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/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
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_VMA(!is_vm_hugetlb_page(vma), 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_VMA(!is_vm_hugetlb_page(vma), vma);
453 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
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_VMA(!is_vm_hugetlb_page(vma), vma);
462 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
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_VMA(!is_vm_hugetlb_page(vma), 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_VMA(!is_vm_hugetlb_page(vma), 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_supported(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)))
610 * common helper functions for hstate_next_node_to_{alloc|free}.
611 * We may have allocated or freed a huge page based on a different
612 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
613 * be outside of *nodes_allowed. Ensure that we use an allowed
614 * node for alloc or free.
616 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
618 nid = next_node(nid, *nodes_allowed);
619 if (nid == MAX_NUMNODES)
620 nid = first_node(*nodes_allowed);
621 VM_BUG_ON(nid >= MAX_NUMNODES);
626 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
628 if (!node_isset(nid, *nodes_allowed))
629 nid = next_node_allowed(nid, nodes_allowed);
634 * returns the previously saved node ["this node"] from which to
635 * allocate a persistent huge page for the pool and advance the
636 * next node from which to allocate, handling wrap at end of node
639 static int hstate_next_node_to_alloc(struct hstate *h,
640 nodemask_t *nodes_allowed)
644 VM_BUG_ON(!nodes_allowed);
646 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
647 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
653 * helper for free_pool_huge_page() - return the previously saved
654 * node ["this node"] from which to free a huge page. Advance the
655 * next node id whether or not we find a free huge page to free so
656 * that the next attempt to free addresses the next node.
658 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
662 VM_BUG_ON(!nodes_allowed);
664 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
665 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
670 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
671 for (nr_nodes = nodes_weight(*mask); \
673 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
676 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
677 for (nr_nodes = nodes_weight(*mask); \
679 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
682 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
683 static void destroy_compound_gigantic_page(struct page *page,
687 int nr_pages = 1 << order;
688 struct page *p = page + 1;
690 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
692 set_page_refcounted(p);
693 p->first_page = NULL;
696 set_compound_order(page, 0);
697 __ClearPageHead(page);
700 static void free_gigantic_page(struct page *page, unsigned int order)
702 free_contig_range(page_to_pfn(page), 1 << order);
705 static int __alloc_gigantic_page(unsigned long start_pfn,
706 unsigned long nr_pages)
708 unsigned long end_pfn = start_pfn + nr_pages;
709 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
712 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
713 unsigned long nr_pages)
715 unsigned long i, end_pfn = start_pfn + nr_pages;
718 for (i = start_pfn; i < end_pfn; i++) {
722 page = pfn_to_page(i);
724 if (PageReserved(page))
727 if (page_count(page) > 0)
737 static bool zone_spans_last_pfn(const struct zone *zone,
738 unsigned long start_pfn, unsigned long nr_pages)
740 unsigned long last_pfn = start_pfn + nr_pages - 1;
741 return zone_spans_pfn(zone, last_pfn);
744 static struct page *alloc_gigantic_page(int nid, unsigned int order)
746 unsigned long nr_pages = 1 << order;
747 unsigned long ret, pfn, flags;
750 z = NODE_DATA(nid)->node_zones;
751 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
752 spin_lock_irqsave(&z->lock, flags);
754 pfn = ALIGN(z->zone_start_pfn, nr_pages);
755 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
756 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
758 * We release the zone lock here because
759 * alloc_contig_range() will also lock the zone
760 * at some point. If there's an allocation
761 * spinning on this lock, it may win the race
762 * and cause alloc_contig_range() to fail...
764 spin_unlock_irqrestore(&z->lock, flags);
765 ret = __alloc_gigantic_page(pfn, nr_pages);
767 return pfn_to_page(pfn);
768 spin_lock_irqsave(&z->lock, flags);
773 spin_unlock_irqrestore(&z->lock, flags);
779 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
780 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
782 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
786 page = alloc_gigantic_page(nid, huge_page_order(h));
788 prep_compound_gigantic_page(page, huge_page_order(h));
789 prep_new_huge_page(h, page, nid);
795 static int alloc_fresh_gigantic_page(struct hstate *h,
796 nodemask_t *nodes_allowed)
798 struct page *page = NULL;
801 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
802 page = alloc_fresh_gigantic_page_node(h, node);
810 static inline bool gigantic_page_supported(void) { return true; }
812 static inline bool gigantic_page_supported(void) { return false; }
813 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
814 static inline void destroy_compound_gigantic_page(struct page *page,
815 unsigned int order) { }
816 static inline int alloc_fresh_gigantic_page(struct hstate *h,
817 nodemask_t *nodes_allowed) { return 0; }
820 static void update_and_free_page(struct hstate *h, struct page *page)
824 if (hstate_is_gigantic(h) && !gigantic_page_supported())
828 h->nr_huge_pages_node[page_to_nid(page)]--;
829 for (i = 0; i < pages_per_huge_page(h); i++) {
830 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
831 1 << PG_referenced | 1 << PG_dirty |
832 1 << PG_active | 1 << PG_private |
835 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
836 set_compound_page_dtor(page, NULL);
837 set_page_refcounted(page);
838 if (hstate_is_gigantic(h)) {
839 destroy_compound_gigantic_page(page, huge_page_order(h));
840 free_gigantic_page(page, huge_page_order(h));
842 arch_release_hugepage(page);
843 __free_pages(page, huge_page_order(h));
847 struct hstate *size_to_hstate(unsigned long size)
852 if (huge_page_size(h) == size)
859 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
860 * to hstate->hugepage_activelist.)
862 * This function can be called for tail pages, but never returns true for them.
864 bool page_huge_active(struct page *page)
866 VM_BUG_ON_PAGE(!PageHuge(page), page);
867 return PageHead(page) && PagePrivate(&page[1]);
870 /* never called for tail page */
871 static void set_page_huge_active(struct page *page)
873 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
874 SetPagePrivate(&page[1]);
877 static void clear_page_huge_active(struct page *page)
879 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
880 ClearPagePrivate(&page[1]);
883 void free_huge_page(struct page *page)
886 * Can't pass hstate in here because it is called from the
887 * compound page destructor.
889 struct hstate *h = page_hstate(page);
890 int nid = page_to_nid(page);
891 struct hugepage_subpool *spool =
892 (struct hugepage_subpool *)page_private(page);
893 bool restore_reserve;
895 set_page_private(page, 0);
896 page->mapping = NULL;
897 BUG_ON(page_count(page));
898 BUG_ON(page_mapcount(page));
899 restore_reserve = PagePrivate(page);
900 ClearPagePrivate(page);
902 spin_lock(&hugetlb_lock);
903 clear_page_huge_active(page);
904 hugetlb_cgroup_uncharge_page(hstate_index(h),
905 pages_per_huge_page(h), page);
907 h->resv_huge_pages++;
909 if (h->surplus_huge_pages_node[nid]) {
910 /* remove the page from active list */
911 list_del(&page->lru);
912 update_and_free_page(h, page);
913 h->surplus_huge_pages--;
914 h->surplus_huge_pages_node[nid]--;
916 arch_clear_hugepage_flags(page);
917 enqueue_huge_page(h, page);
919 spin_unlock(&hugetlb_lock);
920 hugepage_subpool_put_pages(spool, 1);
923 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
925 INIT_LIST_HEAD(&page->lru);
926 set_compound_page_dtor(page, free_huge_page);
927 spin_lock(&hugetlb_lock);
928 set_hugetlb_cgroup(page, NULL);
930 h->nr_huge_pages_node[nid]++;
931 spin_unlock(&hugetlb_lock);
932 put_page(page); /* free it into the hugepage allocator */
935 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
938 int nr_pages = 1 << order;
939 struct page *p = page + 1;
941 /* we rely on prep_new_huge_page to set the destructor */
942 set_compound_order(page, order);
944 __ClearPageReserved(page);
945 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
948 * For gigantic hugepages allocated through bootmem at
949 * boot, it's safer to be consistent with the not-gigantic
950 * hugepages and clear the PG_reserved bit from all tail pages
951 * too. Otherwse drivers using get_user_pages() to access tail
952 * pages may get the reference counting wrong if they see
953 * PG_reserved set on a tail page (despite the head page not
954 * having PG_reserved set). Enforcing this consistency between
955 * head and tail pages allows drivers to optimize away a check
956 * on the head page when they need know if put_page() is needed
957 * after get_user_pages().
959 __ClearPageReserved(p);
960 set_page_count(p, 0);
961 p->first_page = page;
966 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
967 * transparent huge pages. See the PageTransHuge() documentation for more
970 int PageHuge(struct page *page)
972 if (!PageCompound(page))
975 page = compound_head(page);
976 return get_compound_page_dtor(page) == free_huge_page;
978 EXPORT_SYMBOL_GPL(PageHuge);
981 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
982 * normal or transparent huge pages.
984 int PageHeadHuge(struct page *page_head)
986 if (!PageHead(page_head))
989 return get_compound_page_dtor(page_head) == free_huge_page;
992 pgoff_t __basepage_index(struct page *page)
994 struct page *page_head = compound_head(page);
995 pgoff_t index = page_index(page_head);
996 unsigned long compound_idx;
998 if (!PageHuge(page_head))
999 return page_index(page);
1001 if (compound_order(page_head) >= MAX_ORDER)
1002 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1004 compound_idx = page - page_head;
1006 return (index << compound_order(page_head)) + compound_idx;
1009 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1013 page = alloc_pages_exact_node(nid,
1014 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1015 __GFP_REPEAT|__GFP_NOWARN,
1016 huge_page_order(h));
1018 if (arch_prepare_hugepage(page)) {
1019 __free_pages(page, huge_page_order(h));
1022 prep_new_huge_page(h, page, nid);
1028 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1034 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1035 page = alloc_fresh_huge_page_node(h, node);
1043 count_vm_event(HTLB_BUDDY_PGALLOC);
1045 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1051 * Free huge page from pool from next node to free.
1052 * Attempt to keep persistent huge pages more or less
1053 * balanced over allowed nodes.
1054 * Called with hugetlb_lock locked.
1056 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1062 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1064 * If we're returning unused surplus pages, only examine
1065 * nodes with surplus pages.
1067 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1068 !list_empty(&h->hugepage_freelists[node])) {
1070 list_entry(h->hugepage_freelists[node].next,
1072 list_del(&page->lru);
1073 h->free_huge_pages--;
1074 h->free_huge_pages_node[node]--;
1076 h->surplus_huge_pages--;
1077 h->surplus_huge_pages_node[node]--;
1079 update_and_free_page(h, page);
1089 * Dissolve a given free hugepage into free buddy pages. This function does
1090 * nothing for in-use (including surplus) hugepages.
1092 static void dissolve_free_huge_page(struct page *page)
1094 spin_lock(&hugetlb_lock);
1095 if (PageHuge(page) && !page_count(page)) {
1096 struct hstate *h = page_hstate(page);
1097 int nid = page_to_nid(page);
1098 list_del(&page->lru);
1099 h->free_huge_pages--;
1100 h->free_huge_pages_node[nid]--;
1101 update_and_free_page(h, page);
1103 spin_unlock(&hugetlb_lock);
1107 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1108 * make specified memory blocks removable from the system.
1109 * Note that start_pfn should aligned with (minimum) hugepage size.
1111 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1113 unsigned int order = 8 * sizeof(void *);
1117 if (!hugepages_supported())
1120 /* Set scan step to minimum hugepage size */
1122 if (order > huge_page_order(h))
1123 order = huge_page_order(h);
1124 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1125 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1126 dissolve_free_huge_page(pfn_to_page(pfn));
1129 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1134 if (hstate_is_gigantic(h))
1138 * Assume we will successfully allocate the surplus page to
1139 * prevent racing processes from causing the surplus to exceed
1142 * This however introduces a different race, where a process B
1143 * tries to grow the static hugepage pool while alloc_pages() is
1144 * called by process A. B will only examine the per-node
1145 * counters in determining if surplus huge pages can be
1146 * converted to normal huge pages in adjust_pool_surplus(). A
1147 * won't be able to increment the per-node counter, until the
1148 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1149 * no more huge pages can be converted from surplus to normal
1150 * state (and doesn't try to convert again). Thus, we have a
1151 * case where a surplus huge page exists, the pool is grown, and
1152 * the surplus huge page still exists after, even though it
1153 * should just have been converted to a normal huge page. This
1154 * does not leak memory, though, as the hugepage will be freed
1155 * once it is out of use. It also does not allow the counters to
1156 * go out of whack in adjust_pool_surplus() as we don't modify
1157 * the node values until we've gotten the hugepage and only the
1158 * per-node value is checked there.
1160 spin_lock(&hugetlb_lock);
1161 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1162 spin_unlock(&hugetlb_lock);
1166 h->surplus_huge_pages++;
1168 spin_unlock(&hugetlb_lock);
1170 if (nid == NUMA_NO_NODE)
1171 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1172 __GFP_REPEAT|__GFP_NOWARN,
1173 huge_page_order(h));
1175 page = alloc_pages_exact_node(nid,
1176 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1177 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1179 if (page && arch_prepare_hugepage(page)) {
1180 __free_pages(page, huge_page_order(h));
1184 spin_lock(&hugetlb_lock);
1186 INIT_LIST_HEAD(&page->lru);
1187 r_nid = page_to_nid(page);
1188 set_compound_page_dtor(page, free_huge_page);
1189 set_hugetlb_cgroup(page, NULL);
1191 * We incremented the global counters already
1193 h->nr_huge_pages_node[r_nid]++;
1194 h->surplus_huge_pages_node[r_nid]++;
1195 __count_vm_event(HTLB_BUDDY_PGALLOC);
1198 h->surplus_huge_pages--;
1199 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1201 spin_unlock(&hugetlb_lock);
1207 * This allocation function is useful in the context where vma is irrelevant.
1208 * E.g. soft-offlining uses this function because it only cares physical
1209 * address of error page.
1211 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1213 struct page *page = NULL;
1215 spin_lock(&hugetlb_lock);
1216 if (h->free_huge_pages - h->resv_huge_pages > 0)
1217 page = dequeue_huge_page_node(h, nid);
1218 spin_unlock(&hugetlb_lock);
1221 page = alloc_buddy_huge_page(h, nid);
1227 * Increase the hugetlb pool such that it can accommodate a reservation
1230 static int gather_surplus_pages(struct hstate *h, int delta)
1232 struct list_head surplus_list;
1233 struct page *page, *tmp;
1235 int needed, allocated;
1236 bool alloc_ok = true;
1238 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1240 h->resv_huge_pages += delta;
1245 INIT_LIST_HEAD(&surplus_list);
1249 spin_unlock(&hugetlb_lock);
1250 for (i = 0; i < needed; i++) {
1251 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1256 list_add(&page->lru, &surplus_list);
1261 * After retaking hugetlb_lock, we need to recalculate 'needed'
1262 * because either resv_huge_pages or free_huge_pages may have changed.
1264 spin_lock(&hugetlb_lock);
1265 needed = (h->resv_huge_pages + delta) -
1266 (h->free_huge_pages + allocated);
1271 * We were not able to allocate enough pages to
1272 * satisfy the entire reservation so we free what
1273 * we've allocated so far.
1278 * The surplus_list now contains _at_least_ the number of extra pages
1279 * needed to accommodate the reservation. Add the appropriate number
1280 * of pages to the hugetlb pool and free the extras back to the buddy
1281 * allocator. Commit the entire reservation here to prevent another
1282 * process from stealing the pages as they are added to the pool but
1283 * before they are reserved.
1285 needed += allocated;
1286 h->resv_huge_pages += delta;
1289 /* Free the needed pages to the hugetlb pool */
1290 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1294 * This page is now managed by the hugetlb allocator and has
1295 * no users -- drop the buddy allocator's reference.
1297 put_page_testzero(page);
1298 VM_BUG_ON_PAGE(page_count(page), page);
1299 enqueue_huge_page(h, page);
1302 spin_unlock(&hugetlb_lock);
1304 /* Free unnecessary surplus pages to the buddy allocator */
1305 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1307 spin_lock(&hugetlb_lock);
1313 * When releasing a hugetlb pool reservation, any surplus pages that were
1314 * allocated to satisfy the reservation must be explicitly freed if they were
1316 * Called with hugetlb_lock held.
1318 static void return_unused_surplus_pages(struct hstate *h,
1319 unsigned long unused_resv_pages)
1321 unsigned long nr_pages;
1323 /* Uncommit the reservation */
1324 h->resv_huge_pages -= unused_resv_pages;
1326 /* Cannot return gigantic pages currently */
1327 if (hstate_is_gigantic(h))
1330 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1333 * We want to release as many surplus pages as possible, spread
1334 * evenly across all nodes with memory. Iterate across these nodes
1335 * until we can no longer free unreserved surplus pages. This occurs
1336 * when the nodes with surplus pages have no free pages.
1337 * free_pool_huge_page() will balance the the freed pages across the
1338 * on-line nodes with memory and will handle the hstate accounting.
1340 while (nr_pages--) {
1341 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1343 cond_resched_lock(&hugetlb_lock);
1348 * Determine if the huge page at addr within the vma has an associated
1349 * reservation. Where it does not we will need to logically increase
1350 * reservation and actually increase subpool usage before an allocation
1351 * can occur. Where any new reservation would be required the
1352 * reservation change is prepared, but not committed. Once the page
1353 * has been allocated from the subpool and instantiated the change should
1354 * be committed via vma_commit_reservation. No action is required on
1357 static long vma_needs_reservation(struct hstate *h,
1358 struct vm_area_struct *vma, unsigned long addr)
1360 struct resv_map *resv;
1364 resv = vma_resv_map(vma);
1368 idx = vma_hugecache_offset(h, vma, addr);
1369 chg = region_chg(resv, idx, idx + 1);
1371 if (vma->vm_flags & VM_MAYSHARE)
1374 return chg < 0 ? chg : 0;
1376 static void vma_commit_reservation(struct hstate *h,
1377 struct vm_area_struct *vma, unsigned long addr)
1379 struct resv_map *resv;
1382 resv = vma_resv_map(vma);
1386 idx = vma_hugecache_offset(h, vma, addr);
1387 region_add(resv, idx, idx + 1);
1390 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1391 unsigned long addr, int avoid_reserve)
1393 struct hugepage_subpool *spool = subpool_vma(vma);
1394 struct hstate *h = hstate_vma(vma);
1398 struct hugetlb_cgroup *h_cg;
1400 idx = hstate_index(h);
1402 * Processes that did not create the mapping will have no
1403 * reserves and will not have accounted against subpool
1404 * limit. Check that the subpool limit can be made before
1405 * satisfying the allocation MAP_NORESERVE mappings may also
1406 * need pages and subpool limit allocated allocated if no reserve
1409 chg = vma_needs_reservation(h, vma, addr);
1411 return ERR_PTR(-ENOMEM);
1412 if (chg || avoid_reserve)
1413 if (hugepage_subpool_get_pages(spool, 1))
1414 return ERR_PTR(-ENOSPC);
1416 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1418 goto out_subpool_put;
1420 spin_lock(&hugetlb_lock);
1421 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1423 spin_unlock(&hugetlb_lock);
1424 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1426 goto out_uncharge_cgroup;
1428 spin_lock(&hugetlb_lock);
1429 list_move(&page->lru, &h->hugepage_activelist);
1432 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1433 spin_unlock(&hugetlb_lock);
1435 set_page_private(page, (unsigned long)spool);
1437 vma_commit_reservation(h, vma, addr);
1440 out_uncharge_cgroup:
1441 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1443 if (chg || avoid_reserve)
1444 hugepage_subpool_put_pages(spool, 1);
1445 return ERR_PTR(-ENOSPC);
1449 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1450 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1451 * where no ERR_VALUE is expected to be returned.
1453 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1454 unsigned long addr, int avoid_reserve)
1456 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1462 int __weak alloc_bootmem_huge_page(struct hstate *h)
1464 struct huge_bootmem_page *m;
1467 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1470 addr = memblock_virt_alloc_try_nid_nopanic(
1471 huge_page_size(h), huge_page_size(h),
1472 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1475 * Use the beginning of the huge page to store the
1476 * huge_bootmem_page struct (until gather_bootmem
1477 * puts them into the mem_map).
1486 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1487 /* Put them into a private list first because mem_map is not up yet */
1488 list_add(&m->list, &huge_boot_pages);
1493 static void __init prep_compound_huge_page(struct page *page,
1496 if (unlikely(order > (MAX_ORDER - 1)))
1497 prep_compound_gigantic_page(page, order);
1499 prep_compound_page(page, order);
1502 /* Put bootmem huge pages into the standard lists after mem_map is up */
1503 static void __init gather_bootmem_prealloc(void)
1505 struct huge_bootmem_page *m;
1507 list_for_each_entry(m, &huge_boot_pages, list) {
1508 struct hstate *h = m->hstate;
1511 #ifdef CONFIG_HIGHMEM
1512 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1513 memblock_free_late(__pa(m),
1514 sizeof(struct huge_bootmem_page));
1516 page = virt_to_page(m);
1518 WARN_ON(page_count(page) != 1);
1519 prep_compound_huge_page(page, h->order);
1520 WARN_ON(PageReserved(page));
1521 prep_new_huge_page(h, page, page_to_nid(page));
1523 * If we had gigantic hugepages allocated at boot time, we need
1524 * to restore the 'stolen' pages to totalram_pages in order to
1525 * fix confusing memory reports from free(1) and another
1526 * side-effects, like CommitLimit going negative.
1528 if (hstate_is_gigantic(h))
1529 adjust_managed_page_count(page, 1 << h->order);
1533 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1537 for (i = 0; i < h->max_huge_pages; ++i) {
1538 if (hstate_is_gigantic(h)) {
1539 if (!alloc_bootmem_huge_page(h))
1541 } else if (!alloc_fresh_huge_page(h,
1542 &node_states[N_MEMORY]))
1545 h->max_huge_pages = i;
1548 static void __init hugetlb_init_hstates(void)
1552 for_each_hstate(h) {
1553 /* oversize hugepages were init'ed in early boot */
1554 if (!hstate_is_gigantic(h))
1555 hugetlb_hstate_alloc_pages(h);
1559 static char * __init memfmt(char *buf, unsigned long n)
1561 if (n >= (1UL << 30))
1562 sprintf(buf, "%lu GB", n >> 30);
1563 else if (n >= (1UL << 20))
1564 sprintf(buf, "%lu MB", n >> 20);
1566 sprintf(buf, "%lu KB", n >> 10);
1570 static void __init report_hugepages(void)
1574 for_each_hstate(h) {
1576 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1577 memfmt(buf, huge_page_size(h)),
1578 h->free_huge_pages);
1582 #ifdef CONFIG_HIGHMEM
1583 static void try_to_free_low(struct hstate *h, unsigned long count,
1584 nodemask_t *nodes_allowed)
1588 if (hstate_is_gigantic(h))
1591 for_each_node_mask(i, *nodes_allowed) {
1592 struct page *page, *next;
1593 struct list_head *freel = &h->hugepage_freelists[i];
1594 list_for_each_entry_safe(page, next, freel, lru) {
1595 if (count >= h->nr_huge_pages)
1597 if (PageHighMem(page))
1599 list_del(&page->lru);
1600 update_and_free_page(h, page);
1601 h->free_huge_pages--;
1602 h->free_huge_pages_node[page_to_nid(page)]--;
1607 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1608 nodemask_t *nodes_allowed)
1614 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1615 * balanced by operating on them in a round-robin fashion.
1616 * Returns 1 if an adjustment was made.
1618 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1623 VM_BUG_ON(delta != -1 && delta != 1);
1626 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1627 if (h->surplus_huge_pages_node[node])
1631 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1632 if (h->surplus_huge_pages_node[node] <
1633 h->nr_huge_pages_node[node])
1640 h->surplus_huge_pages += delta;
1641 h->surplus_huge_pages_node[node] += delta;
1645 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1646 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1647 nodemask_t *nodes_allowed)
1649 unsigned long min_count, ret;
1651 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1652 return h->max_huge_pages;
1655 * Increase the pool size
1656 * First take pages out of surplus state. Then make up the
1657 * remaining difference by allocating fresh huge pages.
1659 * We might race with alloc_buddy_huge_page() here and be unable
1660 * to convert a surplus huge page to a normal huge page. That is
1661 * not critical, though, it just means the overall size of the
1662 * pool might be one hugepage larger than it needs to be, but
1663 * within all the constraints specified by the sysctls.
1665 spin_lock(&hugetlb_lock);
1666 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1667 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1671 while (count > persistent_huge_pages(h)) {
1673 * If this allocation races such that we no longer need the
1674 * page, free_huge_page will handle it by freeing the page
1675 * and reducing the surplus.
1677 spin_unlock(&hugetlb_lock);
1678 if (hstate_is_gigantic(h))
1679 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1681 ret = alloc_fresh_huge_page(h, nodes_allowed);
1682 spin_lock(&hugetlb_lock);
1686 /* Bail for signals. Probably ctrl-c from user */
1687 if (signal_pending(current))
1692 * Decrease the pool size
1693 * First return free pages to the buddy allocator (being careful
1694 * to keep enough around to satisfy reservations). Then place
1695 * pages into surplus state as needed so the pool will shrink
1696 * to the desired size as pages become free.
1698 * By placing pages into the surplus state independent of the
1699 * overcommit value, we are allowing the surplus pool size to
1700 * exceed overcommit. There are few sane options here. Since
1701 * alloc_buddy_huge_page() is checking the global counter,
1702 * though, we'll note that we're not allowed to exceed surplus
1703 * and won't grow the pool anywhere else. Not until one of the
1704 * sysctls are changed, or the surplus pages go out of use.
1706 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1707 min_count = max(count, min_count);
1708 try_to_free_low(h, min_count, nodes_allowed);
1709 while (min_count < persistent_huge_pages(h)) {
1710 if (!free_pool_huge_page(h, nodes_allowed, 0))
1712 cond_resched_lock(&hugetlb_lock);
1714 while (count < persistent_huge_pages(h)) {
1715 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1719 ret = persistent_huge_pages(h);
1720 spin_unlock(&hugetlb_lock);
1724 #define HSTATE_ATTR_RO(_name) \
1725 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1727 #define HSTATE_ATTR(_name) \
1728 static struct kobj_attribute _name##_attr = \
1729 __ATTR(_name, 0644, _name##_show, _name##_store)
1731 static struct kobject *hugepages_kobj;
1732 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1734 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1736 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1740 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1741 if (hstate_kobjs[i] == kobj) {
1743 *nidp = NUMA_NO_NODE;
1747 return kobj_to_node_hstate(kobj, nidp);
1750 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1751 struct kobj_attribute *attr, char *buf)
1754 unsigned long nr_huge_pages;
1757 h = kobj_to_hstate(kobj, &nid);
1758 if (nid == NUMA_NO_NODE)
1759 nr_huge_pages = h->nr_huge_pages;
1761 nr_huge_pages = h->nr_huge_pages_node[nid];
1763 return sprintf(buf, "%lu\n", nr_huge_pages);
1766 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1767 struct hstate *h, int nid,
1768 unsigned long count, size_t len)
1771 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1773 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1778 if (nid == NUMA_NO_NODE) {
1780 * global hstate attribute
1782 if (!(obey_mempolicy &&
1783 init_nodemask_of_mempolicy(nodes_allowed))) {
1784 NODEMASK_FREE(nodes_allowed);
1785 nodes_allowed = &node_states[N_MEMORY];
1787 } else if (nodes_allowed) {
1789 * per node hstate attribute: adjust count to global,
1790 * but restrict alloc/free to the specified node.
1792 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1793 init_nodemask_of_node(nodes_allowed, nid);
1795 nodes_allowed = &node_states[N_MEMORY];
1797 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1799 if (nodes_allowed != &node_states[N_MEMORY])
1800 NODEMASK_FREE(nodes_allowed);
1804 NODEMASK_FREE(nodes_allowed);
1808 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1809 struct kobject *kobj, const char *buf,
1813 unsigned long count;
1817 err = kstrtoul(buf, 10, &count);
1821 h = kobj_to_hstate(kobj, &nid);
1822 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1825 static ssize_t nr_hugepages_show(struct kobject *kobj,
1826 struct kobj_attribute *attr, char *buf)
1828 return nr_hugepages_show_common(kobj, attr, buf);
1831 static ssize_t nr_hugepages_store(struct kobject *kobj,
1832 struct kobj_attribute *attr, const char *buf, size_t len)
1834 return nr_hugepages_store_common(false, kobj, buf, len);
1836 HSTATE_ATTR(nr_hugepages);
1841 * hstate attribute for optionally mempolicy-based constraint on persistent
1842 * huge page alloc/free.
1844 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1845 struct kobj_attribute *attr, char *buf)
1847 return nr_hugepages_show_common(kobj, attr, buf);
1850 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1851 struct kobj_attribute *attr, const char *buf, size_t len)
1853 return nr_hugepages_store_common(true, kobj, buf, len);
1855 HSTATE_ATTR(nr_hugepages_mempolicy);
1859 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1860 struct kobj_attribute *attr, char *buf)
1862 struct hstate *h = kobj_to_hstate(kobj, NULL);
1863 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1866 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1867 struct kobj_attribute *attr, const char *buf, size_t count)
1870 unsigned long input;
1871 struct hstate *h = kobj_to_hstate(kobj, NULL);
1873 if (hstate_is_gigantic(h))
1876 err = kstrtoul(buf, 10, &input);
1880 spin_lock(&hugetlb_lock);
1881 h->nr_overcommit_huge_pages = input;
1882 spin_unlock(&hugetlb_lock);
1886 HSTATE_ATTR(nr_overcommit_hugepages);
1888 static ssize_t free_hugepages_show(struct kobject *kobj,
1889 struct kobj_attribute *attr, char *buf)
1892 unsigned long free_huge_pages;
1895 h = kobj_to_hstate(kobj, &nid);
1896 if (nid == NUMA_NO_NODE)
1897 free_huge_pages = h->free_huge_pages;
1899 free_huge_pages = h->free_huge_pages_node[nid];
1901 return sprintf(buf, "%lu\n", free_huge_pages);
1903 HSTATE_ATTR_RO(free_hugepages);
1905 static ssize_t resv_hugepages_show(struct kobject *kobj,
1906 struct kobj_attribute *attr, char *buf)
1908 struct hstate *h = kobj_to_hstate(kobj, NULL);
1909 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1911 HSTATE_ATTR_RO(resv_hugepages);
1913 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1914 struct kobj_attribute *attr, char *buf)
1917 unsigned long surplus_huge_pages;
1920 h = kobj_to_hstate(kobj, &nid);
1921 if (nid == NUMA_NO_NODE)
1922 surplus_huge_pages = h->surplus_huge_pages;
1924 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1926 return sprintf(buf, "%lu\n", surplus_huge_pages);
1928 HSTATE_ATTR_RO(surplus_hugepages);
1930 static struct attribute *hstate_attrs[] = {
1931 &nr_hugepages_attr.attr,
1932 &nr_overcommit_hugepages_attr.attr,
1933 &free_hugepages_attr.attr,
1934 &resv_hugepages_attr.attr,
1935 &surplus_hugepages_attr.attr,
1937 &nr_hugepages_mempolicy_attr.attr,
1942 static struct attribute_group hstate_attr_group = {
1943 .attrs = hstate_attrs,
1946 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1947 struct kobject **hstate_kobjs,
1948 struct attribute_group *hstate_attr_group)
1951 int hi = hstate_index(h);
1953 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1954 if (!hstate_kobjs[hi])
1957 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1959 kobject_put(hstate_kobjs[hi]);
1964 static void __init hugetlb_sysfs_init(void)
1969 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1970 if (!hugepages_kobj)
1973 for_each_hstate(h) {
1974 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1975 hstate_kobjs, &hstate_attr_group);
1977 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1984 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1985 * with node devices in node_devices[] using a parallel array. The array
1986 * index of a node device or _hstate == node id.
1987 * This is here to avoid any static dependency of the node device driver, in
1988 * the base kernel, on the hugetlb module.
1990 struct node_hstate {
1991 struct kobject *hugepages_kobj;
1992 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1994 struct node_hstate node_hstates[MAX_NUMNODES];
1997 * A subset of global hstate attributes for node devices
1999 static struct attribute *per_node_hstate_attrs[] = {
2000 &nr_hugepages_attr.attr,
2001 &free_hugepages_attr.attr,
2002 &surplus_hugepages_attr.attr,
2006 static struct attribute_group per_node_hstate_attr_group = {
2007 .attrs = per_node_hstate_attrs,
2011 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2012 * Returns node id via non-NULL nidp.
2014 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2018 for (nid = 0; nid < nr_node_ids; nid++) {
2019 struct node_hstate *nhs = &node_hstates[nid];
2021 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2022 if (nhs->hstate_kobjs[i] == kobj) {
2034 * Unregister hstate attributes from a single node device.
2035 * No-op if no hstate attributes attached.
2037 static void hugetlb_unregister_node(struct node *node)
2040 struct node_hstate *nhs = &node_hstates[node->dev.id];
2042 if (!nhs->hugepages_kobj)
2043 return; /* no hstate attributes */
2045 for_each_hstate(h) {
2046 int idx = hstate_index(h);
2047 if (nhs->hstate_kobjs[idx]) {
2048 kobject_put(nhs->hstate_kobjs[idx]);
2049 nhs->hstate_kobjs[idx] = NULL;
2053 kobject_put(nhs->hugepages_kobj);
2054 nhs->hugepages_kobj = NULL;
2058 * hugetlb module exit: unregister hstate attributes from node devices
2061 static void hugetlb_unregister_all_nodes(void)
2066 * disable node device registrations.
2068 register_hugetlbfs_with_node(NULL, NULL);
2071 * remove hstate attributes from any nodes that have them.
2073 for (nid = 0; nid < nr_node_ids; nid++)
2074 hugetlb_unregister_node(node_devices[nid]);
2078 * Register hstate attributes for a single node device.
2079 * No-op if attributes already registered.
2081 static void hugetlb_register_node(struct node *node)
2084 struct node_hstate *nhs = &node_hstates[node->dev.id];
2087 if (nhs->hugepages_kobj)
2088 return; /* already allocated */
2090 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2092 if (!nhs->hugepages_kobj)
2095 for_each_hstate(h) {
2096 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2098 &per_node_hstate_attr_group);
2100 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2101 h->name, node->dev.id);
2102 hugetlb_unregister_node(node);
2109 * hugetlb init time: register hstate attributes for all registered node
2110 * devices of nodes that have memory. All on-line nodes should have
2111 * registered their associated device by this time.
2113 static void hugetlb_register_all_nodes(void)
2117 for_each_node_state(nid, N_MEMORY) {
2118 struct node *node = node_devices[nid];
2119 if (node->dev.id == nid)
2120 hugetlb_register_node(node);
2124 * Let the node device driver know we're here so it can
2125 * [un]register hstate attributes on node hotplug.
2127 register_hugetlbfs_with_node(hugetlb_register_node,
2128 hugetlb_unregister_node);
2130 #else /* !CONFIG_NUMA */
2132 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2140 static void hugetlb_unregister_all_nodes(void) { }
2142 static void hugetlb_register_all_nodes(void) { }
2146 static void __exit hugetlb_exit(void)
2150 hugetlb_unregister_all_nodes();
2152 for_each_hstate(h) {
2153 kobject_put(hstate_kobjs[hstate_index(h)]);
2156 kobject_put(hugepages_kobj);
2157 kfree(htlb_fault_mutex_table);
2159 module_exit(hugetlb_exit);
2161 static int __init hugetlb_init(void)
2165 if (!hugepages_supported())
2168 if (!size_to_hstate(default_hstate_size)) {
2169 default_hstate_size = HPAGE_SIZE;
2170 if (!size_to_hstate(default_hstate_size))
2171 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2173 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2174 if (default_hstate_max_huge_pages)
2175 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2177 hugetlb_init_hstates();
2178 gather_bootmem_prealloc();
2181 hugetlb_sysfs_init();
2182 hugetlb_register_all_nodes();
2183 hugetlb_cgroup_file_init();
2186 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2188 num_fault_mutexes = 1;
2190 htlb_fault_mutex_table =
2191 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2192 BUG_ON(!htlb_fault_mutex_table);
2194 for (i = 0; i < num_fault_mutexes; i++)
2195 mutex_init(&htlb_fault_mutex_table[i]);
2198 module_init(hugetlb_init);
2200 /* Should be called on processing a hugepagesz=... option */
2201 void __init hugetlb_add_hstate(unsigned int order)
2206 if (size_to_hstate(PAGE_SIZE << order)) {
2207 pr_warning("hugepagesz= specified twice, ignoring\n");
2210 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2212 h = &hstates[hugetlb_max_hstate++];
2214 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2215 h->nr_huge_pages = 0;
2216 h->free_huge_pages = 0;
2217 for (i = 0; i < MAX_NUMNODES; ++i)
2218 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2219 INIT_LIST_HEAD(&h->hugepage_activelist);
2220 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2221 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2222 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2223 huge_page_size(h)/1024);
2228 static int __init hugetlb_nrpages_setup(char *s)
2231 static unsigned long *last_mhp;
2234 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2235 * so this hugepages= parameter goes to the "default hstate".
2237 if (!hugetlb_max_hstate)
2238 mhp = &default_hstate_max_huge_pages;
2240 mhp = &parsed_hstate->max_huge_pages;
2242 if (mhp == last_mhp) {
2243 pr_warning("hugepages= specified twice without "
2244 "interleaving hugepagesz=, ignoring\n");
2248 if (sscanf(s, "%lu", mhp) <= 0)
2252 * Global state is always initialized later in hugetlb_init.
2253 * But we need to allocate >= MAX_ORDER hstates here early to still
2254 * use the bootmem allocator.
2256 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2257 hugetlb_hstate_alloc_pages(parsed_hstate);
2263 __setup("hugepages=", hugetlb_nrpages_setup);
2265 static int __init hugetlb_default_setup(char *s)
2267 default_hstate_size = memparse(s, &s);
2270 __setup("default_hugepagesz=", hugetlb_default_setup);
2272 static unsigned int cpuset_mems_nr(unsigned int *array)
2275 unsigned int nr = 0;
2277 for_each_node_mask(node, cpuset_current_mems_allowed)
2283 #ifdef CONFIG_SYSCTL
2284 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2285 struct ctl_table *table, int write,
2286 void __user *buffer, size_t *length, loff_t *ppos)
2288 struct hstate *h = &default_hstate;
2289 unsigned long tmp = h->max_huge_pages;
2292 if (!hugepages_supported())
2296 table->maxlen = sizeof(unsigned long);
2297 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2302 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2303 NUMA_NO_NODE, tmp, *length);
2308 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2309 void __user *buffer, size_t *length, loff_t *ppos)
2312 return hugetlb_sysctl_handler_common(false, table, write,
2313 buffer, length, ppos);
2317 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2318 void __user *buffer, size_t *length, loff_t *ppos)
2320 return hugetlb_sysctl_handler_common(true, table, write,
2321 buffer, length, ppos);
2323 #endif /* CONFIG_NUMA */
2325 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2326 void __user *buffer,
2327 size_t *length, loff_t *ppos)
2329 struct hstate *h = &default_hstate;
2333 if (!hugepages_supported())
2336 tmp = h->nr_overcommit_huge_pages;
2338 if (write && hstate_is_gigantic(h))
2342 table->maxlen = sizeof(unsigned long);
2343 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2348 spin_lock(&hugetlb_lock);
2349 h->nr_overcommit_huge_pages = tmp;
2350 spin_unlock(&hugetlb_lock);
2356 #endif /* CONFIG_SYSCTL */
2358 void hugetlb_report_meminfo(struct seq_file *m)
2360 struct hstate *h = &default_hstate;
2361 if (!hugepages_supported())
2364 "HugePages_Total: %5lu\n"
2365 "HugePages_Free: %5lu\n"
2366 "HugePages_Rsvd: %5lu\n"
2367 "HugePages_Surp: %5lu\n"
2368 "Hugepagesize: %8lu kB\n",
2372 h->surplus_huge_pages,
2373 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2376 int hugetlb_report_node_meminfo(int nid, char *buf)
2378 struct hstate *h = &default_hstate;
2379 if (!hugepages_supported())
2382 "Node %d HugePages_Total: %5u\n"
2383 "Node %d HugePages_Free: %5u\n"
2384 "Node %d HugePages_Surp: %5u\n",
2385 nid, h->nr_huge_pages_node[nid],
2386 nid, h->free_huge_pages_node[nid],
2387 nid, h->surplus_huge_pages_node[nid]);
2390 void hugetlb_show_meminfo(void)
2395 if (!hugepages_supported())
2398 for_each_node_state(nid, N_MEMORY)
2400 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2402 h->nr_huge_pages_node[nid],
2403 h->free_huge_pages_node[nid],
2404 h->surplus_huge_pages_node[nid],
2405 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2408 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2409 unsigned long hugetlb_total_pages(void)
2412 unsigned long nr_total_pages = 0;
2415 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2416 return nr_total_pages;
2419 static int hugetlb_acct_memory(struct hstate *h, long delta)
2423 spin_lock(&hugetlb_lock);
2425 * When cpuset is configured, it breaks the strict hugetlb page
2426 * reservation as the accounting is done on a global variable. Such
2427 * reservation is completely rubbish in the presence of cpuset because
2428 * the reservation is not checked against page availability for the
2429 * current cpuset. Application can still potentially OOM'ed by kernel
2430 * with lack of free htlb page in cpuset that the task is in.
2431 * Attempt to enforce strict accounting with cpuset is almost
2432 * impossible (or too ugly) because cpuset is too fluid that
2433 * task or memory node can be dynamically moved between cpusets.
2435 * The change of semantics for shared hugetlb mapping with cpuset is
2436 * undesirable. However, in order to preserve some of the semantics,
2437 * we fall back to check against current free page availability as
2438 * a best attempt and hopefully to minimize the impact of changing
2439 * semantics that cpuset has.
2442 if (gather_surplus_pages(h, delta) < 0)
2445 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2446 return_unused_surplus_pages(h, delta);
2453 return_unused_surplus_pages(h, (unsigned long) -delta);
2456 spin_unlock(&hugetlb_lock);
2460 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2462 struct resv_map *resv = vma_resv_map(vma);
2465 * This new VMA should share its siblings reservation map if present.
2466 * The VMA will only ever have a valid reservation map pointer where
2467 * it is being copied for another still existing VMA. As that VMA
2468 * has a reference to the reservation map it cannot disappear until
2469 * after this open call completes. It is therefore safe to take a
2470 * new reference here without additional locking.
2472 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2473 kref_get(&resv->refs);
2476 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2478 struct hstate *h = hstate_vma(vma);
2479 struct resv_map *resv = vma_resv_map(vma);
2480 struct hugepage_subpool *spool = subpool_vma(vma);
2481 unsigned long reserve, start, end;
2483 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2486 start = vma_hugecache_offset(h, vma, vma->vm_start);
2487 end = vma_hugecache_offset(h, vma, vma->vm_end);
2489 reserve = (end - start) - region_count(resv, start, end);
2491 kref_put(&resv->refs, resv_map_release);
2494 hugetlb_acct_memory(h, -reserve);
2495 hugepage_subpool_put_pages(spool, reserve);
2500 * We cannot handle pagefaults against hugetlb pages at all. They cause
2501 * handle_mm_fault() to try to instantiate regular-sized pages in the
2502 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2505 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2511 const struct vm_operations_struct hugetlb_vm_ops = {
2512 .fault = hugetlb_vm_op_fault,
2513 .open = hugetlb_vm_op_open,
2514 .close = hugetlb_vm_op_close,
2517 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2523 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2524 vma->vm_page_prot)));
2526 entry = huge_pte_wrprotect(mk_huge_pte(page,
2527 vma->vm_page_prot));
2529 entry = pte_mkyoung(entry);
2530 entry = pte_mkhuge(entry);
2531 entry = arch_make_huge_pte(entry, vma, page, writable);
2536 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2537 unsigned long address, pte_t *ptep)
2541 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2542 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2543 update_mmu_cache(vma, address, ptep);
2546 static int is_hugetlb_entry_migration(pte_t pte)
2550 if (huge_pte_none(pte) || pte_present(pte))
2552 swp = pte_to_swp_entry(pte);
2553 if (non_swap_entry(swp) && is_migration_entry(swp))
2559 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2563 if (huge_pte_none(pte) || pte_present(pte))
2565 swp = pte_to_swp_entry(pte);
2566 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2572 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2573 struct vm_area_struct *vma)
2575 pte_t *src_pte, *dst_pte, entry;
2576 struct page *ptepage;
2579 struct hstate *h = hstate_vma(vma);
2580 unsigned long sz = huge_page_size(h);
2581 unsigned long mmun_start; /* For mmu_notifiers */
2582 unsigned long mmun_end; /* For mmu_notifiers */
2585 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2587 mmun_start = vma->vm_start;
2588 mmun_end = vma->vm_end;
2590 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2592 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2593 spinlock_t *src_ptl, *dst_ptl;
2594 src_pte = huge_pte_offset(src, addr);
2597 dst_pte = huge_pte_alloc(dst, addr, sz);
2603 /* If the pagetables are shared don't copy or take references */
2604 if (dst_pte == src_pte)
2607 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2608 src_ptl = huge_pte_lockptr(h, src, src_pte);
2609 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2610 entry = huge_ptep_get(src_pte);
2611 if (huge_pte_none(entry)) { /* skip none entry */
2613 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2614 is_hugetlb_entry_hwpoisoned(entry))) {
2615 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2617 if (is_write_migration_entry(swp_entry) && cow) {
2619 * COW mappings require pages in both
2620 * parent and child to be set to read.
2622 make_migration_entry_read(&swp_entry);
2623 entry = swp_entry_to_pte(swp_entry);
2624 set_huge_pte_at(src, addr, src_pte, entry);
2626 set_huge_pte_at(dst, addr, dst_pte, entry);
2629 huge_ptep_set_wrprotect(src, addr, src_pte);
2630 entry = huge_ptep_get(src_pte);
2631 ptepage = pte_page(entry);
2633 page_dup_rmap(ptepage);
2634 set_huge_pte_at(dst, addr, dst_pte, entry);
2636 spin_unlock(src_ptl);
2637 spin_unlock(dst_ptl);
2641 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2646 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2647 unsigned long start, unsigned long end,
2648 struct page *ref_page)
2650 int force_flush = 0;
2651 struct mm_struct *mm = vma->vm_mm;
2652 unsigned long address;
2657 struct hstate *h = hstate_vma(vma);
2658 unsigned long sz = huge_page_size(h);
2659 const unsigned long mmun_start = start; /* For mmu_notifiers */
2660 const unsigned long mmun_end = end; /* For mmu_notifiers */
2662 WARN_ON(!is_vm_hugetlb_page(vma));
2663 BUG_ON(start & ~huge_page_mask(h));
2664 BUG_ON(end & ~huge_page_mask(h));
2666 tlb_start_vma(tlb, vma);
2667 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2669 for (address = start; address < end; address += sz) {
2670 ptep = huge_pte_offset(mm, address);
2674 ptl = huge_pte_lock(h, mm, ptep);
2675 if (huge_pmd_unshare(mm, &address, ptep))
2678 pte = huge_ptep_get(ptep);
2679 if (huge_pte_none(pte))
2683 * Migrating hugepage or HWPoisoned hugepage is already
2684 * unmapped and its refcount is dropped, so just clear pte here.
2686 if (unlikely(!pte_present(pte))) {
2687 huge_pte_clear(mm, address, ptep);
2691 page = pte_page(pte);
2693 * If a reference page is supplied, it is because a specific
2694 * page is being unmapped, not a range. Ensure the page we
2695 * are about to unmap is the actual page of interest.
2698 if (page != ref_page)
2702 * Mark the VMA as having unmapped its page so that
2703 * future faults in this VMA will fail rather than
2704 * looking like data was lost
2706 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2709 pte = huge_ptep_get_and_clear(mm, address, ptep);
2710 tlb_remove_tlb_entry(tlb, ptep, address);
2711 if (huge_pte_dirty(pte))
2712 set_page_dirty(page);
2714 page_remove_rmap(page);
2715 force_flush = !__tlb_remove_page(tlb, page);
2720 /* Bail out after unmapping reference page if supplied */
2729 * mmu_gather ran out of room to batch pages, we break out of
2730 * the PTE lock to avoid doing the potential expensive TLB invalidate
2731 * and page-free while holding it.
2736 if (address < end && !ref_page)
2739 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2740 tlb_end_vma(tlb, vma);
2743 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2744 struct vm_area_struct *vma, unsigned long start,
2745 unsigned long end, struct page *ref_page)
2747 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2750 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2751 * test will fail on a vma being torn down, and not grab a page table
2752 * on its way out. We're lucky that the flag has such an appropriate
2753 * name, and can in fact be safely cleared here. We could clear it
2754 * before the __unmap_hugepage_range above, but all that's necessary
2755 * is to clear it before releasing the i_mmap_mutex. This works
2756 * because in the context this is called, the VMA is about to be
2757 * destroyed and the i_mmap_mutex is held.
2759 vma->vm_flags &= ~VM_MAYSHARE;
2762 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2763 unsigned long end, struct page *ref_page)
2765 struct mm_struct *mm;
2766 struct mmu_gather tlb;
2770 tlb_gather_mmu(&tlb, mm, start, end);
2771 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2772 tlb_finish_mmu(&tlb, start, end);
2776 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2777 * mappping it owns the reserve page for. The intention is to unmap the page
2778 * from other VMAs and let the children be SIGKILLed if they are faulting the
2781 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2782 struct page *page, unsigned long address)
2784 struct hstate *h = hstate_vma(vma);
2785 struct vm_area_struct *iter_vma;
2786 struct address_space *mapping;
2790 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2791 * from page cache lookup which is in HPAGE_SIZE units.
2793 address = address & huge_page_mask(h);
2794 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2796 mapping = file_inode(vma->vm_file)->i_mapping;
2799 * Take the mapping lock for the duration of the table walk. As
2800 * this mapping should be shared between all the VMAs,
2801 * __unmap_hugepage_range() is called as the lock is already held
2803 mutex_lock(&mapping->i_mmap_mutex);
2804 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2805 /* Do not unmap the current VMA */
2806 if (iter_vma == vma)
2810 * Shared VMAs have their own reserves and do not affect
2811 * MAP_PRIVATE accounting but it is possible that a shared
2812 * VMA is using the same page so check and skip such VMAs.
2814 if (iter_vma->vm_flags & VM_MAYSHARE)
2818 * Unmap the page from other VMAs without their own reserves.
2819 * They get marked to be SIGKILLed if they fault in these
2820 * areas. This is because a future no-page fault on this VMA
2821 * could insert a zeroed page instead of the data existing
2822 * from the time of fork. This would look like data corruption
2824 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2825 unmap_hugepage_range(iter_vma, address,
2826 address + huge_page_size(h), page);
2828 mutex_unlock(&mapping->i_mmap_mutex);
2832 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2833 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2834 * cannot race with other handlers or page migration.
2835 * Keep the pte_same checks anyway to make transition from the mutex easier.
2837 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2838 unsigned long address, pte_t *ptep, pte_t pte,
2839 struct page *pagecache_page, spinlock_t *ptl)
2841 struct hstate *h = hstate_vma(vma);
2842 struct page *old_page, *new_page;
2843 int ret = 0, outside_reserve = 0;
2844 unsigned long mmun_start; /* For mmu_notifiers */
2845 unsigned long mmun_end; /* For mmu_notifiers */
2847 old_page = pte_page(pte);
2850 /* If no-one else is actually using this page, avoid the copy
2851 * and just make the page writable */
2852 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2853 page_move_anon_rmap(old_page, vma, address);
2854 set_huge_ptep_writable(vma, address, ptep);
2859 * If the process that created a MAP_PRIVATE mapping is about to
2860 * perform a COW due to a shared page count, attempt to satisfy
2861 * the allocation without using the existing reserves. The pagecache
2862 * page is used to determine if the reserve at this address was
2863 * consumed or not. If reserves were used, a partial faulted mapping
2864 * at the time of fork() could consume its reserves on COW instead
2865 * of the full address range.
2867 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2868 old_page != pagecache_page)
2869 outside_reserve = 1;
2871 page_cache_get(old_page);
2874 * Drop page table lock as buddy allocator may be called. It will
2875 * be acquired again before returning to the caller, as expected.
2878 new_page = alloc_huge_page(vma, address, outside_reserve);
2880 if (IS_ERR(new_page)) {
2882 * If a process owning a MAP_PRIVATE mapping fails to COW,
2883 * it is due to references held by a child and an insufficient
2884 * huge page pool. To guarantee the original mappers
2885 * reliability, unmap the page from child processes. The child
2886 * may get SIGKILLed if it later faults.
2888 if (outside_reserve) {
2889 page_cache_release(old_page);
2890 BUG_ON(huge_pte_none(pte));
2891 unmap_ref_private(mm, vma, old_page, address);
2892 BUG_ON(huge_pte_none(pte));
2894 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2896 pte_same(huge_ptep_get(ptep), pte)))
2897 goto retry_avoidcopy;
2899 * race occurs while re-acquiring page table
2900 * lock, and our job is done.
2905 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2906 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2907 goto out_release_old;
2911 * When the original hugepage is shared one, it does not have
2912 * anon_vma prepared.
2914 if (unlikely(anon_vma_prepare(vma))) {
2916 goto out_release_all;
2919 copy_user_huge_page(new_page, old_page, address, vma,
2920 pages_per_huge_page(h));
2921 __SetPageUptodate(new_page);
2922 set_page_huge_active(new_page);
2924 mmun_start = address & huge_page_mask(h);
2925 mmun_end = mmun_start + huge_page_size(h);
2926 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2929 * Retake the page table lock to check for racing updates
2930 * before the page tables are altered
2933 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2934 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
2935 ClearPagePrivate(new_page);
2938 huge_ptep_clear_flush(vma, address, ptep);
2939 set_huge_pte_at(mm, address, ptep,
2940 make_huge_pte(vma, new_page, 1));
2941 page_remove_rmap(old_page);
2942 hugepage_add_new_anon_rmap(new_page, vma, address);
2943 /* Make the old page be freed below */
2944 new_page = old_page;
2947 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2949 page_cache_release(new_page);
2951 page_cache_release(old_page);
2953 spin_lock(ptl); /* Caller expects lock to be held */
2957 /* Return the pagecache page at a given address within a VMA */
2958 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2959 struct vm_area_struct *vma, unsigned long address)
2961 struct address_space *mapping;
2964 mapping = vma->vm_file->f_mapping;
2965 idx = vma_hugecache_offset(h, vma, address);
2967 return find_lock_page(mapping, idx);
2971 * Return whether there is a pagecache page to back given address within VMA.
2972 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2974 static bool hugetlbfs_pagecache_present(struct hstate *h,
2975 struct vm_area_struct *vma, unsigned long address)
2977 struct address_space *mapping;
2981 mapping = vma->vm_file->f_mapping;
2982 idx = vma_hugecache_offset(h, vma, address);
2984 page = find_get_page(mapping, idx);
2987 return page != NULL;
2990 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2991 struct address_space *mapping, pgoff_t idx,
2992 unsigned long address, pte_t *ptep, unsigned int flags)
2994 struct hstate *h = hstate_vma(vma);
2995 int ret = VM_FAULT_SIGBUS;
3003 * Currently, we are forced to kill the process in the event the
3004 * original mapper has unmapped pages from the child due to a failed
3005 * COW. Warn that such a situation has occurred as it may not be obvious
3007 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3008 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3014 * Use page lock to guard against racing truncation
3015 * before we get page_table_lock.
3018 page = find_lock_page(mapping, idx);
3020 size = i_size_read(mapping->host) >> huge_page_shift(h);
3023 page = alloc_huge_page(vma, address, 0);
3025 ret = PTR_ERR(page);
3029 ret = VM_FAULT_SIGBUS;
3032 clear_huge_page(page, address, pages_per_huge_page(h));
3033 __SetPageUptodate(page);
3034 set_page_huge_active(page);
3036 if (vma->vm_flags & VM_MAYSHARE) {
3038 struct inode *inode = mapping->host;
3040 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3047 ClearPagePrivate(page);
3049 spin_lock(&inode->i_lock);
3050 inode->i_blocks += blocks_per_huge_page(h);
3051 spin_unlock(&inode->i_lock);
3054 if (unlikely(anon_vma_prepare(vma))) {
3056 goto backout_unlocked;
3062 * If memory error occurs between mmap() and fault, some process
3063 * don't have hwpoisoned swap entry for errored virtual address.
3064 * So we need to block hugepage fault by PG_hwpoison bit check.
3066 if (unlikely(PageHWPoison(page))) {
3067 ret = VM_FAULT_HWPOISON |
3068 VM_FAULT_SET_HINDEX(hstate_index(h));
3069 goto backout_unlocked;
3074 * If we are going to COW a private mapping later, we examine the
3075 * pending reservations for this page now. This will ensure that
3076 * any allocations necessary to record that reservation occur outside
3079 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3080 if (vma_needs_reservation(h, vma, address) < 0) {
3082 goto backout_unlocked;
3085 ptl = huge_pte_lockptr(h, mm, ptep);
3087 size = i_size_read(mapping->host) >> huge_page_shift(h);
3092 if (!huge_pte_none(huge_ptep_get(ptep)))
3096 ClearPagePrivate(page);
3097 hugepage_add_new_anon_rmap(page, vma, address);
3099 page_dup_rmap(page);
3100 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3101 && (vma->vm_flags & VM_SHARED)));
3102 set_huge_pte_at(mm, address, ptep, new_pte);
3104 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3105 /* Optimization, do the COW without a second fault */
3106 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3123 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3124 struct vm_area_struct *vma,
3125 struct address_space *mapping,
3126 pgoff_t idx, unsigned long address)
3128 unsigned long key[2];
3131 if (vma->vm_flags & VM_SHARED) {
3132 key[0] = (unsigned long) mapping;
3135 key[0] = (unsigned long) mm;
3136 key[1] = address >> huge_page_shift(h);
3139 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3141 return hash & (num_fault_mutexes - 1);
3145 * For uniprocesor systems we always use a single mutex, so just
3146 * return 0 and avoid the hashing overhead.
3148 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3149 struct vm_area_struct *vma,
3150 struct address_space *mapping,
3151 pgoff_t idx, unsigned long address)
3157 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3158 unsigned long address, unsigned int flags)
3165 struct page *page = NULL;
3166 struct page *pagecache_page = NULL;
3167 struct hstate *h = hstate_vma(vma);
3168 struct address_space *mapping;
3169 int need_wait_lock = 0;
3171 address &= huge_page_mask(h);
3173 ptep = huge_pte_offset(mm, address);
3175 entry = huge_ptep_get(ptep);
3176 if (unlikely(is_hugetlb_entry_migration(entry))) {
3177 migration_entry_wait_huge(vma, mm, ptep);
3179 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3180 return VM_FAULT_HWPOISON_LARGE |
3181 VM_FAULT_SET_HINDEX(hstate_index(h));
3184 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3186 return VM_FAULT_OOM;
3188 mapping = vma->vm_file->f_mapping;
3189 idx = vma_hugecache_offset(h, vma, address);
3192 * Serialize hugepage allocation and instantiation, so that we don't
3193 * get spurious allocation failures if two CPUs race to instantiate
3194 * the same page in the page cache.
3196 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3197 mutex_lock(&htlb_fault_mutex_table[hash]);
3199 entry = huge_ptep_get(ptep);
3200 if (huge_pte_none(entry)) {
3201 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3208 * entry could be a migration/hwpoison entry at this point, so this
3209 * check prevents the kernel from going below assuming that we have
3210 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3211 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3214 if (!pte_present(entry))
3218 * If we are going to COW the mapping later, we examine the pending
3219 * reservations for this page now. This will ensure that any
3220 * allocations necessary to record that reservation occur outside the
3221 * spinlock. For private mappings, we also lookup the pagecache
3222 * page now as it is used to determine if a reservation has been
3225 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3226 if (vma_needs_reservation(h, vma, address) < 0) {
3231 if (!(vma->vm_flags & VM_MAYSHARE))
3232 pagecache_page = hugetlbfs_pagecache_page(h,
3236 ptl = huge_pte_lock(h, mm, ptep);
3238 /* Check for a racing update before calling hugetlb_cow */
3239 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3243 * hugetlb_cow() requires page locks of pte_page(entry) and
3244 * pagecache_page, so here we need take the former one
3245 * when page != pagecache_page or !pagecache_page.
3247 page = pte_page(entry);
3248 if (page != pagecache_page)
3249 if (!trylock_page(page)) {
3256 if (flags & FAULT_FLAG_WRITE) {
3257 if (!huge_pte_write(entry)) {
3258 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3259 pagecache_page, ptl);
3262 entry = huge_pte_mkdirty(entry);
3264 entry = pte_mkyoung(entry);
3265 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3266 flags & FAULT_FLAG_WRITE))
3267 update_mmu_cache(vma, address, ptep);
3269 if (page != pagecache_page)
3275 if (pagecache_page) {
3276 unlock_page(pagecache_page);
3277 put_page(pagecache_page);
3280 mutex_unlock(&htlb_fault_mutex_table[hash]);
3282 * Generally it's safe to hold refcount during waiting page lock. But
3283 * here we just wait to defer the next page fault to avoid busy loop and
3284 * the page is not used after unlocked before returning from the current
3285 * page fault. So we are safe from accessing freed page, even if we wait
3286 * here without taking refcount.
3289 wait_on_page_locked(page);
3293 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3294 struct page **pages, struct vm_area_struct **vmas,
3295 unsigned long *position, unsigned long *nr_pages,
3296 long i, unsigned int flags)
3298 unsigned long pfn_offset;
3299 unsigned long vaddr = *position;
3300 unsigned long remainder = *nr_pages;
3301 struct hstate *h = hstate_vma(vma);
3303 while (vaddr < vma->vm_end && remainder) {
3305 spinlock_t *ptl = NULL;
3310 * Some archs (sparc64, sh*) have multiple pte_ts to
3311 * each hugepage. We have to make sure we get the
3312 * first, for the page indexing below to work.
3314 * Note that page table lock is not held when pte is null.
3316 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3318 ptl = huge_pte_lock(h, mm, pte);
3319 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3322 * When coredumping, it suits get_dump_page if we just return
3323 * an error where there's an empty slot with no huge pagecache
3324 * to back it. This way, we avoid allocating a hugepage, and
3325 * the sparse dumpfile avoids allocating disk blocks, but its
3326 * huge holes still show up with zeroes where they need to be.
3328 if (absent && (flags & FOLL_DUMP) &&
3329 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3337 * We need call hugetlb_fault for both hugepages under migration
3338 * (in which case hugetlb_fault waits for the migration,) and
3339 * hwpoisoned hugepages (in which case we need to prevent the
3340 * caller from accessing to them.) In order to do this, we use
3341 * here is_swap_pte instead of is_hugetlb_entry_migration and
3342 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3343 * both cases, and because we can't follow correct pages
3344 * directly from any kind of swap entries.
3346 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3347 ((flags & FOLL_WRITE) &&
3348 !huge_pte_write(huge_ptep_get(pte)))) {
3353 ret = hugetlb_fault(mm, vma, vaddr,
3354 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3355 if (!(ret & VM_FAULT_ERROR))
3362 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3363 page = pte_page(huge_ptep_get(pte));
3366 pages[i] = mem_map_offset(page, pfn_offset);
3367 get_page_foll(pages[i]);
3377 if (vaddr < vma->vm_end && remainder &&
3378 pfn_offset < pages_per_huge_page(h)) {
3380 * We use pfn_offset to avoid touching the pageframes
3381 * of this compound page.
3387 *nr_pages = remainder;
3390 return i ? i : -EFAULT;
3393 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3394 unsigned long address, unsigned long end, pgprot_t newprot)
3396 struct mm_struct *mm = vma->vm_mm;
3397 unsigned long start = address;
3400 struct hstate *h = hstate_vma(vma);
3401 unsigned long pages = 0;
3403 BUG_ON(address >= end);
3404 flush_cache_range(vma, address, end);
3406 mmu_notifier_invalidate_range_start(mm, start, end);
3407 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3408 for (; address < end; address += huge_page_size(h)) {
3410 ptep = huge_pte_offset(mm, address);
3413 ptl = huge_pte_lock(h, mm, ptep);
3414 if (huge_pmd_unshare(mm, &address, ptep)) {
3419 pte = huge_ptep_get(ptep);
3420 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3424 if (unlikely(is_hugetlb_entry_migration(pte))) {
3425 swp_entry_t entry = pte_to_swp_entry(pte);
3427 if (is_write_migration_entry(entry)) {
3430 make_migration_entry_read(&entry);
3431 newpte = swp_entry_to_pte(entry);
3432 set_huge_pte_at(mm, address, ptep, newpte);
3438 if (!huge_pte_none(pte)) {
3439 pte = huge_ptep_get_and_clear(mm, address, ptep);
3440 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3441 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3442 set_huge_pte_at(mm, address, ptep, pte);
3448 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3449 * may have cleared our pud entry and done put_page on the page table:
3450 * once we release i_mmap_mutex, another task can do the final put_page
3451 * and that page table be reused and filled with junk.
3453 flush_tlb_range(vma, start, end);
3454 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3455 mmu_notifier_invalidate_range_end(mm, start, end);
3457 return pages << h->order;
3460 int hugetlb_reserve_pages(struct inode *inode,
3462 struct vm_area_struct *vma,
3463 vm_flags_t vm_flags)
3466 struct hstate *h = hstate_inode(inode);
3467 struct hugepage_subpool *spool = subpool_inode(inode);
3468 struct resv_map *resv_map;
3471 * Only apply hugepage reservation if asked. At fault time, an
3472 * attempt will be made for VM_NORESERVE to allocate a page
3473 * without using reserves
3475 if (vm_flags & VM_NORESERVE)
3479 * Shared mappings base their reservation on the number of pages that
3480 * are already allocated on behalf of the file. Private mappings need
3481 * to reserve the full area even if read-only as mprotect() may be
3482 * called to make the mapping read-write. Assume !vma is a shm mapping
3484 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3485 resv_map = inode_resv_map(inode);
3487 chg = region_chg(resv_map, from, to);
3490 resv_map = resv_map_alloc();
3496 set_vma_resv_map(vma, resv_map);
3497 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3505 /* There must be enough pages in the subpool for the mapping */
3506 if (hugepage_subpool_get_pages(spool, chg)) {
3512 * Check enough hugepages are available for the reservation.
3513 * Hand the pages back to the subpool if there are not
3515 ret = hugetlb_acct_memory(h, chg);
3517 hugepage_subpool_put_pages(spool, chg);
3522 * Account for the reservations made. Shared mappings record regions
3523 * that have reservations as they are shared by multiple VMAs.
3524 * When the last VMA disappears, the region map says how much
3525 * the reservation was and the page cache tells how much of
3526 * the reservation was consumed. Private mappings are per-VMA and
3527 * only the consumed reservations are tracked. When the VMA
3528 * disappears, the original reservation is the VMA size and the
3529 * consumed reservations are stored in the map. Hence, nothing
3530 * else has to be done for private mappings here
3532 if (!vma || vma->vm_flags & VM_MAYSHARE)
3533 region_add(resv_map, from, to);
3536 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3537 kref_put(&resv_map->refs, resv_map_release);
3541 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3543 struct hstate *h = hstate_inode(inode);
3544 struct resv_map *resv_map = inode_resv_map(inode);
3546 struct hugepage_subpool *spool = subpool_inode(inode);
3549 chg = region_truncate(resv_map, offset);
3550 spin_lock(&inode->i_lock);
3551 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3552 spin_unlock(&inode->i_lock);
3554 hugepage_subpool_put_pages(spool, (chg - freed));
3555 hugetlb_acct_memory(h, -(chg - freed));
3558 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3559 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3560 struct vm_area_struct *vma,
3561 unsigned long addr, pgoff_t idx)
3563 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3565 unsigned long sbase = saddr & PUD_MASK;
3566 unsigned long s_end = sbase + PUD_SIZE;
3568 /* Allow segments to share if only one is marked locked */
3569 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3570 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3573 * match the virtual addresses, permission and the alignment of the
3576 if (pmd_index(addr) != pmd_index(saddr) ||
3577 vm_flags != svm_flags ||
3578 sbase < svma->vm_start || svma->vm_end < s_end)
3584 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3586 unsigned long base = addr & PUD_MASK;
3587 unsigned long end = base + PUD_SIZE;
3590 * check on proper vm_flags and page table alignment
3592 if (vma->vm_flags & VM_MAYSHARE &&
3593 vma->vm_start <= base && end <= vma->vm_end)
3599 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3600 * and returns the corresponding pte. While this is not necessary for the
3601 * !shared pmd case because we can allocate the pmd later as well, it makes the
3602 * code much cleaner. pmd allocation is essential for the shared case because
3603 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3604 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3605 * bad pmd for sharing.
3607 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3609 struct vm_area_struct *vma = find_vma(mm, addr);
3610 struct address_space *mapping = vma->vm_file->f_mapping;
3611 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3613 struct vm_area_struct *svma;
3614 unsigned long saddr;
3619 if (!vma_shareable(vma, addr))
3620 return (pte_t *)pmd_alloc(mm, pud, addr);
3622 mutex_lock(&mapping->i_mmap_mutex);
3623 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3627 saddr = page_table_shareable(svma, vma, addr, idx);
3629 spte = huge_pte_offset(svma->vm_mm, saddr);
3631 get_page(virt_to_page(spte));
3640 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3643 pud_populate(mm, pud,
3644 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3646 put_page(virt_to_page(spte));
3649 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3650 mutex_unlock(&mapping->i_mmap_mutex);
3655 * unmap huge page backed by shared pte.
3657 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3658 * indicated by page_count > 1, unmap is achieved by clearing pud and
3659 * decrementing the ref count. If count == 1, the pte page is not shared.
3661 * called with page table lock held.
3663 * returns: 1 successfully unmapped a shared pte page
3664 * 0 the underlying pte page is not shared, or it is the last user
3666 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3668 pgd_t *pgd = pgd_offset(mm, *addr);
3669 pud_t *pud = pud_offset(pgd, *addr);
3671 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3672 if (page_count(virt_to_page(ptep)) == 1)
3676 put_page(virt_to_page(ptep));
3677 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3680 #define want_pmd_share() (1)
3681 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3682 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3686 #define want_pmd_share() (0)
3687 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3689 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3690 pte_t *huge_pte_alloc(struct mm_struct *mm,
3691 unsigned long addr, unsigned long sz)
3697 pgd = pgd_offset(mm, addr);
3698 pud = pud_alloc(mm, pgd, addr);
3700 if (sz == PUD_SIZE) {
3703 BUG_ON(sz != PMD_SIZE);
3704 if (want_pmd_share() && pud_none(*pud))
3705 pte = huge_pmd_share(mm, addr, pud);
3707 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3710 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3715 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3721 pgd = pgd_offset(mm, addr);
3722 if (pgd_present(*pgd)) {
3723 pud = pud_offset(pgd, addr);
3724 if (pud_present(*pud)) {
3726 return (pte_t *)pud;
3727 pmd = pmd_offset(pud, addr);
3730 return (pte_t *) pmd;
3733 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3736 * These functions are overwritable if your architecture needs its own
3739 struct page * __weak
3740 follow_huge_addr(struct mm_struct *mm, unsigned long address,
3743 return ERR_PTR(-EINVAL);
3746 struct page * __weak
3747 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3748 pmd_t *pmd, int flags)
3750 struct page *page = NULL;
3753 ptl = pmd_lockptr(mm, pmd);
3756 * make sure that the address range covered by this pmd is not
3757 * unmapped from other threads.
3759 if (!pmd_huge(*pmd))
3761 if (pmd_present(*pmd)) {
3762 page = pte_page(*(pte_t *)pmd) +
3763 ((address & ~PMD_MASK) >> PAGE_SHIFT);
3764 if (flags & FOLL_GET)
3767 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3769 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3773 * hwpoisoned entry is treated as no_page_table in
3774 * follow_page_mask().
3782 struct page * __weak
3783 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3784 pud_t *pud, int flags)
3786 if (flags & FOLL_GET)
3789 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
3792 #ifdef CONFIG_MEMORY_FAILURE
3794 /* Should be called in hugetlb_lock */
3795 static int is_hugepage_on_freelist(struct page *hpage)
3799 struct hstate *h = page_hstate(hpage);
3800 int nid = page_to_nid(hpage);
3802 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3809 * This function is called from memory failure code.
3810 * Assume the caller holds page lock of the head page.
3812 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3814 struct hstate *h = page_hstate(hpage);
3815 int nid = page_to_nid(hpage);
3818 spin_lock(&hugetlb_lock);
3819 if (is_hugepage_on_freelist(hpage)) {
3821 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3822 * but dangling hpage->lru can trigger list-debug warnings
3823 * (this happens when we call unpoison_memory() on it),
3824 * so let it point to itself with list_del_init().
3826 list_del_init(&hpage->lru);
3827 set_page_refcounted(hpage);
3828 h->free_huge_pages--;
3829 h->free_huge_pages_node[nid]--;
3832 spin_unlock(&hugetlb_lock);
3837 bool isolate_huge_page(struct page *page, struct list_head *list)
3841 VM_BUG_ON_PAGE(!PageHead(page), page);
3842 spin_lock(&hugetlb_lock);
3843 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
3847 clear_page_huge_active(page);
3848 list_move_tail(&page->lru, list);
3850 spin_unlock(&hugetlb_lock);
3854 void putback_active_hugepage(struct page *page)
3856 VM_BUG_ON_PAGE(!PageHead(page), page);
3857 spin_lock(&hugetlb_lock);
3858 set_page_huge_active(page);
3859 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3860 spin_unlock(&hugetlb_lock);
3864 bool is_hugepage_active(struct page *page)
3866 VM_BUG_ON_PAGE(!PageHuge(page), page);
3868 * This function can be called for a tail page because the caller,
3869 * scan_movable_pages, scans through a given pfn-range which typically
3870 * covers one memory block. In systems using gigantic hugepage (1GB
3871 * for x86_64,) a hugepage is larger than a memory block, and we don't
3872 * support migrating such large hugepages for now, so return false
3873 * when called for tail pages.
3878 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3879 * so we should return false for them.
3881 if (unlikely(PageHWPoison(page)))
3883 return page_count(page) > 0;