __init memfmt(char *buf, unsigned long n)
1389 if (n >= (1UL << 30))
1390 sprintf(buf, "%lu GB", n >> 30);
1391 else if (n >= (1UL << 20))
1392 sprintf(buf, "%lu MB", n >> 20);
1394 sprintf(buf, "%lu KB", n >> 10);
1398 static void __init report_hugepages(void)
1402 for_each_hstate(h) {
1404 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1405 memfmt(buf, huge_page_size(h)),
1406 h->free_huge_pages);
1410 #ifdef CONFIG_HIGHMEM
1411 static void try_to_free_low(struct hstate *h, unsigned long count,
1412 nodemask_t *nodes_allowed)
1416 if (h->order >= MAX_ORDER)
1419 for_each_node_mask(i, *nodes_allowed) {
1420 struct page *page, *next;
1421 struct list_head *freel = &h->hugepage_freelists[i];
1422 list_for_each_entry_safe(page, next, freel, lru) {
1423 if (count >= h->nr_huge_pages)
1425 if (PageHighMem(page))
1427 list_del(&page->lru);
1428 update_and_free_page(h, page);
1429 h->free_huge_pages--;
1430 h->free_huge_pages_node[page_to_nid(page)]--;
1435 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1436 nodemask_t *nodes_allowed)
1442 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1443 * balanced by operating on them in a round-robin fashion.
1444 * Returns 1 if an adjustment was made.
1446 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1451 VM_BUG_ON(delta != -1 && delta != 1);
1454 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1455 if (h->surplus_huge_pages_node[node])
1459 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1460 if (h->surplus_huge_pages_node[node] <
1461 h->nr_huge_pages_node[node])
1468 h->surplus_huge_pages += delta;
1469 h->surplus_huge_pages_node[node] += delta;
1473 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1474 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1475 nodemask_t *nodes_allowed)
1477 unsigned long min_count, ret;
1479 if (h->order >= MAX_ORDER)
1480 return h->max_huge_pages;
1483 * Increase the pool size
1484 * First take pages out of surplus state. Then make up the
1485 * remaining difference by allocating fresh huge pages.
1487 * We might race with alloc_buddy_huge_page() here and be unable
1488 * to convert a surplus huge page to a normal huge page. That is
1489 * not critical, though, it just means the overall size of the
1490 * pool might be one hugepage larger than it needs to be, but
1491 * within all the constraints specified by the sysctls.
1493 spin_lock(&hugetlb_lock);
1494 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1495 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1499 while (count > persistent_huge_pages(h)) {
1501 * If this allocation races such that we no longer need the
1502 * page, free_huge_page will handle it by freeing the page
1503 * and reducing the surplus.
1505 spin_unlock(&hugetlb_lock);
1506 ret = alloc_fresh_huge_page(h, nodes_allowed);
1507 spin_lock(&hugetlb_lock);
1511 /* Bail for signals. Probably ctrl-c from user */
1512 if (signal_pending(current))
1517 * Decrease the pool size
1518 * First return free pages to the buddy allocator (being careful
1519 * to keep enough around to satisfy reservations). Then place
1520 * pages into surplus state as needed so the pool will shrink
1521 * to the desired size as pages become free.
1523 * By placing pages into the surplus state independent of the
1524 * overcommit value, we are allowing the surplus pool size to
1525 * exceed overcommit. There are few sane options here. Since
1526 * alloc_buddy_huge_page() is checking the global counter,
1527 * though, we'll note that we're not allowed to exceed surplus
1528 * and won't grow the pool anywhere else. Not until one of the
1529 * sysctls are changed, or the surplus pages go out of use.
1531 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1532 min_count = max(count, min_count);
1533 try_to_free_low(h, min_count, nodes_allowed);
1534 while (min_count < persistent_huge_pages(h)) {
1535 if (!free_pool_huge_page(h, nodes_allowed, 0))
1538 while (count < persistent_huge_pages(h)) {
1539 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1543 ret = persistent_huge_pages(h);
1544 spin_unlock(&hugetlb_lock);
1548 #define HSTATE_ATTR_RO(_name) \
1549 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1551 #define HSTATE_ATTR(_name) \
1552 static struct kobj_attribute _name##_attr = \
1553 __ATTR(_name, 0644, _name##_show, _name##_store)
1555 static struct kobject *hugepages_kobj;
1556 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1558 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1560 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1564 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1565 if (hstate_kobjs[i] == kobj) {
1567 *nidp = NUMA_NO_NODE;
1571 return kobj_to_node_hstate(kobj, nidp);
1574 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1575 struct kobj_attribute *attr, char *buf)
1578 unsigned long nr_huge_pages;
1581 h = kobj_to_hstate(kobj, &nid);
1582 if (nid == NUMA_NO_NODE)
1583 nr_huge_pages = h->nr_huge_pages;
1585 nr_huge_pages = h->nr_huge_pages_node[nid];
1587 return sprintf(buf, "%lu\n", nr_huge_pages);
1590 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1591 struct kobject *kobj, struct kobj_attribute *attr,
1592 const char *buf, size_t len)
1596 unsigned long count;
1598 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1600 err = kstrtoul(buf, 10, &count);
1604 h = kobj_to_hstate(kobj, &nid);
1605 if (h->order >= MAX_ORDER) {
1610 if (nid == NUMA_NO_NODE) {
1612 * global hstate attribute
1614 if (!(obey_mempolicy &&
1615 init_nodemask_of_mempolicy(nodes_allowed))) {
1616 NODEMASK_FREE(nodes_allowed);
1617 nodes_allowed = &node_states[N_MEMORY];
1619 } else if (nodes_allowed) {
1621 * per node hstate attribute: adjust count to global,
1622 * but restrict alloc/free to the specified node.
1624 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1625 init_nodemask_of_node(nodes_allowed, nid);
1627 nodes_allowed = &node_states[N_MEMORY];
1629 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1631 if (nodes_allowed != &node_states[N_MEMORY])
1632 NODEMASK_FREE(nodes_allowed);
1636 NODEMASK_FREE(nodes_allowed);
1640 static ssize_t nr_hugepages_show(struct kobject *kobj,
1641 struct kobj_attribute *attr, char *buf)
1643 return nr_hugepages_show_common(kobj, attr, buf);
1646 static ssize_t nr_hugepages_store(struct kobject *kobj,
1647 struct kobj_attribute *attr, const char *buf, size_t len)
1649 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1651 HSTATE_ATTR(nr_hugepages);
1656 * hstate attribute for optionally mempolicy-based constraint on persistent
1657 * huge page alloc/free.
1659 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1660 struct kobj_attribute *attr, char *buf)
1662 return nr_hugepages_show_common(kobj, attr, buf);
1665 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1666 struct kobj_attribute *attr, const char *buf, size_t len)
1668 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1670 HSTATE_ATTR(nr_hugepages_mempolicy);
1674 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1675 struct kobj_attribute *attr, char *buf)
1677 struct hstate *h = kobj_to_hstate(kobj, NULL);
1678 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1681 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1682 struct kobj_attribute *attr, const char *buf, size_t count)
1685 unsigned long input;
1686 struct hstate *h = kobj_to_hstate(kobj, NULL);
1688 if (h->order >= MAX_ORDER)
1691 err = kstrtoul(buf, 10, &input);
1695 spin_lock(&hugetlb_lock);
1696 h->nr_overcommit_huge_pages = input;
1697 spin_unlock(&hugetlb_lock);
1701 HSTATE_ATTR(nr_overcommit_hugepages);
1703 static ssize_t free_hugepages_show(struct kobject *kobj,
1704 struct kobj_attribute *attr, char *buf)
1707 unsigned long free_huge_pages;
1710 h = kobj_to_hstate(kobj, &nid);
1711 if (nid == NUMA_NO_NODE)
1712 free_huge_pages = h->free_huge_pages;
1714 free_huge_pages = h->free_huge_pages_node[nid];
1716 return sprintf(buf, "%lu\n", free_huge_pages);
1718 HSTATE_ATTR_RO(free_hugepages);
1720 static ssize_t resv_hugepages_show(struct kobject *kobj,
1721 struct kobj_attribute *attr, char *buf)
1723 struct hstate *h = kobj_to_hstate(kobj, NULL);
1724 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1726 HSTATE_ATTR_RO(resv_hugepages);
1728 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1729 struct kobj_attribute *attr, char *buf)
1732 unsigned long surplus_huge_pages;
1735 h = kobj_to_hstate(kobj, &nid);
1736 if (nid == NUMA_NO_NODE)
1737 surplus_huge_pages = h->surplus_huge_pages;
1739 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1741 return sprintf(buf, "%lu\n", surplus_huge_pages);
1743 HSTATE_ATTR_RO(surplus_hugepages);
1745 static struct attribute *hstate_attrs[] = {
1746 &nr_hugepages_attr.attr,
1747 &nr_overcommit_hugepages_attr.attr,
1748 &free_hugepages_attr.attr,
1749 &resv_hugepages_attr.attr,
1750 &surplus_hugepages_attr.attr,
1752 &nr_hugepages_mempolicy_attr.attr,
1757 static struct attribute_group hstate_attr_group = {
1758 .attrs = hstate_attrs,
1761 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1762 struct kobject **hstate_kobjs,
1763 struct attribute_group *hstate_attr_group)
1766 int hi = hstate_index(h);
1768 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1769 if (!hstate_kobjs[hi])
1772 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1774 kobject_put(hstate_kobjs[hi]);
1779 static void __init hugetlb_sysfs_init(void)
1784 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1785 if (!hugepages_kobj)
1788 for_each_hstate(h) {
1789 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1790 hstate_kobjs, &hstate_attr_group);
1792 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1799 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1800 * with node devices in node_devices[] using a parallel array. The array
1801 * index of a node device or _hstate == node id.
1802 * This is here to avoid any static dependency of the node device driver, in
1803 * the base kernel, on the hugetlb module.
1805 struct node_hstate {
1806 struct kobject *hugepages_kobj;
1807 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1809 struct node_hstate node_hstates[MAX_NUMNODES];
1812 * A subset of global hstate attributes for node devices
1814 static struct attribute *per_node_hstate_attrs[] = {
1815 &nr_hugepages_attr.attr,
1816 &free_hugepages_attr.attr,
1817 &surplus_hugepages_attr.attr,
1821 static struct attribute_group per_node_hstate_attr_group = {
1822 .attrs = per_node_hstate_attrs,
1826 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1827 * Returns node id via non-NULL nidp.
1829 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1833 for (nid = 0; nid < nr_node_ids; nid++) {
1834 struct node_hstate *nhs = &node_hstates[nid];
1836 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1837 if (nhs->hstate_kobjs[i] == kobj) {
1849 * Unregister hstate attributes from a single node device.
1850 * No-op if no hstate attributes attached.
1852 static void hugetlb_unregister_node(struct node *node)
1855 struct node_hstate *nhs = &node_hstates[node->dev.id];
1857 if (!nhs->hugepages_kobj)
1858 return; /* no hstate attributes */
1860 for_each_hstate(h) {
1861 int idx = hstate_index(h);
1862 if (nhs->hstate_kobjs[idx]) {
1863 kobject_put(nhs->hstate_kobjs[idx]);
1864 nhs->hstate_kobjs[idx] = NULL;
1868 kobject_put(nhs->hugepages_kobj);
1869 nhs->hugepages_kobj = NULL;
1873 * hugetlb module exit: unregister hstate attributes from node devices
1876 static void hugetlb_unregister_all_nodes(void)
1881 * disable node device registrations.
1883 register_hugetlbfs_with_node(NULL, NULL);
1886 * remove hstate attributes from any nodes that have them.
1888 for (nid = 0; nid < nr_node_ids; nid++)
1889 hugetlb_unregister_node(node_devices[nid]);
1893 * Register hstate attributes for a single node device.
1894 * No-op if attributes already registered.
1896 static void hugetlb_register_node(struct node *node)
1899 struct node_hstate *nhs = &node_hstates[node->dev.id];
1902 if (nhs->hugepages_kobj)
1903 return; /* already allocated */
1905 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1907 if (!nhs->hugepages_kobj)
1910 for_each_hstate(h) {
1911 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1913 &per_node_hstate_attr_group);
1915 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1916 h->name, node->dev.id);
1917 hugetlb_unregister_node(node);
1924 * hugetlb init time: register hstate attributes for all registered node
1925 * devices of nodes that have memory. All on-line nodes should have
1926 * registered their associated device by this time.
1928 static void hugetlb_register_all_nodes(void)
1932 for_each_node_state(nid, N_MEMORY) {
1933 struct node *node = node_devices[nid];
1934 if (node->dev.id == nid)
1935 hugetlb_register_node(node);
1939 * Let the node device driver know we're here so it can
1940 * [un]register hstate attributes on node hotplug.
1942 register_hugetlbfs_with_node(hugetlb_register_node,
1943 hugetlb_unregister_node);
1945 #else /* !CONFIG_NUMA */
1947 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1955 static void hugetlb_unregister_all_nodes(void) { }
1957 static void hugetlb_register_all_nodes(void) { }
1961 static void __exit hugetlb_exit(void)
1965 hugetlb_unregister_all_nodes();
1967 for_each_hstate(h) {
1968 kobject_put(hstate_kobjs[hstate_index(h)]);
1971 kobject_put(hugepages_kobj);
1972 kfree(htlb_fault_mutex_table);
1974 module_exit(hugetlb_exit);
1976 static int __init hugetlb_init(void)
1980 /* Some platform decide whether they support huge pages at boot
1981 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1982 * there is no such support
1984 if (HPAGE_SHIFT == 0)
1987 if (!size_to_hstate(default_hstate_size)) {
1988 default_hstate_size = HPAGE_SIZE;
1989 if (!size_to_hstate(default_hstate_size))
1990 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1992 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1993 if (default_hstate_max_huge_pages)
1994 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1996 hugetlb_init_hstates();
1997 gather_bootmem_prealloc();
2000 hugetlb_sysfs_init();
2001 hugetlb_register_all_nodes();
2002 hugetlb_cgroup_file_init();
2005 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2007 num_fault_mutexes = 1;
2009 htlb_fault_mutex_table =
2010 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2011 BUG_ON(!htlb_fault_mutex_table);
2013 for (i = 0; i < num_fault_mutexes; i++)
2014 mutex_init(&htlb_fault_mutex_table[i]);
2017 module_init(hugetlb_init);
2019 /* Should be called on processing a hugepagesz=... option */
2020 void __init hugetlb_add_hstate(unsigned order)
2025 if (size_to_hstate(PAGE_SIZE << order)) {
2026 pr_warning("hugepagesz= specified twice, ignoring\n");
2029 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2031 h = &hstates[hugetlb_max_hstate++];
2033 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2034 h->nr_huge_pages = 0;
2035 h->free_huge_pages = 0;
2036 for (i = 0; i < MAX_NUMNODES; ++i)
2037 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2038 INIT_LIST_HEAD(&h->hugepage_activelist);
2039 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2040 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2041 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2042 huge_page_size(h)/1024);
2047 static int __init hugetlb_nrpages_setup(char *s)
2050 static unsigned long *last_mhp;
2053 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2054 * so this hugepages= parameter goes to the "default hstate".
2056 if (!hugetlb_max_hstate)
2057 mhp = &default_hstate_max_huge_pages;
2059 mhp = &parsed_hstate->max_huge_pages;
2061 if (mhp == last_mhp) {
2062 pr_warning("hugepages= specified twice without "
2063 "interleaving hugepagesz=, ignoring\n");
2067 if (sscanf(s, "%lu", mhp) <= 0)
2071 * Global state is always initialized later in hugetlb_init.
2072 * But we need to allocate >= MAX_ORDER hstates here early to still
2073 * use the bootmem allocator.
2075 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2076 hugetlb_hstate_alloc_pages(parsed_hstate);
2082 __setup("hugepages=", hugetlb_nrpages_setup);
2084 static int __init hugetlb_default_setup(char *s)
2086 default_hstate_size = memparse(s, &s);
2089 __setup("default_hugepagesz=", hugetlb_default_setup);
2091 static unsigned int cpuset_mems_nr(unsigned int *array)
2094 unsigned int nr = 0;
2096 for_each_node_mask(node, cpuset_current_mems_allowed)
2102 #ifdef CONFIG_SYSCTL
2103 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2104 struct ctl_table *table, int write,
2105 void __user *buffer, size_t *length, loff_t *ppos)
2107 struct hstate *h = &default_hstate;
2111 tmp = h->max_huge_pages;
2113 if (write && h->order >= MAX_ORDER)
2117 table->maxlen = sizeof(unsigned long);
2118 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2123 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2124 GFP_KERNEL | __GFP_NORETRY);
2125 if (!(obey_mempolicy &&
2126 init_nodemask_of_mempolicy(nodes_allowed))) {
2127 NODEMASK_FREE(nodes_allowed);
2128 nodes_allowed = &node_states[N_MEMORY];
2130 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2132 if (nodes_allowed != &node_states[N_MEMORY])
2133 NODEMASK_FREE(nodes_allowed);
2139 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2140 void __user *buffer, size_t *length, loff_t *ppos)
2143 return hugetlb_sysctl_handler_common(false, table, write,
2144 buffer, length, ppos);
2148 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2149 void __user *buffer, size_t *length, loff_t *ppos)
2151 return hugetlb_sysctl_handler_common(true, table, write,
2152 buffer, length, ppos);
2154 #endif /* CONFIG_NUMA */
2156 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2157 void __user *buffer,
2158 size_t *length, loff_t *ppos)
2160 struct hstate *h = &default_hstate;
2164 tmp = h->nr_overcommit_huge_pages;
2166 if (write && h->order >= MAX_ORDER)
2170 table->maxlen = sizeof(unsigned long);
2171 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2176 spin_lock(&hugetlb_lock);
2177 h->nr_overcommit_huge_pages = tmp;
2178 spin_unlock(&hugetlb_lock);
2184 #endif /* CONFIG_SYSCTL */
2186 void hugetlb_report_meminfo(struct seq_file *m)
2188 struct hstate *h = &default_hstate;
2190 "HugePages_Total: %5lu\n"
2191 "HugePages_Free: %5lu\n"
2192 "HugePages_Rsvd: %5lu\n"
2193 "HugePages_Surp: %5lu\n"
2194 "Hugepagesize: %8lu kB\n",
2198 h->surplus_huge_pages,
2199 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2202 int hugetlb_report_node_meminfo(int nid, char *buf)
2204 struct hstate *h = &default_hstate;
2206 "Node %d HugePages_Total: %5u\n"
2207 "Node %d HugePages_Free: %5u\n"
2208 "Node %d HugePages_Surp: %5u\n",
2209 nid, h->nr_huge_pages_node[nid],
2210 nid, h->free_huge_pages_node[nid],
2211 nid, h->surplus_huge_pages_node[nid]);
2214 void hugetlb_show_meminfo(void)
2219 for_each_node_state(nid, N_MEMORY)
2221 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2223 h->nr_huge_pages_node[nid],
2224 h->free_huge_pages_node[nid],
2225 h->surplus_huge_pages_node[nid],
2226 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2229 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2230 unsigned long hugetlb_total_pages(void)
2233 unsigned long nr_total_pages = 0;
2236 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2237 return nr_total_pages;
2240 static int hugetlb_acct_memory(struct hstate *h, long delta)
2244 spin_lock(&hugetlb_lock);
2246 * When cpuset is configured, it breaks the strict hugetlb page
2247 * reservation as the accounting is done on a global variable. Such
2248 * reservation is completely rubbish in the presence of cpuset because
2249 * the reservation is not checked against page availability for the
2250 * current cpuset. Application can still potentially OOM'ed by kernel
2251 * with lack of free htlb page in cpuset that the task is in.
2252 * Attempt to enforce strict accounting with cpuset is almost
2253 * impossible (or too ugly) because cpuset is too fluid that
2254 * task or memory node can be dynamically moved between cpusets.
2256 * The change of semantics for shared hugetlb mapping with cpuset is
2257 * undesirable. However, in order to preserve some of the semantics,
2258 * we fall back to check against current free page availability as
2259 * a best attempt and hopefully to minimize the impact of changing
2260 * semantics that cpuset has.
2263 if (gather_surplus_pages(h, delta) < 0)
2266 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2267 return_unused_surplus_pages(h, delta);
2274 return_unused_surplus_pages(h, (unsigned long) -delta);
2277 spin_unlock(&hugetlb_lock);
2281 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2283 struct resv_map *resv = vma_resv_map(vma);
2286 * This new VMA should share its siblings reservation map if present.
2287 * The VMA will only ever have a valid reservation map pointer where
2288 * it is being copied for another still existing VMA. As that VMA
2289 * has a reference to the reservation map it cannot disappear until
2290 * after this open call completes. It is therefore safe to take a
2291 * new reference here without additional locking.
2293 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2294 kref_get(&resv->refs);
2297 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2299 struct hstate *h = hstate_vma(vma);
2300 struct resv_map *resv = vma_resv_map(vma);
2301 struct hugepage_subpool *spool = subpool_vma(vma);
2302 unsigned long reserve, start, end;
2304 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2307 start = vma_hugecache_offset(h, vma, vma->vm_start);
2308 end = vma_hugecache_offset(h, vma, vma->vm_end);
2310 reserve = (end - start) - region_count(resv, start, end);
2312 kref_put(&resv->refs, resv_map_release);
2315 hugetlb_acct_memory(h, -reserve);
2316 hugepage_subpool_put_pages(spool, reserve);
2321 * We cannot handle pagefaults against hugetlb pages at all. They cause
2322 * handle_mm_fault() to try to instantiate regular-sized pages in the
2323 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2326 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2332 const struct vm_operations_struct hugetlb_vm_ops = {
2333 .fault = hugetlb_vm_op_fault,
2334 .open = hugetlb_vm_op_open,
2335 .close = hugetlb_vm_op_close,
2338 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2344 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2345 vma->vm_page_prot)));
2347 entry = huge_pte_wrprotect(mk_huge_pte(page,
2348 vma->vm_page_prot));
2350 entry = pte_mkyoung(entry);
2351 entry = pte_mkhuge(entry);
2352 entry = arch_make_huge_pte(entry, vma, page, writable);
2357 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2358 unsigned long address, pte_t *ptep)
2362 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2363 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2364 update_mmu_cache(vma, address, ptep);
2368 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2369 struct vm_area_struct *vma)
2371 pte_t *src_pte, *dst_pte, entry;
2372 struct page *ptepage;
2375 struct hstate *h = hstate_vma(vma);
2376 unsigned long sz = huge_page_size(h);
2377 unsigned long mmun_start; /* For mmu_notifiers */
2378 unsigned long mmun_end; /* For mmu_notifiers */
2381 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2383 mmun_start = vma->vm_start;
2384 mmun_end = vma->vm_end;
2386 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2388 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2389 spinlock_t *src_ptl, *dst_ptl;
2390 src_pte = huge_pte_offset(src, addr);
2393 dst_pte = huge_pte_alloc(dst, addr, sz);
2399 /* If the pagetables are shared don't copy or take references */
2400 if (dst_pte == src_pte)
2403 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2404 src_ptl = huge_pte_lockptr(h, src, src_pte);
2405 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2406 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2408 huge_ptep_set_wrprotect(src, addr, src_pte);
2409 entry = huge_ptep_get(src_pte);
2410 ptepage = pte_page(entry);
2412 page_dup_rmap(ptepage);
2413 set_huge_pte_at(dst, addr, dst_pte, entry);
2415 spin_unlock(src_ptl);
2416 spin_unlock(dst_ptl);
2420 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2425 static int is_hugetlb_entry_migration(pte_t pte)
2429 if (huge_pte_none(pte) || pte_present(pte))
2431 swp = pte_to_swp_entry(pte);
2432 if (non_swap_entry(swp) && is_migration_entry(swp))
2438 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2442 if (huge_pte_none(pte) || pte_present(pte))
2444 swp = pte_to_swp_entry(pte);
2445 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2451 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2452 unsigned long start, unsigned long end,
2453 struct page *ref_page)
2455 int force_flush = 0;
2456 struct mm_struct *mm = vma->vm_mm;
2457 unsigned long address;
2462 struct hstate *h = hstate_vma(vma);
2463 unsigned long sz = huge_page_size(h);
2464 const unsigned long mmun_start = start; /* For mmu_notifiers */
2465 const unsigned long mmun_end = end; /* For mmu_notifiers */
2467 WARN_ON(!is_vm_hugetlb_page(vma));
2468 BUG_ON(start & ~huge_page_mask(h));
2469 BUG_ON(end & ~huge_page_mask(h));
2471 tlb_start_vma(tlb, vma);
2472 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2474 for (address = start; address < end; address += sz) {
2475 ptep = huge_pte_offset(mm, address);
2479 ptl = huge_pte_lock(h, mm, ptep);
2480 if (huge_pmd_unshare(mm, &address, ptep))
2483 pte = huge_ptep_get(ptep);
2484 if (huge_pte_none(pte))
2488 * HWPoisoned hugepage is already unmapped and dropped reference
2490 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2491 huge_pte_clear(mm, address, ptep);
2495 page = pte_page(pte);
2497 * If a reference page is supplied, it is because a specific
2498 * page is being unmapped, not a range. Ensure the page we
2499 * are about to unmap is the actual page of interest.
2502 if (page != ref_page)
2506 * Mark the VMA as having unmapped its page so that
2507 * future faults in this VMA will fail rather than
2508 * looking like data was lost
2510 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2513 pte = huge_ptep_get_and_clear(mm, address, ptep);
2514 tlb_remove_tlb_entry(tlb, ptep, address);
2515 if (huge_pte_dirty(pte))
2516 set_page_dirty(page);
2518 page_remove_rmap(page);
2519 force_flush = !__tlb_remove_page(tlb, page);
2524 /* Bail out after unmapping reference page if supplied */
2533 * mmu_gather ran out of room to batch pages, we break out of
2534 * the PTE lock to avoid doing the potential expensive TLB invalidate
2535 * and page-free while holding it.
2540 if (address < end && !ref_page)
2543 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2544 tlb_end_vma(tlb, vma);
2547 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2548 struct vm_area_struct *vma, unsigned long start,
2549 unsigned long end, struct page *ref_page)
2551 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2554 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2555 * test will fail on a vma being torn down, and not grab a page table
2556 * on its way out. We're lucky that the flag has such an appropriate
2557 * name, and can in fact be safely cleared here. We could clear it
2558 * before the __unmap_hugepage_range above, but all that's necessary
2559 * is to clear it before releasing the i_mmap_mutex. This works
2560 * because in the context this is called, the VMA is about to be
2561 * destroyed and the i_mmap_mutex is held.
2563 vma->vm_flags &= ~VM_MAYSHARE;
2566 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2567 unsigned long end, struct page *ref_page)
2569 struct mm_struct *mm;
2570 struct mmu_gather tlb;
2574 tlb_gather_mmu(&tlb, mm, start, end);
2575 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2576 tlb_finish_mmu(&tlb, start, end);
2580 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2581 * mappping it owns the reserve page for. The intention is to unmap the page
2582 * from other VMAs and let the children be SIGKILLed if they are faulting the
2585 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2586 struct page *page, unsigned long address)
2588 struct hstate *h = hstate_vma(vma);
2589 struct vm_area_struct *iter_vma;
2590 struct address_space *mapping;
2594 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2595 * from page cache lookup which is in HPAGE_SIZE units.
2597 address = address & huge_page_mask(h);
2598 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2600 mapping = file_inode(vma->vm_file)->i_mapping;
2603 * Take the mapping lock for the duration of the table walk. As
2604 * this mapping should be shared between all the VMAs,
2605 * __unmap_hugepage_range() is called as the lock is already held
2607 mutex_lock(&mapping->i_mmap_mutex);
2608 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2609 /* Do not unmap the current VMA */
2610 if (iter_vma == vma)
2614 * Unmap the page from other VMAs without their own reserves.
2615 * They get marked to be SIGKILLed if they fault in these
2616 * areas. This is because a future no-page fault on this VMA
2617 * could insert a zeroed page instead of the data existing
2618 * from the time of fork. This would look like data corruption
2620 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2621 unmap_hugepage_range(iter_vma, address,
2622 address + huge_page_size(h), page);
2624 mutex_unlock(&mapping->i_mmap_mutex);
2630 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2631 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2632 * cannot race with other handlers or page migration.
2633 * Keep the pte_same checks anyway to make transition from the mutex easier.
2635 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2636 unsigned long address, pte_t *ptep, pte_t pte,
2637 struct page *pagecache_page, spinlock_t *ptl)
2639 struct hstate *h = hstate_vma(vma);
2640 struct page *old_page, *new_page;
2641 int outside_reserve = 0;
2642 unsigned long mmun_start; /* For mmu_notifiers */
2643 unsigned long mmun_end; /* For mmu_notifiers */
2645 old_page = pte_page(pte);
2648 /* If no-one else is actually using this page, avoid the copy
2649 * and just make the page writable */
2650 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2651 page_move_anon_rmap(old_page, vma, address);
2652 set_huge_ptep_writable(vma, address, ptep);
2657 * If the process that created a MAP_PRIVATE mapping is about to
2658 * perform a COW due to a shared page count, attempt to satisfy
2659 * the allocation without using the existing reserves. The pagecache
2660 * page is used to determine if the reserve at this address was
2661 * consumed or not. If reserves were used, a partial faulted mapping
2662 * at the time of fork() could consume its reserves on COW instead
2663 * of the full address range.
2665 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2666 old_page != pagecache_page)
2667 outside_reserve = 1;
2669 page_cache_get(old_page);
2671 /* Drop page table lock as buddy allocator may be called */
2673 new_page = alloc_huge_page(vma, address, outside_reserve);
2675 if (IS_ERR(new_page)) {
2676 long err = PTR_ERR(new_page);
2677 page_cache_release(old_page);
2680 * If a process owning a MAP_PRIVATE mapping fails to COW,
2681 * it is due to references held by a child and an insufficient
2682 * huge page pool. To guarantee the original mappers
2683 * reliability, unmap the page from child processes. The child
2684 * may get SIGKILLed if it later faults.
2686 if (outside_reserve) {
2687 BUG_ON(huge_pte_none(pte));
2688 if (unmap_ref_private(mm, vma, old_page, address)) {
2689 BUG_ON(huge_pte_none(pte));
2691 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2692 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2693 goto retry_avoidcopy;
2695 * race occurs while re-acquiring page table
2696 * lock, and our job is done.
2703 /* Caller expects lock to be held */
2706 return VM_FAULT_OOM;
2708 return VM_FAULT_SIGBUS;
2712 * When the original hugepage is shared one, it does not have
2713 * anon_vma prepared.
2715 if (unlikely(anon_vma_prepare(vma))) {
2716 page_cache_release(new_page);
2717 page_cache_release(old_page);
2718 /* Caller expects lock to be held */
2720 return VM_FAULT_OOM;
2723 copy_user_huge_page(new_page, old_page, address, vma,
2724 pages_per_huge_page(h));
2725 __SetPageUptodate(new_page);
2727 mmun_start = address & huge_page_mask(h);
2728 mmun_end = mmun_start + huge_page_size(h);
2729 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2731 * Retake the page table lock to check for racing updates
2732 * before the page tables are altered
2735 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2736 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2737 ClearPagePrivate(new_page);
2740 huge_ptep_clear_flush(vma, address, ptep);
2741 set_huge_pte_at(mm, address, ptep,
2742 make_huge_pte(vma, new_page, 1));
2743 page_remove_rmap(old_page);
2744 hugepage_add_new_anon_rmap(new_page, vma, address);
2745 /* Make the old page be freed below */
2746 new_page = old_page;
2749 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2750 page_cache_release(new_page);
2751 page_cache_release(old_page);
2753 /* Caller expects lock to be held */
2758 /* Return the pagecache page at a given address within a VMA */
2759 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2760 struct vm_area_struct *vma, unsigned long address)
2762 struct address_space *mapping;
2765 mapping = vma->vm_file->f_mapping;
2766 idx = vma_hugecache_offset(h, vma, address);
2768 return find_lock_page(mapping, idx);
2772 * Return whether there is a pagecache page to back given address within VMA.
2773 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2775 static bool hugetlbfs_pagecache_present(struct hstate *h,
2776 struct vm_area_struct *vma, unsigned long address)
2778 struct address_space *mapping;
2782 mapping = vma->vm_file->f_mapping;
2783 idx = vma_hugecache_offset(h, vma, address);
2785 page = find_get_page(mapping, idx);
2788 return page != NULL;
2791 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2792 struct address_space *mapping, pgoff_t idx,
2793 unsigned long address, pte_t *ptep, unsigned int flags)
2795 struct hstate *h = hstate_vma(vma);
2796 int ret = VM_FAULT_SIGBUS;
2804 * Currently, we are forced to kill the process in the event the
2805 * original mapper has unmapped pages from the child due to a failed
2806 * COW. Warn that such a situation has occurred as it may not be obvious
2808 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2809 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2815 * Use page lock to guard against racing truncation
2816 * before we get page_table_lock.
2819 page = find_lock_page(mapping, idx);
2821 size = i_size_read(mapping->host) >> huge_page_shift(h);
2824 page = alloc_huge_page(vma, address, 0);
2826 ret = PTR_ERR(page);
2830 ret = VM_FAULT_SIGBUS;
2833 clear_huge_page(page, address, pages_per_huge_page(h));
2834 __SetPageUptodate(page);
2836 if (vma->vm_flags & VM_MAYSHARE) {
2838 struct inode *inode = mapping->host;
2840 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2847 ClearPagePrivate(page);
2849 spin_lock(&inode->i_lock);
2850 inode->i_blocks += blocks_per_huge_page(h);
2851 spin_unlock(&inode->i_lock);
2854 if (unlikely(anon_vma_prepare(vma))) {
2856 goto backout_unlocked;
2862 * If memory error occurs between mmap() and fault, some process
2863 * don't have hwpoisoned swap entry for errored virtual address.
2864 * So we need to block hugepage fault by PG_hwpoison bit check.
2866 if (unlikely(PageHWPoison(page))) {
2867 ret = VM_FAULT_HWPOISON |
2868 VM_FAULT_SET_HINDEX(hstate_index(h));
2869 goto backout_unlocked;
2874 * If we are going to COW a private mapping later, we examine the
2875 * pending reservations for this page now. This will ensure that
2876 * any allocations necessary to record that reservation occur outside
2879 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2880 if (vma_needs_reservation(h, vma, address) < 0) {
2882 goto backout_unlocked;
2885 ptl = huge_pte_lockptr(h, mm, ptep);
2887 size = i_size_read(mapping->host) >> huge_page_shift(h);
2892 if (!huge_pte_none(huge_ptep_get(ptep)))
2896 ClearPagePrivate(page);
2897 hugepage_add_new_anon_rmap(page, vma, address);
2900 page_dup_rmap(page);
2901 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2902 && (vma->vm_flags & VM_SHARED)));
2903 set_huge_pte_at(mm, address, ptep, new_pte);
2905 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2906 /* Optimization, do the COW without a second fault */
2907 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
2924 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2925 struct vm_area_struct *vma,
2926 struct address_space *mapping,
2927 pgoff_t idx, unsigned long address)
2929 unsigned long key[2];
2932 if (vma->vm_flags & VM_SHARED) {
2933 key[0] = (unsigned long) mapping;
2936 key[0] = (unsigned long) mm;
2937 key[1] = address >> huge_page_shift(h);
2940 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
2942 return hash & (num_fault_mutexes - 1);
2946 * For uniprocesor systems we always use a single mutex, so just
2947 * return 0 and avoid the hashing overhead.
2949 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
2950 struct vm_area_struct *vma,
2951 struct address_space *mapping,
2952 pgoff_t idx, unsigned long address)
2958 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2959 unsigned long address, unsigned int flags)
2966 struct page *page = NULL;
2967 struct page *pagecache_page = NULL;
2968 struct hstate *h = hstate_vma(vma);
2969 struct address_space *mapping;
2971 address &= huge_page_mask(h);
2973 ptep = huge_pte_offset(mm, address);
2975 entry = huge_ptep_get(ptep);
2976 if (unlikely(is_hugetlb_entry_migration(entry))) {
2977 migration_entry_wait_huge(vma, mm, ptep);
2979 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2980 return VM_FAULT_HWPOISON_LARGE |
2981 VM_FAULT_SET_HINDEX(hstate_index(h));
2984 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2986 return VM_FAULT_OOM;
2988 mapping = vma->vm_file->f_mapping;
2989 idx = vma_hugecache_offset(h, vma, address);
2992 * Serialize hugepage allocation and instantiation, so that we don't
2993 * get spurious allocation failures if two CPUs race to instantiate
2994 * the same page in the page cache.
2996 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
2997 mutex_lock(&htlb_fault_mutex_table[hash]);
2999 entry = huge_ptep_get(ptep);
3000 if (huge_pte_none(entry)) {
3001 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3008 * If we are going to COW the mapping later, we examine the pending
3009 * reservations for this page now. This will ensure that any
3010 * allocations necessary to record that reservation occur outside the
3011 * spinlock. For private mappings, we also lookup the pagecache
3012 * page now as it is used to determine if a reservation has been
3015 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3016 if (vma_needs_reservation(h, vma, address) < 0) {
3021 if (!(vma->vm_flags & VM_MAYSHARE))
3022 pagecache_page = hugetlbfs_pagecache_page(h,
3027 * hugetlb_cow() requires page locks of pte_page(entry) and
3028 * pagecache_page, so here we need take the former one
3029 * when page != pagecache_page or !pagecache_page.
3030 * Note that locking order is always pagecache_page -> page,
3031 * so no worry about deadlock.
3033 page = pte_page(entry);
3035 if (page != pagecache_page)
3038 ptl = huge_pte_lockptr(h, mm, ptep);
3040 /* Check for a racing update before calling hugetlb_cow */
3041 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3045 if (flags & FAULT_FLAG_WRITE) {
3046 if (!huge_pte_write(entry)) {
3047 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3048 pagecache_page, ptl);
3051 entry = huge_pte_mkdirty(entry);
3053 entry = pte_mkyoung(entry);
3054 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3055 flags & FAULT_FLAG_WRITE))
3056 update_mmu_cache(vma, address, ptep);
3061 if (pagecache_page) {
3062 unlock_page(pagecache_page);
3063 put_page(pagecache_page);
3065 if (page != pagecache_page)
3070 mutex_unlock(&htlb_fault_mutex_table[hash]);
3074 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3075 struct page **pages, struct vm_area_struct **vmas,
3076 unsigned long *position, unsigned long *nr_pages,
3077 long i, unsigned int flags)
3079 unsigned long pfn_offset;
3080 unsigned long vaddr = *position;
3081 unsigned long remainder = *nr_pages;
3082 struct hstate *h = hstate_vma(vma);
3084 while (vaddr < vma->vm_end && remainder) {
3086 spinlock_t *ptl = NULL;
3091 * Some archs (sparc64, sh*) have multiple pte_ts to
3092 * each hugepage. We have to make sure we get the
3093 * first, for the page indexing below to work.
3095 * Note that page table lock is not held when pte is null.
3097 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3099 ptl = huge_pte_lock(h, mm, pte);
3100 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3103 * When coredumping, it suits get_dump_page if we just return
3104 * an error where there's an empty slot with no huge pagecache
3105 * to back it. This way, we avoid allocating a hugepage, and
3106 * the sparse dumpfile avoids allocating disk blocks, but its
3107 * huge holes still show up with zeroes where they need to be.
3109 if (absent && (flags & FOLL_DUMP) &&
3110 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3118 * We need call hugetlb_fault for both hugepages under migration
3119 * (in which case hugetlb_fault waits for the migration,) and
3120 * hwpoisoned hugepages (in which case we need to prevent the
3121 * caller from accessing to them.) In order to do this, we use
3122 * here is_swap_pte instead of is_hugetlb_entry_migration and
3123 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3124 * both cases, and because we can't follow correct pages
3125 * directly from any kind of swap entries.
3127 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3128 ((flags & FOLL_WRITE) &&
3129 !huge_pte_write(huge_ptep_get(pte)))) {
3134 ret = hugetlb_fault(mm, vma, vaddr,
3135 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3136 if (!(ret & VM_FAULT_ERROR))
3143 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3144 page = pte_page(huge_ptep_get(pte));
3147 pages[i] = mem_map_offset(page, pfn_offset);
3148 get_page_foll(pages[i]);
3158 if (vaddr < vma->vm_end && remainder &&
3159 pfn_offset < pages_per_huge_page(h)) {
3161 * We use pfn_offset to avoid touching the pageframes
3162 * of this compound page.
3168 *nr_pages = remainder;
3171 return i ? i : -EFAULT;
3174 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3175 unsigned long address, unsigned long end, pgprot_t newprot)
3177 struct mm_struct *mm = vma->vm_mm;
3178 unsigned long start = address;
3181 struct hstate *h = hstate_vma(vma);
3182 unsigned long pages = 0;
3184 BUG_ON(address >= end);
3185 flush_cache_range(vma, address, end);
3187 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3188 for (; address < end; address += huge_page_size(h)) {
3190 ptep = huge_pte_offset(mm, address);
3193 ptl = huge_pte_lock(h, mm, ptep);
3194 if (huge_pmd_unshare(mm, &address, ptep)) {
3199 if (!huge_pte_none(huge_ptep_get(ptep))) {
3200 pte = huge_ptep_get_and_clear(mm, address, ptep);
3201 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3202 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3203 set_huge_pte_at(mm, address, ptep, pte);
3209 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3210 * may have cleared our pud entry and done put_page on the page table:
3211 * once we release i_mmap_mutex, another task can do the final put_page
3212 * and that page table be reused and filled with junk.
3214 flush_tlb_range(vma, start, end);
3215 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3217 return pages << h->order;
3220 int hugetlb_reserve_pages(struct inode *inode,
3222 struct vm_area_struct *vma,
3223 vm_flags_t vm_flags)
3226 struct hstate *h = hstate_inode(inode);
3227 struct hugepage_subpool *spool = subpool_inode(inode);
3228 struct resv_map *resv_map;
3231 * Only apply hugepage reservation if asked. At fault time, an
3232 * attempt will be made for VM_NORESERVE to allocate a page
3233 * without using reserves
3235 if (vm_flags & VM_NORESERVE)
3239 * Shared mappings base their reservation on the number of pages that
3240 * are already allocated on behalf of the file. Private mappings need
3241 * to reserve the full area even if read-only as mprotect() may be
3242 * called to make the mapping read-write. Assume !vma is a shm mapping
3244 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3245 resv_map = inode_resv_map(inode);
3247 chg = region_chg(resv_map, from, to);
3250 resv_map = resv_map_alloc();
3256 set_vma_resv_map(vma, resv_map);
3257 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3265 /* There must be enough pages in the subpool for the mapping */
3266 if (hugepage_subpool_get_pages(spool, chg)) {
3272 * Check enough hugepages are available for the reservation.
3273 * Hand the pages back to the subpool if there are not
3275 ret = hugetlb_acct_memory(h, chg);
3277 hugepage_subpool_put_pages(spool, chg);
3282 * Account for the reservations made. Shared mappings record regions
3283 * that have reservations as they are shared by multiple VMAs.
3284 * When the last VMA disappears, the region map says how much
3285 * the reservation was and the page cache tells how much of
3286 * the reservation was consumed. Private mappings are per-VMA and
3287 * only the consumed reservations are tracked. When the VMA
3288 * disappears, the original reservation is the VMA size and the
3289 * consumed reservations are stored in the map. Hence, nothing
3290 * else has to be done for private mappings here
3292 if (!vma || vma->vm_flags & VM_MAYSHARE)
3293 region_add(resv_map, from, to);
3296 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3297 kref_put(&resv_map->refs, resv_map_release);
3301 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3303 struct hstate *h = hstate_inode(inode);
3304 struct resv_map *resv_map = inode_resv_map(inode);
3306 struct hugepage_subpool *spool = subpool_inode(inode);
3309 chg = region_truncate(resv_map, offset);
3310 spin_lock(&inode->i_lock);
3311 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3312 spin_unlock(&inode->i_lock);
3314 hugepage_subpool_put_pages(spool, (chg - freed));
3315 hugetlb_acct_memory(h, -(chg - freed));
3318 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3319 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3320 struct vm_area_struct *vma,
3321 unsigned long addr, pgoff_t idx)
3323 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3325 unsigned long sbase = saddr & PUD_MASK;
3326 unsigned long s_end = sbase + PUD_SIZE;
3328 /* Allow segments to share if only one is marked locked */
3329 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3330 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3333 * match the virtual addresses, permission and the alignment of the
3336 if (pmd_index(addr) != pmd_index(saddr) ||
3337 vm_flags != svm_flags ||
3338 sbase < svma->vm_start || svma->vm_end < s_end)
3344 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3346 unsigned long base = addr & PUD_MASK;
3347 unsigned long end = base + PUD_SIZE;
3350 * check on proper vm_flags and page table alignment
3352 if (vma->vm_flags & VM_MAYSHARE &&
3353 vma->vm_start <= base && end <= vma->vm_end)
3359 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3360 * and returns the corresponding pte. While this is not necessary for the
3361 * !shared pmd case because we can allocate the pmd later as well, it makes the
3362 * code much cleaner. pmd allocation is essential for the shared case because
3363 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3364 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3365 * bad pmd for sharing.
3367 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3369 struct vm_area_struct *vma = find_vma(mm, addr);
3370 struct address_space *mapping = vma->vm_file->f_mapping;
3371 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3373 struct vm_area_struct *svma;
3374 unsigned long saddr;
3379 if (!vma_shareable(vma, addr))
3380 return (pte_t *)pmd_alloc(mm, pud, addr);
3382 mutex_lock(&mapping->i_mmap_mutex);
3383 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3387 saddr = page_table_shareable(svma, vma, addr, idx);
3389 spte = huge_pte_offset(svma->vm_mm, saddr);
3391 get_page(virt_to_page(spte));
3400 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3403 pud_populate(mm, pud,
3404 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3406 put_page(virt_to_page(spte));
3409 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3410 mutex_unlock(&mapping->i_mmap_mutex);
3415 * unmap huge page backed by shared pte.
3417 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3418 * indicated by page_count > 1, unmap is achieved by clearing pud and
3419 * decrementing the ref count. If count == 1, the pte page is not shared.
3421 * called with page table lock held.
3423 * returns: 1 successfully unmapped a shared pte page
3424 * 0 the underlying pte page is not shared, or it is the last user
3426 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3428 pgd_t *pgd = pgd_offset(mm, *addr);
3429 pud_t *pud = pud_offset(pgd, *addr);
3431 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3432 if (page_count(virt_to_page(ptep)) == 1)
3436 put_page(virt_to_page(ptep));
3437 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3440 #define want_pmd_share() (1)
3441 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3442 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3446 #define want_pmd_share() (0)
3447 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3449 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3450 pte_t *huge_pte_alloc(struct mm_struct *mm,
3451 unsigned long addr, unsigned long sz)
3457 pgd = pgd_offset(mm, addr);
3458 pud = pud_alloc(mm, pgd, addr);
3460 if (sz == PUD_SIZE) {
3463 BUG_ON(sz != PMD_SIZE);
3464 if (want_pmd_share() && pud_none(*pud))
3465 pte = huge_pmd_share(mm, addr, pud);
3467 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3470 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3475 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3481 pgd = pgd_offset(mm, addr);
3482 if (pgd_present(*pgd)) {
3483 pud = pud_offset(pgd, addr);
3484 if (pud_present(*pud)) {
3486 return (pte_t *)pud;
3487 pmd = pmd_offset(pud, addr);
3490 return (pte_t *) pmd;
3494 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3495 pmd_t *pmd, int write)
3499 page = pte_page(*(pte_t *)pmd);
3501 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3506 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3507 pud_t *pud, int write)
3511 page = pte_page(*(pte_t *)pud);
3513 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3517 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3519 /* Can be overriden by architectures */
3520 __attribute__((weak)) struct page *
3521 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3522 pud_t *pud, int write)
3528 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3530 #ifdef CONFIG_MEMORY_FAILURE
3532 /* Should be called in hugetlb_lock */
3533 static int is_hugepage_on_freelist(struct page *hpage)
3537 struct hstate *h = page_hstate(hpage);
3538 int nid = page_to_nid(hpage);
3540 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3547 * This function is called from memory failure code.
3548 * Assume the caller holds page lock of the head page.
3550 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3552 struct hstate *h = page_hstate(hpage);
3553 int nid = page_to_nid(hpage);
3556 spin_lock(&hugetlb_lock);
3557 if (is_hugepage_on_freelist(hpage)) {
3559 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3560 * but dangling hpage->lru can trigger list-debug warnings
3561 * (this happens when we call unpoison_memory() on it),
3562 * so let it point to itself with list_del_init().
3564 list_del_init(&hpage->lru);
3565 set_page_refcounted(hpage);
3566 h->free_huge_pages--;
3567 h->free_huge_pages_node[nid]--;
3570 spin_unlock(&hugetlb_lock);
3575 bool isolate_huge_page(struct page *page, struct list_head *list)
3577 VM_BUG_ON_PAGE(!PageHead(page), page);
3578 if (!get_page_unless_zero(page))
3580 spin_lock(&hugetlb_lock);
3581 list_move_tail(&page->lru, list);
3582 spin_unlock(&hugetlb_lock);
3586 void putback_active_hugepage(struct page *page)
3588 VM_BUG_ON_PAGE(!PageHead(page), page);
3589 spin_lock(&hugetlb_lock);
3590 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3591 spin_unlock(&hugetlb_lock);
3595 bool is_hugepage_active(struct page *page)
3597 VM_BUG_ON_PAGE(!PageHuge(page), page);
3599 * This function can be called for a tail page because the caller,
3600 * scan_movable_pages, scans through a given pfn-range which typically
3601 * covers one memory block. In systems using gigantic hugepage (1GB
3602 * for x86_64,) a hugepage is larger than a memory block, and we don't
3603 * support migrating such large hugepages for now, so return false
3604 * when called for tail pages.
3609 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3610 * so we should return false for them.
3612 if (unlikely(PageHWPoison(page)))
3614 return page_count(page) > 0;